text
stringlengths
4
2.78M
meta
dict
--- abstract: 'We prove a version of the implicit function theorem for Lipschitz mappings $f:\mathbb{R}^{n+m}\supset A \to X$ into arbitrary metric spaces. As long as the pull-back of the Hausdorff content $\mathcal{H}_{\infty}^n$ by $f$ has positive upper $n$-density on a set of positive Lebesgue measure, then, there is a local diffeomorphism $G$ in ${\mathbb R}^{n+m}$ and a Lipschitz map $\pi:X\to {\mathbb R}^n$ such that $\pi\circ f\circ G^{-1}$, when restricted to a certain subset of $A$ of positive measure, is the orthogonal projection of ${\mathbb R}^{n+m}$ onto the first $n$-coordinates. This may be seen as a qualitative version of a simlar result of Azzam and Schul [@AzzSch]. The main tool in our proof is the metric change of variables introduced in [@HajMal].' address: - 'P. Haj[ł]{}asz: Department of Mathematics, University of Pittsburgh, 301 Thackeray Hall, Pittsburgh, PA 15260, USA, [[email protected]]{}' - 'S. Zimmerman: Department of Mathematics, University of Connecticut, 341 Mansfield Road U1009, Storrs, CT 06269, USA, [[email protected]]{}' author: - 'Piotr Haj[ł]{}asz and Scott Zimmerman' title: An implicit function theorem for Lipschitz mappings into metric spaces --- *In memoriam: William P. Ziemer (1934-2017)* Introduction ============ The classical implicit function theorem (IFT) ensures that the map is structurally very nice near points where the derivative of the map has a certain rank. In this paper, we present a version of the IFT for Lipschitz mappings $f:\mathbb{R}^{n+m} \supset A \to X$ into arbitrary metric spaces. It turns out that in the case of mappings into metric spaces, the upper density defined below will play a role of the Jacobian of $f$. For a measurable set $A \subset {\mathbb R}^{k}$, and $x\in A$, we define the *lower* and *upper* $n$-*densities* of a mapping $f:A \to X$ as $$\Theta^{*n}(f,x) := \limsup_{r \to 0}\frac{\H_{\infty}^n(f(B(x,r) \cap A))}{\omega_n r^n}, \quad \Theta_*^n(f,x) := \liminf_{r \to 0}\frac{\H_{\infty}^n(f(B(x,r) \cap A))}{\omega_n r^n}.$$ These are simply the upper and the lower $n$-densities of the pull-back of $\mathcal{H}_{\infty}^n$ by $f$ on $A$. Here $\omega_n$ is the volume of the unit ball in ${\mathbb R}^n$ and the $\H_\infty^n$ is the [*Hausdorff content*]{} defined for subsets of $X$ by $$\H^n_\infty(E)=\inf \frac{\omega_n}{2^n}\sum_{i=1}^\infty (\diam A_i)^n,$$ where the infimum is taken over all coverings of $E$, i.e. $E\subset\bigcup_{i=1}^\infty A_i$. Note that the Hausdorff content of [*any*]{} bounded set is finite, and, for an $L$-Lipschitz map $f:A\to X$, $\Theta^{*n}(f,x)\leq L^n$ for [*all*]{} $x\in A$. The reader may want to compare these definitions with the definition (and properties) of the upper and lower densities of measures in [@ambrosiok; @mattila; @simon]. The following observation will be useful throughout the paper: $$\label{1410} \Theta^{*n}(f,x)=0 \quad \text{if and only if} \quad \lim_{d\to 0}\frac{\H_{\infty}^n(f(Q(x,d) \cap A))}{\omega_n d^n} = 0$$ where $Q(x,d)$ is the cube centered at $x$ with side length $d$. (Here and in what follows, a cube has edges parallel to the coordinate axes.) The main result of the paper is as follows: \[HardSard2\] Fix a metric space $X$, a set $A \subset {\mathbb R}^{n+m}$ with positive Lebesgue measure, and a Lipschitz mapping $f:A \to X$. Suppose $\Theta^{*n}(f,x) > 0$ on a subset of $A$ with positive Lebesgue measure. Then - $\H^n(f(A))>0$; - There is a set $K \subset A$ with positive Lebesgue measure, a bi-Lipschitz $C^1$-diffeomorphism $G:U\to G(U)\subset{\mathbb R}^{n+m}$ defined on an open set $U \supset K$ and a $\sqrt{n}$-Lipschitz map $\pi: X\to{\mathbb R}^n$ such that $$\pi\circ f\circ G^{-1}(x_1,\ldots,x_n,y_{1}\ldots,y_m)=(x_1,\ldots,x_n) \quad\text{for all $(x,y)\in G(K)$}$$ is a projection on the first $n$ coordinates when restricted to the set $G(K)$. Moreover the mapping $F=f\circ G^{-1}$ defined on $G(K)$ satisfies - $F^{-1}(F(x,y)) \cap G(K) \subset \{x\} \times {\mathbb R}^m$ for any $(x,y) \in G(K)$; - $F|_{({\mathbb R}^n \times \{y\}) \cap G(K)}$ is bi-Lipschitz for any $y \in {\mathbb R}^m$. It follows from the proof that we can exhaust the set of points where $\Theta^{*n}(f,x)>0$ by sets $K$ as in (B) up to a set of $\H^{n+m}$ measure zero. (See the application of Lemma \[federer\] in the proof of Lemma \[3.1\] as well as Remark \[bluerem\].) \[R145\] The map $\pi:X\to{\mathbb R}^n$ is in fact $1$-Lipschitz as a map from $X$ to $({\mathbb R}^n,\ell^\infty_n)$ where the norm $\ell^\infty_n$ is defined by $\Vert (x_1,\ldots,x_n)\Vert_\infty=\max_i|x_i|$. This will follow from our proof. Statement (C) means that the preimage under $F$ of any point in $F(G(K))=f(K)$ is contained in an $m$-dimensional subspace of ${\mathbb R}^{n+m}$ orthogonal to ${\mathbb R}^n$. For related results about the structure of preimages $f^{-1}(z)$ of Lipschitz maps, see [@karmanova Theorem 1.2], [@reichel Theorem 4.16]. In fact, we will prove a quantitative lower bound in (D): $$\label{404} \Vert x_1-x_2\Vert_\infty\leq d(F(x_1,y),F(x_2,y))$$ for any $y\in{\mathbb R}^m$ and all $(x_1,y),(x_2,y)\in ({\mathbb R}^n \times \{y\}) \cap G(K)$. The classical implicit function theorem is stated using a condition about the rank of the derivative of $f$, and the condition $\Theta^{*n}(f,x)>0$ is a related one. Indeed, in the case $X = \mathbb{R}^n$, we will see in Proposition \[euclem\] that the Jacobian of $f$ defined by $|J^nf|(x) = \sqrt{\det (Df)(Df)^T(x)}$ satisfies $\Theta^{*n}(f,x) = |J^n f|(x)$ almost everywhere. See also Lemma \[elllem\] for the case of mappings $f:A\to\ell^\infty$. In the theorem we cannot replace the density condition $\Theta^{*n}(f,x)>0$ by the simpler measure condition $\H^n(f(A))>0$. Indeed, even in the Euclidean case, Kaufmann [@kaufman] constructed a surjective $C^1$ mapping $f:{\mathbb R}^{n+1}\to{\mathbb R}^n$, $n\geq 2$, satisfying $\rank Df\leq 1$ [*everywhere*]{}. For such a map, condition (B) cannot be satisfied since it would imply that $\rank Df\geq n$ on $K$. Recall that a set $E\subset{\mathbb R}^{n+m}$ is [*countably $\H^m$-rectifiable*]{} if there are Lipschitz mappings $f_i:{\mathbb R}^m\supset E_i\to{\mathbb R}^{n+m}$, $i\in\mathbb{N}$, such that $\H^m(E\setminus\bigcup_{i=1}^\infty f(E_i))=0$. As a corollary of Theorem \[HardSard2\] we obtain \[HSC\] Fix a metric space $X$, a set $A \subset {\mathbb R}^{n+m}$ with positive Lebesgue measure, and a Lipschitz mapping $f:A \to X$. Suppose $\Theta^{*n}(f,\cdot) > 0$ almost everywhere in $A$. Then $f^{-1}(x)$ is countably $\H^m$-rectifiable for $\H^n$-almost all $x\in X$. See Section \[HSCProof\] for the proof. For related results see [@karmanova Theorem 1.2], [@reichel Theorem 4.16]. Our result may be seen as a qualitative version of a theorem proven in 2012 by Azzam and Schul [@AzzSch]. In that paper, the authors proved the following quantitative version of the IFT for Lipschitz mappings into metric spaces: \[AS\] Fix a metric space $X$ and a $1$-Lipschitz mapping $f:{\mathbb R}^{n+m} \to X$. Suppose $0 < \H^n(f([0,1]^{n+m})) \leq 1$ and $$\label{HCap} 0< \delta \leq \mathcal{H}_{\infty}^{n,m}(f,[0,1]^{n+m})$$ for some $\delta > 0$. Then there are constants $\Lambda=\Lambda(n,m,\delta)>1$ and $\eta=\eta(n,m,\delta)>0$, a set $K \subset[0,1]^{n+m}$ with $$\label{KBound} \mathcal{H}^{n+m}(K) \geq \eta,$$ and a $\Lambda$-bi-Lipschitz homeomorphism $G:{\mathbb R}^{n+m} \to {\mathbb R}^{n+m}$ such that $F = f \circ G^{-1}$ satisfies $$F^{-1}(F(x,y)) \cap G(K) \subset \{x\} \times {\mathbb R}^m \quad \text{for any } (x,y)\in G(K) \subset \mathbb{R}^{n+m}$$ and $F|_{ ({\mathbb R}^n \times \{y\}) \cap G(K) }$ is $\Lambda$-bi-Lipschitz for any $y \in {\mathbb R}^m$. The authors of [@AzzSch] call $\H_{\infty}^{n,m}$ the $(n,m)$-[*Hausdorff content of*]{} $f$. It is defined for a Lipschitz map $f:Q\to X$ from a cube $Q\subset{\mathbb R}^{n+m}$ to a metric space by $$\label{AS12} \H_\infty^{n,m}(f,Q)= \inf \sum_{j=1}^\infty \H^n_\infty(f(Q_j))d_j^m,$$ where the infimum is taken over all families of open pairwise disjoint cubes $Q_j\subset Q$ of side length $d_j$ that cover $Q$ up to a set of measure zero. Note that Theorems \[HardSard2\] and \[AS\] provide the same qualitative structure on the vertical and horizontal slices of the preimage of $F$. However, Theorem \[AS\] is a [*quantitative*]{} version of the metric IFT in the sense that it provides the lower bound which depends only on the dimensions $m$, $n$ and $\delta$ from . Moreover, the mapping $G$ is a *globally* defined $C$-bi-Lipschitz homomorphism where $C$ depends only on $m$, $n$, and $\delta$. Our result (Theorem \[HardSard2\]) does not contain these quantitative conclusions. This is because the assumption in Theorem \[AS\] is much stronger than the assumption that $\Theta^{*n}(f,x) > 0$ on a set of positive measure. Indeed, Proposition \[elllem2\] shows that the positivity of $\Theta^{*n}(f,x)$ follows from the assumption . In fact, for any $\eps>0$, one may construct a mapping $f:[0,1]^2 \to \mathbb{R}$ with $\Theta^{*1}(f,x) = 1$ almost everywhere so that the set $K\subset{\mathbb R}^2$ satisfying the conclusion of Theorem \[HardSard2\] (for a global bi-Lipschitz homeomorphism $G$) must satisfy $\mathcal{H}^{2}(K) < \eps$ (and hence cannot hold). See Proposition \[fold\] for the construction and a detailed statement. On the other hand, while the assumptions of Theorem \[HardSard2\] are much weaker than those of Theorem \[AS\], some of the conclusions seem stronger: (1) As we already pointed out, the condition about positivity of $\Theta^{*n}(f,x)$ is much weaker than condition ; (2) Azzam and Schul assume that $0<\H^n(f([0,1]^{n+m}))\leq 1$ while we do not assume anything about the Hausdorff measure of the image. In fact, we prove the lower bound $\H^n(f(A))>0$ in (A) and finiteness of the measure of the image plays no role in our theorem; (3) Our mapping $G$ is a bi-Lipschitz $C^1$ diffeomorphism while their mapping $G$ is only a bi-Lipschitz map. However, their map is defined globally and ours is defined locally only; (4) While parts (C) and (D) are the same as the corresponding statements in Theorem \[AS\], part (B) seems stronger than that. (C) and (D) easily follow from (B), but we do not know if (B) can be concluded from Theorem \[AS\]; (5) We obtain the quantitative lower bound estimate ; (6) At last, but not least, our proof is [*much*]{} simpler than that in [@AzzSch]. The classical IFT states that a $C^1$ mapping has a nice structure near a point where the derivative has rank of a certain order. However, the classical IFT does not provide [*any*]{} estimate for the size of the set where the map is nice. Our result has the same feature as the classical one: we do not obtain any estimate for the size of the set $K$ except that it has a positive measure. The main tool in the proof of Theorem \[HardSard2\] will be the metric change of variables introduced in [@HajMal]. This change of variables has been used to prove versions of Sard’s theorem for Lipschitz mappings and BLD mappings into metric spaces [@HajMal; @HajMalZim]. This paper is organized as follows. In Section \[prelim\] we collect basic definitions and lemmata needed in the proofs of Theorem \[HardSard2\] and Corollary \[HSC\]. In Sections \[SecProof\] and \[HSCProof\] we prove Theorem \[HardSard2\] and Corollary \[HSC\] respectively. Finally, in Section \[SecComp\], we prove some other results that help us compare Theorems \[HardSard2\] and \[AS\], we prove that the condition $\H^{n,m}_\infty(f,Q)>0$ implies positivity of $\Theta^{*n}(f,x)$ on a set of positive measure (Proposition \[elllem2\]), we prove that, if $f:{\mathbb R}^{n+m}\supset A\to{\mathbb R}^n$ is Lipschitz, then $\Theta^{*n}(f,x)=\Theta_*^n(f,x)=|J^nf|(x)$ almost everywhere in $A$ (Proposition \[euclem\]), and we construct an example showing that we cannot obtain any lower bound for $\H^{n+m}(K)$ (Proposition \[fold\]). Notation used in the paper is fairly standard. The $n$-dimensional Hausdorff measure will be denoted by $\H^n$. Note that in ${\mathbb R}^n$, $\H^n$ equals the Lebesgue measure and we will use Hausdorff measure notation in place of the Lebesgue measure. Occasionally we will write $|E|$ to denote the Lebesgue measure of $E$. Notation $\H^n_\infty$ will stand for the Hausdorff content defined above. The constant $\omega_n$ denotes the measure of the unit ball in ${\mathbb R}^n$. The Banach space of bounded [*real valued*]{} sequences will be denoted by $\ell^\infty$. Balls in metric spaces are denoted by $B(x,r)$, and $Q(x,d)$ denotes the Euclidean cube centered at $x$ with side length $d$. All cubes are assumed to have edges parallel to the coordinate axes. Occasionally a $k$-dimensional ball in a Euclidean space will be denoted by $B^k(x,r)$. By a $\Lambda$-bi-Lipschitz homeomorphism $f:(X,d)\to (Y,\rho)$ we mean a homeomorphism satisfying $\Lambda^{-1}d(x,y)\leq \rho(f(x),f(y))\leq \Lambda d(x,y)$. The tangent space to ${\mathbb R}^k$ at $x\in{\mathbb R}^k$ will be denoted by $T_x{\mathbb R}^k$. By $C$ we will denote a general constant whose value may change in a single string of estimates. Writing $C=C(n,m)$, for example, indicates that the constant $C$ depends on $n$ and $m$ only. [**Acknowledgements.**]{} The authors would like to thank the referee for valuable comments that led to an improvement of the paper. Preliminaries {#prelim} ============= In this section we collect basic definitions and results that will be used later on. If $k>n$, then the $\H^n_\infty$ content of subsets of ${\mathbb R}^k$ is very different from their Hausdorff measure. For example $\H^n_\infty(E)<\infty$ for any bounded set $E\subset{\mathbb R}^k$, but $\H^n(B)=\infty$ for any $k$-ball $B\subset{\mathbb R}^k$. However, we have (see [@simon Theorem 2.6]) \[h=h\] $\H^n_\infty(E)=\H^n(E)$ for all sets $E\subset{\mathbb R}^n$. \[kura\] Every separable metric space admits an isometric embedding into $\ell^\infty$. Indeed, given $x_0\in X$ and a dense set $\{x_i\}_{i=1}^\infty$ in a separable metric space $(X,d)$, $$X\ni x \mapsto \kappa(x)=(d(x,x_i)-d(x_i,x_0))_{i=1}^\infty\in\ell^\infty$$ is an isometric embedding. This is the well known Kuratowski embedding for metric spaces. For a proof of the following elementary result, see [@HKST Corollary 4.1.7]. \[2.7\] Let $Y$ be a metric space, let $E\subset Y$ and let $f:E\to\ell^\infty$ be an $L$-Lipschitz mapping. Then there is an $L$-Lipschitz mapping $F:Y\to\ell^\infty$ such that $F|_E=f$. The idea of the proof is very simple. Each component $f_i$ of $f$ is $L$-Lipschitz and we define $F$ by extending each of the components of $f$ using the formula from the McShane extension. Then it is easy to verify that the resulting map is $L$-Lipschitz and it takes values in $\ell^\infty$. Fix an integer $k \geq 1$, and suppose $A \subset \mathbb{R}^k$ is measurable. Recall that a function $f:A\to{\mathbb R}$ is [*approximately differentiable*]{} at $x\in A$ if there is a measurable set $A_x\subset A$ and a linear map $L:{\mathbb R}^n\to{\mathbb R}$ such that $x$ is a density point of $A_x$ and $$\lim_{A_x\ni y\to x}\frac{|f(y)-f(x)-L(y-x)|}{|y-x|}=0.$$ $L$ is called the [*approximate derivative*]{} of $f$ at $x$ and is denoted by $ap\, Df(x)$. Recall also that $x\in E\subset{\mathbb R}^k$ is a [*density point*]{} of $E$ if $\H^k(E\cap B(x,d))/(\omega_kd)^k\to 1$ as $d\to 0$. If in addition $f:A\to{\mathbb R}$ is Lipschitz, then the approximate derivative $ap\, Df(x)$ exists for almost every $x \in A$. This follows from the McShane extension and Rademacher’s theorem. Indeed, if $F:{\mathbb R}^k\to{\mathbb R}$ is a Lipschitz extension of $f$, then $\ap Df(x)$ exists at [*all*]{} points of the set $$E=\{x\in A:\, \text{$x$ is a density point of $A$ and $F$ is differentiable at $x$}\}.$$ Moreover $\ap Df(x)=DF(x)$ at points of the set $E$. For a Lipschitz map $f=(f_1,f_2,\dots):A \to \ell^{\infty}$, we define the [*component-wise approximate derivative*]{} by $$ap \, Df(x) := \left \lceil \begin{array}{c} ap \, Df_1(x) \\ ap \, Df_2(x) \\ \vdots \end{array} \right \rceil$$ Since each component $f_i$ is Lipschitz, $ap\, Df$ exists almost everywhere in $A$. It is easy to see that the row and column ranks of this $\infty \times k$ matrix are equal, and $\rank (ap \, Df(x))$ equals the dimension of the image of $ap \, Df(x)$ in $\ell^{\infty}$. It follows in particular that $\rank (ap \, Df(x)) \leq k$. Let $V$ be a linear space of all real sequences. In particular, $\ell^\infty\subset V$, but we do not equip $V$ with any norm or topology. If all components of a mapping $g=(g_1,g_2,\ldots):{\mathbb R}^k\to V$ are differentiable at a point $x$, we will say that $g$ is [*component-wise differentiable*]{} at $x$ and write $$Dg(x) := \left \lceil \begin{array}{c} Dg_1(x) \\ Dg_2(x) \\ \vdots \end{array} \right \rceil$$ We will also need the following result of Federer (for a proof, see [@MalyZ Theorem 1.69], [@simon Theorem 5.3], [@Whitney]). \[federer\] If $A\subset{\mathbb R}^k$ is measurable and $f:A\to{\mathbb R}$ is Lipschitz, then for any $\eps>0$ there is a function $g\in C^1({\mathbb R}^k)$ such that $$\H^k(\{x\in A:\, f(x)\neq g(x)\})<\eps.$$ It is easy to see that if $x_0$ is a density point of the set $$\label{1456} \{x\in A:\, f(x)=g(x)\},$$ then $\ap Df(x_0)$ exists and $\ap Df(x_0)=Dg(x_0)$. In particular $Dg=\ap Df$ almost everywhere in the set . The next lemma was proven in [@HajMal Proposition 2.3]. \[5790\] Let $D \subset {\mathbb R}^{k}$ be a cube or ball, and let $f:D \to \ell^{\infty}$ be $L$-Lipschitz. Then $$\diam(f(D)) \leq C(k)L\mathcal{H}^{k}(D \setminus A)^{1/k},$$ where $A = \{ x \in D \, : \, Df(x) = 0\}$ and $Df$ is the component-wise derivative of $f$. Finally, in the proof of Corollary \[HSC\] we will need \[coarea\] If $f:X\to Y$ is a Lipschitz mapping between metric spaces and $A\subset X$, $0\leq m\leq n$, then $$\int_Y^*\H^{n-m}(f^{-1}(y)\cap A)\, d\H^m(y)\leq (\lip f)^m \frac{\omega_{n-m}\omega_m}{\omega_n}\, \H^n(A).$$ Here $\int^*$ stands for the upper integral and $\lip f$ is a Lipschitz constant of $f$. Federer [@federer 2.10.25] proved this result under additional assumptions. The general case was obtained by Davies [@davies]. A detailed proof is given in [@reichel Theorem 2.4]. \[zero\] If $f:X\to Y$ is Lipschitz mapping between metric spaces and $A\subset X$, $\H^n(A)=0$, $0\leq m\leq n$, then $\H^{n-m}(f^{-1}(y)\cap A)=0$ for $\H^m$ almost all $y\in Y$. Proof of Theorem \[HardSard2\] {#SecProof} ============================== The proof is based on techniques developed in [@HajMal] (see also [@HajMalZim]). Consider a Lipschitz map $f:A\to\ell^\infty$ defined on a measurable set $A\subset{\mathbb R}^k$. Our first lemma shows that, if the rank of $\ap Df(x)$ is at least $j$ on a set of positive measure, then, up to local diffeomorphisms, $f$ fixes the first $j$ coordinates on some non-null subset. \[3.1\] Suppose $f:A\to\ell^\infty$ is a Lipschitz map defined on a measurable set $A\subset{\mathbb R}^k$. If $\rank (\ap Df(x)) \geq j$ on a subset of $A$ of positive $\H^k$-measure, then there is an open set $U\subset{\mathbb R}^k$, a set $K\subset A\cap U$ of positive $\H^k$-measure, a bi-Lipschitz $C^1$-diffeomorphism $G:U\to G(U)\subset{\mathbb R}^k$, and a permutation of a finite number of coordinates $\Psi:\ell^\infty\to\ell^\infty$ (which is an isometry of $\ell^\infty$) such that $$\label{2020} (\Psi\circ f\circ G^{-1})_i(x)=x_i \quad \text{for $i=1,2,\ldots,j$ and $x\in G(K)$.}$$ That is for $x\in G(K)$ we have $$(\Psi\circ f\circ G^{-1})(x_1,\ldots,x_n)= (x_1,\ldots,x_j,(\Psi\circ f\circ G^{-1})_{j+1}(x),(\Psi\circ f\circ G^{-1})_{j+2}(x),\ldots).$$ By restricting $f$ to the set where $\rank (\ap Df(x)) \geq j$, we may assume that $\rank (\ap Df(x)) \geq j$ a.e. in $A$. Since $f=(f_1,f_2,\dots):A \to \ell^{\infty}$ is Lipschitz, each component $f_i$ of $f$ is Lipschitz. Therefore, by applying Lemma \[federer\] component-wise, we may choose $F \subset A$ with $\H^{k}(F)>0$ and a mapping $g=(g_1,g_2,\dots):\mathbb{R}^{k} \to V$ with $g_j \in C^1({\mathbb R}^{k})$ for every $j \in \mathbb{N}$ and such that $g=f$, $Dg = \ap Df$, and $\rank Dg = \rank \ap Df \geq j$ on $F$. Here, as before, $V$ is the vector space consisting of all real valued sequences. (This is needed since sequences $(g_i(x))_{i=1}^\infty$ are not necessarily bounded.) Fix $x_0 \in F$. Under the above assumptions, there is a bi-Lipschitz $C^1$-diffeomorphism $G:U \to G(U)\subset{\mathbb R}^k$ defined on a neighborhood $U$ of $x_0$ and a permutation $\Psi:V\to V$ of a finite number of coordinates so that $$(\Psi\circ g\circ G^{-1})_i(x)=x_i \quad\text{for $i=1,2,\ldots,j$ and $x\in G(U)$}.$$ That is, $\Psi\circ g\circ G^{-1}$ fixes the first $j$ coordinates on $G(U)$. Since $\rank Dg(x_0)\geq j$, a certain $j\times j$ minor of $Dg(x_0)$ has rank $j$. By precomposing $g$ with a permutation $\tilde{\Psi}$ of $j$ variables in ${\mathbb R}^k$ and postcomposing it with a permutation $\Psi$ of $j$ variables in $V$, we have that $$\tilde{g}=(\tilde{g}_1,\tilde{g}_2,\ldots)=\Psi\circ g\circ\tilde{\Psi}$$ satisfies $$\label{ti382} \det\left[\frac{\partial\tilde{g}_m}{\partial x_\ell}(\tilde{\Psi}^{-1}(x_0))\right]_{1\leq m,\ell\leq j}\neq 0.$$ Let $$H(x)=(\tilde{g}_1(x),\ldots,\tilde{g}_j(x), x_{j+1},\ldots,x_k).$$ It follows from that $\det DH(\tilde{\Psi}^{-1}(x_0))\neq 0$, so $H$ is a diffeomorphism in a neighborhood $\tilde{U}$ of $\tilde{\Psi}^{-1}(x_0)$. Replacing $\tilde{U}$ by a smaller open set, it follows that $H$ is bi-Lipschitz. Now observe that $$(\tilde{g}\circ H^{-1})_i(x)=x_i \quad \text{for $i=1,2,\ldots,j$ and $x\in H(\tilde{U})$}.$$ Therefore, if we write $G=H\circ\tilde{\Psi}^{-1}$, then $\Psi\circ g\circ G^{-1}=\tilde{g}\circ H^{-1}$ satisfies the claim of the lemma on the open set $U=\tilde{\Psi}(\tilde{U})$, $U$ is a neighborhood of $x_0$, and $G(U)=H(\tilde{U})$. Now if $x_0$ is any density point of $F$, then the set $K=F\cap U$ has positive measure. Since $f=g$ on $K$, follows because the permutation of coordinates $\Psi:V\to V$ maps $\ell^\infty\subset V$ to $\ell^\infty\subset V$ in an isometric way. This completes the proof of Lemma \[3.1\]. \[elllem\] Fix a measurable set $A \subset {\mathbb R}^k$ and $n \leq k$. Suppose $f:A \to \ell^{\infty}$ is a Lipschitz map. If $\Theta^{*n}(f,x)>0$ on a subset of $A$ of positive measure, then $\rank (\ap Df(x)) \geq n$ on a set of positive measure. Note that the above lemmata involve Lipschitz mappings into $\ell^{\infty}$. As we will see later, this will be sufficient in the setting of any metric space since the separable metric space $f(A)$ may be embedded isometrically into $\ell^{\infty}$ via the Kuratowski embedding. \[bluerem\] In the following proof, we will see in particular that, for $j\in \{0,1,2,\ldots,n-1\}$, the set of points $x \in A$ where $\Theta^{*n}(f,x)>0$, $\ap Df(x)$ exists, and $\rank (\ap Df(x))=j$ must have measure zero. Suppose to the contrary that $\Theta^{*n}(f,x)>0$ on a set of positive measure and $\rank (\ap Df(x))<n$ almost everywhere in $A$. Then there is $j\in \{0,1,2,\ldots,n-1\}$ and a set $F\subset A$ with $\H^k(F)>0$ such that $\Theta^{*n}(f,x)>0$ for all $x\in F$, $\ap Df(x)$ exists and $\rank (\ap Df(x))=j$ for all $x\in F$. According to Lemma \[3.1\], there is a permutation $\Psi:\ell^\infty\to\ell^\infty$ of a finite number of variables, an open set $U\subset{\mathbb R}^k$, a set $K\subset F\cap U$ with $\H^k(K)>0$ and a bi-Lipschitz $C^1$-diffeomorphism $G:U\to G(U)\subset{\mathbb R}^k$ such that $\hat{f}=\Psi\circ f\circ G^{-1}$ defined on $\hat{A}=G(A \cap U)$ satisfies $$\label{Natalia} \hat{f}_i(x)=x_i \quad \text{for $i=1,2,\ldots,j$ and $x\in \hat{K}$}$$ where $\hat{K} = G(K)$. Note that $\ap D\hat{f}(x)$ exists and $\rank (\ap D\hat{f})(x)=j$ for all $x\in \hat{K}$, because composition with a diffeomorphism and a permutation $\Psi$ preserve approximate differentiability and the rank of the approximate derivative. Assume that $x_0$ is a density point of $K$. Since $\Theta^{*n}(f,x)>0$ for all $x\in K$, in order to arrive to a contradiction, it suffices to show that $$\Theta^{*n}(f,x_0)=0.$$ Note that $y_0=G(x_0)$ is a density point of $\hat{K}=G(K)$ because diffeomorphisms map density points to density points. The next lemma shows that it suffices to prove that $$\label{3.3} \Theta^{*n}(\hat{f},y_0)= \limsup_{d\to 0} \frac{\H^n_\infty(\hat{f}(B(y_0,d)\cap \hat{A}))}{\omega_n d^n}=0.$$ If $\Theta^{*n}(\hat{f},y_0)=0$, then $\Theta^{*n}(f,x_0)=0$. Let $d>0$ be so small that $B(y_0,d)\subset G(U)$. Since the diffeomorphism $G^{-1}$ is bi-Lipschitz on $G(U)$, there is a constant $\Lambda>0$ such that $$B\left(x_0,\frac{d}{\Lambda}\right)\subset G^{-1}(B(y_0,d)).$$ Since the permutation of coordinates $\Psi:\ell^\infty\to\ell^\infty$ is an isometry, it follows that $\H^n_\infty(\hat{f}(E))=\H^n_\infty(f(G^{-1}(E)))$ for any set $E$ in the domain of $\hat{f}$. Therefore $$\H^n_\infty(\hat{f}(B(y_0,d)\cap \hat{A}))= \H^n_\infty(f(G^{-1}(B(y_0,d)\cap \hat{A}))\geq \H^n_\infty(f(B(x_0,d/\Lambda)\cap A)),$$ so $$\begin{split} \Theta^{*n}(\hat{f},y_0) &= \limsup_{d\to 0} \frac{\H^n_\infty(\hat{f}(B(y_0,d)\cap \hat{A}))}{\omega_n d^n}\\ &\geq \Lambda^{-n}\limsup_{d\to 0}\frac{\H^n_\infty(f(B(x_0,d/\Lambda)\cap A))}{\omega_n (d/\Lambda)^n} = \Lambda^{-n}\Theta^{*n}(f,x_0) \end{split}$$ and the lemma follows. To conclude the proof of , we will apply the following lemma. \[HaMaLemma\] Assume $d > 0$ is such that $Q(y_0,d) \subset G(U)$ and $$\H^{k}(Q(y_0,d) \setminus \hat{K}) < \left( \frac{d}{M} \right)^{k}$$ for some positive integer $M$. Then $\hat{f}(Q(y_0,d) \cap \hat{A})$ can be covered by $M^j$ balls of radius $CLdM^{-1}$ for some constant $C = C(k,n) > 0$, where $L$ is the Lipschitz constant of $\hat{f}$. In particular, we have $$\H_{\infty}^n(\hat{f}(Q(y_0,d) \cap \hat{A})) \leq \omega_n(CLd)^n M^{j-n}.$$ Before proving this lemma, we will see how it can be used to prove . Let $\eps > 0$. Fix a positive integer $M$ such that $(CL)^n M^{j-n} < \eps$. (This is possible since $j-n < 0$.) Since $y_0$ is a density point of $\hat{K}$, there is $\delta>0$ such that for $0<d<\delta$, $Q(y_0,d) \subset G(U)$ satisfies $$\H^{k}(Q(y_0,d) \setminus \hat{K}) < \frac{\H^{k}(Q(y_0,d))}{M^k} = \left( \frac{d}{M} \right)^{k}.$$ Hence, by Lemma \[HaMaLemma\], we have $$\frac{\H_{\infty}^n(\hat{f}(Q(y_0,d) \cap \hat{A}))}{\omega_n d^n} \leq (CL)^n M^{j-n} < \eps \quad \text{for $0<d<\delta$}$$ which, along with , implies that $\Theta^{*n}(\hat{f},y_0)=0$. That completes the proof of once Lemma \[HaMaLemma\] has been verified. The proof of Lemma \[HaMaLemma\] is nearly identical to the proof of [@HajMal Lemma 2.7], but we will include it here for completeness. Assume that a positive integer $M>0$ and $d > 0$ satisfy $Q(y_0,d) \subset G(U)$ and $$\H^{k}(Q(y_0,d) \setminus \hat{K}) < \left( \frac{d}{M} \right)^{k}.$$ Since the result is translation invariant, we may assume without loss of generality that $$Q(y_0,d)=Q = [0,d]^j \times [0,d]^{k-j}.$$ According to Lemma \[2.7\], the $L$-Lipschitz mapping $\hat{f}:Q \cap \hat{A} \to \ell^{\infty}$ admits an $L$-Lipschitz extension $\tilde{f}:Q \to \ell^{\infty}$. According to Rademacher’s theorem, $\tilde{f}$ is component-wise differentiable for almost all points in $Q$. Divide $[0,d]^j$ into $M^j$ cubes $\{Q_{\nu}\}_{\nu = 1}^{M^j}$ with pairwise disjoint interiors each of edge length $d/M$. It suffices to show that each set $$\hat{f}((Q_{\nu} \times [0,d]^{k-j}) \cap \hat{A}) \subset \tilde{f}(Q_{\nu} \times [0,d]^{k-j})$$ is contained in an $\ell^{\infty}$-ball of radius $CLdM^{-1}$ for some constant $C=C(k,n)>0$. By our assumptions, for each $\nu$ we have $$\mathcal{H}^k((Q_{\nu} \times [0,d]^{k-j}) \setminus \hat{K}) \leq \mathcal{H}^k(Q \setminus \hat{K}) < \left( \frac{d}{M} \right)^{k}.$$ Hence $$\mathcal{H}^k((Q_{\nu} \times [0,d]^{k-j}) \cap \hat{K}) > \left( M^{-j} - M^{-k} \right)d^k.$$ According to Fubini’s Theorem, we may therefore choose some $\rho \in Q_{\nu}$ such that $$\mathcal{H}^{k-j}((\{\rho\} \times [0,d]^{k-j}) \cap \hat{K}) > \left( 1 - M^{j-k} \right)d^{k-j}$$ and $\tilde{f}$ is component-wise differentiable at almost all points of $\{\rho\} \times [0,d]^{k-j}$. Hence $$\label{2.1} \mathcal{H}^{k-j}((\{\rho\} \times [0,d]^{k-j}) \setminus \hat{K}) < \left( \frac{d}{M} \right)^{k-j}.$$ According to , $\hat{f}$ fixes the first $j$ coordinates in $\hat{K}$. Since $\hat{f}=\tilde{f}$ in $\hat{K}$ and $\rank (\ap D\hat{f}(x))=j$ everywhere in $\hat{K}$, it follows that $\tilde{f}_i(x)=x_i$ for $i=1,2,\ldots,j$ and $x\in \hat{K}$ and $\rank D\tilde{f}(x)=j$ almost everywhere in $\hat{K}$. Therefore, the component-wise derivative of $\tilde{f}$ along $\{\rho\}\times [0,d]^{k-j}$ vanishes at almost all points in $(\{\rho\}\times [0,d]^{k-j})\cap \hat{K}$. That is $$D\big(\tilde{f}\big|_{\{\rho\}\times [0,d]^{k-j}}\big)=0 \quad \text{a.e. in $(\{\rho\}\times [0,d]^{k-j})\cap \hat{K}$.}$$ Therefore Lemma \[5790\] applied to $\tilde{f}:\{\rho\} \times [0,d]^{k-j}\to\ell^\infty$ (with $k$ replaced by $k-j$) together with yield $$\diam(\tilde{f}(\{\rho\} \times [0,d]^{k-j})) \leq CL\mathcal{H}^{k-j}((\{\rho\} \times [0,d]^{k-j}) \setminus \hat{K})^{1/(k-j)} \leq CLdM^{-1}.$$ Since the distance from any point in $Q_{\nu} \times [0,d]^{k-j}$ to the set $\{\rho\} \times [0,d]^{k-j}$ is at most $\diam(Q_\nu ) = \sqrt{j} dM^{-1}$ and $\tilde{f}$ is $L$-Lipschitz, this implies that $$\diam(\tilde{f}(Q_\nu \times [0,d]^{k-j})) \leq CLdM^{-1}$$ (for a larger value of $C$). This proves Lemma \[HaMaLemma\]. This also completes the proof of and hence that of Lemma \[elllem\]. We now can finish the proof of Theorem \[HardSard2\]. Since $f(A)\subset X$ is a separable metric space, there is an isometric embedding $\kappa:f(A)\to\ell^\infty$ (see Lemma \[kura\]). The mapping $\kappa$ is $1$-Lipschitz. According to Lemma \[2.7\], the map $\kappa$ admits a $1$-Lipschitz extension $\K:X\to\ell^\infty$. Then $\bar{f}=\K\circ f=\kappa\circ f:A\to\ell^\infty$ is Lipchitz and $\Theta^{*n}(\bar{f},x) > 0$ on a subset of $A$ with positive measure (composition with an isometric map does not change the upper density). It follows from Lemma \[elllem\] (with $k=n+m$) that $\rank \ap D\bar{f} \geq n$ on a set of positive measure. Therefore, according to Lemma \[3.1\], there is an open set $U\subset{\mathbb R}^{n+m}$, a subset $K\subset A\cap U$ with $\H^{n+m}(K)>0$, a bi-Lipschitz $C^1$-diffeomorphism $G:U\to G(U)\subset{\mathbb R}^{n+m}$ and a permutation of finitely many coordinates $\Psi:\ell^\infty\to\ell^\infty$ such that $$\label{Michal} (\Psi\circ \bar{f}\circ G^{-1})_i(x)=x_i \quad\text{for $i=1,2,\ldots,n$ and $x\in G(K)$.}$$ Let $$P:\ell^\infty\to{\mathbb R}^n, \quad P(x_1,x_2,\ldots)=(x_1,x_2,\ldots,x_n)$$ be the projection onto the first $n$ coordinates. Then $P$ is $1$-Lipschitz as a mapping to ${\mathbb R}^n$ equipped with the $\ell^\infty_n$ norm, $\Vert (x_1,\ldots,x_n)\Vert_\infty=\max_i|x_i|$ and $\sqrt{n}$-Lipschitz as a mapping to ${\mathbb R}^n$ with the Euclidean metric. Therefore, it follows that the mapping $$\pi:X\to{\mathbb R}^n, \quad \pi=P\circ\Psi\circ\K$$ is $1$-Lipschitz as a mapping to ${\mathbb R}^n$ equipped with the norm $\ell^\infty_n$ and $\sqrt{n}$-Lipschitz as a mapping to ${\mathbb R}^n$ with the Euclidean metric (see Remark \[R145\]). If we swith to notation $$(x,y)=(x_1,\ldots,x_n,y_1,\ldots,y_m):= (x_1,\ldots,x_n,x_{n+1},\ldots,x_{n+m}),$$ then clearly, means that $(\pi\circ f\circ G^{-1})(x,y)=x$ for $(x,y)\in G(K)$ which completes the proof of the statement (B). To prove (A), suppose to the contrary that $\H^n(f(A))=0$. Then $\H^n(f(K))=0$ and hence $$\label{Joasia} \H^n((\pi\circ f\circ G^{-1})(G(K))= \H^n(\pi(f(K)))\leq(\sqrt{n})^n\H^n(f(K))=0,$$ because the $\sqrt{n}$-Lipschitz map $\pi$ can increase the $\H^n$-measure no more than by a factor $(\sqrt{n})^n$. On the other hand, $G(K)$ has positive $\H^{n+m}$-measure so it follows from Fubini’s theorem that its projection $(\pi\circ f\circ G^{-1})(G(K))$ onto the first $n$-coordinates has positive $\H^n$-measure which contradicts . Parts (C) and (D) are easy consequences of part (B) as follows. Write $F = f \circ G^{-1}$. Let $(x',y')\in F^{-1}(F(x,y))\cap G(K)$. Then $F(x',y')=F(x,y)$ so $x'=\pi(F(x',y'))=\pi(F(x,y))=x$ and hence $(x',y')=(x,y')\in \{ x\}\times{\mathbb R}^m$ which proves (C). To prove (D), fix $y\in{\mathbb R}^m$ and let $(x_1,y),(x_2,y)\in G(K)$. Let $\Lambda$ be the Lipschitz constant of $F$ on $G(K)$. Since $\pi:X\to({\mathbb R}^n,\ell^\infty_n)$ is $1$-Lipschitz we have $$\begin{split} n^{-1/2}|x_1-x_2|&\leq \Vert x_1-x_2\Vert_\infty= \Vert\pi(F(x_1,y))-\pi(F(x_2,y))\Vert_\infty \\ &\leq d(F(x_1,y),F(x_2,y))\leq\Lambda|x_1-x_2| \end{split}$$ which proves (D) along with the estimate . The proof is complete. Proof of Corollary \[HSC\] {#HSCProof} ========================== Since $\Theta^{*n}(f,x)>0$ almost everywhere in $A$, we can exhaust $A$ up to a set of $\H^{n+m}$ measure zero by a countable family of pairwise disjoint sets of positive $\H^{n+m}$ measure $\{K_i\}$, where each of the sets $K=K_i$ satisfies claim (B) of Theorem \[HardSard2\]. Say $\{ G_i \}$ are the associated bi-Lipschitz $C^1$-diffeomorphisms. Let $W=\bigcup_{i=1}^\infty K_i$ and $Z=A\setminus W$ so $\H^{n+m}(Z)=0$. Let $f_i=f|_{K_i}$ and let $F_i=f_i\circ G_i^{-1}$. Since $F_i$ is defined on $G_i(K_i)$ only, we have from part (C) of Theorem \[HardSard2\] that for [*any*]{} $z\in X$, $F_i^{-1}(z)$ is contained in an $m$-dimensional affine subspace of ${\mathbb R}^{n+m}$ and hence $f_i^{-1}(z)=G_i^{-1}(F_i^{-1}(z))$ is contained in an $m$-dimensional submanifold (of class $C^1$). Therefore, for any $z\in X$, $$f^{-1}(z)\cap W=\bigcup_{i=1}^\infty f_i^{-1}(z)$$ is countably $\H^m$-rectifiable as it is contained in a countable union of $m$-manifolds, and it remains to observe from Corollary \[zero\] that $\H^m (f^{-1}(z)\setminus W) = \H^m (f^{-1}(z)\cap Z)=0$ for $\H^n$ almost all $z\in X$. $\Box$ Comparing Theorems \[HardSard2\] and \[AS\] {#SecComp} =========================================== Recall the $(n,m)$-Hausdorff content which was defined in . As mentioned in the introduction, the assumption that $\Theta^{*n}(f,x)>0$ on a set of positive measure in Theorem \[HardSard2\] is weaker than the assumption of positive $(n,m)$-Hausdorff content of a cube in Theorem \[AS\]. We see this fact in the following proposition, the proof of which follows easily from the Vitali Covering Theorem. \[elllem2\] Suppose $Q \subset \mathbb{R}^{n+m}$ is a cube, $X$ is a metric space, and $f:Q \to X$ is Lipschitz. Then $$\mathcal{H}_{\infty}^{n,m}(f,Q) \leq \frac{\omega_n}{2^n}(n+m)^{n/2}\int_Q \Theta_*^n(f,x) \, dx \leq \frac{\omega_n}{2^n}(n+m)^{n/2}\int_Q \Theta^{*n}(f,x) \, dx.$$ In this proof $Q(x,d)$ and ${\mkern 1.7mu\overline{\mkern-1.7muQ\mkern-1.5mu}\mkern 1.5mu}(x,d)$ will denote open and closed cubes in ${\mathbb R}^{n+m}$ respectively. Note that $Q(x,d)\subset B(x,\lambda d)$, where $\lambda=\frac{\sqrt{n+m}}{2}$. The function $\Theta_*^n(f,\cdot)$ is integrable on $Q$ since it is bounded. Fix $\eps > 0$. Denote by $A$ the set of all points in the interior of $Q$ which are Lebesgue points of the function $\Theta_*^{n}(f,\cdot)$. Fix a point $x \in A$, and choose $d_x>0$ small enough so that $Q(x,d_x)\subset B(x,\lambda d_x) \subset Q$. Choose a sequence $\{ d_x^i \}_{i=1}^{\infty}$ with $d_x > d_x^i \searrow 0$ satisfying the following for each $d = d_x^i$: $$\Theta_*^{n}(f,x) \leq \frac{1}{|Q(x,d)|} \int_{Q(x,d)} \Theta_*^{n}(f,y) \, dy + \frac{\eps}{2}$$ and $$\frac{\mathcal{H}_{\infty}^n(f(Q(x,d)))}{\omega_n (\lambda d)^n} \leq \frac{\mathcal{H}_{\infty}^n(f(B(x,\lambda d)))}{\omega_n (\lambda d)^n}\leq \Theta_*^n(f,x) + \frac{\eps}{2}.$$ Both inequalities imply that $$\H^n_\infty(f(Q(x,d)))d^m\leq \omega_n\lambda^n\left(\int_{Q(x,d)}\Theta_*^n(f,y)\, dy+\eps d^{n+m}\right) \quad \text{for all $x\in A$ and all $d=d_x^i$.}$$ The collection of closed cubes $$\mathcal{Q}=\{ {\mkern 1.7mu\overline{\mkern-1.7muQ\mkern-1.5mu}\mkern 1.5mu}(x,d_x^i):\, x\in A,\ i\in{\mathbb N}\}$$ is a fine Vitali covering of $A$. Thus there is a countable, pairwise disjoint collection of cubes $\{ {\mkern 1.7mu\overline{\mkern-1.7muQ\mkern-1.5mu}\mkern 1.5mu}(x_j,d_j) \}$ in $\mathcal{Q}$ so that $$\H^{n+m}\left(Q\setminus\bigcup_{j}Q(x_j,d_j)\right)= \H^{n+m}\left(A \setminus \bigcup_j {\mkern 1.7mu\overline{\mkern-1.7muQ\mkern-1.5mu}\mkern 1.5mu}(x_j,d_j) \right) = 0.$$ Since the cubes $Q(x_j,d_j)$ are open, pairwise disjoint, contained in $Q$, and they cover $Q$ up to a set of measure zero, the definition of $\H^{n,m}_\infty(f,Q)$ yields $$\begin{aligned} \H_{\infty}^{n,m}(f,Q) \leq \sum_j \H^n_{\infty}(f(Q(x_j,d_j)))d_j^m &\leq \omega_n\lambda^n\left(\sum_j\int_{Q(x_j,d_j)}\Theta_*^n(f,y)\, dy+\eps\sum_j d_j^{n+m}\right)\\ &= \omega_n\lambda^n\left(\int_Q \Theta_*^n(f,y) \, dy + \eps |Q|\right).\end{aligned}$$ Sending $\eps \to 0$ gives the desired result. The next result shows that $\Theta^n_*(f,x)$ is in fact equal to the Jacobian of $f$ when $f$ is a Lipschitz mapping to ${\mathbb R}^n$. This result is related to Lemma \[elllem\]. Consider a mapping $f:\mathbb{R}^{n+m} \to \mathbb{R}^n$ which is differentiable at $x \in {\mathbb R}^{n+m}$. Define the Jacobian $|J^nf|(x)$ at $x$ as follows: $$|J^nf|(x) = \sqrt{\det (Df)(Df)^T(x)}.$$ Geometrically, it follows that, when $\rank Df(x)=n$, the Jacobian satisfies $$\label{wrx} |J^nf|(x) = \frac{\mathcal{H}^n (W_{x,r})}{\omega_n r^n} \quad \text{for any $r>0$,}$$ where $$\label{wrx2} W_{x,r} = f(x) + Df(x) (B(0,r)) \quad \text{for } B(0,r)\subset T_x{\mathbb R}^{n+m}$$ is the ellipsoid approximation (in ${\mathbb R}^n$) of $f(B(x,r))$. This Jacobian plays an important role in the so called co-area formula [@ziemer Theorem 2.7.3]. Observe that if $\pi:T_x{\mathbb R}^{n+m}\to (\ker Df(x))^\perp\subset T_x{\mathbb R}^{n+m}$ is the orthogonal projection onto the $n$-dimensional subspace $(\ker Df(x))^\perp$, then $W_{x,r}=f(x)+Df(x)(\pi(B(0,r))$, so $W_{x,r}$ is (up to a translation by the vector $f(x)$) the image of the $n$-dimensional ball $\pi(B(0,r))\subset (\ker Df(x))^\perp$ of radius $r$ under the linear map $Df(x)$. That is, $|J^nf|(x)$ is the ratio of the volume of the ellipsoid $W_{x,r}$ to the volume of $\pi(B(0,r))$. If the rank of $Df(x)$ is less than $n$, we have $|J^nf|(x)=0$. Therefore $|J^n f|(x)>0$ if and only if $\rank Df(x)=n$. We similarly define the Jacobian of any Lipschitz mapping $f:{\mathbb R}^{n+m} \supset A \to{\mathbb R}^n$ using the approximate derivative. \[euclem\] Let $f:A\to{\mathbb R}^n$ be a Lipschitz map defined on a measurable set $A\subset{\mathbb R}^{n+m}$. Then $$\label{TTJ} \Theta_*^n(f,x)=\Theta^{*n}(f,x) =|J^nf|(x)$$ for almost every $x \in A$. Note that combining this result with Proposition \[elllem2\] gives the following for any cube $Q \subset \mathbb{R}^{n+m}$ and any Lipschitz $f:Q\to{\mathbb R}^n$: $$\label{asasas} \mathcal{H}_{\infty}^{n,m}(f,Q) \leq \frac{\omega_n}{2^n}(n+m)^{n/2} \int_{Q} |J^nf|(x) \, dx.$$ This inequality is essentially Lemma 6.13 in [@AzzSch]. Assume first that $f:{\mathbb R}^{n+m}\to{\mathbb R}^n$ is an $L$-Lipschitz mapping defined on all of ${\mathbb R}^{n+m}$. It suffices to prove that holds true at all points of differentiability of $f$. Let $x\in{\mathbb R}^{n+m}$ be a point of differentiability of $f$. Given $L>\eps>0$, there is $\delta>0$ such that $$\label{eq4.3} |f(y)-f(x)-Df(x)(y-x)|<\eps r \quad \text{for all $0<r<\delta$ and $y\in B(x,r)$.}$$ Assume first that $|J^nf|(x)=0$. We will show that $\Theta_*^n(f,x)=\Theta^{*n}(f,x) =0$. Let $W_x=f(x)+Df(x)(T_x{\mathbb R}^{n+m})$ be an affine space through $f(x)$ (which is the image of the derivative in ${\mathbb R}^n$). Since $|J^nf|(x)=0$, we have that $\dim W_x\leq n-1$ and hence $$\label{7654} f(B(x,r))\subset B(f(x),Lr)\cap \{z\in{\mathbb R}^n:\,\dist (z,W_x)<\eps r\} \quad \text{for $0<r<\delta$.}$$ Since $\dim W_x=k\leq n-1$ we have that $$\H^n_\infty(f(B(x,r)))\leq C(n)\eps L^{n-1}r^{n}.$$ Indeed, the $k$-dimensional affine ball $B(f(x),Lr)\cap W_x\subset{\mathbb R}^n$ can be covered by $$C\left(\frac{Lr}{\eps r}\right)^k\leq C\left(\frac{L}{\eps}\right)^{n-1}$$ balls in ${\mathbb R}^n$ of radius $\eps r$ and centered at the points of $B(f(x),Lr)\cap W_x$. Then the balls with radii $2\eps r$ and the same centers cover the set on the right hand side of , and hence they also cover $f(B(x,r))$. Since a ball of radius $2\eps r$ has diameter $4\eps r$ we have that $$\H^n_\infty(f(B(x,r)))\leq \frac{\omega_n}{2^n}(4\eps r)^n C\left(\frac{L}{\eps}\right)^{n-1} = C(n) \omega_n \eps r^n L^{n-1}.$$ Therefore, $$\frac{\H^n_\infty(f(B(x,r)))}{\omega_n r^n}\leq C\eps L^{n-1} \quad\text{for $0<r<\delta$}$$ which readily yields $\Theta_*^n(f,x)=\Theta^{*n}(f,x) =0$. Assume now that $|J^nf|(x)>0$. Let $W_{x,r}=f(x)+Df(x)(B(0,r))$ be the ellipsoid considered in . Let $0<\lambda_1\leq\lambda_2\leq\ldots\leq\lambda_n$ be the singular values of $Df(x)$ i.e., the lengths of the semiaxes of $W_{x,r}$ are $0<\lambda_1 r\leq\lambda_2 r\leq\ldots\leq\lambda_n r$. ($\lambda_1>0$ because $|J^nf|(x)>0$). Consider the three concentric and homothetic ellipsoids (we further assume $0<\eps<\lambda_1$ so $1-\eps/\lambda_1>0$) $$W_{x,(1-\eps/\lambda_1)r}\subset W_{x,r}\subset W_{x,(1+\eps/\lambda_1)r}.$$ The distance between the boundary of the ellipsoid $W_{x,r}$ and the boundaries of each of the other two ellipsoids equals $\eps r$ since the distance between the homothetic ellipsoids is measured along the shortest semiaxes (as an easy exercise for the Lagrange multipliers). Therefore it follows from that $$\label{rudy102} W_{x,(1-\eps/\lambda_1)r}\subset f(B(x,r))\subset W_{x,(1+\eps/\lambda_1)r} \quad \text{for $0<r<\delta$.}$$ Indeed, the right inclusion follows immediately from . The proof of the left inclusion is more intricate. Suppose to the contrary that $$z\in W_{x,(1-\eps/\lambda_1)r}\setminus f(B(x,r)).$$ Then using a ‘radial’ projection from $z$ and estimate one can construct a retraction of the ellipsoid $W_{x,r}$ to its boundary which is a contradiction. We leave details of a construction of a retraction to the reader. It follows from Lemma \[h=h\] and that for any $R>0$ $$\H^n_\infty(W_{x,R})=\H^n(W_{x,R})=|J^n f|(x)\omega_n R^n$$ so implies that for $0<r<\delta$ we have $$|J^nf|(x)\left(1-\frac{\eps}{\lambda_1}\right)^n\leq \frac{\H^n_\infty(f(B(x,r)))}{\omega_n r^n}\leq |J^nf|(x)\left(1+\frac{\eps}{\lambda_1}\right)^n$$ and letting $\eps\to 0$ yields . Note that the proof presented above is enough to establish . We can now proceed to the proof of the result in the general case when $f:{\mathbb R}^{n+m} \supset A \to{\mathbb R}^n$ is Lipschitz. Let $\tilde{f}:{\mathbb R}^{n+m}\to{\mathbb R}^n$ be a Lipschitz extension of $f$. Assume that $L$ is the Lipschitz constant of $\tilde{f}$. Note that $|J^n f|=|J^n \tilde{f}|$ at almost all points of $A$, and, by the proof presented above, $|J^n\tilde{f}|(x)=\Theta_*^n(\tilde{f},x)=\Theta^{*n}(\tilde{f},x)$ for almost all $x\in{\mathbb R}^{n+m}$. Note also that $\Theta^{*n}(\tilde{f},x)\geq \Theta^{*n}({f},x)$, because in the case of $\Theta^{*n}(\tilde{f},x)$ we consider the Hausdorff content of $\tilde{f}(B(x,r))$ while in the case of $\Theta^{*n}({f},x)$ we only consider the Hausdorff content of $f(B(x,r)\cap A)=\tilde{f}(B(x,r)\cap A)$. Since for almost all $x\in A$ we have $$|J^n f|(x)=|J^n\tilde{f}|(x)=\Theta_*^n(\tilde{f},x)= \Theta^{*n}(\tilde{f},x)\geq \Theta^{*n}(f,x)\geq \Theta_*^n(f,x),$$ it suffices to show that $$\label{L10} \Theta_*^n(f,x)\geq |J^n\tilde{f}|(x) \quad \text{for almost all $x\in A$.}$$ For almost all $x\in A$ such that $|J^nf|(x)=0$, this is particularly easy. Indeed, we have $$\Theta_*^n(f,x)\geq 0=|J^n f|(x)=|J^n\tilde{f}|(x),$$ so is obvious. We are left with the case when $|J^nf|(x)>0$. Since we want to prove almost everywhere, we can assume that $x$ is a density point of $A$ and $\tilde{f}$ is differentiable at $x$. Then $|J^n\tilde{f}|(x)=\Theta^{*n}(\tilde{f},x)=\Theta_*^n(\tilde{f},x)$, $\ap Df(x)=D\tilde{f}(x)$, and $|J^n\tilde{f}|(x)=|J^nf|(x)>0$. In particular, we have $\rank D\tilde{f}(x) = n$. The idea of the rest of the proof is simple. Since $x$ is a density point of $A$, for small $r>0$, the content $\H^n_\infty(f(B(x,r)\cap A))=\H^n(\tilde{f}(B(x,r)\cap A))$ [**is not much smaller**]{} than $\H^n(\tilde{f}(B(x,r)))=\H^n_\infty(\tilde{f}(B(x,r)))$. Therefore dividing by $\omega_nr^n$ and passing to the liminf as $r\to 0$ gives $$\liminf_{r\to 0}\frac{\H^n_\infty(f(B(x,r)\cap A))}{\omega_n r^n} +\eps \geq \liminf_{r\to 0}\frac{\H^n_\infty(\tilde{f}(B(x,r)))}{\omega_n r^n}=\Theta_*^n(\tilde{f},x)= |J^n\tilde{f}|(x)$$ for all $\eps > 0$. Thus the main focus in the argument presented below is proving the phrase “[**is not much smaller**]{}”. While the idea of the proof presented below is very geometric and relatively simple, the details are not. By translating the coordinate system we may assume that $x=0$. The ellipsoid $W_{0,r}=\tilde{f}(0)+D\tilde{f}(0)(B(0,r))$ is the image of the ball $B^{n+m}(0,r)\subset T_0{\mathbb R}^{n+m}$. By abusing notation we will identify the tangent space $T_0{\mathbb R}^{n+m}$ with ${\mathbb R}^{n+m}$. For example the same notation will be used for the ball $B^{n+m}(0,r)$ in the tangent space $T_0{\mathbb R}^{n+m}$, and for the ball $B^{n+m}(0,r)=0+B^{n+m}(0,r)$ in ${\mathbb R}^{n+m}$. Since $\rank D\tilde{f}(0)=n$, we have $\dim\ker D\tilde{f}(0)=m$. Rotating the coordinate system in ${\mathbb R}^{n+m}$ we may assume that $${\mathbb R}^{n+m}=T_0{\mathbb R}^{n+m}=(\ker D\tilde{f}(0))^\perp\oplus (\ker D\tilde{f}(0))={\mathbb R}^n\oplus {\mathbb R}^m.$$ Let $$\pi:{\mathbb R}^n\oplus{\mathbb R}^m\to{\mathbb R}^n\oplus\{ 0\}\subset{\mathbb R}^n\oplus{\mathbb R}^m$$ be the orthogonal projection. Note that the $n$-dimensional ball in the tangent space $$B^n_0(r):=\pi(B^{n+m}(0,r)) =({\mathbb R}^n\times\{ 0\})\cap B^{n+m}(0,r)\subset T_0{\mathbb R}^{n+m}$$ has radius $r$ and $$W_{0,r}=\tilde{f}(0)+D\tilde{f}(0)(B^n_0(r)).$$ Let $\eps>0$ be given, then there is a positive integer $M$ such that $$\label{1492} |J^n\tilde{f}|(0)\left(1-\frac{\sqrt{2}}{\lambda_1 M}\right)^n\left(1-\frac{1}{2M}\right)- \frac{L^n}{2M} \geq |J^n\tilde{f}|(0) - \eps,$$ where $0<\lambda_1\leq\lambda_2\leq\ldots\leq\lambda_n$ are the singular values of $D\tilde{f}(0)$. For any $r>0$ and any $0<t<1$ let $$V_{t,r}=({\mathbb R}^n\times B^m(0,tr))\cap B^{n+m}(0,r)$$ be the $tr$-cylinder around $B^n_0(r)$ inside of the ball $B^{n+m}(0,r)$. Clearly $\H^{n+m}(V_{t,r})<\omega_n r^n\cdot\omega_m(tr)^m$ because $V_{t,r}\subset B^n(0,r)\times B^m(0,tr)$. Also, when $t$ is small, the volume of $V_{t,r}$ must be close to the volume of this product of balls in the following sense: $$\lim_{t\to 0}\frac{\H^{n+m}(V_{t,r})}{\omega_n r^n\cdot\omega_m (tr)^m}=1.$$ Thus there is $0<t_M<(1-\frac{1}{2M})^{1/n}$ such that $$\label{L1} \left(1-\frac{1}{4M}\right)\omega_n\omega_mr^{n+m}t_M^m< \H^{n+m}(V_{t_M,r})< \omega_n\omega_m r^{n+m} t_M^m.$$ Note that $t_M$ depends on $M$ but not on $r$ because $V_{t,r} = rV_{t,1}$ (where $rE := \{rx \, : \, x \in E\}$ for $E \subset {\mathbb R}^{n+m}$). Since $0\in A$ is a density point of $A$, we may choose $\delta>0$ depending on $M$ so that for $0<r<\delta$ we have $$\label{L4} \H^{n+m}(V_{t_M,r}\setminus A)\leq \H^{n+m}(B^{n+m}(0,r)\setminus A)<\frac{1}{4M}\omega_n\omega_m r^{n+m} t_M^m$$ and hence $$\label{L2} \H^{n+m}(V_{t_M,r}\cap A)= \H^{n+m}(V_{t_M,r})-\H^{n+m}(V_{t_M,r}\setminus A)> \left(1-\frac{1}{2M}\right)\omega_n\omega_m r^{n+m} t_M^m.$$ Since $\tilde{f}$ is differentiable at $0$, we may also assume (by taking, if necessary, a smaller $\delta>0$ depending on $M$) that $$\label{L5} |\tilde{f}(y)-\tilde{f}(0)-D\tilde{f}(0)y|< \frac{r}{M} \quad \text{for all $0<r<\delta$ and $y\in B^{n+m}(0,r)$}.$$ Let $0<r<\delta$. For $b\in{\mathbb R}^m$ we define $$B^n_b(r)=({\mathbb R}^n\times\{ b\})\cap B^{n+m}(0,r).$$ If we regard $V_{t_M,r}$ as a cylinder with base $B^m(0,t_M r)$ (and with spherical caps), then the fibers (orthogonal to the base) are the $n$-balls $B^n_b(r)$ where $b$ ranges over $B^m(0,t_M r)$. We claim that the set of $b\in B^m(0,t_M r)$ which satisfy $$\label{L3} \H^n(B^n_b(r)\cap A)>\left(1-\frac{1}{2M}\right)\, \omega_n r^n$$ has positive $\H^m$-measure. Indeed, suppose to the contrary that $$\H^n(B^n_b(r)\cap A)\leq\left(1-\frac{1}{2M}\right)\, \omega_n r^n \quad \text{for } \H^m \text{-almost all $b\in B^m(0,t_M r)$}.$$ Then it follows from Fubini’s theorem that $$\H^{n+m}(V_{t_M,r}\cap A)\leq\left(1-\frac{1}{2M}\right)\, \omega_nr^n\cdot \omega_m (t_M r)^m$$ which contradicts . In other words, we have shown that the set of fibers of $V_{t_M,r}$ which see a “large” part of $A$ has positive $\H^m$-measure. Let $b\in B^m(0, t_M r)$ be such that is satisfied. Then the radius $R$ of the ball $B^n_b(r)$ satisfies $$r\geq R>\left(1-\frac{1}{2M}\right)^{1/n}r,$$ and since $D \tilde{f}(0)$ vanishes in the direction of $(0,b)$ $$\tilde{f}(0)+D\tilde{f}(0)(B^n_b(r))= \tilde{f}(0)+D\tilde{f}(0)(\pi(B^n_b(r)))= \tilde{f}(0)+D\tilde{f}(0)(B^n_0(R))=W_{0,R},$$ $$\H^n(W_{0,R})=|J^n\tilde{f}|(0)\omega_n R^n.$$ Recall that $$0<t_M<\left(1-\frac{1}{2M}\right)^{1/n} \quad \text{and} \quad b\in B^m(0,t_Mr).$$ Therefore, $|b| < t_M r < R$. Thus by and the Pythagorean theorem, we have $$\label{L100} |\tilde{f}(y)-(\tilde{f}(0)+D\tilde{f}(0)y)| \leq M^{-1}\sqrt{(t_M r)^2+R^2} < \sqrt{2} M^{-1}R \quad \text{for $y\in \partial B^n_b(r)$.}$$ Since the distance between the boundaries of the ellipsoids (we assume that $M$ is so large that $\sqrt{2} /M<\lambda_1$) $$W_{0,(1-\sqrt{2} M^{-1}/\lambda_1)R}\subset W_{0,R}$$ equals $\sqrt{2}M^{-1} R$, it follows from (as in ) that $$W_{0,(1-\sqrt{2}M^{-1}/\lambda_1)R}\subset \tilde{f}(B^n_b(r)).$$ Therefore $$\label{L6} \begin{split} \H^n(\tilde{f}(B^n_b(r))) &\geq \H^n(W_{0,(1-\sqrt{2} M^{-1}/\lambda_1)R})= |J^n\tilde{f}|(0)\omega_n\left(1-\frac{\sqrt{2}}{\lambda_1 M}\right)^n R^n\\ &> |J^n\tilde{f}|(0)\omega_n\left(1-\frac{\sqrt{2}}{\lambda_1 M}\right)^n\left(1-\frac{1}{2M}\right)r^n. \end{split}$$ Inequality also implies that $$\H^n(B^n_b(r)\setminus A)= \H^n(B^n_b(r))-\H^n(B^n_b(r)\cap A)<\frac{1}{2M}\,\omega_n r^n.$$ Therefore $$\label{L7} \H^n(\tilde{f}(B^n_b(r)\setminus A))\leq \frac{L^n}{2M}\, \omega_n r^n.$$ We have $$\label{L8} \tilde{f}(B^n_b(r))=\tilde{f}(B^n_b(r)\cap A)\cup\tilde{f}(B^n_b(r)\setminus A)$$ so , , and yield $$\begin{split} \H^n(\tilde{f}(B^n_b(r)\cap A)) &\geq \H^n(\tilde{f}(B^n_b(r)))-\H^n(\tilde{f}(B^n_b(r)\setminus A)) \\ &\geq \omega_n r^n\left(|J^n\tilde{f}|(0)\left(1-\frac{\sqrt{2}}{\lambda_1M}\right)^n\left(1-\frac{1}{2M}\right)- \frac{L^n}{2M}\right) \end{split}$$ and hence yields $$\begin{split} \frac{\H^n_\infty(f(B^{n+m}(0,r)\cap A))}{\omega_n r^n} &= \frac{\H^n(\tilde{f}(B^{n+m}(0,r)\cap A))}{\omega_n r^n}\geq \frac{\H^n(\tilde{f}(B^n_b(r)\cap A))}{\omega_n r^n}\\ &\geq |J^n\tilde{f}|(0)\left(1-\frac{\sqrt{2}}{\lambda_1 M}\right)^n\left(1-\frac{1}{2M}\right)- \frac{L^n}{2M}\\ &\geq |J^n\tilde{f}|(0)-\eps \end{split}$$ for any $0<r<\delta$. Therefore $$\Theta^n_*(f,0)= \liminf_{r\to 0} \frac{\H^n_\infty(f(B^{n+m}(0,r)\cap A))}{\omega_n r^n}\\ \geq |J^n\tilde{f}|(0)$$ which completes the proof of and hence that of Proposition \[euclem\]. The following example provides evidence that, if the assumption is replaced by the assumptions of Theorem \[HardSard2\], then the bound and global bi-Lipschitz homeomorphism $G$ cannot be recovered. In other words, even if the $n$-density of $f$ satisfies $\Theta^{*n}(f,x)> 0$ on a set of positive measure, there is no universal constant $\eta > 0$ depending only on $m$, $n$, and $\delta$ so that $\H^{n+m}(K) > \eta$. \[fold\] Fix a constant $\Lambda>1$. For any $\eps > 0$, there is a mapping $f:\mathbb{R}^{1+1} \supset [0,1]^2 \to \mathbb{R}$ with $\Theta^{*1}(f,x)=\Theta_*^1(f,x)=|J^1f|(x) = 1$ a.e. satisfying the following: for any measurable set $K \subset [0,1]^2$ and any $\Lambda$-bi-Lipschitz homeomorphism $G:{\mathbb R}^2 \to {\mathbb R}^2$ such that $(f \circ G^{-1})|_{ ({\mathbb R}\times \{y\}) \cap G(K) }$ is $\Lambda$-bi-Lipschitz for any $y \in {\mathbb R}$, we have $\H^2(K) < \eps$. Fix $\eps > 0$ and choose $N \in {\mathbb N}$ large enough so that $\Lambda^4 2^{1-N} \sqrt{2} < \eps$. For any $n \in {\mathbb N}$, define $f_n:[0,2^{-(n-1)}]^2 \to [0,2^{-n}]^2$ as follows: $$f_n(x,y) = \left\{ \begin{array}{ll} (x,y) & \text{if } (x,y) \in [0,2^{-n}] \times [0,2^{-n}] \\ (2^{-(n-1)}-x,y) & \text{if } (x,y) \in [2^{-n},2^{-(n-1)}] \times [0,2^{-n}] \\ (x,2^{-(n-1)}-y) & \text{if } (x,y) \in [0,2^{-n}] \times [2^{-n},2^{-(n-1)}] \\ (2^{-(n-1)}-x,2^{-(n-1)}-y) & \text{if } (x,y) \in [2^{-n},2^{-(n-1)}] \times [2^{-n},2^{-(n-1)}] \\ \end{array} \right. .$$ That is, we divide $[0,2^{-(n-1)}]^2$ into four squares of equal size. On the lower left square, $f_n$ is the identity mapping. On the upper left and lower right squares, $f_n$ is a reflection over an edge onto the lower left square. On the upper right square, $f_n$ is a reflection over both the bottom and left edges onto the lower left square. Define $f:[0,1]^2 \to [0,2^{-N}]$ to be a composition of $N$ of these reflections together with the projection $\pi:\mathbb{R}^2 \to \mathbb{R}$ onto the first coordinate: $\pi(x,y) = x$. That is, we set $$f := \pi \circ f_N \circ f_{N-1} \circ \cdots \circ f_2 \circ f_1$$ Clearly, $f$ is Lipschitz. Divide $[0,1]^2$ into $(2^N)^2$ squares $\{Q_i\}$ of side length $2^{-N}$. Note that in each of the squares $f$ is a composition of an isometry of ${\mathbb R}^2$ and the orthogonal projection to ${\mathbb R}$ so $|J^1f|=1$ and hence $\Theta^{*1}(f,x)=\Theta_*^1(f,x)=|J^1f|(x) = 1$ a.e. Let $G$ be any $\Lambda$-bi-Lipschitz homeomorphism of ${\mathbb R}^{2}$ and $K \subset [0,1]^2$ be a measurable set such that $(f \circ G^{-1})|_{ ({\mathbb R}\times \{y\}) \cap G(K) }$ is $\Lambda$-bi-Lipschitz for any $y \in {\mathbb R}$. Write $F = f \circ G^{-1}$. For each $y \in {\mathbb R}$, we have $$\H^1(({\mathbb R}\times \{y\}) \cap G(K)) \leq \Lambda \H^1(F(({\mathbb R}\times \{y\}) \cap G(K))) \leq \Lambda \H^1([0,2^{-N}]) =\Lambda 2^{-N}.$$ Indeed, the first inequality is a consequence of the fact that $F|_{ ({\mathbb R}\times \{y\}) \cap G(K) }$ is $\Lambda$-bi-Lipschitz and the second inequality follows simply from the fact that the image of $F$ is contained in $[0,2^{-N}]$. Note also that $\diam(G(K)) \leq \Lambda\diam(K) \leq \Lambda\sqrt{2}$. In particular, $G(K)$ is contained in some square $Q=I_1 \times I_2$ where $I_1$ and $I_2$ are intervals of length $2\Lambda \sqrt{2}$. Thus $$\begin{aligned} \H^2(K) \leq \Lambda^2 \H^2(G(K))= \Lambda^2 \int_{Q} \chi_{G(K)} &= \Lambda^2 \int_{I_2} \H^1(({\mathbb R}\times \{y\}) \cap G(K)) \, dy \\ &\leq \Lambda^2 \int_{I_2} \Lambda 2^{-N} \, dy =\Lambda^4 2^{1-N} \sqrt{2} < \eps .\end{aligned}$$ [888]{} Rectifiable sets in metric and Banach spaces. *Math. Ann.* 318 (2000), 527–555. Hard Sard: quantitative implicit function and extension theorems for Lipschitz maps. *Geom. Funct. Anal.* 22 (2012), 1062–1123. Increasing sequences of sets and Hausdorff measure, [*Proc. London Math. Soc.*]{} 20 (1970), 222-236. *Measure theory and fine properties of functions.* Studies in Advanced Mathematics. CRC Press, Boca Raton, FL, 1992. Die Grundlehren der mathematischen Wissenschaften, Band 153 Springer-Verlag New York Inc., New York 1969. On conditions for unrectifiability of a metric space. *Anal. Geom. Metr. Spaces* 3 (2015), 1–14. Weak BLD mappings and Hausdorff measure. *Nonlinear Anal.* 177 (2018), 524-531. *Sobolev spaces on metric measure spaces. An approach based on upper gradients.* New Mathematical Monographs, 27. Cambridge University Press, Cambridge, 2015. Rectifiable sets and coarea formula for metric-valued mappings. [*J. Funct. Anal.*]{} 254 (2008), 1410-1447. A singular map of a cube onto a square. [*J. Differential Geom.*]{} 14 (1979), 593-594 (1981). Mathematical Surveys and Monographs, 51. American Mathematical Society, Providence, RI, 1997. *Geometry of sets and measures in Euclidean spaces. Fractals and rectifiability.* Cambridge Studies in Advanced Mathematics, 44. Cambridge University Press, Cambridge, 1995. The coarea formula for metric space valued maps, Ph.D. thesis, ETH Zürich, 2009. *Lectures on geometric measure theory.* Proceedings of the Centre for Mathematical Analysis, Australian National University, 3. Australian National University, Centre for Mathematical Analysis, Canberra, 1983. On totally differentiable and smooth functions. *Pacific J. Math.* 1 (1951), 143-–159. *Weakly differentiable functions. Sobolev spaces and functions of bounded variation.* Graduate Texts in Mathematics, 120. Springer-Verlag, New York, 1989.
{ "pile_set_name": "ArXiv" }
--- author: - 'Yoshihide [Tanaka]{}, Akira [Oguri]{} and Hiroumi [Ishii]{}' title: ' Perturbation Study of the Conductance through an Interacting Region Connected to Multi-Mode Leads ' --- \#1\#2 Introduction ============ Low-dimensional electron systems have been one of the current interests in the fields of the condensed matter physics and materials science. For instance, in some of the organic conductors, the electron correlation has been considered to be important to understand the physical properties. [@ishiguro] The Kondo effect in quantum dots [@NL; @GR; @Kawabata] has also been studied intensively from theoretical [@MWL1-2; @HDW2] and experimental [@RalphBuhrman; @Goldharber-Gordon; @Kouwenhoven] sides. When the average number of the electrons in a quantum dot is odd, the perfect transmission due to the Kondo resonance situated at the Fermi energy occurs at low temperatures. Recently, artificial molecules which are realized by arranging two or more quantum dots have also been studied. [@oosterkamp; @tokura] Theoretically, the crossover from the high-temperature Coulomb-blockade to low-temperature Fermi-liquid behaviors of the quantum dots has been studied using advanced numerical methods such as the numerical renormalization group [@Izumida1-3; @Izumida4-5] and quantum Monte Carlo methods. [@Sakai; @ao; @ao6] In a previous work, one of the authors has studied the conductance of small interacting systems connected to two single-mode leads, [@ao7; @ao9] and has calculated the conductance of a Hubbard chain of finite size $N$ using the order $U^2$ self-energy. The results obtained in the electron-hole symmetric case depend strongly on whether $N$ is even or odd. In the even cases, the conductance decreases with increasing $N$ showing a tendency toward a Mott-Hubbard insulator. On the other hand, in the odd cases the perfect transmission due to the Kondo resonance occurs.[@ao9] The purpose of the present work is to generalize the formulation to the multi-mode systems where the interacting system is connceted to noninteracting leads with a number of channels. As in the single-mode case, at $T=0$ the contributions of the vertex corrections on the dc conductance $g$ vanish. Then the conductance is determined by the value of the single-particle Green’s function at the Fermi energy $\omega=0$, and thus $g$ can be written in terms of the transmission coefficient of the free quasiparticles described by an effective Hamiltonian also in the multi-mode case. We apply the formulation to a two-dimensional Hubbard model of finite size consisting of $N_{\rm C}=N \times M$ sites, where $N$ and $M$ are the size in the $x$- and $y$-direction, respectively. Two noninteracting leads of $M$ channels are attached to this cluster along the $x$-direction. This system may be considered as a model for two-dimensional materials such as an array of quantum dots and a carbon nanotube. For instance, for $M=2$ it may also be considered as a model for a ladder system of nanometer size. Since there is no translational symmetry in the systems we are considering, the self-energy has $N_{\rm C} \times N_{\rm C}$ matrix elements. We calculate all these elements within the second order perturbation expansion in $U$ in the electron-hole symmetric case, and determine the parameters of the effective Hamiltonian up to terms of order $U^2$. The results of the conductance show maximums at finite $U$ for some values of $t_y/t_x$, where $t_x$ ($t_y$) is the nearest-neighbor transfer integral in the $x$-direction ($y$-direction). This is caused by resonances occurring in some of the subbands. These behaviors can also be understood through the $U$ dependence of the eigenvalues of the effective Hamiltonian. We note that the second-order perturbation theory has been applied by a number of groups to study transport through the single Anderson impurity[@HDW2; @YMF; @MiiMakoshi; @TakagiSaso] and systems consisting of a number of sites.[@PFA; @KKNO] Also, the band calculations has been applied to obtain the conductance of nanometer systems. [@Lang] Our approach can, in principle, be extended to realistic systems by evaluating the order $U^2$ self-energy using the orbitals which are obtained with the band calculations. In §2, we describe the outline of the general formulation and introduce the effective Hamiltonian. In §3, we apply the method to a two-dimensional Hubbard model, and present the results of the conductance. Summary is given in §4. In Appendix, we provide the derivation of the expressions of the conductance and the total charge displacement. Formulation {#sec:Model} =========== In this section, we give the outline of a general formulation of the conductance through interacting systems connected to Fermi-liquid reservoirs with a number of channels. Our formulation is applicable for various systems which have the time reversal symmetry, i.e., eqs. (\[eq:g\_multi\]) and (\[eq:Friedel\_multi\]) hold for a wide range of the systems described by the Hamiltonian eq. (\[eq:H\]). We provide the details of the derivations in Appendix, and demonstrate the application to a Hubbard model in the next section. We start with a system which consists of three regions as illustrated in Fig. \[fig:multi\]; a finite interacting region (${\rm C}$) situated in the middle, and two noninteracting reservoirs on the left (${\rm L}$) and right (${\rm R}$). The central region contains $N_{\rm C}$ resonant levels, and the inter-electron interaction $U_{j_4 j_3; j_2 j_1}$ is switched on in the central region. Each of the two reservoirs is infinitely large and has a continuous energy spectrum. The central region and two reservoirs are connected by $M_{\rm L}$ and $M_{\rm R}$ channels which are described by the mixing matrix elements $v_{{\rm L},m}^{\phantom{\dagger}}$ and $v_{{\rm R},m}^{\phantom{\dagger}}$, respectively. The complete Hamiltonian is given by $$\begin{aligned} {\cal H} \ &=& \ {\cal H}_{\rm L} + {\cal H}_{\rm R} + {\cal H}_{\rm C}^0 + {\cal H}_{\rm C}^{\rm int} + {\cal H}_{\rm mix} \label{eq:H} \;, \\ {\cal H}_{\rm L} &=& \sum_{ij\in {\rm L}} \sum_{\sigma} \left(\,-t_{ij}^{\rm L} - \mu\, \delta_{ij} \,\right) c^{\dagger}_{i \sigma} c^{\phantom{\dagger}}_{j \sigma} \;, \label{eq:H_L} \\ {\cal H}_{\rm R} &=& \sum_{ij\in {\rm R}} \sum_{\sigma} \left(\,-t_{ij}^{\rm R} - \mu\, \delta_{ij} \,\right) c^{\dagger}_{i \sigma} c^{\phantom{\dagger}}_{j \sigma} \;, \label{eq:H_R} \\ {\cal H}_{\rm C}^{0} &=& \sum_{ij\in {\rm C}} \sum_{\sigma} \left(\,-t_{ij}^{\rm C} - \mu\, \delta_{ij} \,\right) c^{\dagger}_{i \sigma} c^{\phantom{\dagger}}_{j \sigma} \;, \\ {\cal H}_{\rm C}^{\rm int} &=& {1 \over 2} \sum_{\{j\} \in {\rm C}}\sum_{\sigma \sigma'} U_{j_4 j_3; j_2 j_1}\, c^{\dagger}_{j_4 \sigma} c^{\dagger}_{j_3 \sigma'} c^{\phantom{\dagger}}_{j_2 \sigma'} c^{\phantom{\dagger}}_{j_1 \sigma} \;, \label{eq:H_int} \\ {\cal H}_{\rm mix} &=& - \sum_{m=1}^{M_{\rm L}} \sum_{\sigma} v_{{\rm L},m}^{\phantom{\dagger}} \left(\, c^{\dagger}_{{\cal I}_m \sigma}\, c^{\phantom{\dagger}}_{{\cal L}_m \sigma} + c^{\dagger}_{{\cal L}_m \sigma}\, c^{\phantom{\dagger}}_{{\cal I}_m \sigma} \, \right) \nonumber \\ & & - \sum_{m=1}^{M_{\rm R}} \sum_{\sigma} v_{{\rm R},m}^{\phantom{\dagger}} \left(\, c^{\dagger}_{{\cal R}_m \sigma}\, c^{\phantom{\dagger}}_{{\cal N}_m \sigma} + c^{\dagger}_{{\cal N}_m\, \sigma} c^{\phantom{\dagger}}_{{\cal R}_m \sigma} \, \right) . \label{eq:H_mix_multi}\end{aligned}$$ Here $c^{\dagger}_{j \sigma}$ ($c^{\phantom{\dagger}}_{j \sigma}$) creates (destroys) an electron with spin $\sigma$ at site $j$, and $\mu$ is the chemical potential. $t_{ij}^{\rm L}$, $t_{ij}^{\rm R}$, and $t_{ij}^{\rm C}$ are the intra-region hopping matrix elements in each of the regions ${\rm L}$, ${\rm R}$, and ${\rm C}$, respectively. The labels $1$, $2$, $\ldots$, $N_{\rm C}$ are assigned to the sites in the central region. In eq. (\[eq:H\_mix\_multi\]), ${\cal I}_m$ (${\cal L}_m$) is the label assigned to the $m$-th site at the sample side (lead side) of the interface on the left. ${\cal N}_m$ (${\cal R}_m$) is the label assigned to the $m$-th site at the sample side (lead side) of the interface on the right. Note that the number of the channels is less than the number of the interacting sites, i.e, $M_{\rm L} \leq N_{\rm C}$ and $M_{\rm R}\leq N_{\rm C}$. For instance, in the case of $M_{\rm L}=M_{\rm R}=N_{\rm C}$, all the sites in the central region are connected to both of the reservoirs, i.e., ${\cal I}_m = {\cal N}_m$ for $m=1,\,2,\,\ldots,\, N_{\rm C}$. We assume that all the hopping matrix elements are real, and the interaction has the time reversal symmetry: $U_{4 3; 2 1}$ is real and $U_{4 3; 2 1}=U_{3 4; 1 2}=U_{1 2; 3 4 }=U_{4 2; 3 1}=U_{1 3; 2 4}$. We will be using units $\hbar=1$ unless otherwise noted. The single-particle Green’s function is defined by $$G_{jj'}({\rm i}\varepsilon_n) = - \int_0^{\beta} \! {\rm d}\tau \left \langle T_{\tau} \, c^{\phantom{\dagger}}_{j \sigma} (\tau) \, c^{\dagger}_{j' \sigma} (0) \right \rangle \, {\rm e}^{{\rm i}\, \varepsilon_n \tau} . \label{eq:G_Matsubara}$$ Here $\beta= 1/T$, $\varepsilon_n = (2n+1)\pi/\beta$, $c_{j \sigma}(\tau) = {\rm e}^{\tau {\cal H}} c_{j \sigma} {\rm e}^{- \tau {\cal H}}$, and $\langle \cdots \rangle$ denotes the thermal average $\mbox{Tr} \left[ \, {\rm e}^{-\beta {\cal H} }\, {\cdots} \,\right]/\mbox{Tr} \, {\rm e}^{-\beta {\cal H} }$. The spin index has been omitted from the left-hand side of eq. (\[eq:G\_Matsubara\]) assuming the expectation value to be independent of whether spin is up or down. Since the interaction is switched on only for the electrons in the central region, the Dyson equation can be written as $$G_{ij}(z) = G^0_{ij}(z) + \sum_{i'j' \in {\rm C}}\,G^0_{ii'}(z)\, \Sigma_{i'j'}(z) \, G_{j'j}(z) \;. \label{eq:Dyson}$$ Here $G^0_{ij}(z)$ is the unperturbed Green’s function corresponding to the noninteracting Hamiltonian ${\cal H}^0 \equiv {\cal H}_{\rm L} + {\cal H}_{\rm R} + {\cal H}_{\rm C}^0 + {\cal H}_{\rm mix}$. The summations with respect to $i'$ and $j'$ run over the sites in the central region, and $\Sigma_{i'j'}(z)$ is the self-energy due to the interaction ${\cal H}_{\rm C}^{\rm int}$. Because of the time reversal symmetry of ${\cal H}$, these functions are symmetric against the interchange of the site indices: $G_{ij}(z) = G_{ji}(z)$ and $\Sigma_{ij}(z) = \Sigma_{ji}(z)$. Note that at $T=0$ the single-particle excitation at the Fermi energy $z= {\rm i}0^+$ does not decay, i.e., $\mbox{Im}\,\Sigma_{ij}^+ (0)=0$.[@LangerAmbegaokar] In what follows, we will treat $z$ as a complex variable, and use the symbol $+$ ($-$) in the superscript as a label for the retarded (advanced) function: $\Sigma_{ij}^{\pm}(\varepsilon) \equiv \Sigma_{ij}(\varepsilon \pm {\rm i}0^+)$. As in the single-mode case, the dc conductance at $T=0$ can be expressed in terms of the single-particle Green’s function at $\omega=0$ \[see Appendix \[sec:CONDUCTANCE\]\]; $$g = {2 e^2 \over h} \, \mbox{Tr} \left[\,4\, \mbox{\boldmath $\Gamma$}_{\rm R} (0) \, \mbox{\boldmath $G$}_{{\cal N}{\cal I}}^{+}(0)\, \mbox{\boldmath $\Gamma$}_{\rm L}(0) \, \mbox{\boldmath $G$}_{{\cal I}{\cal N}}^{-}(0) \,\right] . \label{eq:g_multi}$$ Here $\mbox{\boldmath $\Gamma$}_{\rm L}(\omega)$ and $\mbox{\boldmath $\Gamma$}_{\rm R}(\omega)$ are $M_{\rm L} \times M_{\rm L}$ and $M_{\rm R} \times M_{\rm R}$ matrices, respectively. These two matrices are caused by the coupling with the left (${\rm L}$) and right (${\rm R}$) leads, and the elements are given by $\Gamma_{\alpha;mm'} (\omega) = - \mbox{Im} \left[ v_{\alpha,m}\, F_{\alpha,mm'}^{+}(\omega) \,v_{\alpha,m'} \right]$ with $F_{\alpha,mm'}(z)$ being the Green’s functions at the interface of the isolated lead ($\alpha ={\rm L},{\rm R}$). In eq. (\[eq:g\_multi\]), $\mbox{\boldmath $G$}_{{\cal N}{\cal I}}^+$ and $\mbox{\boldmath $G$}_{{\cal I}{\cal N}}^-$ are $M_{\rm R} \times M_{\rm L}$ and $M_{\rm L} \times M_{\rm R}$ matrices the elements of which are given by $G_{{\cal N}_{l}{\cal I}_{m}}^+$ and $G_{{\cal I}_{l}{\cal N}_{m}}^-$, respectively. Another quantity which can be related to the scattering coefficients is the displacement of the total charge defined by [@LangerAmbegaokar; @Langreth] $$\begin{aligned} \Delta N_{\rm tot} \ &=& \ \sum_{i\in {\rm C}} \sum_{\sigma} \langle c^{\dagger}_{i \sigma} c^{\phantom{\dagger}}_{i \sigma} \rangle \nonumber\\ & & + \sum_{i\in {\rm L}} \sum_{\sigma} \left[\, \langle c^{\dagger}_{i \sigma} c^{\phantom{\dagger}}_{i \sigma} \rangle - \langle c^{\dagger}_{i \sigma} c^{\phantom{\dagger}}_{i \sigma} \rangle_{\rm L}^{\phantom{0}} \,\right] \nonumber\\ & & + \sum_{i\in {\rm R}} \sum_{\sigma} \left[\, \langle c^{\dagger}_{i \sigma} c^{\phantom{\dagger}}_{i \sigma} \rangle - \langle c^{\dagger}_{i \sigma} c^{\phantom{\dagger}}_{i \sigma} \rangle_{\rm R}^{\phantom{0}} \,\right]. \label{eq:dn_def}\end{aligned}$$ Here $\,\langle \cdots \rangle_{\rm L}^{\phantom{0}}$ and $\,\langle \cdots \rangle_{\rm R}^{\phantom{0}}$ denote the ground-state average of isolated leads described by ${\cal H}_{\rm L}$ and ${\cal H}_{\rm R}$, respectively. At $T=0$, following the derivation of the Friedel sum rule by Langer and Ambegaokar,[@LangerAmbegaokar] $\Delta N_{\rm tot}$ can be written in terms of a $(M_{\rm L} +M_{\rm R}) \times (M_{\rm L} +M_{\rm R})$ matrix $\mbox{\boldmath $S$}$, $$\begin{aligned} & &\Delta N_{\rm tot} = {1 \over \pi {\rm i}}\, \log [\, \det \mbox{\boldmath $S$} \,] \;, \label{eq:Friedel_multi} \\ \nonumber \\ & &\mbox{\boldmath $S$} \, = \, \mbox{\boldmath $1$} - \, 2 {\rm i} \left [ \matrix { \mbox{\boldmath $\Gamma$}_{\rm L}(0) & \mbox{\boldmath $0$} \cr \mbox{\boldmath $0$} & \mbox{\boldmath $\Gamma$}_{\rm R}(0) \cr } \right ] \left [ \matrix { \mbox{\boldmath $G$}_{{\cal I}{\cal I}}^{+}(0) & \mbox{\boldmath $G$}_{{\cal I}{\cal N}}^{+}(0) \cr \mbox{\boldmath $G$}_{{\cal N}{\cal I}}^{+}(0) & \mbox{\boldmath $G$}_{{\cal N}{\cal N}}^{+}(0) \cr } \right ] . \nonumber \\ \label{eq:S_multi}\end{aligned}$$ Here $\mbox{\boldmath $G$}_{{\cal I}{\cal I}}^+$ and $\mbox{\boldmath $G$}_{{\cal N}{\cal N}}^+$ are $M_{\rm L} \times M_{\rm L}$ and $M_{\rm R} \times M_{\rm R}$ matrices, and the elements are given by $G_{{\cal I}_{l}{\cal I}_{m}}^+$ and $G_{{\cal N}_{l}{\cal N}_{m}}^+$, respectively. The outline of the Friedel sum rule in the single-mode case is given in Appendix \[sec:Friedel\]. The extension to multi-mode case eq. (\[eq:Friedel\_multi\]) is straightforward because in the case of the Friedel sum rule we do not have to take into account the contributions vertex corrections. For convenience, in eq. (\[eq:S\_multi\]) we have assumed that there is no common element in the sets ${\cal I}$ and ${\cal N}$. This is not an important assumption, and the expression without the assumption can be obtained in the similar way. Note that eq. (\[eq:Friedel\_multi\]) is also written in terms of the Green’s function at $T=0$ and $\omega=0$. Thus, due to the property $\mbox{Im}\,\Sigma_{ij}^+ (0)=0$, the values of $g$ and $\Delta N_{\rm tot}$ at $T=0$ can be expressed in terms of the transmission and reflection coefficients defined with respect to a one-particle Hamiltonian[@ao9] $$\widetilde{\cal H}_{\rm qp} = {\cal H}_{\rm L} + {\cal H}_{\rm R} + {\cal H}_{\rm C}^{\rm eff} + {\cal H}_{\rm mix}\;. \label{eq:H_eff}$$ Here ${\cal H}_{\rm C}^{\rm eff} = \sum_{ij\in {\rm C},\sigma} \left(\,-\widetilde{t}_{ij}^{\rm C} - \mu\, \delta_{ij} \,\right) c^{\dagger}_{i \sigma} c^{\phantom{\dagger}}_{j \sigma}$ with $-\widetilde{t}_{ij}^{\rm C} = -t_{ij}^{\rm C} + \mbox{Re}\, \Sigma_{ij}^+ (0)$. This effective Hamiltonian describes free quasiparticles of the local Fermi-liquid. As shown in the next section, the eigenvalues of the partial Hamiltonian ${\cal H}_{\rm C}^{\rm eff}$ have important information about the ground-state properties. Application to a Two-dimensional Hubbard Model {#sec:Hubbard2D} ============================================== In this section, we apply the formulation to a two-dimensional Hubbard model connected to reservoirs and calculate the conductance using the order $U^2$ self-energy in the electron-hole symmetric case. Model ----- The schematic picture of the model is illustrated in Fig. \[fig:2dHubbard\]. The system size in the $y$-direction, $M$, is finite. In the $x$-direction, the central region consists of $N$ columns, and two noninteracting leads are connected at $x=1$ and $x=N$. Thus, the total number of the interacting sites is $N_{\rm C} = N\times M$. The parameters of the Hamiltonian eq. (\[eq:H\]) are specified as follows. The mixing matrix element is assumed to be uniform in the $y$-direction: $v_{{\rm L},m}^{\phantom{0}} =v_{\rm L}^{\phantom{0}}$ and $v_{{\rm R},m}^{\phantom{0}} = v_{\rm R}^{\phantom{0}}$ for $m=1,2,\ldots,M$. We assume that the off diagonal part of $t_{ij}^{\rm C}$ describes the nearest-neighbor hopping; $t_x$ and $t_y$ in the $x$- and $y$-directions, respectively. Along the $y$-direction, we use the periodic boundary condition. Furthermore, we assume $U_{j_4 j_3; j_2 j_1}$ to be an onsite repulsion $U$, and concentrate on the electron-hole symmetric case taking the parameters to be $\mu =0$ and $\epsilon_d + U/2 =0$, where $\epsilon_d$ is the onsite energy of the interacting sites $-t_{ii}^{\rm C}=\epsilon_d$. Then the Dyson equation can be written in a $N_{\rm C}\times N_{\rm C}$ matrix form: $ \left\{ \widehat{\mbox{\boldmath ${\cal G}$}}(z) \right\}^{-1} = \left\{ \widehat{\mbox{\boldmath ${\cal G}$}}^0(z) \right\}^{-1} - \widehat{\mbox{\boldmath $\Sigma$}}(z)$ with $$\begin{aligned} \left\{ \widehat{\mbox{\boldmath ${\cal G}$}}^0(z) \right\}^{-1} &=& z \, \widehat{\mbox{\boldmath $1$}} - \widehat{\mbox{\boldmath ${\cal H}$}}_{\rm C}^0 - \widehat{\mbox{\boldmath ${\cal V}$}}_{\rm mix}(z) \;. \label{eq:81}\end{aligned}$$ Here $\widehat{\mbox{\boldmath $1$}}$ is the $N_{\rm C} \times N_{\rm C}$ unit matrix, and the matrices can be written in the partitioned forms; $$\begin{aligned} \widehat{\mbox{\boldmath ${\cal H}$}}_{\rm C}^0 &=& \left[ \matrix{ \mbox{\boldmath $h$}_y^0& -t_x \mbox{\boldmath $1$} & & \mbox{\boldmath $0$} \cr -t_x \mbox{\boldmath $1$} & \ddots & \ddots & \cr & \ddots & \ddots & -t_x \mbox{\boldmath $1$} \cr \mbox{\boldmath $0$} & & -t_x \mbox{\boldmath $1$} & \mbox{\boldmath $h$}_y^0 \cr } \right], \label{eq:85} \\ \widehat{\mbox{\boldmath ${\cal V}$}}_{\rm mix}(z) &=& \left[ \matrix{ v_{\rm L}^2 \mbox{\boldmath $F$}_{\rm L}(z) & & & & \cr & & & & \cr & & \mbox{\boldmath $0$} & & \cr & & & & \cr & & & & v_{\rm R}^2 \mbox{\boldmath $F$}_{\rm R}(z) \cr } \right], \label{eq:87} \\ \widehat{\mbox{\boldmath $\Sigma$}}(z) &=& \left[ \matrix{ \mbox{\boldmath $\Sigma$}_{11}(z) & \mbox{\boldmath $\Sigma$}_{12}(z) & \ldots & \mbox{\boldmath $\Sigma$}_{1N}(z) \cr \mbox{\boldmath $\Sigma$}_{21}(z) & \ddots & \ddots & \vdots \cr \vdots & \ddots & \ddots & \vdots \cr \mbox{\boldmath $\Sigma$}_{N1}(z) & \ldots & \ldots & \mbox{\boldmath $\Sigma$}_{NN}(z) \cr } \right], \label{eq:86}\end{aligned}$$ where $\mbox{\boldmath $1$}$ is the $M\times M$ unit matrix, and $\mbox{\boldmath $h$}_y^0$ is the $M\times M$ hopping matrix in the $y$-direction; $$\mbox{\boldmath $h$}_y^0 \,= \left [ \matrix{ 0 & -t_y & & & & -t_y \cr -t_y & 0 & -t_y & & \mbox{\boldmath $0$} & \cr & -t_y & 0 & \ddots & & \cr & & \ddots & \ddots & \ddots & \cr & \mbox{\boldmath $0$} & & \ddots & \ddots & -t_y \cr -t_y & & & & -t_y & 0 \cr } \right]. \label{eq:59}$$ We note that the Hartree-Fock term of the self-energy is already included in the unperturbed Green’s function $\widehat{\mbox{\boldmath ${\cal G}$}}^0(z)$ defined by eq. (\[eq:81\]). Therefore, $\widehat{\mbox{\boldmath $\Sigma$}}(z)$ is the self-energy correction beyond the mean-field theory, which is described by the many-body perturbation theory with respect to $ {\cal H}_{\rm C}^{\rm int} = U \sum_{i=1}^{N_{\rm C}} \left[\, n_{i \uparrow}\, n_{i \downarrow} - ( n_{i \uparrow} + n_{i \downarrow} )/2 \,\right] $ where $n_{i \sigma} = c^{\dagger}_{i \sigma} c^{\phantom{\dagger}}_{i \sigma}$. We note that the mixing with the noninteracting leads is included in the unperturbed Hamiltonian, and thus $\widehat{\mbox{\boldmath $\Sigma$}}(z)$ depends on $v_{\rm L}$ and $v_{\rm R}$ through $\widehat{\mbox{\boldmath ${\cal G}$}}^0(z)$. In eq. (\[eq:86\]) the partitioned element $\mbox{\boldmath $\Sigma$}_{ll'}(z)$ is a $M \times M$ matrix the $(m,m')$ element of which corresponds to the self-energy correction between the sites located at $\mbox{\boldmath $r$}=(l, m)$ and $\mbox{\boldmath $r$}'=(l', m')$, where $l$ and $m$ correspond to the $x$ and $y$ coordinates, respectively. The local Green’s functions at the interfaces of the isolated leads, $\mbox{\boldmath $F$}_{\rm L}(z)$ and $\mbox{\boldmath $F$}_{\rm R}(z)$ in eq. (\[eq:87\]), depend on the excitation spectrum of the leads, i.e., ${\cal H}_{\rm L}$ and ${\cal H}_{\rm R}$. We concentrate on the case $\mbox{\boldmath $F$}_{\rm L}=\mbox{\boldmath $F$}_{\rm R}$ ($\equiv \mbox{\boldmath $F$}$) assuming the same excitation spectrum for the left and right leads. Then the matrices of the level width are given by $\mbox{\boldmath $\Gamma$}_{\rm L}(\omega) = -v_{\rm L}^2\, \mbox{Im}\, \mbox{\boldmath $F$}(\omega+{\rm i} 0^+)$ and $\mbox{\boldmath $\Gamma$}_{\rm R}(\omega) = -v_{\rm R}^2 \,\mbox{Im}\,\mbox{\boldmath $F$}(\omega+{\rm i} 0^+)$. For the function $\mbox{\boldmath $F$}(z)$, we consider two types models: I) semi-infinite tight-binding leads, and II) leads of a constant density of states. A schematic picture of the type I lead is illustrated in Fig. \[fig:2dHubbard\_2\]: the hopping matrix element is given by the nearest neighbor one, $t_x$ and $t_y$, as that in the central region. Therefore the Green’s function for the type I lead $\mbox{\boldmath $F$}^{\rm I}$ satisfies a $M\times M$ matrix equation $\mbox{\boldmath $F$}^{\rm I}(z) = \left[\, z \mbox{\boldmath $1$} - \mbox{\boldmath $h$}_y^0 - t_x^2\, \mbox{\boldmath $F$}^{\rm I}(z) \,\right]^{-1}$. In the case of the type II leads, we assume that the local density of states at the interfaces $\rho$ is a constant and the bandwidth is infinity. Then the corresponding retarded Green’s function becomes pure imaginary independent of the frequency $\omega$. Specifically, we consider a simple diagonal matrix of the form $\mbox{\boldmath $F$}^{\rm II}(\omega+{\rm i} 0^+) = -{\rm i}\,\pi \rho \,\mbox{\boldmath $1$}$. Thus for the type II leads, the effects of the mixing are parametrized by the constant $\Gamma_{\alpha} = \pi \rho \, v^2_{\alpha}$ for $\alpha ={\rm L}$ and ${\rm R}$. The subband structure of the system is determined by the eigenstates of eq. (\[eq:59\]): $\mbox{\boldmath $h$}_y^0 \, \mbox{\boldmath $\chi$}_n = \epsilon_n \, \mbox{\boldmath $\chi$}_n$ for $n=1,2,\ldots, M$. Due to the translational symmetry in the $y$-direction, the self-energy $\mbox{\boldmath $\Sigma$}_{ll'}(z)$ and the $M \times M$ matrix Green’s function $\mbox{\boldmath $G$}_{ll'}(z)$, which are the $(l,l')$ partitioned element of $\widehat{\mbox{\boldmath ${\cal G}$}}(z)$, can be diagonalized using the eigenstates; $$\begin{aligned} \mbox{\boldmath $\Sigma$}_{ll'}(z) &=& \sum_{n=1}^M \mbox{\boldmath $\chi$}_n \, \Sigma_{ll'}^{(n)}(z) \, \mbox{\boldmath $\chi$}_n^{\dagger} \;, \label{eq:self_mode} \\ \mbox{\boldmath $G$}_{ll'}(z) &=& \sum_{n=1}^M \mbox{\boldmath $\chi$}_n \, G_{ll'}^{(n)}(z) \, \mbox{\boldmath $\chi$}_n^{\dagger} \;. \label{eq:Green_mode}\end{aligned}$$ Therefore, the conductance eq. (\[eq:g\_multi\]) can be decomposed into the sum of the contributions of the subbands: $$g \, = \, {2 e^2 \over h} \, \sum_{n=1}^M 4\, \Gamma_{\rm R}^{(n)} (0) \, G_{N1}^{(n)+}(0)\, \Gamma_{\rm L}^{(n)}(0) \, G_{1N}^{(n)-}(0) \;. \label{eq:g_multi_mode}$$ Here $\Gamma_{\alpha}^{(n)}(0)$ is defined by $\mbox{\boldmath $\Gamma$}_{\alpha}(z) = \sum_{n=1}^M \mbox{\boldmath $\chi$}_n \, \Gamma_{\alpha}^{(n)}(0) \, \mbox{\boldmath $\chi$}_n^{\dagger} $ for $\alpha={\rm L},{\rm R}$. Similarly, the Friedel sum rule eq. (\[eq:Friedel\_multi\]) can be rewritten as $$\begin{aligned} & &\Delta N_{\rm tot} = {2 \over \pi} \sum_{n=1}^{M} {1 \over 2 {\rm i}} \,\log \left[ \det \mbox{\boldmath $S$}^{(n)} \right] \;, \label{eq:Friedel_mode} \\ \nonumber \\ & &\mbox{\boldmath $S$}^{(n)} = \left [ \, \matrix { 1 & 0 \cr 0 & 1 \cr } \, \right ] \nonumber \\ & & \qquad - \, 2 {\rm i} \left [ \matrix { \Gamma_{\rm L}^{(n)}(0) & 0 \cr 0 & \Gamma_{\rm R}^{(n)}(0) \cr } \right ] \left [ \matrix { G_{11}^{(n)+}(0) & G_{1N}^{(n)+}(0) \cr G_{N1}^{(n)+}(0) & G_{NN}^{(n)+}(0) \cr } \right ] . \nonumber \\ \label{eq:S_mode}\end{aligned}$$ Here $\delta^{(n)} \equiv 1/(2{\rm i})\, \log \left[ \det\mbox{\boldmath $S$}^{(n)} \right]$ corresponds to the phase shift of the $n$-th subband, and the charge displacement in each subband is given by $\Delta N_{\rm tot}^{(n)} = 2\, \delta^{(n)}/\pi$. As mentioned, at $T=0$ the conductance is determined by the value of the Green’s function at $\omega=0$, and $\mbox{Im}\,\widehat{\mbox{\boldmath $\Sigma$}}^+(0) = \widehat{\mbox{\boldmath $0$}}$ due to the Fermi-liquid property. Thus the effective Hamiltonian $\widehat{\mbox{\boldmath ${\cal H}$}}_{\rm C}^{\rm eff} =\widehat{\mbox{\boldmath ${\cal H}$}}_{\rm C}^0 + \mbox{Re}\, \widehat{\mbox{\boldmath $\Sigma$}}^+(0)$, which was introduced in the previous section, can be used to calculate the conductance and the displacement of the total charge.[@ao9] We now consider the $M\times M$ matrix $\mbox{\boldmath ${\cal H}$}_{{\rm C};ll'}^{\rm eff}$, which is the partitioned element of $\widehat{\mbox{\boldmath ${\cal H}$}}_{\rm C}^{\rm eff}$. The matrix $\mbox{\boldmath ${\cal H}$}_{{\rm C};ll'}^{\rm eff}$ can also be diagonalized as $$\begin{aligned} \mbox{\boldmath ${\cal H}$}_{{\rm C};ll'}^{\rm eff} \, &=& \, - \sum_{n=1}^M \mbox{\boldmath $\chi$}_n \, \widetilde{t}^{(n)}_{ll'} \, \mbox{\boldmath $\chi$}_n^{\dagger} \;, \label{eq:H_eff_part} \\ -\widetilde{t}^{(n)}_{ll'} \, &=& \, \epsilon_n \, \delta_{ll'} \, -t_x \left[\,\delta_{l,l'+1} \, + \, \delta_{l+1,l'} \,\right] + \mbox{Re}\,\Sigma_{ll'}^{(n)+}(0) \;, \nonumber \\ \label{eq:25}\end{aligned}$$ where $1\leq l, l' \leq N$. Thus each of the modes can be mapped onto a tight-biding model in one-dimension with the renormalized hopping matrix element $\widetilde{t}^{(n)}_{ll'}$. As it will be seen later, the behavior of eigenvalues of the $N \times N$ effective Hamiltonian defined by $\widetilde{\mbox{\boldmath ${\cal H}$}}_{\rm C}^{(n)} = \{ -\widetilde{t}^{(n)}_{ll'} \}$ is related to the transport properties. We calculate the conductance using the perturbation expansion with respect to $U$. In the electron-hole symmetric case, the order $U^2$ self-energy can be described by the diagram shown in Fig. \[fig:diagramSelf\]; $$\begin{aligned} & &\Sigma_{jj'}^+(0) \, = \nonumber \\ & & \quad - U^2 \int_{-\infty}^{\infty} \! \int_{-\infty}^{\infty} \frac{{\rm d}\varepsilon\, {\rm d}\varepsilon'}{(2\pi)^2} \, G^{0}_{jj'}({\rm i} \varepsilon) \, G^{0}_{jj'}({\rm i} \varepsilon') \, G^{0}_{j'j}({\rm i} \varepsilon + {\rm i} \varepsilon') \;, \nonumber \\ \label{eq:Self_2v} \end{aligned}$$ where $1\leq j, j' \leq N_{\rm C}$. The explicit form of the unperturbed Green’s function $G^{0}_{jj'}({\rm i} \varepsilon)$ can be obtained by taking the inverse of eq. (\[eq:81\]). Since ${\cal H}_{\rm mix}$ is included in the unperturbed part, $G^{0}_{jj'}$ and $\Sigma_{jj'}^+$ depend on the mixing matrix elements $v_{\rm L}^{\phantom{0}}$ and $v_{\rm R}^{\phantom{0}}$. Note that the retarded function at $\omega=0$ can be obtained from the Matsubara function, i.e., $\Sigma_{jj'}^+(0) = \Sigma_{jj'}({\rm i}\varepsilon)|_{\varepsilon \to 0^+}$. The imaginary part of eq. (\[eq:Self\_2v\]) vanishes, i.e., $\mbox{Im} \Sigma_{jj'}^+(0)=0$ owing to the Fermi-liquid property. Furthermore, due to the electron-hole symmetry, $\mbox{Re}\, \Sigma_{jj'}^+(0) = 0$ if $j$ and $j'$ belong to the same sublattice. We calculate all the nonzero elements of $\mbox{Re}\, \Sigma_{jj'}^+(0)$ carrying out the integrations in eq. (\[eq:Self\_2v\]) numerically. Then the full Green’s function $\mbox{\boldmath $G$}_{N1}^{+}(0)$ is evaluated substituting the order $U^2$ self-energy into the Dyson equation. In what follows we will assume the inversion symmetry $v_{\rm L}^{\phantom{0}} =v_{\rm R}^{\phantom{0}}$ ($\equiv v$), and take the parameter to be $v/t_x= 0.9$ in the case of the type I leads. We have actually done some calculations also for $v/t_x= 0.7$, but the results are similar to those for $v/t_x= 0.9$ qualitatively. Correspondingly, in the case of the type II leads we set $\Gamma_{\rm L} = \Gamma_{\rm R}$ ( $\equiv\ \Gamma$ ) and take the value to be $\Gamma/t_x= 0.75$. We will mainly discuss the results obtained for the leads of the width $M=4$. Noninteracting systems ---------------------- We now discuss some properties in the noninteracting case in order to show the situations we are considering clear. Fig. \[fig:40\] shows the conductance $g_{\rm 1d}^0(E_F)$ for noninteracting electrons in the one-dimensional systems connected to the leads of the type I (solid line) and II (dashed line). In this figure, $E_F$ is the Fermi-energy, and the size of the central region is taken to be $N=4$. The coupling with the leads are taken to be $v/t_x=0.9$ for the type I leads, and $\Gamma/t_x= 0.75$ for the type II leads. There are $N$ ($=4$) peaks of the resonance, and the peaks become sharp if the mixing matrix element $v/t_x$ or $\Gamma/t_x$ decreases. In two dimensional lattice with the periodic boundary condition in the $y$-direction, the conductance for noninteracting electrons is given by $g_{\rm 2d}^0 = \sum_{n=1}^M \, g_{\rm 1d}^0(\epsilon_n)$ in the electron-hole symmetric case. For instance, in the case of $M=4$, the eigenvalues $\epsilon_n$ are given by $\epsilon_1= -\epsilon_4 = 2t_y$, and $\epsilon_2=\epsilon_3=0$. Therefore, when $t_y$ decreases, the conductance shows peaks corresponding to the resonance occurring in the one-dimensional system $g_{\rm 1d}^0(\epsilon_n)$. As an example of the the subband structure in the case of the type I leads, we show the dispersion relation of an ideal system $E_{nk_x} = -2t_x \cos k_x + \epsilon_n$ in Fig. \[fig:09\]. Here the size along the $y$-direction is taken to be $M=4$, and the hopping matrix elements are taken to be $t_y/t_x=1.0$ (solid line) and $t_y/t_x=0.5$ (dashed line). In the case of $t_y/t_x =1.0$ the highest and lowest subbands, i.e., modes $1$ and $4$, do not contribute to the conductance because the Fermi energy is situated at the band edge of the subbands. These two subband become conducting for $0.0<t_y/t_x <1.0$. We note that the modes $2$ and $3$ are degenerate and several curves are overlaps on the line corresponding to $-2t_x \cos k_x$ in Fig. \[fig:09\]. Hubbard model connected to the type I leads ------------------------------------------- In this subsection, we discuss the results for the Hubbard model connected to the type I leads. The conductance $g$ for the isotropic hopping $t_y/t_x=1.0$ is shown as a function of $U$ in Fig. \[fig:30\], where the size of the interacting region in the $y$-direction is $M=4$ and that in the $x$-direction is $N=4,6,8$, and $10$. The conductance decreases with increasing $U$. The $N$ dependence is similar to that in the one-dimensional case, i.e., $g$ decreases monotonically with increasing $N$ showing a tendency toward the Mott-Hubbard insulator.[@ao9] As mentioned in the above, the modes $1$ and $4$ are not conducting in the case of $t_y/t_x=1.0$, and the conductance is determined by the contributions of the modes $2$ and $3$. This situation is changed in the anisotropic cases $t_y/t_x <1.0$. In Fig. \[fig:31\], the conductance is plotted for several values of $t_y/t_x$, where the size of the interacting region is taken to be $N=4$ and $M=4$. There is a broad peak at finite $U$ for $0.65 \lesssim t_y/t_x \lesssim 0.75$. This shows that there are parameter regions where the total conductance increases with $U$ even in the half-filled case. To clarify this behavior, in Fig. \[fig:35\] the contributions of each of the conducting modes are plotted separately taking $t_y/t_x$ to be $0.7$. The resonant tunneling occurs in the modes $1$ and $4$ at $U/(2 \pi t_x) \simeq 3.3$, while the contributions of the modes $2$ and $3$ decreases monotonically with increasing $U$. Note that the pair of the subbands whose wavenumber in the $y$-direction are $k_y$ and $-k_y$ give the same contributions to the conductance. Furthermore in the electric-hole symmetric case, the contributions of the subbands whose eigenvalues are $\epsilon_n$ and $-\epsilon_n$ are the same. The occurrence of the resonant tunneling is linked with the behavior of the eigenvalues of effective Hamiltonian which is defined in terms of the renormalized hopping matrix element $-\widetilde{t}^{(n)}_{ll'}$ given by eq. (\[eq:25\]). In Fig. \[fig:50\], the eigenvalue of the $N\times N$ matrix $\widetilde{\mbox{\boldmath ${\cal H}$}}_{\rm C}^{(n)}$ is plotted as a function of $U$ for the mode $1$ (solid line), where $t_y/t_x=0.7$, $N=4$ and $M=4$. The second lowest eigenvalue crosses the Fermi energy at $U/(2 \pi t_x) \simeq 4.6$. This corresponds to the peak seen at $U/(2 \pi t_x) \simeq 3.3$ in Fig. \[fig:35\], although the values of $U$ do not coincide. This difference is mainly due to the contribution of the real part of the mixing term eq. (\[eq:87\]): $\mbox{Re} \widehat{\mbox{\boldmath ${\cal V}$}}_{\rm mix}^+(0)$ causes the energy shift, and the position of the resonance corresponding to the second eigenvalue moves toward the low-energy side. While the transmission probability is defined with respect to the effective Hamiltonian of the whole system $\widetilde{\cal H}_{\rm qp}$ in eq. (\[eq:H\_eff\]), the eigenvalues of the partial Hamiltonian $\widetilde{\mbox{\boldmath ${\cal H}$}}_{\rm C}^{(n)}$ are useful to investigate the behaviors of the conductance. Note that $\widetilde{t}^{(n)}_{ll'}$ depends on the mixing matrix elements $v_{\rm L}^{\phantom{0}}$ and $v_{\rm R}^{\phantom{0}}$ because the unperturbed Green’s functions used to calculate the self-energy eq. (\[eq:Self\_2v\]) are defined with respect to the whole system. In Fig. \[fig:50\], the eigenvalues for the modes $2$ and $3$ are also plotted (dashed lines). In the present case, the eigenvalues of the modes $2$ and $3$ are the same, and the eigenvalues of the modes $1$ and $4$ are symmetric with respect to the Fermi energy $\omega=0$. The energy gap between the second and third eigenvalues of the mode $2$ (or $3$) becomes large with increasing $U$, and it seems to show the tendency toward the Mott-Hubbard insulator. We have also examined the conductance in the case of $M=6$, where the subbands are classified into two groups; $\epsilon_1=-\epsilon_6=2t_y$ and $\epsilon_2=\epsilon_3=-\epsilon_4=-\epsilon_5=t_y$. In Fig. \[fig:57\], the total conductance (solid line) and the contributions of the two groups of subbands (dashed line) are shown for $t_y/t_x=0.65$, where the size of the interacting region is taken to be $N=6$ and $M=6$. The total conductance show two bump like behaviors at $U/(2\pi t_x)\simeq 1.5$ and $5.1$, which correspond to the resonance occurring in each group of the subbands. The resonance occurs at different values of $U$ for the different groups of the subbands. This seems to show a general tendency that the resonance does not occur simultaneously in the different groups of the subbands. Thus, when the resonance occurs at one group, the remaining $M-2$ or $M-4$ subbands behave monotonically. Therefore the presence of the visible maximum in the $U$ dependence of the total conductance is expected only for small $M$ where the contributions of the resonating subbands are comparable to the those of the remaining subbands. The number of the eigenvalues which cross the Fermi energy is expected to increase with the size in the $x$-direction $N$. To clarify this feature, we calculate the conductance for $N=10$ keeping the width $M=4$ unchanged. The results are shown in Fig. \[fig:24\] for several values of $t_y/t_x$. In each of the cases, at least one peak is visible at finite $U$. Specifically in the case of $t_y/t_x=0.5$, a shoulder like structure is seen at $U/2\pi t_x \simeq 0.5$ in addition to the peak at $U/2\pi t_x \simeq 3.8$. To see this behavior precisely, the contributions of each of the subbands are plotted in Fig. \[fig:51\]. Two resonant peaks are seen at $U/(2 \pi t_x) \simeq 0.5$ and $3.8$ in the modes $1$ and $4$. The $U$ dependence of the eigenvalues of $\widetilde{\mbox{\boldmath ${\cal H}$}}_{\rm C}^{(n)}$ for $n=1$ (mode $1$) is shown in Fig. \[fig:52\]. The 4th and 5th lowest eigenvalues cross the Fermi energy at $U/(2 \pi t_x) \simeq 0.7$ and $4.1$, respectively. These correspond to the two resonant peaks seen in Fig. \[fig:51\] as in the case of $N=4$. The energy shift caused by the mixing with the leads seems to decrease with increasing $N$. Hubbard model connected to the type II leads -------------------------------------------- We next discuss the conductance of the Hubbard model connected to the type II leads. In this lead, the local density of states at the interface of the lead is a constant and the bandwidth is infinity. Then $\widehat{\mbox{\boldmath ${\cal V}$}}_{\rm mix}^+(\omega)$ becomes pure imaginary and independent of $\omega$, i.e., $\mbox{Re}\, \widehat{\mbox{\boldmath ${\cal V}$}}_{\rm mix}^+(\omega) \equiv 0$ and $\partial \widehat{\mbox{\boldmath ${\cal V}$}}_{\rm mix}^+(\omega) /\,\partial \omega \equiv 0$. Therefore the energy shift caused by the mixing with the leads is expected to be small compared to that in the case of the leads I. In Fig. \[fig:54\] (a), the conductance in the case of $t_y/t_x=0.6$ and $0.78$ is shown as a function of $U$, where the size of the interacting region is chosen to be $N=4$ and $M=4$. There is a resonance peak at finite $U$ in each line; at $U/(2 \pi t_x) \simeq 4.3$ for $t_y/t_x=0.6$ and $U/(2 \pi t_x) \simeq 0.6$ for $t_y/t_x=0.78$. The contributions of each of the subbands are plotted separately in Fig. \[fig:54\] (b). The peak seen in the total conductance is due to the resonance tunneling in the modes $1$ and $4$. Qualitatively, these results are similar to those of the type I leads. In Fig. \[fig:55\], the eigenvalues of the effective Hamiltonian $\widetilde{\mbox{\boldmath ${\cal H}$}}_{\rm C}^{(n)}$ for $n=1$ (mode $1$) are plotted for (a) $t_y/t_x=0.6$ and (b) $t_y/t_x=0.78$. The second lowest eigenvalue for $t_y/t_x=0.6$ becomes zero at $U/(2 \pi t_x) \simeq 4.1$. This is close to the position of the resonance peak seen in Fig. \[fig:54\] (b). The lowest eigenvalue for $t_y/t_x=0.78$ does not cross the Fermi energy for any values of $U$, although it is nearly zero for small $U$. Nevertheless, this eigenstate corresponds to the resonance peak seen at $U/(2 \pi t_x) \simeq 0.6$ in Fig. \[fig:54\] (b). This is because the lowest eigenvalue is shifted slightly up due to the mixing, and it becomes to cross the Fermi energy. We next discuss the phase shift $\delta^{(n)}$. As mentioned, the charge displacement of the $n$-th subband can be deduced from the phase shift using the Friedel sum rule $\Delta N_{\rm tot}^{(n)} = 2\, \delta^{(n)}/\pi$, where the factor $2$ corresponds to the spin degeneracy. Specifically, in the case of the type II leads, the compensation theorem holds.[@Anderson] Then the second and the third term in the right-hand of eq. (\[eq:dn\_def\]) vanish because $\widehat{\mbox{\boldmath ${\cal V}$}}_{\rm mix}^+(\omega)$ is independent of $\omega$ (see also Appendix \[sec:Friedel\]). Thus the number of electrons in the type II leads is unchanged when the leads are connected to the interacting region, and the left-hand side of eq. (\[eq:Friedel\_mode\]), $\Delta N_{\rm tot}$, becomes equal to the number of electrons in the interacting region $\sum_{i\in {\rm C},\sigma} \langle c^{\dagger}_{i \sigma} c^{\phantom{\dagger}}_{i \sigma} \rangle$. In Fig. \[fig:56\], the phase shift $\delta^{(n)}/\pi$ is plotted as a function of $U$ for the modes $1$ and $4$, where the parameters are taken to be $N=4$, $M=4$, and $t_y/t_x=0.78$. Here the principal value of the phase shift is determined in the limit of $U \to 0$ by comparing with the charge displacement obtained independently. Reflecting the resonance seen in Fig. \[fig:54\] (b), $\delta^{(n)}$ for the mode $1$ starts to increase rapidly at $U/(2 \pi t_x) \simeq 0.6$. Correspondingly in the mode $4$, which is the lowest subband in this case, the phase shift $\delta^{(4)}$ decreases. These behaviors show how the filling of the subband changes when the resonance peak crosses the Fermi energy. In Fig. \[fig:56\], the sum of the two phase shifts is a constant, i.e., $2\,\delta^{(1)}/\pi + 2\,\delta^{(4)}/\pi = 2\,N$ with $N=4$. The phase shifts for the modes $2$ and $3$ are the same, and it is a constant independent of $U$, i.e., $2\,\delta^{(n)}/\pi= N$ for $n=2$ and $3$ with $N=4$. This means that these two subbands are half filled. Note that the total charge displacement is given by $\Delta N_{\rm tot}=M\times N$ in the electron-hole symmetric case. Remarks ------- The results presented in this section are obtained using the order $U^2$ self-energy given by eq. (\[eq:Self\_2v\]). Therefore, the results plotted for rather large values of $U$ may not be sufficient in quantitative sense. Nevertheless, some typical features of the results are seen even for small $U$. For instance, the resonance at finite $U$ occurs for small value of $U/(2 \pi t_x) \lesssim 1$ in the case of $t_y/t_x=0.78$ as shown in Fig. \[fig:54\] (a). Therefore we believe that the qualitative features of the results, such as the relation between the resonance and the eigenvalues of the effective Hamiltonian, hold also for large $U$. The perturbation approach gives us a correct, in principle, description of the conductance and the total charge displacement with respect to the local Fermi-liquid ground state. Summary ======= We have studied the conductance through small interacting systems connected to multi-mode leads based on a local Fermi-liquid approach. At $T=0$, the conductance and the total charge displacement are determined by the value of the Green’s function at the Fermi energy. Since the excitations at the Fermi energy do not decay at $T=0$, there exists the one-particle Hamiltonian which reproduces these two quantities exactly. We have applied this formulation to a two-dimensional Hubbard model of finite size in the electron-hole symmetric case. We have calculated all the matrix elements of the order $U^2$ self-energy and determine the parameters of the effective Hamiltonian up to terms of order $U^2$. Specifically, we have examined two types of the noninteracting leads: I) semi-infinite tight-binding leads, and II) leads of a constant density of states. The results are similar, qualitatively, in both types of the leads. The conductance shows maximums in the $U$ dependence for some ranges of $t_y/t_x$, where $t_x$ and $t_y$ are the hopping matrix element in the $x$- and $y$-directions, respectively. This means that there exists parameter regions, where the total conductance increases with $U$, even in the half-filled case. By decomposing the total conductance into the sum of the contributions of the subbands, it is clarified that the peaks of the conductance are caused by the resonance occurring in some group of the subbands which have the similar symmetric properties. The phase shift of the subbands obtained from the Friedel sum rule shows a typical changes when the resonance occurs. The resonance generally does not occur simultaneously in different groups of the subbands, and the subbands of off-resonance behave monotonically. Therefore the maximum in the $U$ dependence of the total conductance is expected only for the mesoscopic systems in which the number of the conducting modes is small enough and the contributions of the subbands of on-resonance are comparable to the those of the remaining subbands. The perturbation approach examined in this work can be applied to interacting electrons in disordered systems. Especially, the analysis of the eigen values of the effective Hamiltonian, which is used in the present study to investigate the behaviors of the resonance, may be combined with the Thouless-number[@LicciThouless] and finite-size scaling[@MackinnonKramer] methods. The application along this line seems to be interesting in relation to the metal-insulator transition observed in two-dimensional systems.[@Kravchenko] Furthermore, extensions to the finite temperatures[@ao11] and nonequilibrium situations[@ao10] are left for future studies. Acknowledgements {#acknowledgements .unnumbered} ================ We would like to thank K. Murata, K. Tanigaki, and S. Nonoyama for valuable discussions. Numerical computation was partly performed at computation center of Nagoya University and at Yukawa Institute Computer Facility. This work was supported by the Grant-in-Aid for Scientific Research from the Ministry of Education, Science and Culture, Japan. Conductance {#sec:CONDUCTANCE} =========== In this appendix, we provide the derivation of the dc conductance in the multi-mode case eq. (\[eq:g\_multi\]) by generalizing the derivation given in the single-mode case.[@ao6] The current operator corresponding to the mixing term eq. (\[eq:H\_mix\_multi\]) is given by $$\begin{aligned} J_{\rm L} &=& {\rm i}\, e \sum_{m=1}^{M_{\rm L}} \sum_{\sigma} v_{{\rm L},m}^{\phantom{\dagger}} \left(\, c^{\dagger}_{{\cal I}_m \sigma}\, c^{\phantom{\dagger}}_{{\cal L}_m \sigma} - c^{\dagger}_{{\cal L}_m \sigma}\, c^{\phantom{\dagger}}_{{\cal I}_m\sigma} \, \right) \label{eq:J_L} \;,\\ J_{\rm R} &=& {\rm i}\, e \sum_{m=1}^{M_{\rm R}} \sum_{\sigma} v_{{\rm R},m}^{\phantom{\dagger}} \left(\, c^{\dagger}_{{\cal R}_m \sigma}\, c^{\phantom{\dagger}}_{{\cal N}_m \sigma} - c^{\dagger}_{{\cal N}_m \sigma}\, c^{\phantom{\dagger}}_{{\cal R}_m \sigma} \, \right) \label{eq:J_R} \;.\end{aligned}$$ Here $J_{\rm L}$ is the total current flowing into the sample from the left lead, and $J_{\rm R}$ is the current flowing out to the right lead from the sample. These currents and the total charge in the sample $\rho_{\rm C}^{\phantom{0}} = e \sum_{j\in {\rm C}, \sigma} c^{\dagger}_{j \sigma} c^{\phantom{\dagger}}_{j \sigma}$ satisfy the equation of continuity $\partial \rho_{\rm C}^{\phantom{0}} / \partial t \,+ J_{\rm R} - J_{\rm L} = 0$. In the linear response theory, the dc conductance is given by $$\begin{aligned} g \ &=& \ \lim_{\omega \to 0} { K_{\alpha\alpha'}^+(\omega) - K_{\alpha\alpha'}^+(0) \over {\rm i} \omega } \label{eq:Kubo_M} \;, \\ K_{\alpha\alpha'}({\rm i} \nu_l) &=& \int_0^{\beta} d\tau \, \langle\, T_{\tau}\, J_{\alpha}(\tau)\, J_{\alpha'}(0) \, \rangle \, {\rm e}^{{\rm i}\, \nu_l \tau} \;, \end{aligned}$$ where $\alpha, \alpha' = {\rm L}, {\rm R}$, and $K_{\alpha\alpha'}^+(\omega)$ is the retarded function which is obtained from the Matsubara function $K_{\alpha\alpha'}({\rm i} \nu_l)$ by the analytic continuation ${\rm i}\nu_l \to \omega + {\rm i}0^+$. The dc conductance eq. (\[eq:Kubo\_M\]) corresponds to the $\omega$-linear imaginary part of $K_{\alpha\alpha'}^+(\omega)$, and it is independent of $\alpha$ and $\alpha'$ owing to the current conservation.[@Fisher; @Lee] Moreover, $K_{\alpha'\alpha}(z) = K_{\alpha\alpha'}(z)$ because of the time reversal symmetry, and thus the imaginary part of $K_{\alpha\alpha'}^+(\omega)$ corresponds to the discontinuity of $K_{\alpha\alpha'}(z)$ along the real axis in the complex $z$-plane. Therefore, the dc conductance is equal to the coefficient of the $\nu\, \mbox{sgn}\, \nu$ term of $K_{\alpha\alpha'}({\rm i}\nu)$,[@Shiba] where $\mbox{sgn}\, \nu$ is the sign function. In what follows, we extract this singular term from $K_{\alpha\alpha'}({\rm i}\nu)$ taking the current operators to be $\alpha={\rm R}$ and $\alpha'={\rm L}$. At $T=0$, $K_{\rm RL}({\rm i}\nu)$ is written, treating the Matsubara frequencies to be continuous variables, as \[see also Fig. \[fig:multi\]\] $$\begin{aligned} K_{\rm RL}({\rm i} \nu) &=& K_{\rm RL}^{(a)}({\rm i} \nu) \ + \ K_{\rm RL}^{(b)}({\rm i} \nu) \; , \label{eq:K_nu_M} \\ K_{\rm RL}^{(a)}({\rm i} \nu) &=& e^2 \sum_{\sigma} \int_{-\infty}^{+\infty} \! {{\rm d}\varepsilon \over 2 \pi} \nonumber \\ & & \times \mbox{Tr} \Bigl[\ \mbox{\boldmath $v$}_{\rm R}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal R}{\cal L}}({\rm i}\varepsilon + i\nu) \, \mbox{\boldmath $v$}_{\rm L}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal I}{\cal N}}({\rm i}\varepsilon) + \mbox{\boldmath $v$}_{\rm R}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal N}{\cal I}}({\rm i}\varepsilon + {\rm i}\nu) \, \mbox{\boldmath $v$}_{\rm L}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal L}{\cal R}}({\rm i}\varepsilon) \nonumber \\ & & \quad - \mbox{\boldmath $v$}_{\rm R}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal R}{\cal I}}({\rm i}\varepsilon + {\rm i}\nu) \, \mbox{\boldmath $v$}_{\rm L}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal L}{\cal N}}({\rm i}\varepsilon) - \mbox{\boldmath $v$}_{\rm R}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal N}{\cal L}}({\rm i}\varepsilon + {\rm i}\nu) \, \mbox{\boldmath $v$}_{\rm L}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal I}{\cal R}}({\rm i}\varepsilon) \ \Bigr] \;, \label{eq:K_nu_a_M} \\ K_{\rm RL}^{(b)}({\rm i} \nu) &=& e^2 \sum_{\sigma \sigma '} \sum_{ \{j\} \in {\rm C}} \int_{-\infty}^{+\infty} \! {{\rm d}\varepsilon {\rm d}\varepsilon' \over (2 \pi)^2} \; \Gamma_{\sigma \sigma ' ;\, \sigma ' \sigma}^{j_1 j_2;\, j_3 j_4} ({\rm i} \varepsilon + {\rm i} \nu, {\rm i} \varepsilon'+ {\rm i} \nu ; \, {\rm i} \varepsilon' , {\rm i} \varepsilon ) \nonumber \\ & & \times \Bigl\{ \mbox{\boldmath $G$}_{{\rm C}{\cal L}}({\rm i}\varepsilon + i\nu) \, \mbox{\boldmath $v$}_{\rm L}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal I}{\rm C}}({\rm i}\varepsilon) \, - \mbox{\boldmath $G$}_{{\rm C}{\cal I}}({\rm i}\varepsilon + {\rm i}\nu) \, \mbox{\boldmath $v$}_{\rm L}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal L}{\rm C}}({\rm i}\varepsilon) \Bigr\}_{j_1j_4} \nonumber \\ & & \times \Bigl\{ \mbox{\boldmath $G$}_{{\rm C}{\cal N}}({\rm i}\varepsilon') \, \mbox{\boldmath $v$}_{\rm R}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal R}{\rm C}}({\rm i}\varepsilon' + {\rm i}\nu) - \mbox{\boldmath $G$}_{{\rm C}{\cal R}}({\rm i}\varepsilon') \, \mbox{\boldmath $v$}_{\rm R}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal N}{\rm C}}({\rm i}\varepsilon' + {\rm i}\nu) \Bigr\}_{j_3j_2} \;. \label{eq:K_nu_b_M}\end{aligned}$$ Here $\mbox{\boldmath $v$}_{\rm L}^{\phantom{\dagger}}$ and $\mbox{\boldmath $v$}_{\rm R}^{\phantom{\dagger}}$ are diagonal matrices corresponding to $v_{{\rm L},m}^{\phantom{\dagger}}$ and $v_{{\rm R},m}^{\phantom{\dagger}}$, respectively. The subscript ${\rm C}$ denotes a set consisting of $N_{\rm C}$ sites in the central region. $\mbox{\boldmath $G$}_{{\rm C}{\cal L}}(z)$ is a $N_{\rm C} \times M_{\rm L}$ matrix, and the $(j,m)$ element is given by $G_{j,{\cal L}_m}(z)$ with $j \in {\rm C}$. Also, $\mbox{\boldmath $G$}_{{\cal R}{\rm C}}(z)$ is a $M_{\rm R} \times N_{\rm C}$ matrix, and the $(m,j)$ element is given by $G_{{\cal R}_m,j}(z)$ with $j \in {\rm C}$. The matrix Green’s functions in eqs. (\[eq:K\_nu\_a\_M\]) and (\[eq:K\_nu\_b\_M\]) are defined in this way. $\Gamma_{\sigma \sigma ' ;\, \sigma ' \sigma}^{j_1 j_2;\, j_3 j_4} ({\rm i} \varepsilon_1, {\rm i} \varepsilon_2 ;\, {\rm i} \varepsilon_{3}, {\rm i} \varepsilon_{4} )$ is the vertex corrections due to the inter-electron interaction, and illustrated in Fig. \[fig:vertex\]. Since we are considering the interaction which is switched on only in the central region, the Green’s function satisfies following relations at the interfaces; $$\begin{aligned} \left\{ \!\!\! \begin{array}{ll} \mbox{\boldmath $G$}_{{\cal R} \gamma}(z) = - \, \mbox{\boldmath $F$}_{\rm R}(z) \, \mbox{\boldmath $v$}_{\rm R}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal N}\gamma }(z) &\mbox{for} \ \ \gamma = {\cal L},\,{\cal I},\,{\rm C},\,{\cal N} \\ \mbox{\boldmath $G$}_{\gamma{\cal L}}(z) = - \, \mbox{\boldmath $G$}_{\gamma{\cal I}}(z) \, \mbox{\boldmath $v$}_{\rm L}^{\phantom{\dagger}} \, \mbox{\boldmath $F$}_{\rm L}(z) &\mbox{for} \ \ \gamma = {\cal I},\,{\rm C},\,{\cal N},\, {\cal R} \end{array} \right. \nonumber \\ . \label{eq:rec_lead_M}\end{aligned}$$ Using these relations, eqs. (\[eq:K\_nu\_a\_M\]) and (\[eq:K\_nu\_b\_M\]) are rewritten as $$\begin{aligned} K_{\rm RL}^{(a)}({\rm i} \nu) &=& e^2 \sum_{\sigma} \int_{-\infty}^{+\infty} \! {{\rm d}\varepsilon \over 2 \pi} \, \mbox{Tr} \biggl[\, \mbox{\boldmath $v$}_{\rm R}^{\phantom{\dagger}} \left[\, \mbox{\boldmath $F$}_{\rm R}({\rm i}\varepsilon + {\rm i}\nu) - \mbox{\boldmath $F$}_{\rm R}({\rm i}\varepsilon) \,\right] \, \mbox{\boldmath $v$}_{\rm R}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal N}{\cal I}}({\rm i}\varepsilon + {\rm i}\nu) \nonumber \\ & & \qquad \qquad \qquad \qquad \times \, \mbox{\boldmath $v$}_{\rm L}^{\phantom{\dagger}} \, \left[\, \mbox{\boldmath $F$}_{\rm L}({\rm i}\varepsilon + {\rm i}\nu) - \mbox{\boldmath $F$}_{\rm L}({\rm i}\varepsilon) \,\right] \, \mbox{\boldmath $v$}_{\rm L}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal I}{\cal N}}({\rm i}\varepsilon) \, \biggr] \;, \label{eq:bubble_M} \\ K_{\rm RL}^{(b)}({\rm i} \nu) &=& e^2 \sum_{\sigma \sigma '} \sum_{\{j\} \in {\rm C}} \int_{-\infty}^{+\infty} \! {{\rm d}\varepsilon {\rm d}\varepsilon' \over (2 \pi)^2} \; \Gamma_{\sigma \sigma ' ;\, \sigma ' \sigma}^{j_1 j_2;\, j_3 j_4} ({\rm i} \varepsilon + {\rm i} \nu, {\rm i} \varepsilon'+ {\rm i} \nu ;\, {\rm i} \varepsilon' , {\rm i} \varepsilon ) \nonumber \\ & & \times \Bigl\{ \mbox{\boldmath $G$}_{{\rm C}{\cal I}}({\rm i}\varepsilon + {\rm i}\nu) \, \mbox{\boldmath $v$}_{\rm L}^{\phantom{\dagger}} \, \left[\, \mbox{\boldmath $F$}_{\rm L}({\rm i}\varepsilon + {\rm i}\nu) - \mbox{\boldmath $F$}_{\rm L}({\rm i}\varepsilon) \,\right] \, \mbox{\boldmath $v$}_{\rm L}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal I}{\rm C}}({\rm i}\varepsilon) \Bigr\}_{j_1j_4} \nonumber \\ & & \times \Bigl\{ \mbox{\boldmath $G$}_{{\rm C}{\cal N}}({\rm i}\varepsilon') \, \mbox{\boldmath $v$}_{\rm R}^{\phantom{\dagger}} \, \left[\, \mbox{\boldmath $F$}_{\rm R}({\rm i}\varepsilon' + {\rm i}\nu) - \mbox{\boldmath $F$}_{\rm R}({\rm i}\varepsilon') \,\right] \, \mbox{\boldmath $v$}_{\rm R}^{\phantom{\dagger}} \, \mbox{\boldmath $G$}_{{\cal N}{\rm C}}({\rm i}\varepsilon' + {\rm i}\nu) \Bigr\}_{j_3j_2} \;. \label{eq:vertex_M}\end{aligned}$$ We can now extract the singular $\nu\, \mbox{sgn}\, \nu$ term of $K_{\rm RL}({\rm i}\nu)$ using eqs. (\[eq:bubble\_M\]) and (\[eq:vertex\_M\]), as in the case of the single Anderson impurity.[@Shiba; @ao3] It can be shown that there is no singular $\nu\, \mbox{sgn}\, \nu$ term in $K_{\rm RL}^{(b)}({\rm i}\nu)$ due the two factors $[ \mbox{\boldmath $F$}_{\rm L}({\rm i}\varepsilon + {\rm i}\nu) - \mbox{\boldmath $F$}_{\rm L}({\rm i}\varepsilon) ]$ and $[ \mbox{\boldmath $F$}_{\rm R}({\rm i}\varepsilon' + {\rm i}\nu) - \mbox{\boldmath $F$}_{\rm R}({\rm i}\varepsilon') ]$ which have different frequencies. Thus, the $\nu\, \mbox{sgn}\, \nu$ term comes only from $K_{\rm RL}^{(a)}({\rm i}\nu)$. It can be obtained from eq. (\[eq:bubble\_M\]) and yields eq. (\[eq:g\_multi\]). Friedel sum rule {#sec:Friedel} ================ We provide here the outline of the derivation of the Friedel sum rule in the single-mode case eq. (\[eq:Friedel\]) following Langer and Ambegaokar.[@LangerAmbegaokar] The Green’s function in the left lead $ij \in {\rm L}$ is written as $$\begin{aligned} G_{ij}^{+}(\omega) &=& F^{\rm L}_{ij}(\omega) \ + \ F^{\rm L}_{i0}(\omega)\, v_{\rm L}^{\phantom{\dagger}} \, G_{11}^{+}(\omega)\, v_{\rm L}^{\phantom{\dagger}} \, F^{\rm L}_{0j}(\omega) \;, \\ F_{ij}^{\rm L}(\omega) &=& \sum_n { \phi_{n,{\rm L}}^{\phantom{\dagger}}(i) \,\phi_{n,{\rm L}}^*(j) \over \omega - \epsilon_{n,{\rm L}} + {\rm i}0^+ } \;.\end{aligned}$$ Here $F_{ij}^{\rm L}(\omega)$ is the retarded Green’s function of the isolated lead, and $\epsilon_{n,{\rm L}}$ and $\phi_{n,{\rm L}}^{\phantom{\dagger}}(i)$ are the eigenvalue and the corresponding eigenfunction of the one-particle Hamiltonian ${\cal H}_{\rm L}$. Using the orthogonality relation $\,\sum_{i \in {\rm L}} \phi_{n,{\rm L}}^{\phantom{\dagger}}(i)\, \phi_{n',{\rm L}}^*(i) = \delta_{nn'}$, we obtain $$\begin{aligned} \sum_{i \in {\rm L}} \left[\, G_{ii}^{+} - F^{\rm L}_{ii}\,\right] \ = \ - \, {\partial F^{\rm L}_{00} \over \partial \omega} \, v_{\rm L}^{\phantom{\dagger}} \, G_{11}^{+} \, v_{\rm L}^{\phantom{\dagger}} \;. \label{eq:dn_L}\end{aligned}$$ Here $F^{\rm L}_{00} \equiv F_{\rm L}^{+}$, which is the Green’s function at the interface. Using eq. (\[eq:dn\_L\]) and the corresponding relation for the right lead, the displacement of the total charge defined by eq. (\[eq:dn\_def\]) is written, at $T=0$, as $$\begin{aligned} \Delta N_{\rm tot} &=& - \, {2 \over \pi} \ \mbox{Im} \, \int_{-\infty}^0 {\rm d}\omega \Biggl(\, \, \sum_{i \in {\rm C}} G_{ii}^+ \ + \ \sum_{i \in {\rm L}} \left[\, G_{ii}^+ - F^{\rm L}_{ii}\,\right] \ + \ \sum_{i \in {\rm R}} \left[\, G_{ii}^+ - F^{\rm R}_{ii}\,\right] \, \Biggr) \nonumber\\ &=& - \, {2 \over \pi} \ \mbox{Im} \, \int_{-\infty}^0 {\rm d}\omega \ \mbox{Tr} \left[\, \mbox{\boldmath ${\cal G}$}^+ \, - \, \mbox{\boldmath ${\cal G}$}^+ \, {\partial \mbox{\boldmath ${\cal V}$}_{\rm mix}^+ \over \partial \omega} \,\right] \nonumber\\ &=& - \, {2 \over \pi} \ \mbox{Im} \, \int_{-\infty}^0 {\rm d}\omega \left(\, {\partial \over \partial \omega} \ \mbox{Tr} \left[\, \log \left\{ \mbox{\boldmath ${\cal G}$}^+ \right\}^{-1} \,\right] \ + \ \mbox{Tr} \left[\, \mbox{\boldmath ${\cal G}$}^+ \, {\partial \mbox{\boldmath $\Sigma$}^+ \over \partial \omega} \, \right] \, \right) . \label{eq:dn}\end{aligned}$$ Here we have used a matrix notation as that was used in the text eqs. (\[eq:81\])-(\[eq:86\]) for the multi-mode case: the Green’s function in the central region is denoted by $\mbox{\boldmath ${\cal G}$}(z) = \{G_{ij}(z)\}$ with $ij \in {\rm C}$ and the Dyson equation is given by $ \left\{\mbox{\boldmath ${\cal G}$}(z)\right\}^{-1} = z \, \mbox{\boldmath $1$} - \mbox{\boldmath ${\cal H}$}_{\rm C}^0 - \mbox{\boldmath ${\cal V}$}_{\rm mix}(z) - \mbox{\boldmath $\Sigma$}(z)$, where $\mbox{\boldmath ${\cal H}$}_{\rm C}^0 = \{ -t_{ij}^{\rm C}-\mu \,\delta_{ij} \}$, and $\mbox{\boldmath ${\cal V}$}_{\rm mix}(z)$ and $\mbox{\boldmath $\Sigma$}(z)$ are the single-mode version of eqs. (\[eq:87\]) and (\[eq:86\]), respectively. It has been shown that the contribution of the second term in the third line of eq. (\[eq:dn\]) vanishes.[@LuttingerWard; @Luttinger] Thus, eq. (\[eq:dn\]) can be rewritten as $$\begin{aligned} \Delta N_{\rm tot} &=& {1\over \pi {\rm i}}\, \log \left[ \det\left\{ \mbox{\boldmath ${\cal G}$}^-(0) \right\}^{-1} / \det\left\{ \mbox{\boldmath ${\cal G}$}^+(0) \right\}^{-1} \right] \label{eq:detS} \\ &=& {1\over \pi {\rm i}}\, \log \, \det\left[ \mbox{\boldmath $1$} + 2 {\rm i} \,\mbox{Im} \left\{ \mbox{\boldmath ${\cal V}$}_{\rm mix}^+(0) \right\} \mbox{\boldmath ${\cal G}$}^+(0) \right] . \label{eq:dn_det}\end{aligned}$$ For obtaining eq. (\[eq:dn\_det\]), we have used the Fermi-liquid property $\mbox{Im} \mbox{\boldmath $\Sigma$}^+(0)$. The matrix $\mbox{\boldmath ${\cal V}$}_{\rm mix}$ has only two non-zero elements as eq. (\[eq:87\]). Therefore, the determinant in eq. (\[eq:dn\_det\]) can be written in terms of a $2\times 2$ matrix as $$\begin{aligned} & &\Delta N_{\rm tot} = {1 \over \pi {\rm i}}\, \log [\, \det \mbox{\boldmath $S$} \,] \;, \label{eq:Friedel} \\ \nonumber \\ & & \mbox{\boldmath $S$} = \left [ \matrix { 1 & 0 \cr 0 & 1 \cr } \right ] - \, 2 {\rm i} \left [ \matrix { \Gamma_{\rm L}(0) & 0 \cr 0 & \Gamma_{\rm R}(0) \cr } \right ] \left [ \matrix { G_{1 1}^{+}(0) & G_{1 N}^{+}(0) \cr G_{N 1}^{+}(0) & G_{N N}^{+}(0) \cr } \right ] . \nonumber\\ \label{eq:S}\end{aligned}$$ [99]{} See, for instance, T. Ishiguro, K. Yamaji and G. Saito: [*Organic Superconductors Second Edition*]{}, (Springer-Verlag, Berlin, 1998). T. K. Ng and P. A. Lee: Phys. Rev. Lett. [**61**]{} (1988) 1768. L. I. Glazman and M. E. Raikh: Pis’ma Zh.  Eksp. Teor. Fiz. [**47**]{} (1988) 378, \[JETP Lett. [**47**]{} 452 (1988)\]. A. Kawabata: J. Phys. Soc. Jpn. [**60**]{} (1991) 3222. Y. Meir, N. S. Wingreen and P. A. Lee: Phys. Rev. Lett. [**66**]{} (1991) 3048; [*ibid.*]{} [**70**]{} (1993) 2601. S. Hershfield, J. H. Davies and J. W. Wilkins: Phys. Rev. B [**46**]{} (1992) 7046. D. C. Ralph and R. A. Buhrman: Phys. Rev. Lett. [**72**]{} (1994) 3401. D. Goldharber-Gordon, H. Shtrikman, D. Mahalu, D. Abusch-Magder, U. Meirav and M. A. Kastner: Nature [**391**]{} (1998) 156. S. M. Cronenwett, T. H. Oosterkamp and L. P. Kouwenhoven: Science [**281**]{} (1998) 540. T. H. Oosterkamp, T. Fujisawa, W.G. van der Wiel, K. Ishibashi, R. V. Hijman, S. Tarucha and L. P. Kouwenhoven: Nature [**395**]{} (1998) 873. Y. Tokura, D. G. Austing and S. Tarucha: J. Phys.: Condens. Matter [**56**]{} (1999) 6023. W. Izumida, O. Sakai and Y. Shimizu: J. Phys. Soc. Jpn. [**66**]{} (1997) 717; [*ibid.*]{} [**67**]{} (1998) 2444. W. Izumida and O. Sakai: Phys. Rev. B [**62**]{} (2000) 10260; W. Izumida, O. Sakai and S. Suzuki: J. Phys. Soc. Jpn. [**70**]{} (2001) 1045. O. Sakai, S. Suzuki, W. Izumida and A. Oguri: J. Phys. Soc. Jpn. [**68**]{} (1999) 1640. A. Oguri, H. Ishii and T. Saso: Phys. Rev. B [**51**]{} (1995) 4715. A. Oguri: Phys. Rev. B [**56**]{} (1997) 13422 \[Errata: [**58**]{} (1998) 1690\]. A. Oguri: Phys. Rev. B [**59**]{} (1999) 12240. A. Oguri: Phys. Rev. B [**63**]{} (2001) 115305 \[Errata: [**63**]{} (2001) 249901\]. A. Yeyati, A. Martín-Rodero and F. Flores: Phys. Rev. Lett. [**71**]{} (1993) 2991. T. Mii and K. Makoshi: Jpn. J. Appl. Phys. [**35**]{} (1996) 3706. O. Takagi and T. Saso: J. Phys. Soc. Jpn. [**68**]{} (1999) 1997. P. L. Pernas, F. Flores and E. V. Anda: J. Phys.: Condens. Matter [**4**]{} (1992) 5309. Y. Kawahito, H. Kasai, H. Nakanishi and A. Okiji: J. Appl. Phys. [**85**]{} (1999) 947. N. D. Lang and Ph. Avouris: Phys. Rev. Lett. [**84**]{} (1999) 358. P. W. Anderson: Phys. Rev. [**124**]{} (1961) 41. D. C. Licciardello and D. J. Thouless: J. Phys. C [**8**]{} (1975) 4157. A. MacKinnon and B. Kramer: Z. Phys. B [**53**]{} (1983) 1. S. V. Kravchenko, G. V. Kravchenko, J. E. Furneaux, V. M. Pudalov and M. D’Iorio: Phys. Rev. B [**50**]{} (1994) 8039. A. Oguri: J. Phys. Soc. Jpn. [**70**]{}, (2001) 2666. A. Oguri: Phys. Rev. B [**64**]{}, (2001) 153305. D. S. Fisher and P. A. Lee: Phys. Rev. B [**23**]{} (1981) 6851. P. A. Lee and D. S. Fisher: Phys. Rev.  Lett. [**47**]{} (1981) 882. J. S. Langer and V. Ambegaokar: Phys. Rev. [**121**]{} (1961) 1090. D. C. Langreth: Phys. Rev. [**150**]{} (1966) 516. H. Shiba: Prog. Theor. Phys. [**54**]{} (1975) 967. A. Oguri: Phys. Rev. B [**52**]{} (1995) 16727. J. M. Luttinger and J. C. Ward: Phys. Rev. [**118**]{} (1960) 1417. J. M. Luttinger: Phys. Rev. [**119**]{} (1960) 1153.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We prove an asymptotic formula for the number of scattering resonances in a strip near the real axis when the trapped set is $r$-normally hyperbolic with $r$ large and a pinching condition on the normal expansion rates holds. Our dynamical assumptions are stable under smooth perturbations and motivated by the setting of black holes. The key tool is a Fourier integral operator which microlocally projects onto the resonant states in the strip. In addition to Weyl law, this operator provides new information about microlocal concentration of resonant states.' address: 'Department of Mathematics, University of California, Berkeley, CA 94720, USA' author: - Semyon Dyatlov --- For a semiclassical Schrödinger operator $h^2\Delta_g+V(x)$, $V\in C^\infty(X;\mathbb R)$, on a compact Riemannian manifold $(X,g)$ the Weyl law (see for example [@d-sj Theorem 10.1]) provides an asymptotic for the number of eigenvalues $\lambda_j(h)$ as $h\to 0$: $$\label{e:weyl-compact} \#(\lambda_j(h)\in [\alpha_0,\alpha_1])=(2\pi h)^{-n}\big(\operatorname{Vol}_{\sigma}(p_V^{-1}([\alpha_0,\alpha_1]))+\mathcal O(h)\big).$$ Here $n$ is the dimension of $X$, $p_V(x,\xi)=|\xi|_{g}^2+V(x)$ is the (semiclassical) principal symbol of our Schrödinger operator, defined on the cotangent bundle $T^*X$, and $\operatorname{Vol}_\sigma$ is the symplectic volume on $T^*X$. A natural generalization of eigenvalues to noncompact manifolds are *resonances*, the poles of the meromorphic continuation of the resolvent to the lower half-plane $\{\Im\omega\leq 0\}\subset \mathbb C$, see  and §§\[s:framework-schrodinger\], \[s:framework-ah\]. However, there are very few results giving Weyl asymptotics of resonances in the style of . The first one is probably due to Regge [@regge], with some of the following results including [@zw-u; @sj-v; @sj-z-bands; @sj-new; @f-t] – see the discussion of related work below. This paper provides a new Weyl asymptotic formula for resonances, under the assumption that the trapped set is *$r$-normally hyperbolic* and expansion rates satisfy a pinching condition – see Theorems \[t:gaps\] and \[t:weyl-law\]. These dynamical assumptions are motivated by the study of black holes, see [@k-s]; this continues the previous work of the author [@skds; @xpd; @zeeman], and the application to stationary perturbations of Kerr–de Sitter black holes will be given in [@thesis]. See also [@GSWW] for applications of normally hyperbolic trapping to molecular dynamics. Because the imaginary part of a resonance can be interpreted as the exponential decay rate of the corresponding linear wave, we study *long-living resonances*, i.e. those in strips of size $Ch$ around the real axis. More precisely, we establish an asymptotic formula for the number of resonances in a band located between two resonance free strips. To illustrate the results, we consider here the setting of semiclassical Schrödinger operators on $X=\mathbb R^n$, studied in detail in §\[s:framework-schrodinger\]: $$\label{e:P-V} P_V:=h^2\Delta+V(x),\quad V\in C_0^\infty(\mathbb R^n;\mathbb R).$$ Here $\Delta=-\sum_j \partial_{x_j}^2$ is the Euclidean Laplacian. The results apply under the more general assumptions of §§\[s:framework-assumptions\] and \[s:dynamics\], in particular in the setting of even asymptotically hyperbolic manifolds – see §\[s:framework-ah\] and Appendix \[s:example\]. *Resonances* are the poles of the meromorphic continuation of the resolvent $$\label{e:r-v} R_V(\omega)=(P_V-\omega^2)^{-1}:L^2(\mathbb R^n)\to H^2(\mathbb R^n),\quad \Im\omega>0,$$ across the ray $(0,\infty)\subset\mathbb C$, as a family of operators $L^2_{\operatorname{comp}}(\mathbb R^n)\to H^2_{\operatorname{loc}}(\mathbb R^n)$. For the proofs, it is convenient to consider a different operator with the same set of poles $$\label{e:r-cal-omega} \mathcal R(\omega)=\mathcal P(\omega)^{-1}:\mathcal H_2\to\mathcal H_1,$$ where $\mathcal H_1=H^2_h(\mathbb R^n)$ is a semiclassical Sobolev space, $\mathcal H_2=L^2(\mathbb R^n)$, and $\mathcal P(\omega):\mathcal H_1\to\mathcal H_2$ is constructed from $P_V$ using the method of complex scaling (see §\[s:framework-schrodinger\]). To formulate dynamical assumptions, let $p_V(x,\xi)=|\xi|^2+V(x)$, fix energy intervals $[\alpha_0,\alpha_1]\Subset [\beta_0,\beta_1]\subset (0,\infty)$, put $p=\sqrt{p_V}$ on $p_V^{-1}([\beta_0^2,\beta_1^2])$ (see  for the general case) and define the *incoming/outgoing tails* $\Gamma_\pm$ and the *trapped set* $K$ as $$\Gamma_\pm:=\{\rho\in p_V^{-1}([\beta_0^2,\beta_1^2])\mid \exp(tH_p)(\rho)\not\to \infty\text{ as }t\to\mp\infty\},\quad K:=\Gamma_+\cap\Gamma_-.$$ Here $\exp(tH_{p})$ denotes the Hamiltonian flow of $p$. We assume that (see §\[s:dynamics\] for details) $\Gamma_\pm$ are sufficiently smooth codimension one submanifolds intersecting transversely at $K$, which is symplectic, and the flow is *$r$-normally hyperbolic* for large $r$ in the sense that the minimal expansion rate $\nu_{\min}$ of the flow $\exp(tH_p)$ in the directions transverse to $K$ is much greater than the maximal expansion rate $\mu_{\max}$ along $K$ – see , , . These assumptions are stable under small smooth perturbations of the symbol $p$, using the results of [@HPS] – see §\[s:stability\]. ![(a) An illustration of Theorem \[t:weyl-law\], with  counting resonances in the outlined box. The unshaded regions above and below the box are the resonance-free regions of Theorem \[t:gaps\]. (b) The canonical relation $\Lambda^\circ$, with the flow lines of $\mathcal V_\pm$ dashed.[]{data-label="f:Lambda"}](nhp.4 "fig:") ![(a) An illustration of Theorem \[t:weyl-law\], with  counting resonances in the outlined box. The unshaded regions above and below the box are the resonance-free regions of Theorem \[t:gaps\]. (b) The canonical relation $\Lambda^\circ$, with the flow lines of $\mathcal V_\pm$ dashed.[]{data-label="f:Lambda"}](nhp.3 "fig:") to Let $\nu_{\max}$ be the maximal expansion rate of the flow $\exp(tH_p)$ in the directions transverse to the trapped set, see . The following theorem provides a resonance free region with a polynomial resolvent bound: \[t:gaps\] Let the assumptions of §§\[s:framework-assumptions\] and \[s:dynamics\] hold and fix $\varepsilon>0$. Then for $$\label{e:omega-gaps} \Re\omega\in [\alpha_0,\alpha_1],\quad \Im\omega\in [-(\nu_{\min}-\varepsilon)h,0]\setminus \textstyle{1\over 2}(-(\nu_{\max}+\varepsilon)h,-(\nu_{\min}-\varepsilon)h),$$ $\omega$ is not a resonance and we have the bound[^1] $$\label{e:gaps} \|\mathcal R(\omega)\|_{\mathcal H_2\to\mathcal H_1}\leq Ch^{-2}.$$ In particular, we get a resonance free strip $\{\Im\omega>-{\nu_{\min}-\varepsilon\over 2}h\}$, recovering in our situation the results of [@ge-sj; @w-z; @n-z-new] – see below for a detailed discussion. Under the pinching condition $$\label{e:pinching} \nu_{\max}<2\nu_{\min},$$ we get a *second* resonance free strip $\{\Im\omega\in [-(\nu_{\min}-\varepsilon)h,-(\nu_{\max}+\varepsilon)h/2]\}$. We can then count the resonances in the band between the two strips, see Figure \[f:Lambda\](a): \[t:weyl-law\] Let the assumptions of §§\[s:framework-assumptions\] and \[s:dynamics\] and the condition  hold. Fix $\varepsilon>0$ such that $\nu_{\max}+\varepsilon<2(\nu_{\min}-\varepsilon)$. Then, with $\operatorname{Res}$ denoting the set of resonances counted with multiplicities (see ), $$\label{e:weyl-law} \begin{gathered} \#\big(\operatorname{Res}\cap \{\Re \omega\in [\alpha'_0,\alpha'_1],\ \Im\omega\in{\textstyle{1\over 2}}[-(\nu_{\max}+\varepsilon)h,-(\nu_{\min}-\varepsilon)h]\}\big)\\ =(2\pi h)^{1-n}(\operatorname{Vol}_\sigma(K\cap p^{-1}([\alpha'_0,\alpha'_1]))+o(1)), \end{gathered}$$ as $h\to 0$, for every $[\alpha'_0,\alpha'_1]\subset(\alpha_0,\alpha_1)$ such that $p^{-1}(\alpha'_j)\cap K$ has zero measure in $K$. Here $\operatorname{Vol}_\sigma$ denotes the symplectic volume on $K$, defined by $d\operatorname{Vol}_\sigma=\sigma_S^{n-1}/(n-1)!$. A band structure similar to the one exhibited in Theorems \[t:gaps\] and \[t:weyl-law\], with Weyl laws in each band, has been obtained in [@f-t] for a related setting of Anosov diffeomorphisms, see the discussion below. The key tool in proving Theorems \[t:gaps\] and \[t:weyl-law\] is a microlocal projector $\Pi$ corresponding to resonances in the band . We construct it as a *Fourier integral operator* (see §\[s:prelim-fio\]), associated to the canonical relation $\Lambda^\circ\subset T^*X\times T^*X$ defined as follows. Let $\mathcal V_\pm\subset T\Gamma_\pm$ be the symplectic complements of $T\Gamma_\pm$ in $T_{\Gamma_\pm}(T^*X)$. For some neighborhoods $\Gamma_\pm^\circ, K^\circ$ of $K\cap p^{-1}([\alpha_0,\alpha_1])$ in $\Gamma_\pm,K$, respectively, we can define the projections $\pi_\pm:\Gamma_\pm^\circ\to K^\circ$ along the flow lines of $\mathcal V_\pm$ – see §\[s:projections\]. We define (see also [@bfrz]) $$\label{e:intro-Lambda} \Lambda^\circ:=\{(\rho_-,\rho_+)\in \Gamma_-^\circ\times\Gamma_+^\circ\mid \pi_-(\rho_-)=\pi_+(\rho_+)\}.$$ Then $\Lambda^\circ$ is a canonical relation, see §\[s:projections\]; it is pictured on Figure \[f:Lambda\](b). We now construct an operator $\Pi$ with the following properties (see Theorem \[t:our-Pi\] in §\[s:construction-1\] for details, including a uniqueness statement): 1. \[c:1\] $\Pi$ is a compactly supported Fourier integral operator associated to $\Lambda^\circ$; 2. \[c:2\] $\Pi^2=\Pi+\mathcal O(h^\infty)$ microlocally near $K\cap p^{-1}([\alpha_0,\alpha_1])$; 3. \[c:3\] $[P,\Pi]=\mathcal O(h^\infty)$ microlocally near $K\cap p^{-1}([\alpha_0,\alpha_1])$. Here $P$ is a pseudodifferential operator equal to $\sqrt{P_V}$ microlocally in $p_V^{-1}([\beta_0^2,\beta_1^2])$ (see Lemma \[l:resolution\] for the general case). Conditions  and  mimic idempotency and commutation properties of spectral projectors of self-adjoint operators. The operator $\Pi$ is constructed iteratively, solving a degenerate transport equation on each step, with regularity of resulting functions guaranteed by $r$-normal hyperbolicity. The obtained operator provides a rich microlocal structure, which makes it possible to locally relate our situation to the Taylor expansion, ultimately proving Theorems \[t:gaps\] and \[t:weyl-law\]. See §\[s:ideas\] for a more detailed explanation of the ideas behind the proofs. A particular consequence of Theorem \[t:gaps\] is a resonance free strip $\{\Im\omega>-{\nu_{\min}-\varepsilon\over 2}h\}$. For *normally hyperbolic* trapped sets, such strips (also called spectral gaps) have been obtained by Gérard–Sjöstrand [@g-sj-gap] for operators with analytic coefficients and possibly non-smooth $\Gamma_\pm$; Wunsch–Zworski [@w-z] for sufficiently smooth $\Gamma_\pm$, without specifying the size of the gap; and Dolgopyat [@dolgopyat], Liverani [@liverani], and Tsujii [@tsujii] for contact Anosov flows. The recent preprint of Nonnenmacher and Zworski [@n-z-new] gives a gap of optimal size for a variety of normally hyperbolic trapped sets with very weak assumptions on the regularity of $\Gamma_\pm$; in our special case, the gap of [@n-z-new] coincides with the one given by Theorem \[t:gaps\]. For a related, yet quite different, case of *hyperbolic* trapped sets (where the flow is hyperbolic in all directions, but no assumptions are made on the regularity of $\Gamma_\pm$ and $K$), such gaps are known under a pressure condition, see [@n-z-acta] and the references given there. Upper bounds for the number of resonances in strips near the real axis have been proved in different situations, both for normally hyperbolic and for hyperbolic trapping, by Sjöstrand [@sj90], Guillopé–Lin–Zworski [@GLZ], Sjöstrand–Zworski [@sj-z], Nonnenmacher–Sjöstrand–Zworski [@n-sj-z; @n-sj-z2], Faure–Sjöstrand [@f-sj], Datchev–Dyatlov [@fwl], and Datchev–Dyatlov–Zworski [@ddz]; see [@n-sj-z2] or [@fwl] for a more detailed overview. The optimal known bounds follow the *fractal Weyl law*, $$\label{e:fwl} \#(\operatorname{Res}\cap \{\Re\omega\in[\alpha_0,\alpha_1],\ |\Im\omega|\leq C_0h\})\leq Ch^{-1-\delta}.$$ Here $C_0$ is any fixed number and $2\delta+2$ is bigger than the upper Minkowski dimension of the trapped set $K$ (inside $T^*X$), or equal to it if $K$ is of pure dimension. In our case, $\dim K=2n-2$, therefore the Weyl law  saturates the bound . Much less is known about lower bounds for hyperbolic or normally hyperbolic trapped sets – some special completely integrable cases were studied in particular by Gérard–Sjöstrand [@ge-sj], Sá Barreto–Zworski [@sb-z], and the author [@zeeman], a lower bound with a smaller power of $h^{-1}$ than  for certain hyperbolic surfaces was proved by Jakobson–Naud [@j-n], and Weyl laws have been established in some situations in [@sj-z-bands; @sj-v; @f-t; @f-t2; @f-t3] – see below. It has been conjectured [@non Definition 6.1] that for $C_0$ large enough, a lower bound matching  holds, but no such bound for non-integer $\delta$ has been proved so far. There also exists a Weyl asymptotic for surfaces with cusps, see Müller [@muller]; in this case, the infinite ends of the manifold are so narrow that almost all trajectories are trapped, and the Weyl law in strips coincides with the Weyl law in disks, with a power $h^{-n}$. Other Weyl asymptotics in large regions in the complex plane have been obtained by Zworski [@zw-u] for one-dimensional potential scattering and by Sjöstrand [@sj-new] for Schrödinger operators with randomly perturbed potentials. Finally, some situations where resonances form several bands of different depth were studied in [@sj-z-bands; @st-v; @sj-v; @f-t; @f-t2; @f-t3]. Sjöstrand–Zworski [@sj-z-bands] showed existence of cubic bands of resonances for strictly convex obstacles, under a pinching condition on the curvature, with a Weyl law in each band. Stefanov–Vodev [@st-v] studied the elasticity problem outside of a convex obstacle with Neumann boundary condition and showed existence of resonances $\mathcal O((\Re\omega)^{-\infty})$ close to the real line and a gap below this set of resonances; a Weyl law for resonances close to the real line was proved by Sjöstrand–Vodev [@sj-v]. A case bearing some similarities to the one considered here, namely contact Anosov diffeomorphisms, has been studied by Faure–Tsujii [@f-t]; their upcoming work [@f-t2; @f-t3] will handle contact Anosov flows – the latter can be put in the framework of §\[s:framework-assumptions\] using the work of Faure–Sjöstrand [@f-sj]. The results of [@f-t; @f-t2; @f-t3] for the dynamical setting include, under a pinching condition, the band structure of resonances (with the first band analogous to the one in Theorem \[t:weyl-law\]) and Weyl asymptotics in each band; the trapped set has to be normally hyperbolic, symplectic, and smooth, however the manifolds $\Gamma_\pm$ need only have Hölder regularity, and no assumption of $r$-normal hyperbolicity is made. These considerably weaker assumptions on regularity are crucial for Anosov flows and maps, as one cannot even expect $\Gamma_\pm$ to be $C^2$ in most cases. The lower regularity is in part handled by conjugating $\mathcal P(\omega)$ by the exponential of an escape function, similar to the one in [@ddz Lemma 4.2] – this reduces the analysis to an $\mathcal O(h^{1/2})$ sized neighborhood of the trapped set. It then suffices to construct only the principal part of the projector $\Pi$ to first order on the trapped set; such projector is uniquely defined locally on $K$ (by putting the principal symbol to be equal to 1 on $K$), without the need for the global construction of §\[s:construction-1\] or the transport equation . The present paper however was motivated by resonance expansions on perturbations of slowly rotating black holes, where the more restrictive $r$-normal hyperbolicity assumption is satisfied and it is important to have an operator $\Pi$ defined to all orders in $h$ and away, as well as on, the trapped set. Another advantage of such a global operator is the study of resonant states, see §\[s:resonant-states\]. Outline of the paper {#s:outline} ==================== In this section, we explain informally the ideas behind the construction of the projector $\Pi$ and the proofs of Theorems \[t:gaps\] and \[t:weyl-law\], list some directions in which the results could possibly be improved, and describe the structure of the paper. Ideas of the proofs and concentration of resonant states {#s:ideas} -------------------------------------------------------- An important tool is the model case (see §\[s:model\]) $$\label{e:intro-model} X=\mathbb R^n,\quad \Gamma_-^0=\{x_n=0\},\quad \Gamma_+^0=\{\xi_n=0\},\quad \Pi^0 f(x',x_n)=f(x',0).$$ Any operator satisfying properties  and  of $\Pi$ listed in the introduction can be microlocally conjugated to $\Pi^0$ (see Proposition \[l:reduce-model\] and part 2 of Proposition \[l:idempotents\]). However, there is no canonical way of doing this, and to construct $\Pi$ globally, we need to use property , which eventually reduces to solving the transport equation on $\Gamma_\pm$ $$\label{e:intro-transport} H_p a=f,\quad a|_K=0,$$ where $f$ is a given smooth function on $\Gamma_\pm$ with $f|_K=0$. The solution to  exists and is unique for any normally hyperbolic trapped set, by representing $a(\rho)$ as an exponentially converging integral of $f$ over the forward ($\Gamma_-$) or backward ($\Gamma_+$) flow line of $H_p$ starting at $\rho$. However, to know that $a$ lies in $C^r$ we need $r$-normal hyperbolicity (see Lemma \[l:ode\]). This explains why $r$-normal hyperbolicity, and not just normal hyperbolicity, is needed to construct the operator $\Pi$. The proof in §\[s:resolvent-bounds\] is based on positive commutator arguments, with additional microlocal structure coming from the projector $\Pi$ and the annihilating operators $\Theta_\pm$ discussed below. However, here we present a more intuitive (but harder to make rigorous) argument based on propagation by $$U(t)=e^{-itP/h},$$ which is a Fourier integral operator quantizing the Hamiltonian flow $e^{tH_p}$ (see Proposition \[l:schrodinger\]). Note that we use not the original operator $\mathcal P(\omega)$, but the operator $P$ constructed in Lemma \[l:resolution\], equal to $\sqrt{P_V}$ for the case ; this means that $U(t)$ is the wave, rather than the Schrödinger, propagator. We will only care about the behavior of $U(t)$ near the trapped set; for this purpose, we introduce a pseudodifferential cutoff $\mathcal X$ microlocalized in a neighborhood of $K$. For a family of functions $f=f(h)$ whose semiclassical wavefront set (as discussed in §\[s:prelim-basics\]) is contained in a small neighborhood of $K\cap p^{-1}([\alpha_0,\alpha_1])$, Theorem \[t:gaps\] follows from the following two estimates (a rigorous analog of  is Proposition \[l:estimate-kernel\], and of , Proposition \[l:estimate-image\]): for $t>0$, $$\begin{gathered} \label{e:ee-kernel} \|\mathcal X U(t)(1-\Pi)f\|_{L^2}\leq (Ch^{-1}e^{-(\nu_{\min}-\varepsilon/2)t}+\mathcal O(h^\infty))\|f\|_{L^2},\\ \label{e:ee-image} \begin{aligned} C^{-1}e^{-{(\nu_{\max}+\varepsilon/2)t\over 2}}\|\mathcal X\Pi f\|_{L^2}-\mathcal O(h^\infty)\|f\|_{L^2} \leq \|\mathcal X U(t) \Pi f\|_{L^2} \\\leq Ce^{-{(\nu_{\min}-\varepsilon/2)t\over 2}}\|\mathcal X\Pi f\|_{L^2}+\mathcal O(h^\infty)\|f\|_{L^2}. \end{aligned}\end{gathered}$$ The estimates  and  are of independent value, as they give information about the long time behavior of solutions to the wave equation, resembling resonance expansions of linear waves; an application to black holes will be given in [@thesis]. Note however that these estimates are nontrivial only when $t=\mathcal O(\log(1/h))$, because of the $\mathcal O(h^\infty)$ error term. The resonance free region  of Theorem \[t:gaps\] is derived from here as follows. Assume that $\omega$ is a resonance in . Then there exists a *resonant state*, namely a function $u\in\mathcal H_1$ such that $\mathcal P(\omega)u=0$ and $\|u\|_{\mathcal H_1}\sim 1$. We formally have $U(t)u=e^{-it\omega/h}u$. Also, $u$ is microlocalized on the outgoing tail $\Gamma_+$, which is propagated by the flow $e^{tH_p}$ towards infinity; this means that if $f:=\mathcal X_1 u$ for a suitably chosen pseudodifferential cutoff $\mathcal X_1$, then $\Pi u=\Pi f+\mathcal O(h^\infty)$ and for $t>0$, $$U(t)f=e^{-it\omega/h}f+\mathcal O(h^\infty)\quad\text{microlocally near }\WFh(\mathcal X).$$ Since $\Pi$ commutes with $P$ modulo $\mathcal O(h^\infty)$, it also commutes with $U(t)$, which gives $$\begin{gathered} \mathcal X U(t)(1-\Pi)f=e^{-it\omega/h}\mathcal X(1-\Pi)f+\mathcal O(h^\infty),\\ \mathcal X U(t)\Pi f=e^{-it\omega/h}\mathcal X\Pi f+\mathcal O(h^\infty). \end{gathered}$$ Since $\Im\omega\geq-(\nu_{\min}-\varepsilon)h$, we take $t=N\log(1/h)$ for arbitrarily large constant $N$ in  to get $\|\mathcal X (1-\Pi)f\|_{L^2}=\mathcal O(h^\infty)$. Since $\Im\omega\not\in (-(\nu_{\max}+\varepsilon)h/2,-(\nu_{\min}-\varepsilon)h/2)$, by  we get $\|\mathcal X \Pi f\|_{L^2}=\mathcal O(h^\infty)$. Together, they give $\|\mathcal X f\|_{L^2}=\mathcal O(h^\infty)$, implying by standard outgoing estimates (see Lemma \[l:smart-bound\]) that $\|u\|_{\mathcal H_1}=\mathcal O(h^\infty)$, a contradiction. We now give an intuitive explanation for  and . We start by considering the model case , with the pseudodifferential cutoff $\mathcal X$ replaced by the multiplication operator by some $\chi\in C_0^\infty(\mathbb R^n)$. For the operator $P$, we consider the model (somewhat inappropriate since the actual Hamiltonian vector field $H_p$ is typically nonvanishing on $K$, contrary to the model case, but reflecting the nature of the flow in the transverse directions) $P=x_n\cdot hD_{x_n}-ih/2$; here the term $-ih/2$ makes $P$ symmetric. We then have in the model case, $p=x_n\xi_n$, $e^{tH_p}(x,\xi)=(x',e^tx_n,\xi',e^{-t}\xi_n)$, $\nu_{\min}=\nu_{\max}=1$, and $$U(t)f(x',x_n)=e^{-t/2}f(x',e^{-t}x_n).$$ Then  (in fact, a better estimate with $e^{-3t/2}$ in place of $e^{-t}$ – see the possible improvements subsection below) follows by Taylor expansion at $x_n=0$. More precisely, we use the following form of this expansion: for $f\in C_0^\infty(\mathbb R^n)$, $$\label{e:intro-Xi} (1-\Pi^0) f=x_n\cdot g,\quad g(x',x_n):={f(x',x_n)-f(x',0)\over x_n},$$ and one can show that $\|g\|_{L^2}\leq Ch^{-1}\|f\|_{H^1_h}$, the factor $h^{-1}$ coming from taking one nonsemiclassical derivative to obtain $g$ from $f$ (see Lemma \[l:Xi-0\]). Then $\chi U(t)(1-\Pi^0)f=\chi U(t)x_n U(-t)U(t)g$, where (by a special case of Egorov’s theorem following by direct computation) $\chi U(t)x_n U(-t)$ is a multiplication operator by $$\label{e:zz-kernel} \chi U(t)x_n U(-t)=\chi(x)e^{-t}x_n=\mathcal O(e^{-t});$$ this shows that $\|\chi U(t)(1-\Pi^0)f\|_{L^2}\leq Ce^{-t}\|g\|_{L^2} \leq Ch^{-1}e^{-t}\|f\|_{H^1_h}$ and  follows. To show  in the model case, we start with the identity $$\|\chi U(t)\Pi^0 f\|_{L^2}=\|\chi_t\Pi^0 f\|_{L^2},\quad \chi_t:=U(-t)\chi U(t).$$ If $\chi\in C_0^\infty(\mathbb R^n)$, then $\chi_t(x)=\chi(x',e^tx_n)$ has shrinking support as $t\to\infty$. To compare $\|\chi_t\Pi^0 f\|_{L^2}$ to $\|\chi\Pi^0 f\|_{L^2}$, we use the following fact: $$\label{e:zz-image} hD_{x_n}\Pi^0 f=0.$$ This implies that for each $a(x)\in C_0^\infty(\mathbb R^n)$, the inner product $\langle a\Pi^0 f,\Pi^0 f\rangle$ depends only on the function $b(x')=\int_{\mathbb R}a(x',x_n)\,dx_n$; writing $\|\chi\Pi^0 f\|_{L^2}^2$ and $\|\chi_t\Pi^0 f\|_{L^2}^2$ as inner products, we get $\|\chi_t\Pi^0 f\|_{L^2}^2=e^{-t}\|\chi \Pi^0 f\|_{L^2}^2$ and  follows. The proofs of  and  in the general case work as in the model case, once we find appropriate replacements for differential operators $x_n$ and $hD_{x_n}$ in  and . It turns out that one needs to take pseudodifferential operators $\Theta_\pm$ solving, microlocally near $K\cap p^{-1}([\alpha_0,\alpha_1])$, $$\label{e:intro-ideals} \Pi\Theta_-=\mathcal O(h^\infty),\quad \Theta_+\Pi=\mathcal O(h^\infty),$$ then $\Theta_-$ is a replacement for $x_n$ and $\Theta_+$, for $hD_{x_n}$. Note that $\Theta_\pm$ are not unique, in fact solutions to  form one-sided ideals in the algebra of pseudodifferential operators – see §§\[s:ideals\] and \[s:construction-2\]. The principal symbols of $\Theta_\pm$ are defining functions of $\Gamma_\pm$. As a byproduct of the discussion above, we obtain new information about microlocal concentration of resonant states, that is, functions $u\in\mathcal H_1$ such that $\mathcal P(\omega)u=0$ and $\|u\|_{\mathcal H_1}\sim 1$. It is well-known (see for example [@n-z-acta Theorem 4]) that the wavefront set of $u$ is contained in $\Gamma_+\cap p^{-1}(\Re\omega)$. The new information we obtain is that if $\omega$ is a resonance in the band given by Theorem \[t:weyl-law\] (that is, $\Im\omega>-(\nu_{\min}-\varepsilon)h$), then by , $u=\Pi u+\mathcal O(h^\infty)$ microlocally near $K$. Then by , $\Theta_+u=\mathcal O(h^\infty)$ near $K$, that is, $u$ solves a pseudodifferential equation; note that the Hamiltonian flow lines of the principal symbol of $\Theta_+$ are transverse to the trapped set. This implies in particular that any corresponding semiclassical defect measure is determined uniquely by a measure on the trapped set which is conditionally invariant under $H_p$, similarly to the damped wave equation. See Theorem \[t:resonant-states\] in §\[s:resonant-states\] for details. We start with constructing a well-posed Grushin problem, representing resonances as zeroes of a certain Fredholm determinant $F(\omega)$. Using complex analysis (essentially the argument principle), we reduce counting resonances to computing a contour integral of the logarithmic derivative $F'(\omega)/F(\omega)$, which, taking $\nu_-=-(\nu_{\max}+\varepsilon)/2$, $\nu_+=-(\nu_{\min}-\varepsilon)/2$, is similar to (see §\[s:trace\] for the actual expression) $${1\over 2\pi i}(\mathcal I_--\mathcal I_+),\quad \mathcal I_\pm:=\int_{\Im\omega=h\nu_\pm} \tilde\chi(\omega)\operatorname{Tr}(\Pi \mathcal R(\omega))\,d\omega$$ for some cutoff function $\tilde\chi(\omega)$. The integration is over the region where Theorem \[t:gaps\] gives polynomial bounds on the resolvent $\mathcal R(\omega)$, and we can use the methods developed for the proof of this theorem to evaluate both integrals, yielding Theorem \[t:weyl-law\]. An important additional tool, explaining in particular why the two integrals do not cancel each other, is microlocal analysis in the spectral parameter $\omega$, or equivalently a study of the essential support of the Fourier transform of $\Pi\mathcal R(\omega)$ in $\omega$ – see §§\[s:estimate-spectral\] and \[s:trace\]. Possible improvements {#s:improvements} --------------------- First of all, it would be interesting to see if one could construct further bands of resonances, lying below the one in Theorem \[t:weyl-law\]. One expects these bands to have the form $$\{\Im\omega\in [-(k+1/2)(\nu_{\max}+\varepsilon)h,-(k+1/2)(\nu_{\min}-\varepsilon)h]\},\quad k\in \mathbb Z,\ k\geq 0,$$ and to have a Weyl law in the $k$-th band under the pinching condition $(k+1/2)\nu_{\max}<(k+3/2)\nu_{\min}$. Note that the presence of the second band of resonances improves the size of the second resonance free strip in Theorem \[t:gaps\] and gives a weaker pinching condition $\nu_{\max}<3\nu_{\min}$ for the Weyl law in the first band. The proofs are expected to work similarly to the present paper, if one constructs a family of operators $\Pi_0=\Pi,\Pi_1,\dots,\Pi_k$ such that $\Pi_j$ is $h^{-j}$ times a Fourier integral operator associated to $\Lambda^\circ$, $\Pi_j\Pi_k=\mathcal O(h^\infty)$, and $[P,\Pi_j]=\mathcal O(h^\infty)$ (microlocally near $K\cap p^{-1}([\alpha_0,\alpha_1])$). However, the method of §\[s:construction-1\] does not apply directly to construct $\Pi_k$ for $k>0$, since one cannot conjugate all $\Pi_j$ to the model case, which is the base of the crucial Proposition \[l:idempotents\]. Another direction would be to consider the case when the operator $P$ is quantum completely integrable on the trapped set (a notion that needs to be made precise), and derive a quantization condition for resonances like the one for the special case of black holes [@sb-z; @zeeman]. The author also believes that the results of the present paper should be adaptable to the situation when $\Gamma_\pm$ have codimension higher than 1, which makes it possible to revisit the distribution of resonances generated by one closed hyperbolic trajectory, studied in [@ge-sj]. An interesting special case lying on the intersection of the current work and [@f-t; @f-t2; @f-t3] is given by geodesic flows on compact manifolds of constant negative curvature; the corresponding manifolds $\Gamma_\pm$ and $K$ are smooth in this situation. While $r$-normal hyperbolicity does not hold (in fact, $\mu_{\max}=\nu_{\min}=\nu_{\max}$), the rigid algebraic structure of hyperbolic quotients suggests that one could still look for the projector $\Pi$ as a (smooth) Fourier integral operator – in terms of the construction of §\[s:construction-1\], the transport equation , while not yielding a smooth solution for an arbitrary choice of the right-hand side $f$, will have a smooth solution for the specific functions $f$ arising in the construction. Finally, a natural question is improving the $o(1)$ remainder in the Weyl law . Obtaining an $\mathcal O(h^\delta)$ remainder for $\delta<1$ does not seem to require conceptual changes to the microlocal structure of the argument; however, for the $\mathcal O(h)$ remainder of Hörmander [@hohoho] or the $o(h)$ remainder of Duistermaat–Guillemin [@du-gu], one would need a finer analysis of the interaction of the operator $\Pi$ with the Schrödinger propagator, and more assumptions on the flow on the trapped set might be needed. Moreover, the complex analysis argument of §\[s:weyl-law\] does not work in the case of an $\mathcal O(h)$ remainder; a reasonable replacement would be to adapt to the considered case the work of Sjöstrand [@sj-dwe] on the damped wave equation. Structure of the paper {#s:structure} ---------------------- - In §\[s:prelim\], we review the tools we need from semiclassical analysis. - In §\[s:framework\], we present a framework which makes it possible to handle resonances and the spatial infinity in an abstract fashion. The assumptions we make are listed in §\[s:framework-assumptions\], followed by some useful lemmas (§\[s:framework-corollaries\]) and applications to Schrödinger operators (§\[s:framework-schrodinger\]) and even asymptotically hyperbolic manifolds (§\[s:framework-ah\]). - In §\[s:nh\], we study $r$-normally hyperbolic trapped sets, stating the dynamical assumptions (§\[s:dynamics\]), discussing their stability under perturbations (§\[s:stability\]), and deriving some corollaries (§§\[s:phi-pm\]–\[s:transport\]). - In §\[s:calculus\], we study in detail Fourier integral operators associated to $\Lambda^\circ$, and in particular properties of operators solving $\Pi^2=\Pi+\mathcal O(h^\infty)$. - In §\[s:global-construction\], we construct the projector $\Pi$ and the annihilating operators $\Theta_\pm$. - In §\[s:resolvent-bounds\], we prove Theorem \[t:gaps\], establish microlocal estimates on the resolvent, and study the microlocal concentration of resonant states (§\[s:resonant-states\]). - In §\[s:grushin\], we formulate a well-posed Grushin problem for $\mathcal P(\omega)$, representing resonances as zeroes of a certain Fredholm determinant. - In §\[s:trace\], we prove a trace formula for $\mathcal R(\omega)$ microlocally on the image of $\Pi$. - In §\[s:weyl-law\], we prove the Weyl asymptotic for resonances (Theorem \[t:weyl-law\]). - In Appendix \[s:example\], we provide an example of an asymptotically hyperbolic manifold satisfying the dynamical assumptions of §\[s:dynamics\]. Semiclassical preliminaries {#s:prelim} =========================== In this section, we review semiclassical pseudodifferential operators, wavefront sets, and Fourier integral operators; the reader is directed to [@e-z; @d-sj] for a detailed treatment and [@ho3; @ho4; @gr-sj] for the closely related microlocal case. Pseudodifferential operators and microlocalization {#s:prelim-basics} -------------------------------------------------- Let $X$ be a manifold without boundary. Following [@e-z §9.3 and 14.2], we consider the symbol classes $S^k(T^*X)$, $k\in\mathbb R$, consisting of smooth functions $a$ on the cotangent bundle $T^*X$ satisfying in local coordinates $$\sup_h\sup_{x\in K}|\partial^\alpha_x\partial^\beta_\xi a(x,\xi;h)|\leq C_{\alpha\beta K}\langle\xi\rangle^{k-|\beta|},$$ for each multiindices $\alpha,\beta$ and each compact set $K\subset X$. The corresponding class of semiclassical pseudodifferential operators is denoted $\Psi^k(X)$. The residual symbol class $h^\infty S^{-\infty}$ consists of symbols decaying rapidly in $h$ and $\xi$ over compact subsets of $X$; the operators in the corresponding class $h^\infty\Psi^{-\infty}$ have Schwartz kernels in $h^\infty C^\infty(X\times X)$. Operators in $\Psi^k$ are bounded, uniformly in $h$, between the semiclassical Sobolev spaces $H^s_{h,\operatorname{comp}}(X)\to H^{s-k}_{h,\operatorname{loc}}(X)$, see [@e-z (14.2.3)] for the definition of the latter. Note that for noncompact $X$, we impose no restrictions on the behavior of symbols as $x\to\infty$. Accordingly, we cannot control the behavior of operators in $\Psi^k(X)$ near spatial infinity; in fact, a priori we only require them to act $C_0^\infty(X)\to C^\infty(X)$ and on the spaces of distributions $\mathcal E'(X)\to \mathcal D'(X)$. However, each $A\in\Psi^k(X)$ can be written as the sum of an $h^\infty\Psi^{-\infty}$ remainder and an operator properly supported uniformly in $h$ – see for example [@ho3 Proposition 18.1.22]. Properly supported pseudodifferential operators act $C_0^\infty\to C_0^\infty$ and $C^\infty\to C^\infty$ and therefore can be multiplied with each other, giving an algebra structure on the whole $\Psi^k$, modulo $h^\infty\Psi^{-\infty}$. To study the behavior of symbols near fiber infinity, we use the fiber-radial compactified cotangent bundle $\overline T^*X$, a manifold with boundary whose interior is diffeomorphic to $T^*X$ and whose boundary $\partial\overline T^*X$ is diffeomorphic to the cosphere bundle over $X$ – see for example [@v1 §2.2]. We will restrict ourselves to the space of *classical symbols*, i.e. those having an asymptotic expansion $$a(x,\xi;h)\sim\sum_{j\geq 0}h^ja_j(x,\xi),$$ with $a_j\in S^{k-j}$ classical in the sense that $\langle\xi\rangle^{j-k}a_j$ extends to a smooth function on $\overline T^*X$. The principal symbol $\sigma(A):=a_0\in S^k$ of an operator is defined independently of quantization. We say that $A\in\Psi^k$ is elliptic at some $(x,\xi)\in\overline T^*X$ if $\langle \xi\rangle^{-k}\sigma(A)$ does not vanish at $(x,\xi)$. Another invariant object associated to $A\in\Psi^k(X)$ is its wavefront set $\WFh(A)$, which is a closed subset of $\overline T^*X$; a point $(x,\xi)\in \overline T^*X$ does not lie in $\WFh(A)$ if and only if there exists a neighborhood $U$ of $(x,\xi)$ in $\overline T^*X$ such that the full symbol of $A$ (in any quantization) is in $h^\infty S^{-\infty}$ in this neighborhood. Note that $\WFh(A)=\emptyset$ if and only if $A=\mathcal O(h^\infty)_{\Psi^{-\infty}}$. We say that $A_1=A_2+\mathcal O(h^\infty)$ *microlocally* in some $U\subset\overline T^*X$, if $\WFh(A-B)\cap U=\emptyset$. We denote by $\Psi^{\operatorname{comp}}(X)$ the space of all operators $A\in\Psi^0(X)$ such that $\WFh(A)$ is a compact subset of $T^*X$, in particular not intersecting the fiber infinity $\partial \overline T^*X$. Note that $\Psi^{\operatorname{comp}}(X)\subset \Psi^k(X)$ for all $k\in \mathbb R$. Let $u=u(h)$ be an $h$-dependent family of distributions in $\mathcal D'(X)$. We say that $u$ is $h$-tempered (or polynomially bounded), if for each $\chi\in C_0^\infty(X)$, there exists $N$ such that $\|\chi u\|_{H^{-N}_h}=\mathcal O(h^{-N})$. The class of $h$-tempered distributions is closed under properly supported pseudodifferential operators. For an $h$-tempered $u$, define the wavefront set $\WFh(u)$, a closed subset of $\overline T^*X$, as follows: $(x,\xi)\in \overline T^*X$ does not lie in $\WFh(u)$ if and only if there exists a neighborhood $U$ of $(x,\xi)$ in $\overline T^*X$ such that for each properly supported $A\in\Psi^0(X)$ with $\WFh(A)\subset U$, we have $Au=\mathcal O(h^\infty)_{C^\infty}$. We have $\WFh(u)=\emptyset$ if and only if $u=\mathcal O(h^\infty)_{C^\infty}$. We say that $u=v+\mathcal O(h^\infty)$ microlocally on some $U\subset \overline T^*X$ if $\WFh(u-v)\cap U=\emptyset$. Let $X_1$ and $X_2$ be two manifolds. An operator $B:C_0^\infty(X_1)\to\mathcal D'(X_2)$ is identified with its Schwartz kernel $\mathcal K_B(y,x)\in\mathcal D'(X_2\times X_1)$: $$\label{e:op-kernel} Bf(y)=\int_{X_1} \mathcal K_B(y,x)u(x)\,dx,\quad u\in \mathcal C_0^\infty(X_1).$$ Here we assume that $X_1$ is equipped with some smooth density $dx$; later, we will also assume that densities on our manifolds are specified when talking about adjoints. We say that $B$ is $h$-tempered if $\mathcal K_B$ is, and define the wavefront set of $B$ as $$\label{e:op-wf} \WFh(B):=\{(x,\xi,y,\eta)\in \overline T^*(X_1\times X_2)\mid (y,\eta,x,-\xi)\in\WFh(K_B)\}.$$ If $B\in\Psi^k(X)$, then the wavefront set of $B$ as an $h$-tempered operator is equal to its wavefront set as a pseudodifferential operator, under the diagonal embedding $\overline T^*X\to \overline T^*(X\times X)$. Lagrangian distributions and Fourier integral operators {#s:prelim-fio} ------------------------------------------------------- We now review the theory of Lagrangian distributions; for details, the reader is directed to [@e-z Chapters 10–11], [@gu-st1 Chapter 6], or [@svn §2.3], and to [@ho4 Chapter 25] or [@gr-sj Chapters 10–11] for the closely related microlocal setting. Here, we only present the relatively simple local part of the theory; geometric constructions of invariant symbols will be done by hand when needed, without studying the structure of the bundles obtained (see §\[s:general\]). For a more complete discussion, see for example [@qeefun §3]. A semiclassical Lagrangian distribution locally takes the form $$\label{e:lagrangian} u(x;h)=(2\pi h)^{-m/2} \int_{X\times\mathbb R^m} e^{{i\over h}\Phi(x,\theta)}a(x,\theta;h)\,d\theta.$$ Here $\Phi$ is a nondegenerate phase function, i.e. a real-valued function defined on an open subset of $X\times \mathbb R^m$, for some $m$, such that the differentials $d(\partial_{\theta_1}\Phi),\dots,d(\partial_{\theta_m}\Phi)$ are linearly independent on the critical set $$\mathcal C_\Phi := \{(x,\theta)\mid \partial_\theta \Phi(x,\theta)=0\}.$$ The amplitude $a(x,\theta;h)$ is a classical symbol (that is, having an asymptotic expansion in nonnegative integer powers of $h$ as $h\to 0$) compactly supported inside the domain of $\Phi$. The resulting function $u(x;h)$ is smooth, compactly supported, $h$-tempered, and$$\label{e:lagrangian-wf} \WFh(u)\subset \{(x,\partial_x\Phi(x,\theta))\mid (x,\theta)\in \mathcal C_\Phi\cap\operatorname{supp}a\}.$$ We say that $\Phi$ generates the (immersed, and we shrink the domain of $\Phi$ to make it embedded) Lagrangian submanifold $$\Lambda_\Phi:=\{(x,\partial_x\Phi(x,\theta))\mid (x,\theta)\in \mathcal C_\Phi\};$$ note that $\WFh(u)\subset\Lambda_\Phi$. Moreover, if we restrict $\Phi$ to $\mathcal C_\Phi$ and pull it back to $\Lambda_\Phi$, then $d\Phi$ equals the canonical 1-form $\xi\,dx$ on $\Lambda_\Phi$. In general, assume that $\Lambda$ is an embedded Lagrangian submanifold of $T^*X$ which is moreover *exact* in the sense that the canonical form $\xi\,dx$ is exact on $\Lambda$; we fix an *antiderivative* on $\Lambda$, namely a function $F$ such that $\xi\,dx=dF$ on $\Lambda$. (This is somewhat similar to the notion of Legendre distributions, see [@m-z §11].) Then we say that a compactly supported $h$-tempered family of distributions $u$ is a (compactly microlocalized) Lagrangian distribution associated to $\Lambda$, if $u$ can be written as a finite sum of expressions , with phase functions $\Phi_j$ generating open subsets of $\Lambda$, plus an $\mathcal O(h^\infty)_{C_0^\infty}$ remainder, where $\Phi_j$ are normalized (by adding a constant) so that the pull-back to $\Lambda$ of the restriction of $\Phi_j$ to $\mathcal C_{\Phi_j}$ equals $F$. (Without such normalization, passing from one phase function to the other produces a factor $e^{is\over h}$ for some constant $s$, which does not preserve the class of classical symbols – this is an additional complication of the theory compared to the nonsemiclassical case.) Denote by $I_{\operatorname{comp}}(\Lambda)$ the class of all Lagrangian distributions associated to $\Lambda$. For $u\in I_{\operatorname{comp}}(\Lambda)$, we have $\WFh(u)\subset\Lambda$; in particular, $\WFh(u)$ does not intersect the fiber infinity $\partial\overline T^*X$. If now $X_1,X_2$ are two manifolds of dimensions $n_1,n_2$ respectively, and $\Lambda\subset T^*X_1\times T^*X_2$ is an exact canonical relation (with some fixed antiderivative), then an operator $B:C^\infty(X_1)\to C_0^\infty(X_2)$ is called a (compactly microlocalized) Fourier integral operator associated to $\Lambda$, if its Schwartz kernel $\mathcal K_B(y,x)$ is $h^{-(n_1+n_2)/4}$ times a Lagrangian distribution associated to $$\{(y,\eta,x,-\xi)\in T^*(X_1\times X_2)\mid (x,\xi,y,\eta)\in\Lambda\}.$$ We write $B\in I_{\operatorname{comp}}(\Lambda)$; note that $\WFh(B)\subset\Lambda$. A particular case is when $\Lambda$ is the graph of a canonical transformation $\varkappa:U_1\to U_2$, with $U_j$ open subsets in $T^*X_j$. Operators associated to canonical transformations (but not general relations!) are bounded $H^s_h\to H^{s'}_h$ uniformly in $h$, for each $s,s'$. Compactly microlocalized Fourier integral operators associated to the identity transformation are exactly compactly supported pseudodifferential operators in $\Psi^{\operatorname{comp}}(X)$. Another example of Fourier integral operators is given by Schrödinger propagators, see for instance [@e-z Theorem 10.4][^2] or [@qeefun Proposition 3.8]: \[l:schrodinger\] Assume that $P\in\Psi^{\operatorname{comp}}(X)$ is compactly supported, $\WFh(P)$ is contained in some compact subset $V\subset T^*X$, and $p=\sigma(P)$ is real-valued. Then for $t\in\mathbb R$ bounded by any fixed constant, the operator $e^{-itP/h}:L^2(X)\to L^2(X)$ is the sum of the identity and a compactly supported operator microlocalized in $V\times V$. Moreover, for each compactly supported $A\in\Psi^{\operatorname{comp}}(X)$, $Ae^{-itP/h}$ and $e^{-itP/h}A$ are smooth families of Fourier integral operators associated to the Hamiltonian flow $e^{tH_p}:T^*X\to T^*X$. Here we put the antiderivative $F$ for the identity transformation to equal zero, and extend it to the antiderivative $F_t$ on the graph of $e^{tH_p}$ by putting $$F_t(\gamma(0),\gamma(t)):=tp(\gamma(0))-\int_{\gamma([0,t])}\xi\,dx$$ for each flow line $\gamma$ of $H_p$. The corresponding phase function is produced by a solution to the Hamilton–Jacobi equation [@e-z Lemma 10.5]. We finally discuss products of Fourier integral operators. Assume that $B_j\in I_{\operatorname{comp}}(\Lambda_j)$, $j=1,2$, where $\Lambda_1\subset T^*X_1\times T^*X_2$ and $\Lambda_2\subset T^*X_2\times T^*X_3$ are exact canonical relations. Assume moreover that $\Lambda_1,\Lambda_2$ satisfy the following transversality assumption: the manifolds $\Lambda_1\times\Lambda_2$ and $T^*X_1\times \Delta(T^*X_2)\times T^*X_3$, where $\Delta(T^*X_2)\subset T^*X_2\times T^*X_2$ is the diagonal, intersect transversely inside $T^*X_1\times T^*X_2\times T^*X_2\times T^*X_3$, and their intersection projects diffeomorphically onto $T^*X_1\times T^*X_3$. Then $B_2B_1\in I_{\operatorname{comp}}(\Lambda_2\circ\Lambda_1)$, where $$\label{e:comp-fio} \Lambda_2\circ\Lambda_1:=\{(\rho_1,\rho_3)\mid \exists \rho_2\in T^*X_2: (\rho_1,\rho_2)\in\Lambda_1,\ (\rho_2,\rho_3)\in\Lambda_2\},$$ and, if $F_j$ is the antiderivative on $\Lambda_j$, then $F_1(\rho_1,\rho_2)+F_2(\rho_2,\rho_3)$ is the antiderivative on $\Lambda_2\circ\Lambda_1$. See for example [@ho4 Theorem 25.2.3] or [@gr-sj Theorem 11.12] for the closely related microlocal case, which is adapted directly to the semiclassical situation. The transversality condition is always satisfied when at least one of the $\Lambda_j$ is the graph of a canonical transformation. In particular, one can always multiply a pseudodifferential operator by a Fourier integral operator, and obtain a Fourier integral operator associated to the same canonical relation. Basic estimates --------------- In this section, we review some standard semiclassical estimates, parametrices, and microlocalization statements. Throughout the section, we assume that $k,s\in\mathbb R$, $P,Q\in\Psi^k(X)$ are properly supported and $u,f$ are $h$-tempered distributions on $X$, in the sense of §\[s:prelim-basics\]. We start with the elliptic estimate, see for instance [@zeeman Proposition 2.2]: (Elliptic estimate) \[l:elliptic\] Assume that $Pu=f$. Then: 1\. If $A,B\in\Psi^0(X)$ are compactly supported and $P,B$ are elliptic on $\WFh(A)$, then $$\label{e:ell-est} \|Au\|_{H^s_h}\leq C\|Bf\|_{H^{s-k}_h}+\mathcal O(h^\infty).$$ 2\. We have $$\label{e:ell-wf} \WFh(u)\subset \WFh(f)\cup \{\langle\xi\rangle^{-k}\sigma(P)=0\}.$$ Proposition \[l:elliptic\] is typically proved using the following fact, which is of independent interest: (Elliptic parametrix) \[l:eparametrix\] If $V\subset \overline T^*X$ is compact and $P$ is elliptic on $V$, then there exists a compactly supported operator $P'\in\Psi^{-k}(X)$ such that $PP'=1+\mathcal O(h^\infty),P'P=1+\mathcal O(h^\infty)$ microlocally near $V$. Moreover, $\sigma(P')=\sigma(P)^{-1}$ near $V$. We next give a version of propagation of singularities which allows for a complex absorbing operator $Q$, see for instance [@v1 §2.3]: (Propagation of singularities) \[l:microhyperbolic\] Assume that $\sigma(P)$ is real-valued, $\sigma(Q)\geq 0$, and $(P\pm iQ)u=f$. Then: 1\. If $A_1,A_2,B\in\Psi^0(X)$ are compactly supported and for each flow line $\gamma(t)$ of the Hamiltonian field $\pm\langle\xi\rangle^{1-k}H_{\sigma(P)}$ such that $\gamma(0)\in\WFh(A_1)$, there exists $t\geq 0$ such that $A_2$ is elliptic at $\gamma(t)$ and $B$ is elliptic on the segment $\gamma([0,t])$, then $$\label{e:mhp-est} \|A_1u\|_{H^s_h}\leq C\|A_2u\|_{H^s_h}+Ch^{-1}\|Bf\|_{H^{s-k+1}_h}+\mathcal O(h^\infty).$$ 2\. If $\gamma(t)$, $0\leq t\leq T$, is a flow line of $\pm\langle\xi\rangle^{1-k}H_{\sigma(P)}$, then $$\gamma([0,T])\cap \WFh(f)=\emptyset,\ \gamma(T)\not\in\WFh(u)\ \Longrightarrow\ \gamma(0)\not\in\WFh(u).$$ For $Q=0$, Proposition \[l:microhyperbolic\] can be viewed as a microlocal version of uniqueness of solutions to the Cauchy problem for hyperbolic equations; a corresponding microlocal existence fact is given by (Hyperbolic parametrix) \[l:hparametrix\] Assume that $\sigma(P)$ is real-valued, $\WFh(f)\subset T^*X$ is compact, $U,V\subset T^*X$ are compactly contained open sets, and for each flow line $\gamma(t)$ of the Hamiltonian field $H_{\sigma(P)}$ such that $\gamma(0)\in\WFh(f)$, there exists $t\in\mathbb R$ such that $\gamma(t)\in U$ and $\gamma(s)\in V$ for all $s$ between $0$ and $t$. Then there exists an $h$-tempered family $v(h)\in C_0^\infty(X)$ such that $\WFh(v)\subset V$ and $$\|v\|_{L^2}\leq Ch^{-1}\|f\|_{L^2},\quad \|Pv\|_{L^2}\leq C\|f\|_{L^2},\quad \WFh(Pv-f)\subset U.$$ By applying a microlocal partition of unity to $f$, we may assume that there exists $T>0$ (the case $T<0$ is considered similarly and the case $T=0$ is trivial by putting $v=0$) such that for each flow line $\gamma(t)$ of $H_{\sigma(P)}$ such that $\gamma(0)\in\WFh(f)$, we have $\gamma(T)\in U$ and $\gamma([0,T])\in V$. Take $\varepsilon\in (0,T)$ such that $\gamma([T-\varepsilon,T])\subset U$ for each such $\gamma$. Since $V$ is compactly contained in $T^*X$, we may assume that $P$ is compactly supported and $P\in\Psi^{\operatorname{comp}}(X)$. We then take $\chi\in C_0^\infty(-\infty,T)$ such that $\chi=1$ near $[0,T-\varepsilon]$ and put $$v:={i\over h}\int_0^T \chi(t)e^{-itP/h}f\,dt.$$ Then $\|v\|_{L^2}\leq Ch^{-1}\|f\|_{L^2}$ and $\WFh(v)\subset V$ by Proposition \[l:schrodinger\]. Integrating by parts, we compute $$Pv=-\int_0^T \chi(t)\partial_t e^{-itP/h}f\,dt=f+\int_0^T(\partial_t\chi(t))e^{-itP/h}f\,dt;$$ therefore, $\|Pv\|_{L^2}\leq C\|f\|_{L^2}$ and by Proposition \[l:schrodinger\], $\WFh(Pv-f)\subset U$. We also need the following version of the sharp Garding inequality, see [@e-z Theorem 4.32] or [@skds Proposition 5.2]: (Sharp Garding inequality) \[l:garding\] Assume that $A\in\Psi^{\operatorname{comp}}(X)$ is compactly supported and $\Re\sigma(A)\geq 0$ near $\WFh(u)$. Assume also that $B\in\Psi^{\operatorname{comp}}(X)$ is compactly supported and elliptic on $\WFh(A)\cap\WFh(u)$. Then $$\Re\langle Au,u\rangle\geq -Ch\|Bu\|_{L^2}^2-\mathcal O(h^\infty).$$ Abstract framework near infinity {#s:framework} ================================ In this section, we provide an abstract microlocal framework for studying resonances; the general assumptions are listed in §\[s:framework-assumptions\]. Rather than considering resonances as poles of the meromoprhic continuation of the cutoff resolvent, we define them as solutions of a nonselfadjoint eigenvalue problem featuring a holomorphic family of Fredholm operators, $\mathcal P(\omega)$. We assume that the dependence of the principal symbol of $\mathcal P(\omega)$ on $\omega$ can be resolved in a convex neighborhood $\mathcal U$ of the trapped set, yielding the $\omega$-independent symbol $p$ (and the operator $P$ later in Lemma \[l:resolution\]). Finally, we require the existence of a semiclassically outgoing parametrix for $\mathcal P(\omega)$, resolving it modulo an operator microlocalized near the trapped set. In §\[s:framework-corollaries\], we derive several useful corollaries of our assumptions, making it possible to treat spatial infinity as a black box in the following sections. Finally, in §§\[s:framework-schrodinger\] and \[s:framework-ah\], we provide two examples of situations when the assumptions of §\[s:framework-assumptions\] (but not necessarily the dynamical assumptions of §\[s:dynamics\]) are satisfied: Schrödinger operators on $\mathbb R^n$, studied using complex scaling, and Laplacians on even asymptotically hyperbolic manifolds, handled using [@v1; @v2]. General assumptions {#s:framework-assumptions} ------------------- Assume that: 1. \[a:basic\] $X$ is a smooth $n$-dimensional manifold without boundary, possibly noncompact, with a prescribed volume form; 2. $\mathcal P(\omega)\in\Psi^k(X)$ is a family of properly supported semiclassical pseudodifferential operators depending holomorphically on $\omega$ lying in an open simply connected set $\Omega\subset\mathbb C$ such that $\mathbb R\cap\Omega$ is connected, with principal symbol $\mathbf p(x,\xi,\omega)$; 3. \[a:spaces\] $\mathcal H_1,\mathcal H_2$ are $h$-dependent Hilbert spaces such that $H^{N}_{h,\operatorname{comp}}(X)\subset \mathcal H_j\subset H^{-N}_{h,\operatorname{loc}}(X)$ for some $N$, with norms of embeddings $\mathcal O(h^{-N})$, and $\mathcal P(\omega)$ is bounded $\mathcal H_1\to\mathcal H_2$ with norm $\mathcal O(1)$; 4. \[a:fredholm\] for some fixed $[\alpha_0,\alpha_1]\subset \mathbb R\cap\Omega$ and $C_0>0$, the operator $\mathcal P(\omega):\mathcal H_1\to\mathcal H_2$ is Fredholm of index zero in the region $$\label{e:omega-region} \Re\omega\in [\alpha_0,\alpha_1],\quad |\Im\omega|\leq C_0h.$$ Together with invertibility of $\mathcal P(\omega)$ in a subregion of  proved in Theorem \[t:gaps\], by Analytic Fredholm Theory [@e-z Theorem D.4] our assumptions imply that $$\label{e:resolvent} \mathcal R(\omega):=\mathcal P(\omega)^{-1}:\mathcal H_2\to\mathcal H_1$$ is a meromorphic family of operators with poles of finite rank for $\omega$ satisfying . *Resonances* are defined as poles of $\mathcal R(\omega)$. Following [@go-si Theorem 2.1], we define the multiplicity of a resonance $\omega_0$ as $$\label{e:multiplicity} {1\over 2\pi i}\operatorname{Tr}\oint_{\omega_0} \mathcal P(\omega)^{-1}\partial_\omega\mathcal P(\omega)\,d\omega.$$ Here $\oint_{\omega_0}$ stands for the integral over a contour enclosing $\omega_0$, but no other poles of $\mathcal R(\omega)$. Since $\mathcal R(\omega)$ has poles of finite rank, we see that the integral in  yields a finite dimensional operator on $\mathcal H_1$ and thus one can take the trace. The fact that the resulting multiplicity is a positive integer will follow for example from the representation of resonances as zeroes of a Fredholm determinant, in part 1 of Proposition \[l:grushin-ultimate\]. See also [@sj-dwe Appendix A]. We next fix a ‘physical region’ $\mathcal U$ in phase space, where most of our analysis will take place, in particular the intersection of the trapped set with the relevant energy shell will be contained in $\mathcal U$. The region $\mathcal U$ will be contained in a larger region $\mathcal U'$, which is used to determine when trajectories have escaped from $\mathcal U$. (See  and  for the definitions of $\mathcal U,\mathcal U'$ for the examples we consider.) We assume that: 1. \[a:u’\] $\mathcal U'\subset T^*X$ is open and bounded, and each compactly supported $A\in\Psi^{\operatorname{comp}}(X)$ with $\WFh(A)\subset\mathcal U'$ is bounded $L^2\to\mathcal H_j, \mathcal H_j\to L^2$, $j=1,2$, with norm $\mathcal O(1)$; 2. \[a:s-a\] $\mathcal P(\omega)^*= \mathcal P(\omega)+\mathcal O(h^\infty)$ microlocally in $\mathcal U'$ when $\omega\in \mathbb R\cap \Omega$; 3. \[a:p-res\] for each $(x,\xi)\in \mathcal U'$, the equation $\mathbf p(x,\xi,\omega)=0$, $\omega\in\Omega$ has unique solution $$\label{e:p} \omega=p(x,\xi).$$ Moreover, $p(x,\xi)\in\mathbb R$ and $\partial_\omega \mathbf p(x,\xi,p(x,\xi))<0$ for $(x,\xi)\in \mathcal U'$; 4. \[a:U\] $\mathcal U\subset \mathcal U'$ is a compactly contained open subset, whose closure $\overline{\mathcal U}$ is relatively convex with respect to the Hamiltonian flow of $p$, i.e. if $\gamma(t),0\leq t\leq T$, is a flow line of $H_p$ in $\mathcal U'$ and $\gamma(0),\gamma(T)\in \overline{\mathcal U}$, then $\gamma([0,T])\subset \overline{\mathcal U}$; Note that for $\omega\in \mathbb R\cap\Omega$, Hamiltonian flow lines of $p$ in $\mathcal U'\cap p^{-1}(\omega)$ are rescaled Hamiltonian flow lines of $\mathbf p(\,\cdot\,,\omega)$ in $\{\rho\in\mathcal U'\mid \mathbf p(\rho,\omega)=0\}$. The symbol $p$ is typically the square root of the principal symbol of the original Laplacian or Schrödinger operator, see  and . We can now define the *incoming/outgoing tails* $\Gamma_\pm\subset\overline{\mathcal U}$ as follows: $\rho\in\overline{\mathcal U}$ lies in $\Gamma_\pm$ if and only if $e^{\mp tH_p}(\rho)$ stays in $\overline{\mathcal U}$ for all $t\geq 0$. Define the *trapped set* as $$\label{e:trapped-set} K:=\Gamma_+\cap\Gamma_-.$$ Note that $\Gamma_\pm$ and $K$ are closed subsets of $\overline{\mathcal U}$ (and thus the sets $\Gamma_\pm$ defined here are smaller than the original $\Gamma_\pm$ defined in the introduction), and $e^{tH_p}(\Gamma_\pm)\subset\Gamma_\pm$ for $\mp t\geq 0$, thus $e^{tH_p}(K)=K$ for all $t$. We assume that, with $\alpha_0,\alpha_1$ defined in , 1. \[a:k-compact\] $K\cap p^{-1}([\alpha_0,\alpha_1])$ is a nonempty compact subset of $\mathcal U$. Finally, we assume the existence of a semiclassically outgoing parametrix, which will make it possible to reduce our analysis to a neighborhood of the trapped set in §\[s:framework-corollaries\]: 1. \[a:parametrix\] $\mathcal Q\in\Psi^{\operatorname{comp}}(X)$ is compactly supported, $\WFh(\mathcal Q)\subset \mathcal U$, and the operator $$\label{e:r'} \mathcal R'(\omega):=(\mathcal P(\omega)-i \mathcal Q)^{-1}:\mathcal H_2\to \mathcal H_1$$ satisfies, for $\omega$ in , $$\label{e:r'-bound} \|\mathcal R'(\omega)\|_{\mathcal H_2\to \mathcal H_1}\leq Ch^{-1};$$ 2. \[a:outgoing\] for $\omega$ in , $\mathcal R'(\omega)$ is *semiclassically outgoing* in the following sense: if $(\rho,\rho')\in\WFh(\mathcal R'(\omega))$ and $\rho,\rho'\in \mathcal U'$, there exists $t\geq 0$ such that $e^{tH_p}(\rho)=\rho'$ and $e^{sH_p}(\rho)\in \mathcal U'$ for $0\leq s\leq t$. (See Figure \[f:outgoing\](a) below.) Some consequences of general assumptions {#s:framework-corollaries} ---------------------------------------- In this section, we derive several corollaries of the assumptions of §\[s:framework-assumptions\], used throughout the rest of the paper. We start with two technical lemmas: \[l:the-flow\] Assume that $\rho\in\Gamma_\pm$. Then as $t\to\mp\infty$, the distance $d(e^{tH_p}(\rho),K)$ converges to zero. We consider the case $\rho\in\Gamma_-$. Put $\gamma(t):=e^{tH_p}(\rho)$, then $\gamma(t)\in\Gamma_-$ for all $t\geq 0$. Assume that $d(\gamma(t),K)$ does not converge to zero as $t\to+\infty$, then there exists a sequence of times $t_j\to +\infty$ such that $\gamma(t_j)$ does not lie in a fixed neighborhood of $K$. By passing to a subsequence, we may assume that $\gamma(t_j)$ converge to some $\rho_\infty\in\Gamma_-\setminus K$. Then $\rho_\infty\not\in\Gamma_+$; therefore, there exists $T\geq 0$ such that $e^{-TH_p}(\rho_\infty)\not\in\overline{\mathcal U}$. For $j$ large enough, we have $\gamma(t_j-T)=e^{-TH_p}(\gamma(t_j))\not\in\overline{\mathcal U}$ and $t_j\geq T$; this contradicts convexity of $\overline{\mathcal U}$ (assumption ). \[l:the-flow-2\] Assume that $U_1$ is a neighborhood of $K$ in $\overline{\mathcal U}$. Then there exists a neighborhood $U_2$ of $K$ in $\overline{\mathcal U}$ such that for each flow line $\gamma(t)$, $0\leq t\leq T$ of $H_p$ in $\overline{\mathcal U}$, if $\gamma(0),\gamma(T)\in U_2$, then $\gamma([0,t])\subset U_1$. Assume the contrary, then there exist flow lines $\gamma_j(t)$, $0\leq t\leq T_j$, in $\overline{\mathcal U}$, such that $d(\gamma_j(0),K)\to 0$, $d(\gamma_j(T_j),K)\to 0$, yet $\gamma_j(t_j)\not\in U_1$ for some $t_j\in [0,T_j]$. Passing to a subsequence, we may assume that $\gamma_j(t_j)\to\rho_\infty\in\overline {\mathcal U}\setminus K$. Without loss of generality, we assume that $\rho_\infty\not\in\Gamma_+$. Then there exists $T>0$ such that $e^{-TH_p}(\rho_\infty)\in\mathcal U'\setminus\overline{\mathcal U}$, and thus $e^{-TH_p}(\gamma_j(t_j))\not\in\overline{\mathcal U}$ for $j$ large enough. Since $\gamma_j([0,T_j])\subset \overline{\mathcal U}$, we have $t_j\leq T$. By passing to a subsequence, we may assume that $t_j\to t_\infty\in [0,T]$. However, then $\gamma_j(0)\to e^{-t_\infty H_p}(\rho_\infty)$, which implies that $e^{-t_\infty H_p}(\rho_\infty)\in \Gamma_+$, contradicting the fact that $\rho_\infty\not\in \Gamma_+$. We reduce the operator $\mathcal P(\omega)$ microlocally near $\mathcal U$ to an operator of the form $P-\omega$, see also [@isz §4]: \[l:resolution\] There exist: - a compactly supported $P\in\Psi^{\operatorname{comp}}(X)$ such that $P^*=P$ and $\sigma(P)=p$ near $\mathcal U$, where $p$ is defined in , and - a family of compactly supported operators $\mathcal S(\omega)\in\Psi^{\operatorname{comp}}(X)$, holomorphic in $\omega\in\Omega$, with $\mathcal S(\omega)^*=\mathcal S(\omega)$ for $\omega\in\mathbb R\cap\Omega$ and $\mathcal S(\omega)$ elliptic near $\mathcal U$, such that $$\label{e:resolution} \mathcal P(\omega)=\mathcal S(\omega)(P-\omega)\mathcal S(\omega)+\mathcal O(h^\infty) \quad\text{microlocally near }\mathcal U.$$ We argue by induction, constructing compactly supported operators $P_j,\mathcal S_j(\omega)\in\Psi^{\operatorname{comp}}(X)$, such that $P_j^*=P_j$, $\mathcal S_j^*(\omega)=\mathcal S_j(\omega)$ for $\omega\in\mathbb R\cap\Omega$, and $\mathcal P(\omega)=\mathcal S_j(\omega)(P_j-\omega)\mathcal S_j(\omega)+\mathcal O(h^{j+1})$ microlocally near $\mathcal U$. It will remain to take the asymptotic limit. For $j=0$, it suffices to take any $P_0,\mathcal S_0(\omega)$ such that $\sigma(P_0)=p$ and $\sigma(\mathcal S_0(\omega))(\rho)=s_0(\rho,\omega)$ near $\mathcal U$, where (with $\mathbf p(\cdot,\omega)$ denoting the principal symbol of $\mathcal P(\omega)$) $$\mathbf p(\rho,\omega)=s_0(\rho,\omega)^2(p(\rho)-\omega),\quad \rho\in \mathcal U';$$ the existence of such $s_0$ and the fact that it is real-valued for real $\omega$ follows from assumption . Now, given $P_j,\mathcal S_j(\omega)$ for some $j\geq 0$, we construct $P_{j+1},\mathcal S_{j+1}(\omega)$. We have $\mathcal P(\omega)=\mathcal S_j(\omega)(P_j-\omega)\mathcal S_j(\omega)+h^{j+1} R_j(\omega)$ microlocally near $\mathcal U$, where $R_j(\omega)\in\Psi^{\operatorname{comp}}$ is a holomorphic family of operators and, by assumption , $R_j(\omega)^*=R_j(\omega)+\mathcal O(h^\infty)$ microlocally near $\mathcal U$ when $\omega\in\mathbb R\cap\Omega$. We then put $P_{j+1}=P_j+h^{j+1}A_j$, $\mathcal S_{j+1}(\omega)=\mathcal S_j(\omega)+h^{j+1}B_j(\omega)$, where $\sigma(A_j)=p_j,\sigma(B_j(\omega))(\rho)=s_j(\rho,\omega)$ near $\mathcal U$ and $$\sigma(R_j)(\rho,\omega)=2s_0(\rho,\omega)s_j(\rho,\omega)(p(\rho)-\omega) +s_0(\rho,\omega)^2p_j(\rho),\quad \rho\in\mathcal U'.$$ The existence of $s_j(\rho,\omega),p_j(\rho)$ and the fact that $p_j(\rho)\in\mathbb R$ and $s_j(\rho,\omega)\in\mathbb R$ for $\rho$ near $\mathcal U$ and $\omega\in\mathbb R\cap\Omega$ follow from assumption . In particular, we put $p_j(\rho)=\sigma(R_j)(\rho,p(\rho))/s_0(\rho,p(\rho))^2$. Note that, if $u(h)\in \mathcal H_1,f(h)\in \mathcal H_2$ have norms polynomially bounded in $h$ (and in light of assumption  are $h$-tempered in the sense of §\[s:prelim-basics\]), and $\mathcal P(\omega)u=f$, then $$\label{e:conjugated} (P-\omega)\mathcal S(\omega)u=\mathcal S'(\omega)f+\mathcal O(h^\infty)\quad \text{microlocally near }\mathcal U,$$ where $\mathcal S'(\omega)\in\Psi^{\operatorname{comp}}(X)$ is an elliptic parametrix of $\mathcal S(\omega)$ microlocally near $\mathcal U$, constructed in Proposition \[l:eparametrix\]. ![(a) Assumption , with the undashed part of the flow line of $\rho$ corresponding to $\rho'\in\mathcal U'$ such that $(\rho,\rho')\in\WFh(\mathcal R'(\omega))$.(b) An illustration of Lemma \[l:propagate-outgoing\], with $\WFh(f)$ the shaded set and $\WFh(u)$ containing undashed parts of the flow lines.[]{data-label="f:outgoing"}](nhp.6 "fig:") ![(a) Assumption , with the undashed part of the flow line of $\rho$ corresponding to $\rho'\in\mathcal U'$ such that $(\rho,\rho')\in\WFh(\mathcal R'(\omega))$.(b) An illustration of Lemma \[l:propagate-outgoing\], with $\WFh(f)$ the shaded set and $\WFh(u)$ containing undashed parts of the flow lines.[]{data-label="f:outgoing"}](nhp.7 "fig:") to Next, we use the semiclassically outgoing parametrix $\mathcal R'(\omega)$ from  to derive a key restriction on the wavefront set of functions in the image of $\mathcal R(\omega)$, see Figure \[f:outgoing\](b): \[l:propagate-outgoing\] Assume that $u(h)\in \mathcal H_1,f(h)\in \mathcal H_2$ have norms polynomially bounded in $h$, $\mathcal P(\omega)u=f$ for some $\omega=\omega(h)$ satisfying , and $\WFh(f)\subset \mathcal U$. Then for each $\rho\in\WFh(u)\cap \mathcal U$, if $\gamma(t)=e^{tH_p}(\rho)$ is the corresponding maximally extended flow line in $\mathcal U'$, then either $\gamma(t)\in \overline{\mathcal U}$ for all $t\leq 0$ or $\gamma(t)\in\WFh(f)$ for some $t\leq 0$. By propagation of singularities (Proposition \[l:microhyperbolic\]) applied to , we see that either $\gamma(t)\in\overline{\mathcal U}$ for all $t\leq 0$, or $\gamma(t)\in\WFh(f)$ for some $t\leq 0$, or there exists $t\leq 0$ such that $\gamma(t)\in\WFh(u)\cap (\mathcal U'\setminus \overline{\mathcal U})$; we need to exclude the third case. However, in this case by convexity of $\overline{\mathcal U}$ (assumption ), $\gamma(t-s)\not\in\overline{\mathcal U}$ for all $s\geq 0$; by assumption , and since $u=\mathcal R'(\omega)(f-i \mathcal Qu)$ with $\WFh(f-i\mathcal Q u)\subset \mathcal U$, we see that $\gamma(t)\not\in\WFh(u)$, a contradiction. It follows from Lemma \[l:propagate-outgoing\] that any resonant state, i.e. a function $u$ such that $\|u\|_{\mathcal H_1}\sim 1$ and $\mathcal P(\omega)u=0$, has to satisfy $\WFh(u)\cap \mathcal U\subset\Gamma_+$. The next statement improves on the parametrix $\mathcal R'(\omega)$, inverting the operator $\mathcal P(\omega)$ outside of any given neighborhood of the trapped set. One can see this as a geometric control statement (see for instance [@b-z Theorem 3]). \[l:smart-parametrix\] Let $W\subset \mathcal U$ be a neighborhood of $K\cap p^{-1}([\alpha_0,\alpha_1])$ (which is a compact subset of $\mathcal U$ by assumption ), and assume that $f(h)\in\mathcal H_2$ has norm bounded polynomially in $h$ and each $\omega=\omega(h)$ is in . Then there exists $v(h)\in\mathcal H_1$, with $f-\mathcal P(\omega)v$ compactly supported in $X$ and $$\|v\|_{\mathcal H_1}\leq Ch^{-1}\|f\|_{\mathcal H_2},\quad \|\mathcal P(\omega)v\|_{\mathcal H_2}\leq C\|f\|_{\mathcal H_2},\quad \WFh(f-\mathcal P(\omega)v)\subset W.$$ First of all, take compactly supported $\mathcal Q'\in\Psi^{\operatorname{comp}}(X)$ such that $\WFh(\mathcal Q')\subset \mathcal U$ and $\mathcal Q'=1$ microlocally near $\WFh(\mathcal Q)$ (with $\mathcal Q$ defined in assumption ), and put $$v_1:=(1-\mathcal Q')\mathcal R'(\omega)f.$$ Then by , $\|v_1\|_{\mathcal H_1}\leq Ch^{-1}\|f\|_{\mathcal H_2}$ and $\mathcal P(\omega)v_1=f_1$, where $$f_1=(1-\mathcal Q'-[\mathcal P(\omega),\mathcal Q']\mathcal R'(\omega)+ (1-\mathcal Q')i \mathcal Q \mathcal R'(\omega))f.$$ Since $(1-\mathcal Q')i \mathcal Q=\mathcal O(h^\infty)_{\Psi^{-\infty}}$, by  we find $\|f_1\|_{\mathcal H_2}\leq C\|f\|_{\mathcal H_2}$, $f-f_1$ is compactly supported, and $\WFh(f-f_1)\subset \WFh(\mathcal Q')$. It is now enough to prove our statement for $f-f_1$ in place of $f$; therefore, we may assume that $f$ is compactly supported and $$\WFh(f)\subset \WFh(\mathcal Q').$$ Since $\WFh(\mathcal Q')$ is compact, by a microlocal partition of unity we may assume that $\WFh(f)$ is contained in a small neighborhood of some fixed $\rho\in\WFh(\mathcal Q')\subset\mathcal U$. We now consider three cases: **Case 1**: $\rho\not\in p^{-1}([\alpha_0,\alpha_1])$. Then the operator $\mathcal P(\omega)$ is elliptic at $\rho$, therefore we may assume it is elliptic on $\WFh(f)$. The function $v$ is then obtained by applying to $f$ an elliptic parametrix of $\mathcal P(\omega)$ given in Proposition \[l:eparametrix\]; we have $f-\mathcal P(\omega)v=\mathcal O(h^\infty)_{C_0^\infty}$. **Case 2**: $\rho\in\Gamma_-\cap p^{-1}([\alpha_0,\alpha_1])$. By Lemma \[l:the-flow\], there exists $t\geq 0$ such that $e^{tH_p}(\rho)\in W$. We may then assume that $e^{tH_p}(\WFh(f))\subset W$, and $v$ is then constructed by Proposition \[l:hparametrix\], using ; we have $\WFh(v)\subset\mathcal U$ and $\WFh(f-\mathcal P(\omega) v)\subset W$. **Case 3**: $\rho\not\in\Gamma_-$. Then there exists $t\geq 0$ such that $e^{tH_p}(\rho)\in\mathcal U'\setminus\overline{\mathcal U}$. As in case 2, subtracting from $v$ the parametrix of Proposition \[l:hparametrix\], we may assume that $f$ is instead microlocalized in a neighborhood of $e^{tH_p}(\rho)$. Now, put $v=\mathcal R'(\omega)f$, with $\mathcal R'(\omega)$ defined in ; then $\|v\|_{\mathcal H_1}\leq Ch^{-1}\|f\|_{\mathcal H_2}$ by  and $$f-\mathcal P(\omega)v=-i \mathcal Qv.$$ However, by assumption , and by convexity of $\overline{\mathcal U}$ (assumption ), we have $\WFh(\mathcal Q)\cap \WFh(v)=\emptyset$ and thus $f-\mathcal P(\omega) v=\mathcal O(h^\infty)_{C_0^\infty}$. Finally, we can estimate the norm of $u\in\mathcal H_1$ by the norm of $\mathcal P(\omega)u$ and the norm of $u$ microlocally near the trapped set. This can be viewed as an observability statement (see for instance [@b-z Theorem 2]). \[l:smart-bound\] Let $A\in\Psi^{\operatorname{comp}}(X)$ be compactly supported and elliptic on $K\cap p^{-1}([\alpha_0,\alpha_1])$. Then we have for any $u\in \mathcal H_1$ and any $\omega$ in , $$\label{e:smart-bound} \|u\|_{\mathcal H_1}\leq C\|Au\|_{L^2}+Ch^{-1}\|\mathcal P(\omega)u\|_{\mathcal H_2}.$$ By rescaling, we may assume that $u=u(h)$ has $\|u\|_{\mathcal H_1}=1$ and put $f=\mathcal P(\omega)u$. Take a neighborhood $W$ of $K\cap p^{-1}([\alpha_0,\alpha_1])$ such that $A$ is elliptic on $W$. Replacing $u$ by $u-v$, where $v$ is constructed from $f$ in Lemma \[l:smart-parametrix\], we may assume that $\WFh(f)\subset W$. Take $\mathcal Q',\mathcal Q''\in\Psi^{\operatorname{comp}}(X)$ compactly supported, with $\WFh(\mathcal Q'')\subset\mathcal U$, $\mathcal Q''=1+\mathcal O(h^\infty)$ microlocally near $\WFh(\mathcal Q')$, and $\mathcal Q'=1+\mathcal O(h^\infty)$ microlocally near $\WFh(\mathcal Q)$ (with $\mathcal Q$ defined in assumption ). Then by the elliptic estimate (Proposition \[l:elliptic\]), $$\begin{aligned} \label{e:toronto-1} \|\mathcal Q' u\|_{\mathcal H_1}&\leq C\|\mathcal Q''u\|_{L^2}+\mathcal O(h^\infty),\\ \label{e:toronto-2} \|[\mathcal P(\omega),\mathcal Q']u\|_{\mathcal H_2}&\leq Ch\|\mathcal Q'' u\|_{L^2} +\mathcal O(h^\infty).\end{aligned}$$ Now, $$(1-\mathcal Q')u=\mathcal R'(\omega)((1-Q')f-[\mathcal P(\omega),\mathcal Q']u-i\mathcal Q(1-\mathcal Q')u);$$ since $i\mathcal Q(1-\mathcal Q')=\mathcal O(h^\infty)_{\Psi^{-\infty}}$, we get by  and , $$\|(1-\mathcal Q')u\|_{\mathcal H_1}\leq C\|\mathcal Q'' u\|_{L^2}+Ch^{-1}\|f\|_{\mathcal H_2} +\mathcal O(h^\infty);$$ by , it then remains to prove that $$\|\mathcal Q'' u\|_{L^2}\leq C\|Au\|_{L^2}+Ch^{-1}\|f\|_{\mathcal H_2}+\mathcal O(h^\infty).$$ By a microlocal partition of unity, it suffices to estimate $\|Bu\|_{L^2}$ for $B\in\Psi^{\operatorname{comp}}(X)$ compactly supported with $\WFh(B)$ in a small neighborhood of some $\rho\in\WFh(Q'')\subset\mathcal U$. We now consider three cases: **Case 1**: $\rho\not\in p^{-1}([\alpha_0,\alpha_1])$. Then $\mathcal P(\omega)$ is elliptic at $\rho$, therefore we may assume it is elliptic on $\WFh(B)$. By Proposition \[l:elliptic\], we get $\|Bu\|_{L^2}\leq C\|f\|_{\mathcal H_2}+\mathcal O(h^\infty)$. **Case 2**: there exists $t\leq 0$ such that $e^{tH_p}(\rho)\in W$, therefore we may assume that $e^{tH_p}(\WFh(B))\subset W$. Since $A$ is elliptic on $W$, by Proposition \[l:microhyperbolic\] together with , we get $\|Bu\|_{L^2}\leq C\|Au\|_{L^2}+Ch^{-1}\|f\|_{\mathcal H_2}+\mathcal O(h^\infty)$. **Case 3**: if $\gamma(t)=e^{tH_p}(\rho)$ is the maximally extended trajectory of $H_p$ in $\mathcal U'$, then $\rho\in p^{-1}([\alpha_0,\alpha_1])$ and $\gamma(t)\not\in W$ for all $t\leq 0$. By Lemma \[l:the-flow\], we have $\rho\not\in\Gamma_+$. Since $\WFh(f)\subset W$, Lemma \[l:propagate-outgoing\] implies that $\rho\not\in\WFh(u)$. We may then assume that $\WFh(B)\cap\WFh(u)=\emptyset$ and thus $\|Bu\|_{L^2}=\mathcal O(h^\infty)$. Example: Schrödinger operators on Rn {#s:framework-schrodinger} ------------------------------------ In this section, we consider the case described the introduction, namely a Schrödinger operator on $X=\mathbb R^n$ with $$P_V=h^2\Delta+V(x),$$ where $\Delta$ is the Euclidean Laplacian and $V\in C_0^\infty(\mathbb R^n;\mathbb R)$. We will explain how this case fits into the framework of §\[s:framework-assumptions\]. To define resonances for $P_0$, we use the method of *complex scaling* of Aguilar–Combes [@ag-co], which also applies to more general operators and potentials – see [@sj-z-91], [@sj], and the references given there. Take $R>0$ large enough so that $$\operatorname{supp}V\subset \{|x|<R/2\}.$$ Fix the deformation angle $\theta\in (0,\pi/2)$ and consider a deformation $\Gamma_{\theta,R}\subset\mathbb C^n$ of $\mathbb R^n$ defined by $$\Gamma_{\theta,R}:=\{x+iF_{\theta,R}(x)\mid x\in\mathbb R^n\},$$ where $F_{\theta,R}:\mathbb R^n\to\mathbb R^n$ is defined in polar coordinates $(r,\varphi)\in [0,\infty)\times \mathbb S^{n-1}$ by $$F_{\theta,R}(r,\varphi)=(f_{\theta,R}(r),\varphi),$$ and the function $f_{\theta,R}\in C^\infty([0,\infty))$ is chosen so that (see Figure \[f:complex-scaling\](a)) $$\begin{aligned} f_{\theta,R}(r)=0,\quad r\leq R;&\quad f_{\theta,R}(r)=r\tan\theta,\quad r\geq 2R;\\ f'_{\theta,R}(r)\geq 0,\quad r\geq 0;&\quad \{f'_{\theta,R}=0\}=\{f_{\theta,R}=0\}. \end{aligned}$$ Note that $$\begin{aligned} \Gamma_{\theta,R}\cap \{|\Re z|\leq R\}&=\mathbb R^n\cap \{|\Re z|\leq R\};\\ \Gamma_{\theta,R}\cap \{|\Re z|\geq 2R\}&=e^{i\theta}\mathbb R^n\cap \{|\Re z|\geq 2R\}. \end{aligned}$$ Define the deformed differential operator $\widetilde P_V$ on $\Gamma_{\theta,R}$ it as follows: $\widetilde P_V=P_V$ on $\mathbb R^n\cap\Gamma_{\theta,R}$, and on the complementing region $\{|\Re z|>R\}$, it is defined by the formula $$\widetilde P_V(v)=\sum_{j=1}^n (hD_{z_j})^2 \tilde v|_{\Gamma_{\theta,R}},$$ for each $v\in C_0^\infty(\Gamma_{\theta,R}\cap \{|\Re z|>R\})$ and each almost analytic continuation $\tilde v$ of $v$ (that is, $\tilde v|_{\Gamma_{\theta,R}}=v$ and $\partial_{\bar z}\tilde v$ vanishes to infinite order on $\Gamma_{\theta,R}$ – the existence of such continuation follows from the fact that $\Gamma_{\theta,R}$ is totally real, that is for each $z\in \Gamma_{\theta,R}$, $T_z\Gamma_{\theta,R}\cap iT_z\Gamma_{\Theta,R}=0$). We identify $\Gamma_{\theta,R}$ with $\mathbb R^n$ by the map $$\iota:\mathbb R^n\to\Gamma_{\theta,R}\subset\mathbb C^n,\quad \iota(x)=x+iF_{\theta,R}(x),$$ so that $\widetilde P_V$ can be viewed as a second order differential operator on $\mathbb R^n$. Then in polar coordinates $(r,\varphi)$, we can write for $r>R$, $$\widetilde P_V=\bigg({1\over 1+if'_{\theta,R}(r)}hD_r\bigg)^2-{(n-1)i\over (r+if_{\theta,R}(r))(1+if'_{\theta,R}(r))}h^2D_r +{\Delta_\varphi\over (r+if_{\theta,R}(r))^2},$$ with $\Delta_\varphi$ denoting the Laplacian on the round sphere $\mathbb S^{n-1}$. We have $$\label{e:cs-symbol} \sigma(\widetilde P_V)={|\xi_r|^2\over (1+i f'_{\theta,R}(r))^2}+{|\xi_\varphi|^2\over (r+if_{\theta,R}(r))^2}+V(r,\varphi).$$ Fix a range of energies $[\alpha_0,\alpha_1]\subset (0,\infty)$ and a bounded open set $\Omega\subset\mathbb C$ such that (see Figure \[f:complex-scaling\](b)) ![(a) The graph of $f_{\theta,r}$. (b) The region where complex scaling provides meromorphic continuation of the resolvent.[]{data-label="f:complex-scaling"}](nhp.8 "fig:") ![(a) The graph of $f_{\theta,r}$. (b) The region where complex scaling provides meromorphic continuation of the resolvent.[]{data-label="f:complex-scaling"}](nhp.9 "fig:") to $$[\alpha_0,\alpha_1]\subset\Omega,\quad \overline\Omega\subset \{-\theta<\arg \omega<\pi-\theta\}.$$ For $\omega\in\Omega$, define the operator $$\mathcal P(\omega)=\widetilde P_V-\omega^2:\mathcal H_1\to \mathcal H_2,\qquad \mathcal H_1:=H^2_h(\mathbb R^n),\quad \mathcal H_2:=L^2(\mathbb R^n).$$ Then $\mathcal P(\omega)$ is Fredholm $\mathcal H_1\to\mathcal H_2$ for $\omega\in\Omega$. Indeed, $$\mathcal P(\omega)=\cos^2\theta e^{-2i\theta}h^2\Delta-\omega^2\quad\text{on } \{|x|\geq 2R\},$$ thus $\mathcal P(\omega)$ is elliptic on $\{|x|\geq 2R\}$, as well as for $|\xi|$ large enough, in the class $S(\langle\xi\rangle^2)$ of [@e-z §4.4.1] (this class incorporates the behavior of symbols as $x\to\infty$, in contrast with those used in §\[s:prelim-basics\]). Using a construction similar to Lemma \[l:eparametrix\], but with symbols in the class $S(\langle\xi\rangle^{-2})$, we can define a parametrix near (both spatial and fiber) infinity, $\mathcal R_\infty(\omega)$, with $\|\mathcal R_\infty\|_{L^2(\mathbb R^n)\to H^2_h(\mathbb R^n)}=\mathcal O(1)$ and $$\label{e:cs-par} \begin{aligned} \mathcal R_\infty(\omega)\mathcal P(\omega)&=1+Z(\omega)+\mathcal O(h^\infty)_{H^2_h(\mathbb R^n)\to H^2_h(\mathbb R^n)},\\ \mathcal P(\omega)\mathcal R_\infty(\omega)&=1+Z'(\omega)+\mathcal O(h^\infty)_{L^2(\mathbb R^n)\to L^2(\mathbb R^n)}, \end{aligned}$$ where $Z(\omega),Z'(\omega)\in\Psi^{\operatorname{comp}}(\mathbb R^n)$ are compactly supported inside $\{|x|<2R+1\}$. Since $1+\mathcal O(h^\infty)$ is invertible and $Z(\omega),Z'(\omega)$ are compact, we see that $\mathcal P(\omega)$ is indeed Fredholm $\mathcal H_1\to\mathcal H_2$. We have thus verified assumptions – of §\[s:framework-assumptions\]. The identification of the poles of $\mathcal R(\omega)$ with the poles of the meromorphic continuation of the resolvent $R_V(\omega)=(P_V-\omega^2)^{-1}$ defined in  from $ \{\Im \omega>0\}$ to $\Omega$, and in fact, the existence of such a continuation, follows from the following formula (implicit in [@sj], and discussed in [@TaZw]): if $ \chi \in C_0^\infty ( \mathbb R^n ) $, $ \operatorname{supp}\chi \Subset B ( 0 , R ) $, then $$\label{eq:two_r} \chi \mathcal R(\omega) \chi = \chi R_V(\omega) \chi.$$ This is initially valid in $ \Omega\cap\{\Im \omega > 0\} $ so that the right-hand side is well-defined, and then by analytic continuation in the region where the left hand side is meromorphic. Now, we take intervals $$[\alpha_0,\alpha_1]\Subset [\beta_0,\beta_1]\Subset [\beta'_0,\beta'_1]\subset\Omega\cap (0,\infty)$$ and put $$\label{e:sets-schrodinger} \begin{aligned} \mathcal U'&:=\{|x|<R,\ |\xi|^2+V(x)\in ((\beta'_0)^2,(\beta'_1)^2)\},\\ \mathcal U&:=\{|x|<3R/4,\ |\xi|^2+V(x)\in (\beta_0^2,\beta_1^2)\}. \end{aligned}$$ Note that $\mathcal P(\omega)=P_V-\omega^2$ in $\mathcal U'$; this verifies assumptions  and . Assumption  is also satisfied, with $$\label{e:p-schrodinger} p(x,\xi)=\sqrt{|\xi|^2+V(x)},\quad (x,\xi)\in\mathcal U'.$$ The operators $P$ and $\mathcal S(\omega)$ from Lemma \[l:resolution\] take the form, microlocally near $\mathcal U$, $$\label{e:sqrt} P=\sqrt{P_V},\quad \mathcal S(\omega)=\sqrt{\sqrt{P_V}+\omega}.$$ Here the square root is understood in the microlocal sense: for an operator $A\in\Psi^k(X)$ with $\sigma(A)>0$ on $\mathcal U'$, we define the microlocal square root $\sqrt{A}\in\Psi^{\operatorname{comp}}(X)$ of $A$ in $\mathcal U'$ as the (unique modulo $\mathcal O(h^\infty)$ microlocally in $\mathcal U'$) operator such that $(\sqrt{A})^2=A+\mathcal O(h^\infty)$ microlocally in $\mathcal U'$ and $\sigma(\sqrt{A})=\sqrt{\sigma(A)}$. See for example [@gr-sj Lemma 4.6] for details of the construction of the symbol. Assumption , namely convexity of $\overline{\mathcal U}$, is satisfied since for each $(x,\xi)\in\mathcal U'$, if $|x|\geq R/2$ and $H_p|x|^2=0$ at $(x,\xi)$, then $H_p^2|x|^2>0$ at $(x,\xi)$; therefore, the function $|x|^2$ cannot attain a local maximum on a trajectory of $e^{tH_p}$ in $\mathcal U'\setminus\overline{\mathcal U}$. Same observation shows assumption ; in fact, $K\subset \{|x|\leq R/2\}$. Finally, for assumptions  and , we take any compactly supported $\mathcal Q\in\Psi^{\operatorname{comp}}(X)$ such that $\WFh(\mathcal Q)\subset\mathcal U$ and $$\sigma(\mathcal Q)\geq 0\quad\text{everywhere};\quad \sigma(\mathcal Q)>0\quad\text{on } p^{-1}([\alpha_0,\alpha_1])\cap \{|x|\leq R/2\}.$$ To verify assumption , consider an arbitrary family $u=u(h)\in H^2_h(\mathbb R^n)$, with norm bounded polynomially in $h$, and put $$f=(\mathcal P(\omega)-i\mathcal Q)u,$$ where $\omega$ satisfies . By , and since $\Im\omega=\mathcal O(h)$, we find $$\begin{gathered} \Im\sigma(\mathcal P(\omega))\leq 0\quad\text{everywhere};\\ \{\langle\xi\rangle^{-2}\sigma(\mathcal P(\omega))=0\}\subset \{F_{\theta,R}(x)=0\}. \end{gathered}$$ Note also that $\sigma(\mathcal P(\omega))=|\xi|^2+V(x)-\omega^2$ on $\{F_{\theta,R}(x)=0\}$. Together with the convexity property of $|x|^2$ mentioned above, we see that for each $\rho\in T^*X$, there exists $t\leq 0$ such that $\mathcal P(\omega)-i\mathcal Q$ is elliptic at $\exp(tH_{\Re\sigma(\mathcal P(\omega))})(\rho)$. Since $\Im\sigma(\mathcal P(\omega)-i\mathcal Q)\leq 0$ everywhere, by propagation of singularities with a complex absorbing term (Proposition \[l:microhyperbolic\]) and the elliptic estimate (Proposition \[l:elliptic\]) we get $$\|Z(\omega)u\|_{H^2_h}\leq Ch^{-1}\|f\|_{L^2}+\mathcal O(h^\infty),$$ where $Z(\omega)$ is defined in . Then by , $$\|u\|_{H^2_h(\mathbb R^n)}\leq C\|f\|_{L^2(\mathbb R^n)}+\|Z(\omega)u\|_{H^2_h}+\mathcal O(h^\infty) \leq Ch^{-1}\|f\|_{L^2(\mathbb R^n)}+\mathcal O(h^\infty),$$ proving the estimate  of assumption . Assumption  is proved in a similar fashion: assume that $\WFh(f)\subset\mathcal U'$ and $\rho'\in\WFh(u)\cap \mathcal U'$. Denote $\gamma(t)=\exp(tH_{\Re\sigma(\mathcal P(\omega))})(\rho')$. Then there exists $t_0\geq 0$ such that $\mathcal P(\omega)-i \mathcal Q$ is elliptic at $\gamma(-t_0)$. By Proposition \[l:microhyperbolic\], we see that either $\exp(-tH_{\Re\sigma(\mathcal P(\omega))})(\rho')\in\WFh(f)$ for some $t\in [0,t_0]$ or $\exp(-t_0H_{\Re\sigma(\mathcal P(\omega))})(\rho')\in\WFh(u)$, in which case this point also lies in $\WFh(f)$ by Proposition \[l:elliptic\]; therefore, $\gamma(-t)\in\WFh(f)$ for some $t\geq 0$. Let $t_1$ be the minimal nonnegative number such that $\gamma(-t_1)\in\WFh(f)$; we may assume that $t_1>0$. Since $\gamma((-t_1,0])$ does not intersect $\WFh(f)$, it also does not intersect the elliptic set of $\mathcal P(\omega)$; therefore, $\gamma([-t_1,0])\subset \{F_{\theta,R}(x)=0\}$ and thus $\sigma(\mathcal P(\omega))=p^2-\omega^2$ on $\gamma([-t_1,0])$. It follows that $e^{-tH_p}(\rho')\in\WFh(f)$ for some $t\geq 0$, as required. Example: even asymptotically hyperbolic manifolds {#s:framework-ah} ------------------------------------------------- In this section, we define resonances, in the framework of §\[s:framework-assumptions\], for an $n$-dimensional complete noncompact Riemannian manifold $(M,g)$ which is *asymptotically hyperbolic* in the following sense: $M$ is diffeomorphic to the interior of a smooth manifold with boundary $\overline M$, and for some choice of the boundary defining function $\tilde x\in C^\infty(\overline M)$ and the product decomposition $\{\tilde x<\varepsilon\}\sim [0,\varepsilon)\times\partial\overline M$, the metric $g$ takes the following form in $\{0<\tilde x<\varepsilon\}$: $$\label{e:ah-metric} g={d\tilde x^2+g_1(\tilde x,\tilde y,d\tilde y)\over \tilde x^2}.$$ Here $g_1$ is a family of Riemannian metrics on $\partial\overline M$ depending smoothly on $\tilde x\in [0,\varepsilon)$. We moreover require that the metric is *even* in the sense that $g_1$ is a smooth function of $\tilde x^2$. To put the Laplacian $\Delta_g$ on $M$ into the framework of §\[s:framework-assumptions\], we use the recent construction of Vasy [@v2]. We follow in part [@fwl §4.1], see also [@fwl Appendix B] for a detailed description of the phase space properties of the resulting operator in a model case. Take the space $\overline M_{\operatorname{even}}$ obtained from $\overline M$ by taking the new boundary defining function $\mu=\tilde x^2$ and put (see [@v2 §3.1]) $$P_1(\omega)=\mu^{-{1\over 2}-{n+1\over 4}}e^{i\omega\phi\over h}(h^2(\Delta_g-(n-1)^2/4)-\omega^2) e^{-{i\omega\phi\over h}}\mu^{-{1\over 2}+{n+1\over 4}}.$$ Here $\phi$ is a smooth real-valued function on $M$ such that $$e^\phi=\mu^{1/2}(1+\mu)^{-1/4}\quad\text{on }\{0<\mu<\delta_0\},$$ where $\delta_0>0$ is a small constant; the values of $\phi$ on $\{\mu\geq\delta_0\}$ are chosen as in the paragraph preceding [@v2 (3.14)]. We can furthermore choose $e^\phi$ and $\mu$ to be equal to 1 near the set $\{\tilde x>\varepsilon_0/2\}$, for any fixed $\varepsilon_0>0$ (and $\delta_0$ chosen small depending on $\varepsilon_0$) so that $$\label{e:same} P_1(\omega)=h^2(\Delta_g-(n-1)^2/4)-\omega^2\quad\text{on }\{\tilde x>\varepsilon_0/2\}.$$ The differential operator $P_1(\omega)$ has coefficients smooth up to the boundary of $\overline M_{\operatorname{even}}$; then it is possible to find a compact $n$-dimensional manifold $X$ without boundary such that $\overline M_{\operatorname{even}}$ embeds into $X$ as $\{\mu\geq 0\}$ and extend $P_1(\omega)$ to an operator $P_2(\omega)\in\Psi^2(X)$, see [@v2 §3.5] or [@fwl Lemma 4.1]. Finally, we fix a complex absorbing operator $Q\in\Psi^2(X)$, with Schwartz kernel supported in the nonphysical region $\{\mu<0\}$, satisfying the assumptions of [@v2 §3.5]. We now fix an interval $[\alpha_0,\alpha_1]\subset (0,\infty)$, take $\Omega\subset\mathbb C$ a small neighborhood of $[\alpha_0,\alpha_1]$, and put $$\mathcal P(\omega):=P_2(\omega)-iQ,\quad \omega\in\Omega.$$ Fix $C_0>0$, take $s>C_0+1/2$, and put $\mathcal H_2=H^{s-1}_h(X)$ and $$\mathcal H_1=\{u\in H_h^s(X)\mid P_2(1)u\in H_h^{s-1}(X)\},\quad \|u\|_{\mathcal H_1}^2=\|u\|_{H_h^s(X)}^2+\|P_2(1)u\|_{H_h^{s-1}(X)}^2.$$ It is proved in [@v2 Theorem 4.3] that for $\omega$ satisfying , the operator $\mathcal P(\omega):\mathcal H_1\to\mathcal H_2$ is Fredholm of index zero; therefore, we have verified assumptions – of §\[s:framework-assumptions\]. The poles of $\mathcal R(\omega)=\mathcal P(\omega)^{-1}$ coincide with the poles of the meromorphic continuation of the Schwartz kernel of the resolvent $$R_g(\omega):=(h^2(\Delta_g-(n-1)^2/4)-\omega^2)^{-1}:L^2(M)\to L^2(M),\quad \Im\omega>0,$$ to the entire $\mathbb C$, first constructed in [@m-m] with improvements by [@gui] – see [@v2 Theorem 5.1]. We can now proceed similarly to §\[s:framework-schrodinger\], using that the regions $\{\tilde x>\varepsilon_0\}$ are geodesically convex for $\varepsilon_0>0$ small enough (see for instance [@qeefun Lemma 7.1]). Fix small $\varepsilon_0>0$, take any intervals $$[\alpha_0,\alpha_1]\Subset[\beta_0,\beta_1]\Subset [\beta'_0,\beta'_1]\subset\Omega\cap (0,\infty),$$ and define $$\label{e:sets-ah} \mathcal U':=\{\tilde x>\varepsilon_0/2,\ |\xi|_g\in (\beta'_0,\beta'_1)\},\quad \mathcal U:=\{\tilde x>\varepsilon_0,\ |\xi|_g\in (\beta_0,\beta_1)\}.$$ As in §\[s:framework-schrodinger\], assumptions – hold, with $$\label{e:p-ah} p(x,\xi)=|\xi|_g.$$ The operators $P$ and $\mathcal S(\omega)$ constructed in Lemma \[l:resolution\] are given microlocally near $\mathcal U$ by $$P=\sqrt{h^2\Delta_g-(n-1)^2/4},\quad S(\omega)=\sqrt{\sqrt{h^2\Delta_g-(n-1)^2/4}+\omega},$$ with the square roots defined as in . Finally, for assumptions  and , take $\mathcal Q\in\Psi^{\operatorname{comp}}(X)$ with $\WFh(\mathcal Q)\subset\mathcal U$ and $$\sigma(\mathcal Q)\geq 0\quad\text{everywhere};\quad \sigma(\mathcal Q)>0\quad\text{on }p^{-1}([\alpha_0,\alpha_1])\cap \{\tilde x\geq 2\varepsilon_0\}.$$ Then assumption  follows from [@v2 Theorem 4.8]. To verify assumption , we modify the proof of [@v2 Theorem 4.9] as follows: assume that $f=f(h)\in \mathcal H_2$ has norm bounded polynomially in $h$ and put $u=\mathcal R'(\omega)f$, for $\omega=\omega(h)$ satisfying . Assume also that $\WFh(f)\subset\mathcal U'$ and take $\rho'\in\WFh(u)\cap\mathcal U'$. We may assume that $P_2(\omega)$ is not elliptic at $\rho'$, since otherwise $\rho'\in\WFh(f)$. If $\gamma(t)$ is the bicharacteristic of $\sigma(P_2(\omega))$ starting at $\rho'$, then (see [@v2 (3.32) and the end of §3.5]) either $\gamma(t)$ converges to the set $L_+\subset \partial \overline T^*X\cap\{\mu=0\}$ of radial points as $t\to -\infty$, or $\mathcal Q$ is elliptic at $\gamma(-t_0)$ for some $t_0>0$. In the first case, $\gamma(-t_0)\not\in\WFh(u)$ for $t_0>0$ large enough by the radial points argument [@v2 Proposition 4.5]; in the second case, by Proposition \[l:elliptic\] we see that if $\gamma(-t_0)\in\WFh(u)$, then $\gamma(-t_0)\in\WFh(f)$. Combining this with Proposition \[l:microhyperbolic\], we see that there exists $t_1\geq 0$ such that $\gamma(-t_1)\in\WFh(f)$. Since $\gamma(0),\gamma(-t_1)\in\mathcal U'$, and $\mathcal U'$ is convex with respect to the bicharacteristic flow of $\sigma(P_2(\omega))$ (the latter being just a rescaling of the geodesic flow pulled back by a certain diffeomorphism), we see that $\gamma([-t_1,0])\subset\mathcal U'$. Now, by , $\gamma([-t_1,0])$ is a flow line of $H_{p^2}$; therefore, for some $t\geq 0$, $e^{-tH_p}(\rho')\in\WFh(f)$, as required. r-normally hyperbolic trapped sets {#s:nh} ================================== In this section, we state the dynamical assumptions on the flow near the trapped set $K$, namely $r$-normal hyperbolicity, and define the expansion rates $\nu_{\min},\nu_{\max}$ (§\[s:dynamics\]). We next establish some properties of $r$-normally hyperbolic trapped sets: existence of special defining functions $\varphi_\pm$ of the incoming/outgoing tails $\Gamma_\pm$ near $K$ (§\[s:phi-pm\]), existence of the canonical projections $\pi_\pm$ from open subsets $\Gamma_\pm^\circ\subset\Gamma_\pm$ to $K$ and the canonical relation $\Lambda^\circ$ (§\[s:projections\]), and regularity of solutions to the transport equations (§\[s:transport\]). Dynamical assumptions {#s:dynamics} --------------------- Let $\mathcal U\subset \mathcal U'$ be the open sets from §\[s:framework-assumptions\], and $p\in C^\infty(\mathcal U';\mathbb R)$ be the function defined in . Consider also the incoming/outgoing tails $\Gamma_\pm\subset\overline {\mathcal U}$ and the trapped set $K=\Gamma_+\cap\Gamma_-$ defined in . We assume that, for a large fixed integer $r$ depending only on the dimension $n$ (see Figure \[f:basic-dynamics\](a)), 1. \[aa:basic\] $\Gamma_\pm$ are equal to the intersections of $\overline{\mathcal U}$ with codimension 1 orientable $C^r$ submanifolds of $T^*X$; 2. \[aa:symplectic\] $\Gamma_+$ and $\Gamma_-$ intersect transversely, and the symplectic form $\sigma_S$ is nondegenerate on $TK$; that is, $K$ extends to a symplectic submanifold of $T^*X$ of codimension two. Consider one-dimensional subbundles $\mathcal V_\pm\subset T\Gamma_\pm$ defined as the symplectic complements of $T\Gamma_\pm$ in $T_{\Gamma_\pm}(T^*X)$ (see Figure \[f:basic-dynamics\](b)); they are invariant under the flow $e^{tH_p}$. By assumption , we have $T_K\Gamma_\pm=\mathcal V_\pm|_K\oplus TK$. Define the minimal expansion rate in the normal direction, $\nu_{\min}$, as the supremum of all $\nu$ for which there exists a constant $C$ such that $$\label{e:nu-min} \sup_{\rho\in K}\|de^{\mp tH_p}(\rho)|_{\mathcal V_\pm}\|\leq Ce^{-\nu t},\quad t>0.$$ Here $\|\cdot\|$ denotes the operator norm with respect to any smooth inner product on the fibers of $T(T^*X)$. Similarly we define the maximal expansion rate in the normal direction, $\nu_{\max}$, as the infimum of all $\nu$ for which there exists a constant $c>0$ such that $$\label{e:nu-max} \inf_{\rho\in K}\|de^{\mp t H_p}(\rho)|_{\mathcal V_\pm}\|\geq ce^{-\nu t},\quad t>0.$$ Since $e^{tH_p}$ preserves the symplectic form $\sigma_S$, which is nondegenerate on $\mathcal V_+|_K\oplus\mathcal V_-|_K$, it is enough to require  and  for a specific choice of sign. ![(a) Dynamics of $e^{tH_p}$ in the directions transverse to the trapped set. (b) Dynamics on $\Gamma_\pm$; the flow lines of $\mathcal V_\pm$ are dashed.[]{data-label="f:basic-dynamics"}](nhp.1 "fig:") ![(a) Dynamics of $e^{tH_p}$ in the directions transverse to the trapped set. (b) Dynamics on $\Gamma_\pm$; the flow lines of $\mathcal V_\pm$ are dashed.[]{data-label="f:basic-dynamics"}](nhp.2 "fig:") to We assume *$r$-normal hyperbolicity*: 1. \[aa:r-nh\] Let $\mu_{\max}$ be the maximal expansion rate of the flow along $K$, defined as the infimum of all $\mu$ for which there exists a constant $C$ such that $$\label{e:mu-max} \sup_{\rho\in K}\|de^{tH_p}(\rho)|_{TK}\|\leq Ce^{\mu|t|},\quad t\in \mathbb R.$$ Then $$\label{e:r-nh} \nu_{\min}>r\mu_{\max}.$$ Assumption , rather than a weaker assumption of *normal hyperbolicity* $\nu_{\min}>0$, is needed for regularity of solutions to the transport equations, see Lemma \[l:ode\] below. The number $r$ depends on how many derivatives of the symbols constructed below are needed for the semiclassical arguments to work. In the proofs, we will often take $r=\infty$, keeping in mind that a large fixed $r$ is always enough. Stability {#s:stability} --------- We now briefly discuss stability of our dynamical assumptions under perturbations; a more detailed presentation, with applications to general relativity, will be given in [@thesis]. Assume that $p_s$, where $s\in\mathbb R$ varies in a neighborhood of zero, is a family of real-valued functions on $\mathcal U'$ such that $p_0=p$ and $p_s$ is continuous at $s=0$ with values in $C^\infty(\mathcal U')$. Assume moreover that conditions  and  of §\[s:framework-assumptions\] are satisfied with $p$ replaced by any $p_s$. Here $\Gamma_\pm$ and $K$ are replaced by the sets $\Gamma_\pm(s)$ and $K(s)$ defined using $p_s$ instead of $p$. We claim that assumptions – of §\[s:dynamics\] are satisfied for $p_s,\Gamma_\pm(s),K(s)$ when $s$ is small enough. We use the work of Hirsch–Pugh–Shub [@HPS] on stability of $r$-normally hyperbolic invariant manifolds. Assumptions – imply that the flow $e^{tH_p}$ is eventually absolutely $r$-normally hyperbolic on $K$ in the sense of [@HPS Definition 4]. Then by [@HPS Theorem 4.1], for $s$ small enough, $\Gamma_\pm(s)$ and $K(s)$ are $C^r$ submanifolds of $T^*X$, which converge to $\Gamma_\pm$ and $K$ in $C^r$ as $s\to 0$. It follows immediately that conditions  and  are satisfied for small $s$. To see that condition  is satisfied for small $s$, as well as stability of the pinching condition  under perturbations, it suffices to show that, with $\nu_{\min}(s),\nu_{\max}(s),\mu_{\max}(s)$ defined using $e^{tH_{p_s}},\Gamma_\pm(s),K(s)$, $$\begin{gathered} \label{e:stab-1} \liminf_{s\to 0}\nu_{\min}(s)\geq \nu_{\min},\\ \label{e:stab-2} \limsup_{s\to 0}\nu_{\max}(s)\leq \nu_{\max},\\ \label{e:stab-3} \limsup_{s\to 0}\mu_{\max}(s)\leq \mu_{\max}.\end{gathered}$$ We show ; the other two inequalities are proved similarly. Fix a smooth metric on the fibers of $T(T^*X)$. Take arbitrary $\varepsilon>0$, then for $T>0$ large enough, we have $$\sup_{\rho\in K}\|de^{\mp TH_p}(\rho)|_{\mathcal V_\pm}\|\leq e^{-(\nu_{\min}-\varepsilon)T}.$$ Fix $T$; since $p_s$, $\Gamma_\pm(s)$, $K(s)$, and the corresponding subbundles $\mathcal V_\pm(s)$ depend continuously on $s$ at $s=0$, we have for $s$ small enough, $$\sup_{\rho\in K(s)}\|de^{\mp TH_{p_s}}(\rho)|_{\mathcal V_\pm(s)}\|\leq e^{-(\nu_{\min}-\varepsilon/2)T}.$$ Since $e^{tH_{p_s}}$ is a one-parameter group of diffeomorphisms, we get $$\sup_{\rho\in K(s)}\|de^{\mp tH_{p_s}}(\rho)|_{\mathcal V_\pm(s)}\|\leq Ce^{-(\nu_{\min}-\varepsilon/2)t},\quad t\geq 0;$$ therefore, $\nu_{\min}(s)\geq \nu_{\min}-\varepsilon/2$ for $s$ small enough and  follows. Adapted defining functions {#s:phi-pm} -------------------------- In this section, we construct special defining functions $\varphi_\pm$ of $\Gamma_\pm$ near $K$. We will assume below that $\Gamma_\pm$ are smooth; however, if $\Gamma_\pm$ are $C^r$ with $r\geq 1$, we can still obtain $\varphi_\pm\in C^r$. A similar construction can be found in [@w-z Lemma 4.1]. \[l:phi-pm\] Fix $\varepsilon>0$.[^3] Then there exist smooth functions $\varphi_\pm$, defined in a neighborhood of $K$ in $\mathcal U'$, such that for $\delta>0$ small enough, the set $$\label{e:u-delta} U_\delta:=\overline{\mathcal U}\cap \{|\varphi_+|\leq\delta,\ |\varphi_-|\leq\delta\},$$ is a compact subset of $\mathcal U$ when intersected with $p^{-1}([\alpha_0,\alpha_1])$, and: 1. \[c:defining\] $\Gamma_\pm\cap U_\delta=\{\varphi_\pm=0\}\cap U_\delta$, and $d\varphi_\pm\neq 0$ on $U_\delta$; 2. \[c:c-pm\] $H_p\varphi_\pm=\mp c_\pm\varphi_\pm$ on $U_\delta$, where $c_\pm$ are smooth functions on $U_\delta$ and, with $\nu_{\min},\nu_{\max}$ defined in , , $$\label{e:nu-bound} \nu_{\min}-\varepsilon<c_\pm<\nu_{\max}+\varepsilon\quad\text{on }U_\delta;$$ 3. the Hamiltonian field $H_{\varphi_\pm}$ spans the subbundle $\mathcal V_\pm$ on $\Gamma_\pm\cap U_\delta$ defined before ; 4. \[c:poisson\] $\{\varphi_+,\varphi_-\}>0$ on $U_\delta$; 5. \[c:convex\] $U_\delta$ is convex, namely if $\gamma(t)$, $0\leq t\leq T$, is a Hamiltonian flow line of $p$ in $\overline{\mathcal U}$ and $\gamma(0),\gamma(T)\in U_\delta$, then $\gamma([0,T])\subset U_\delta$. Since $\Gamma_\pm$ are orientable, there exist defining functions $\tilde\varphi_\pm$ of $\Gamma_\pm$ near $K$; that is, $\tilde\varphi_\pm$ are smooth, defined in some neighborhood $U$ of $K$, and $d\tilde\varphi_\pm\neq 0$ on $U$ and $\Gamma_\pm\cap U=\overline{\mathcal U}\cap \{\tilde\varphi_\pm=0\}$. Since $K$ is symplectic, by changing the sign of $\tilde\varphi_-$ if necessary, we can moreover assume that $\{\tilde\varphi_+,\tilde\varphi_-\}>0$ on $K$. Since $e^{tH_p}(\Gamma_\pm)\subset\Gamma_\pm$ for $\mp t\geq 0$, we have $H_p\tilde\varphi_\pm=0$ on $\Gamma_\pm$; therefore, $$H_p\tilde\varphi_\pm=\mp \tilde c_\pm\tilde\varphi_\pm,$$ where $\tilde c_\pm$ are smooth functions on $U$. The functions $\tilde c_\pm$ control how fast $\tilde\varphi_\pm$ decays along the flow as $t\to\pm\infty$. The constants $\nu_{\min}$ and $\nu_{\max}$ control the average decay rate; to construct $\varphi_\pm$, we will modify $\tilde\varphi_\pm$ by averaging along the flow for a large time. For any $\rho\in \Gamma_\pm\cap U$, the kernel of $d\tilde\varphi_\pm(\rho)$ is equal to $T_\rho\Gamma_\pm$; therefore, the Hamiltonian fields $H_{\tilde\varphi_\pm}$ span $\mathcal V_\pm$ on $\Gamma_\pm\cap U$. We then see from the definitions , of $\nu_{\min},\nu_{\max}$ that there exists a constant $C$ such that, with $(e^{\mp tH_p})_*H_{\tilde\varphi_\pm}\in\mathcal V_\pm$ denoting the push-forward of the vector field $H_{\tilde\varphi_\pm}$ by the diffeomorphism $e^{\mp tH_p}$, $$C^{-1}e^{-(\nu_{\max}+\varepsilon/2)t}\leq {(e^{\mp tH_p})_*H_{\tilde\varphi_\pm}\over H_{\tilde\varphi_\pm}}\leq Ce^{-(\nu_{\min}-\varepsilon/2)t}\quad\text{on } K,\quad t\geq 0.$$ Now, we calculate on $K$, $$\begin{gathered} \partial_t ((e^{\mp tH_p})_* H_{\tilde\varphi_\pm}) =\pm (e^{\mp tH_p})_*[H_p,H_{\tilde\varphi_\pm}]\\ =-(e^{\mp tH_p})_*H_{\tilde c_\pm\tilde\varphi_\pm} =-(\tilde c_\pm\circ e^{\pm tH_p})(e^{\mp tH_p})_* H_{\tilde\varphi_\pm}. \end{gathered}$$ Combining these two facts, we get for $T>0$ large enough, $$\nu_{\min}-\varepsilon<\langle \tilde c_\pm\rangle_T<\nu_{\max}+\varepsilon\quad\text{on }K,$$ where $\langle \cdot\rangle_T$ stands for the ergodic average on $K$: $$\langle f\rangle_T:={1\over T}\int_0^T f\circ e^{tH_p}\,dt.$$ Fix $T$. We now put $\varphi_\pm:=e^{\mp f_\pm}\cdot\tilde\varphi_\pm$, where $f_\pm$ are smooth functions on $U$ with $$f_\pm={1\over T}\int_0^T (T-t)\tilde c_\pm\circ e^{tH_p}\,dt\quad \text{on }K,$$ so that $H_p f_\pm=\langle \tilde c_\pm\rangle_T-\tilde c_\pm$ on $K$. Then $\varphi_\pm$ satisfy conditions –, with $$c_\pm=\mp{H_p \varphi_\pm\over \varphi_\pm}=\langle\tilde c_\pm \rangle_T\in (\nu_{\min}-\varepsilon,\nu_{\max}+\varepsilon)$$ on $K$, and thus on $U_\delta$ for $\delta$ small enough. To verify condition , fix $\delta_0>0$ small enough so that $\pm H_p\varphi_\pm^2\leq 0$ on $U_{\delta_0}$. By Lemma \[l:the-flow-2\], for $\delta$ small enough depending on $\delta_0$, for each Hamiltonian flow line $\gamma(t)$, $0\leq t\leq T$, of $p$ in $\overline{\mathcal U}$, if $\gamma(0),\gamma(T)\in U_\delta$, then $\gamma([0,T])\subset U_{\delta_0}$. Since $\pm \partial_t\varphi_\pm(\gamma(t))^2\leq 0$ for $0\leq t\leq T$ and $|\varphi_\pm(\gamma(t))|\leq \delta$ for $t=0,T$, we see that $\gamma([0,T])\subset U_\delta$. The canonical relation Lambda {#s:projections} ----------------------------- We next construct the projections $\pi_\pm$ from subsets $\Gamma_\pm^\circ\subset\Gamma_\pm$ to $K$. Fix $\delta_0,\delta_1>0$ small enough so that Lemma \[l:phi-pm\] holds with $\delta_0$ in place of $\delta$ and $K\cap p^{-1}([\alpha_0-\delta_1,\alpha_1+\delta_1])$ is a compact subset of $\mathcal U$ (the latter is possible by assumption  in §\[s:framework-assumptions\]), consider the functions $\varphi_\pm$ from Lemma \[l:phi-pm\] and put $$\label{e:k-circ} \Gamma_\pm^\circ:=\Gamma_\pm\cap p^{-1}(\alpha_0-\delta_1,\alpha_1+\delta_1)\cap \{|\varphi_\mp|<\delta_0\},\quad K^\circ:=K\cap p^{-1}(\alpha_0-\delta_1,\alpha_1+\delta_1),$$ so that $K^\circ=\Gamma_+^\circ\cap\Gamma_-^\circ$ and, for $\delta_0$ small enough, $\Gamma_\pm^\circ\subset\mathcal U$. Note that, by part  of Lemma \[l:phi-pm\], the level sets of $p$ on $\Gamma_\pm$ are invariant under $H_{\varphi_\pm}$ and $e^{tH_p}(\Gamma_\pm^\circ)\subset\Gamma_\pm^\circ$ for $\mp t\geq 0$. By part  of Lemma \[l:phi-pm\], $\Gamma_\pm^\circ$ is foliated by trajectories of $H_{\varphi_\pm}$ (or equivalently, by trajectories of $\mathcal V_\pm$), moreover each trajectory intersects $K$ at a single point. This defines projection maps $$\pi_\pm:\Gamma^\circ_\pm\to K^\circ,$$ mapping each trajectory to its intersection with $K$. The flow $e^{tH_p}$ preserves the subbundle $\mathcal V_\pm$ generated by $H_{\varphi_\pm}$, therefore $$\label{e:pi-pm-commute} \pi_\pm\circ e^{\mp tH_p}=e^{\mp tH_p}\circ\pi_\pm,\quad t\geq 0.$$ Now, define the $2n$-dimensional submanifold $\Lambda^\circ\subset T^*X\times T^*X$ by $$\label{e:the-Lambda} \Lambda^\circ:=\{(\rho_-,\rho_+)\in\Gamma_-^\circ\times\Gamma_+^\circ\mid \pi_-(\rho_-)=\pi_+(\rho_+)\}.$$ We claim that $\Lambda^\circ$ is a canonical relation. Indeed, it is enough to prove that $\sigma_S|_{T\Gamma^\circ_\pm}=\pi_\pm^* (\sigma_S|_{TK^\circ})$, where $\sigma_S$ is the symplectic form on $T^*M$. This is true since the Hamiltonian flow $e^{tH_{\varphi_\pm}}$ preserves $\sigma_S$ and $\mathcal V_\pm|_K$ is symplectically orthogonal to $TK$. The transport equations {#s:transport} ----------------------- Finally, we use $r$-normal hyperbolicity to establish existence of solutions to the transport equations, needed in the construction of the projector $\Pi$ in §\[s:construction-1\]. We start by estimating higher derivatives of the flow. Take $\delta_0,\Gamma_\pm^\circ,K^\circ$ from §\[s:projections\] and identify $\Gamma_\pm^\circ\sim K^\circ\times (-\delta_0,\delta_0)$ by the map $$\label{e:identify} \rho_\pm\in \Gamma_\pm^\circ\mapsto (\pi_\pm(\rho_\pm),\varphi_\mp(\rho_\pm)).$$ Denote elements of $K^\circ\times (-\delta_0,\delta_0)$ by $(\theta,s)$ and the flow $e^{tH_p}$ on $\Gamma_\pm^\circ$, $\mp t\geq 0$, by (recall ) $$e^{tH_p}:(\theta,s)\mapsto (e^{tH_p}(\theta),\psi^t_\pm(\theta,s)).$$ Note that $\psi^t_\pm(\theta,0)=0$. We have the following estimate on higher derivatives of the flow on $K^\circ$ (in any fixed coordinate system), see for example [@qeefun Lemma C.1] (which is stated for geodesic flows, but the proof applies to any smooth flow): $$\label{e:k-ders} \sup_{\theta\in K^\circ}|\partial^\alpha_\theta e^{tH_p}(\theta)|\leq C_\alpha e^{(|\alpha|\mu_{\max}+\tilde\varepsilon)|t|},\quad t\in\mathbb R.$$ Here $\mu_{\max}$ is defined by , $\tilde\varepsilon>0$ is any fixed constant, and $C_\alpha$ depends on $\tilde\varepsilon$. We choose $\tilde\varepsilon$ small enough in  below and the constant $\varepsilon>0$ in Lemma \[l:phi-pm\] is small depending on $\tilde\varepsilon$. Next, we estimate the derivatives of $\psi^t_\pm$. We have, with $c_\pm$ defined in part  of Lemma \[l:phi-pm\], $$\partial_t\psi^t_\pm(\theta,s)=\pm c_\mp(e^{tH_p}(\theta),\psi^t_\pm(\theta,s))\psi^t_\pm(\theta,s).$$ Then $$\partial_t (\partial^k_s\partial^\alpha_\theta\psi^t_\pm(\theta,s)) =\pm c_\mp(e^{tH_p}(\theta),0)\partial^k_s\partial^\alpha_\theta\psi^t_\pm(\theta,s) +\dots,$$ where $\dots$ is a linear combination, with uniformly bounded variable coefficients depending on the derivatives of $c_\mp$, of expressions of the form $$\partial^{\beta_1}_\theta e^{tH_p}(\theta)\cdots\partial_\theta^{\beta_m}e^{tH_p}(\theta)\, \partial^{\gamma_1}_\theta\partial^{k_1}_s\psi^t_\pm(\theta,s)\cdots\partial_\theta^{\gamma_l}\partial^{k_l}_s\psi^t_\pm(\theta,s),$$ where $\beta_1+\dots+\beta_m+\gamma_1+\dots+\gamma_l=\alpha$, $k_1+\dots+k_l=k$, and $|\beta_j|,|\gamma_j|+k_j>0$. Moreover, if $l=0$ or $l+m=1$, then the corresponding coefficient is a bounded multiple of $\psi^t_\pm(\theta,s)$. It now follows by induction from  that $$\label{e:s-ders} \sup_{\theta\in K^\circ,\, |s|<\delta_0}|\partial^k_s\partial^\alpha_\theta \psi^{\mp t}_\pm(\theta,s)|\leq C_{\alpha k}e^{(|\alpha|\mu_{\max}-\nu_{\min}+\tilde\varepsilon)t},\quad t\geq 0.$$ We can now prove the following \[l:ode\] Assume that  is satisfied, with some integer $r>0$. Let $f\in C^{r+1}(\Gamma_\pm^\circ)$ be such that $f|_K=0$. Then there exists unique solution $a\in C^r(\Gamma_\pm^\circ)$ to the equation $$\label{e:transport} H_p a=f,\quad a|_{K^\circ}=0.$$ Using , choose $\tilde\varepsilon>0$ so that $$\label{e:r-nh-2} r\mu_{\max}-\nu_{\min}+\tilde\varepsilon<0.$$ Any solution to  satisfies for each $T>0$, $$a=a\circ e^{\mp TH_p}\pm\int_0^T f\circ e^{\mp tH_p}\,dt.$$ Since $a|_{K^\circ}=0$, by letting $T\to +\infty$ we see that the unique solution to  is $$\label{e:transport-sol} a=\pm\int_0^\infty f\circ e^{\mp tH_p}\,dt.$$ The integral  converges exponentially, as $$|f\circ e^{\mp tH_p}(\theta,s)|\leq C|\psi^{\mp t}_\pm(\theta,s)|\leq Ce^{-(\nu_{\min}-\varepsilon)t}.$$ To show that $a\in C^r$, it suffices to prove that when $|\alpha|+k\leq r$, the integral $$\int_0^\infty \partial^k_s\partial^\alpha_\theta(f\circ e^{\mp tH_p})\,dt$$ converges uniformly in $s,\theta$. Given , it is enough to show that $$\label{e:tr-int} \sup_{\theta,s}|\partial^k_s\partial^\alpha_\theta (f\circ e^{\mp tH_p})(\theta,s)|\leq C_{\alpha k}e^{(|\alpha|\mu_{\max} -\nu_{\min}+\tilde\varepsilon)t},\quad t>0.$$ To see , we use the chain rule to estimate the left-hand side by a sum of terms of the form $$\partial^m_\theta\partial^l_s f(e^{\mp tH_p}(\theta,s)) \partial_\theta^{\beta_1}e^{\mp tH_p}(\theta)\cdots\partial_\theta^{\beta_m}e^{\mp tH_p}(\theta) \partial^{\gamma_1}_\theta\partial^{k_1}_s\psi^{\mp t}_\pm(\theta,s)\cdots\partial_\theta^{\gamma_l}\partial^{k_l}_s\psi^{\mp t}_\pm(\theta,s)$$ where $\beta_1+\dots+\beta_m+\gamma_1+\dots+\gamma_l=\alpha$, $k_1+\dots+k_l=k$, and $|\beta_j|,|\gamma_j|+k_j>0$. For $l=0$, we have $|\partial^m_\theta f\circ e^{\mp tH_p}|=\mathcal O(e^{-(\nu_{\min}-\varepsilon)t})$ and  follows from . For $l>0$, follows from  and . Calculus of microlocal projectors {#s:calculus} ================================= In this section, we develop tools for handling Fourier integral operators associated to the canonical relation $\Lambda^\circ$ introduced in §\[s:projections\]. We will not use theoperator $P$ or the global dynamics of the flow $e^{tH_p}$; we will only assume that $X$ is an $n$-dimensional manifold and - $\Gamma_\pm^\circ\subset T^*X$ are smooth orientable hypersurfaces; - $\Gamma_\pm^\circ$ intersect transversely and $K^\circ:=\Gamma_+^\circ\cap\Gamma_-^\circ$ is symplectic; - if $\mathcal V_\pm\subset T\Gamma_\pm^\circ$ is the symplectic complement of $T\Gamma_\pm^\circ$ in $T(T^*X)$, then each maximally extended flow line of $\mathcal V_\pm$ on $\Gamma_\pm^\circ$ intersects $K^\circ$ at precisely one point, giving rise to the projection maps $\pi_\pm:\Gamma_\pm^\circ\to K^\circ$; - the canonical relation $\Lambda^\circ\subset T^*(X\times X)$ is defined by $$\Lambda^\circ=\{(\rho_-,\rho_+)\in\Gamma_-^\circ\times\Gamma_+^\circ\mid \pi_-(\rho_-)=\pi_+(\rho_+)\};$$ - the projections $\tilde\pi_\pm:\Lambda^\circ\to\Gamma_\pm^\circ$ are defined by $$\label{e:tilde-pi} \tilde\pi_\pm(\rho_-,\rho_+)=\rho_\pm.$$ If we only consider a bounded number of terms in the asymptotic expansions of the studied symbols, and require existence of a fixed number of derivatives of these symbols, then the smoothness requirement above can be replaced by $C^r$ for $r$ large enough depending only on $n$. We will study the operators in the class $\II$ considered in §\[s:prelim-fio\]. The antiderivative on $\Lambda^\circ$ (see §\[s:prelim-fio\]) is fixed so that it vanishes on the image of the embedding $$\label{e:j-k} j_K:K^\circ\to\Lambda^\circ,\quad j_K(\rho)=(\rho,\rho);$$ this is possible since $j_K^*(\eta\,dy-\xi\,dx)=0$ and the image of $j_K$ is a deformation retract of $\Lambda^\circ$. We are particularly interested in defining invariantly the principal symbol $\sigma_\Lambda(A)$ of an operator $A\in\II$. This could be done using the global theory of Fourier integral operators; we take instead a more direct approach based on the model case studied in §\[s:model\]. The principal symbols on a neighborhood $\widetilde\Lambda$ of a compact subset $\widehat K\subset K^\circ$ are defined as sections of certain vector bundles in §\[s:general\]. We are also interested in the symbol of a product of two operators in $\II$. Note that such a product lies again in $\II$, since $\Lambda^\circ$ satisfies the transversality condition with itself and, with the composition defined as in , $\Lambda^\circ\circ\Lambda^\circ=\Lambda^\circ$. To study the principal symbol of the product, we again use the model case – see Proposition \[l:calculus\]. Next, in §\[s:idempotents\], we study idempotents in $\II$, microlocally near $\widehat K$, proving technical lemmas need in the construction of the microlocal projector $\Pi$ in §\[s:global-construction\]. Finally, in §\[s:ideals\], we consider left and right ideals of pseudodifferential operators annihilating a microlocal idempotent, which are key for proving resolvent estimates in §\[s:resolvent-bounds\]. Model case {#s:model} ---------- We start with the model case $$\label{e:model-gamma} X:=\mathbb R^n,\quad \Gamma^0_+:=\{\xi_n=0\},\quad \Gamma^0_-:=\{x_n=0\}.$$ Then $K^0=\{x_n=\xi_n=0\}$ is canonically diffeomorphic to $T^*\mathbb R^{n-1}$. If we denote elements of $\mathbb R^{2n}\simeq T^*\mathbb R^n$ by $(x',x_n,\xi',\xi_n)$, with $x',\xi'\in\mathbb R^{n-1}$, then the projection maps $\pi_\pm:\Gamma_\pm^0\to K^0$ take the form $$\pi_+(x,\xi',0)=(x',0,\xi',0),\quad \pi_-(x',0,\xi)=(x',0,\xi',0),$$ and the map $$\label{e:phi} \phi:(x,\xi)\mapsto (x',0,\xi;x,\xi',0)\in T^*(\mathbb R^n\times\mathbb R^n)$$ gives a diffeomorphism of $\mathbb R^{2n}$ onto the corresponding canonical relation $\Lambda^0$. For a Schwartz function $a(x,\xi)\in\mathscr S(\mathbb R^{2n})$, define its $\Lambda^0$-quantization $\operatorname{Op}_h^\Lambda(a):\mathscr S'(\mathbb R^n)\to \mathscr S(\mathbb R^n)$ by the formula $$\label{e:lambda-quant} \operatorname{Op}_h^\Lambda(a)u(x)=(2\pi h)^{-n} \int_{\mathbb R^{2n}}e^{{i\over h}(x'\cdot\xi'-y\cdot\xi)}a(x,\xi)u(y)\,dyd\xi.$$ The operator $\operatorname{Op}_h^\Lambda(a)$ will be a Fourier integral operator associated to $\Lambda^0$, see below for details. We also use the standard quantization for pseudodifferential operators [@e-z §4.1.1], where $a(x,\xi;h)\in C^\infty(\mathbb R^{2n})$ and all derivatives of $a$ are bounded uniformly in $h$ by a fixed power of $1+|x|^2+|\xi|^2$: $$\label{e:op-h} \operatorname{Op}_h(a)u(x)=(2\pi h)^{-n}\int_{\mathbb R^{2n}}e^{{i\over h}(x-y)\cdot\xi}a(x,\xi)u(y)\,dyd\xi.$$ The symbol $a$ can be extracted from $\operatorname{Op}_h^\Lambda(a)$ or $\operatorname{Op}_h(a)$ by the following oscillatory testing formulas, see [@e-z Theorem 4.19]: $$\begin{gathered} \label{e:osc-test} \operatorname{Op}_h^\Lambda(a)(e^{{i\over h}x\cdot\xi})=e^{{i\over h}x'\cdot\xi'}a(x,\xi),\quad \xi\in\mathbb R^n,\\ \label{e:osc-test-2} \operatorname{Op}_h(a)(e^{{i\over h}x\cdot\xi})=e^{{i\over h}x\cdot\xi}a(x,\xi),\quad \xi\in\mathbb R^n.\end{gathered}$$ From here, using stationary phase expansions similarly to [@e-z Theorems 4.11 and 4.12], we get (where the symbols quantized by $\operatorname{Op}_h^\Lambda$ are Schwartz) $$\begin{gathered} \label{e:calc-model-1} \operatorname{Op}_h^\Lambda(a)\operatorname{Op}_h^\Lambda(b)=\operatorname{Op}_h^\Lambda(a\#^\Lambda b),\\ \label{e:calc-model-2} \operatorname{Op}_h^\Lambda(a)\operatorname{Op}_h(b)=\operatorname{Op}_h^\Lambda(a_{\#b}),\\ \label{e:calc-model-3} \operatorname{Op}_h(b)\operatorname{Op}_h^\Lambda(a)=\operatorname{Op}_h^\Lambda(a_{b\#}),\end{gathered}$$ where the symbols $a\#^\Lambda b,a_{\#b},a_{b\#}\in \mathscr S(\mathbb R^{2n})$ have asymptotic expansions $$\begin{gathered} \label{e:exp-model-1} a\#^\Lambda b(x,\xi) \sim\sum_{\alpha} {(-ih)^{|\alpha|}\over\alpha!}\partial_\xi^\alpha a(x,\xi',0)\partial_x^\alpha b(x',0,\xi),\\ \label{e:exp-model-2} a_{\#b}(x,\xi)\sim\sum_{\alpha}{(-ih)^{|\alpha|}\over\alpha!} \partial_\xi^\alpha a(x,\xi) \partial_x^\alpha b(x',0,\xi),\\ \label{e:exp-model-3} a_{b\#}(x,\xi)\sim \sum_{\alpha}{(-ih)^{|\alpha|}\over\alpha!} \partial_\xi^\alpha b(x,\xi',0)\partial_x^\alpha a(x,\xi).\end{gathered}$$ Finally, the operators $\operatorname{Op}_h^\Lambda(a)$ are bounded $L^2\to L^2$ with norm $\mathcal O(h^{-1/2})$: \[l:bund\] If $a\in\mathscr S(\mathbb R^{2n})$, then there exists a constant $C$ such that $$\|\operatorname{Op}_h^\Lambda(a)\|_{L^2(\mathbb R^n)\to L^2(\mathbb R^n)}\leq Ch^{-1/2}.$$ Define the semiclassical Fourier transform $$\hat u(\xi):=(2\pi h)^{-n/2}\int_{\mathbb R^n} e^{-{i\over h}y\cdot\xi}u(y)\,dy,$$ then $\|\hat u\|_{L^2}=\|u\|_{L^2}$ and $$\operatorname{Op}_h^\Lambda(a)u(x)=(2\pi h)^{-1/2}\int_{\mathbb R} v(x,\xi_n)\,d\xi_n,$$ where $$v(x,\xi_n):=(2\pi h)^{-(n-1)/2}\int_{\mathbb R^{n-1}} e^{{i\over h}x'\cdot\xi'}a(x,\xi',\xi_n)\hat u(\xi',\xi_n)\,d\xi'.$$ Using the $L^2$-boundedness of pseudodifferential operators on $\mathbb R^{n-1}$, we see that for each $(x_n,\xi_n)\in \mathbb R^2$, $$\|v(\cdot,x_n,\xi_n)\|_{L^2_{x'}}\leq F(x_n,\xi_n)\|\hat u(\cdot,\xi_n)\|_{L^2_{\xi'}},$$ where $F(x_n,\xi_n)$ is bounded by a certain $\mathscr S(\mathbb R^{2n-2})$ seminorm of $a(\cdot,x_n,\cdot,\xi_n)$. Then $F$ is rapidly decaying on $\mathbb R^2$ and for any $N$, $$\|v(\cdot,\xi_n)\|_{L^2_x}\leq C\langle \xi_n\rangle^{-N}\|\hat u(\cdot,\xi_n)\|_{L^2_{\xi'}}.$$ Therefore, $$\|\operatorname{Op}_h^\Lambda(a)u(x)\|_{L^2}\leq Ch^{-1/2}\int_{\mathbb R}\|v(\cdot,\xi_n)\|_{L^2_x}\,d\xi_n \leq Ch^{-1/2}\|u\|_{L^2}$$ as required. For $a\in\mathscr S(\mathbb R^{2n})$, the operator $\operatorname{Op}_h^\Lambda(a)$ is $h$-tempered as defined in Section \[s:prelim-basics\]. Moreover, the following analog of  follows from  and : $$\label{e:model-wf} \WFh(\operatorname{Op}_h^\Lambda(a))\subset \phi(\operatorname{supp}a)\subset\Lambda^0,$$ with $\phi$ defined by . For $a\in C_0^\infty(\mathbb R^{2n})$, we use  to check that $\operatorname{Op}_h^\Lambda(a)$ is, modulo an $\mathcal O(h^\infty)_{\mathscr S'\to\mathscr S}$ remainder, a Fourier integral operator in the class $I_{\operatorname{comp}}(\Lambda^0)$ defined in §\[s:prelim-basics\]. We will also use the operator $\operatorname{Op}_h^\Lambda(1):C^\infty(\mathbb R^n)\to C^\infty(\mathbb R^n)$ defined by $$\label{e:model-projector} \operatorname{Op}_h^\Lambda(1)f(x)=f(x',0),\quad f\in C^\infty(\mathbb R^n).$$ Since  was defined only for Schwartz symbols, we understand  as follows: if $a\in C_0^\infty(\mathbb R^{2n})$ is equal to 1 near some open set $U\subset \mathbb R^{2n}$, then the operator $\operatorname{Op}_h^\Lambda(1)$ defined in  is equal to the operator $\operatorname{Op}_h^\Lambda(a)$ defined in , microlocally near $\phi(U)\subset T^*(\mathbb R^n\times\mathbb R^n)$. Moreover, $\WFh(\operatorname{Op}_h^\Lambda(1))\cap T^*(\mathbb R^n\times\mathbb R^n)\subset\Lambda^0$. To see this, it is enough to note that for $a\in C_0^\infty(\mathbb R^{2n})$ and $\chi\in C_0^\infty(\mathbb R^n)$, we have $\chi\operatorname{Op}_h^\Lambda(1)\operatorname{Op}_h(a)=\operatorname{Op}_h^\Lambda(\tilde a)$, where $\tilde a(x,\xi)=\chi(x)a(x',0,\xi)\in C_0^\infty(\mathbb R^{2n})$ and $\operatorname{Op}_h^\Lambda(\tilde a)$ is defined using . We now study how $\operatorname{Op}^\Lambda_h(a)$ changes under quantized canonical transformations preserving its canonical relation (see §\[s:prelim-fio\]). Let $U,V\subset\mathbb R^{2n}$ be two bounded open sets and $\varkappa:U\to V$ a symplectomorphism such that $$\varkappa(\Gamma^0_\pm\cap U)=\Gamma^0_\pm\cap V,$$ with $\Gamma^0_\pm$ given by . We further assume that for each $(x',\xi')\in T^*\mathbb R^{n-1}$, the sets $\{x_n\mid (x',x_n,\xi',0)\in U\}$ and $\{\xi_n\mid (x',0,\xi',\xi_n)\in U\}$, and the corresponding sets for $V$, are either empty or intervals containing zero, so that the maps $\pi_\pm:U\cap\Gamma_\pm^0\to U\cap K^0$ are well-defined. Since $\varkappa$ preserves the subbundles $\mathcal V_\pm$, it commutes with the maps $\pi_\pm$ and thus preserves $\Lambda^0$; using the map $\phi$ from , we define the open sets $\widehat U,\widehat V\subset\mathbb R^{2n}$ and the diffeomorphism $\widehat\varkappa:\widehat U\to\widehat V$ by $$\begin{gathered} \widehat U:=\phi^{-1}(U\times U),\ \widehat V:=\phi^{-1}(V\times V),\quad \phi\circ\widehat\varkappa=\varkappa\circ\phi. \end{gathered}$$ \[l:change-of-variables\] Let $B,B':C^\infty(\mathbb R^{n})\to C_0^\infty(\mathbb R^n)$ be two compactly microlocalized Fourier integral operators associated to $\varkappa$ and $\varkappa^{-1}$, respectively,[^4] such that $$\label{e:b-inv} \begin{gathered} BB'=1+\mathcal O(h^\infty)\quad\text{microlocally near }V',\\ B'B=1+\mathcal O(h^\infty)\quad\text{microlocally near }U', \end{gathered}$$ for some open $U'\Subset U$, $V'\Subset V$ such that $\varkappa(U')=V'$. Then for each $a\in C_0^\infty(\widehat V)$, $$B'\operatorname{Op}^\Lambda_h(a)B=\operatorname{Op}^\Lambda_h(a_\varkappa)+\mathcal O(h^\infty)_{\mathscr S'\to\mathscr S},$$ for some classical symbol $a_\varkappa$ compactly supported in $\widehat U$, and $$\label{e:a-change} a_\varkappa(x,\xi)=\gamma^+_\varkappa(x,\xi')\gamma^-_\varkappa(x',\xi) a(\widehat\varkappa(x,\xi))+\mathcal O(h) \quad\text{on }\phi^{-1}(U'\times U'),$$ where $\gamma^\pm_\varkappa$ are smooth functions on $U\cap\Gamma_\pm$ depending on $\varkappa,B,B'$ with $\gamma^\pm_\varkappa|_{K^0\cap U'}=1$. Assume first that $\varkappa$ has a generating function $S(x,\eta)$: $$\varkappa(x,\xi)=(y,\eta)\Longleftrightarrow \xi=\partial_x S(x,\eta),\ y=\partial_\eta S(x,\eta).$$ If $\mathscr D_S\subset\mathbb R^{2n}$ is the domain of $S$, then for each $(x',\eta')\in T^*\mathbb R^{n-1}$, the sets $\{x_n\mid (x',x_n,\eta',0)\in\mathscr D_S\}$ and $\{\eta_n\mid (x',0,\eta',\eta_n)\in\mathscr D_S\}$ are either empty or intervals containing zero. Since $\varkappa$ preserves $\Gamma_\pm$, we find $\partial_{\eta_n}S(x',0,\eta)=\partial_{x_n}S(x,\eta',0)=0$ and thus $$\label{e:S-eq} S(x,\eta',0)=S(x',0,\eta)=S(x',0,\eta',0).$$ We can write, modulo $\mathcal O(h^\infty)_{\mathscr S'\to\mathscr S}$ errors, $$\begin{gathered} Bu(y)=(2\pi h)^{-n}\int e^{{i\over h}(y\cdot\eta-S(x,\eta))}b(x,\eta;h)u(x)\,dxd\eta,\\ B'u(x)=(2\pi h)^{-n}\int e^{{i\over h}(S(x,\eta)-y\cdot\eta)}b'(x,\eta;h)u(y)\,dyd\eta, \end{gathered}$$ where $b,b'$ are compactly supported classical symbols and by  the principal symbols $b_0$ and $b'_0$ have to satisfy for $(x,\xi)\in U'$, $$\label{e:b-eq} b_0(x,\eta)b'_0(x,\eta)=|\det \partial^2_{x\eta}S(x,\eta)|.$$ We can now use oscillatory testing  to get $$\begin{gathered} a_\varkappa(x,\xi):=e^{-{i\over h}x'\cdot\xi'}B'\operatorname{Op}_h^\Lambda(a) B(e^{{i\over h}x\cdot\xi})\\ = (2\pi h)^{-2n}\int_{\mathbb R^{4n}} e^{{i\over h}( -x'\cdot\xi'+S(x,\tilde\eta)-y\cdot\tilde\eta+y'\cdot \eta'-S(\tilde x,\eta)+\tilde x\cdot\xi)} b'(x,\tilde\eta;h)a(y,\eta)b(\tilde x,\eta;h)\,dyd\tilde \eta d\eta d\tilde x. \end{gathered}$$ We analyse this integral by the method of stationary phase; this will yield that $a_\varkappa$ is a classical symbol in $h$, compactly supported in $\widehat U$ modulo an $\mathcal O(h^\infty)_{\mathscr S(\mathbb R^{2n})}$ error, and thus $B'\operatorname{Op}_h^\Lambda(a)B=\operatorname{Op}_h^\Lambda(a_\varkappa)$. The stationary points are given by $$\tilde\eta=(\eta',0),\ \tilde x=(x',0),\ (y,\eta)=\widehat\varkappa(x,\xi).$$ The value of the phase at stationary points is zero due to . To compute the Hessian, we make the change of variables $\tilde\eta=\check\eta+(\eta',0)$. We can then remove the variables $y,\check\eta$ and pass from the original Hessian to $\partial^2_{\eta'\eta'} S(x,\eta',0)-\partial^2 S(x',0,\eta)$, where the first matrix is padded with zeros. Since $\partial_{\eta_n}S(x',0,\eta)=0$, we have $\partial^2_{\eta_n\eta_n}S=\partial^2_{\eta_n x'}S=\partial^2_{\eta_n \eta'}S=0$ at $(x',0,\eta)$, therefore we can remove the $x_n,\eta_n$ variables, with a multiplicand of $(\partial^2_{x_n\eta_n}S(x',0,\eta))^2$ in the determinant. Next, by  $\partial^2_{\eta'\eta'}S(x',0,\eta)=\partial^2_{\eta'\eta'}S(x,\eta',0)$; therefore, the Hessian has signature zero and determinant $$(\partial^2_{x_n\eta_n}S(x',0,\eta)\det \partial^2_{x'\eta'}S(x',0,\eta))^2.$$ Since $\partial^2_{x'\eta_n}S(x',0,\eta)=0$, this is equal to $(\det \partial^2_{x\eta}S(x',0,\eta))^2$. Therefore, we get  with $$\gamma^+_\varkappa(x,\xi')\gamma^-_\varkappa(x',\xi)={b_0'(x,\eta',0)b_0(x',0,\eta)\over |\det \partial^2_{x\eta}S(x',0,\eta)|} ={b'_0(x,\eta',0)\over b'_0(x',0,\eta)};$$ here $(y,\eta)=\widehat\varkappa(x,\xi)$ and the last equality follows from . We then find $$\label{e:transition-coeff} \gamma^+_\varkappa(x,\xi')=b'_0(x,\eta',0)/b'_0(x',0,\eta',0),\ \gamma^-_\varkappa(x',\xi)=b'_0(x',0,\eta',0)/b'_0(x',0,\eta).$$ We now consider the case of general $\varkappa$. Using a partition of unity for $a$, we may assume that the intersection $U\cap K^0$ is arbitrary small. We now represent $\varkappa$ as a product of several canonical relations, each of which satisfies the conditions of this Proposition and has a generating function; this will finish the proof. First of all, consider a canonical transformation of the form $$\label{e:tan-can} (x,\xi)\mapsto (y,\eta),\quad (y',\eta')=\widetilde\varkappa(x',\xi'),\ (y_n,\eta_n)=(x_n,\xi_n),$$ with $\widetilde\varkappa$ a canonical transformation on $T^*\mathbb R^{n-1}\simeq K^0$. We can write $\widetilde\varkappa$ locally as a product of canonical transformations close to the identity, each of which has a generating function – see [@e-z Theorems 10.4 and 11.4]. If $\widetilde S(x',\eta')$ is a generating function for $\widetilde\varkappa$, then $\widetilde S(x',\eta')+x_n\eta_n$ is a generating function for . Multiplying our $\varkappa$ by a transformation of the form  with $\widetilde\varkappa=(\varkappa|_{K^0})^{-1}$, we reduce to the case $$\varkappa(x',0,\xi',0)=(x',0,\xi',0)\quad\text{for } (x',0,\xi',0)\in U\cap K^0.$$ If $\varkappa(x,\xi)=(y(x,\xi),\eta(x,\xi))$, since $\varkappa$ commutes with $\pi_\pm$ we have $$\label{e:can-int} \begin{gathered} y'(x,\xi',0)=y'(x',0,\xi)=x',\\ \eta'(x,\xi',0)=\eta'(x',0,\xi)=\xi'. \end{gathered}$$ We now claim that $\varkappa$ has a generating function, if we shrink $U$ to be a small neighborhood of $U\cap (\Gamma^0_+\cup\Gamma^0_-)$ (which does not change anything since $\operatorname{Op}_h^\Lambda(a)$ is microlocalized in $\Gamma^0_-\times\Gamma^0_+$). For that, it is enough to show that the map $$\psi:(x,\xi)\mapsto (x,\eta(x,\xi))$$ is a diffeomorphism from $U$ onto some open subset $\mathscr D_S\subset\mathbb R^{2n}$. We first show that $\psi$ is a local diffeomorphism near $\Gamma^0_\pm$; that is, the differential $\partial_\xi\eta$ is nondegenerate on $\Gamma^0_\pm$. By , $\partial_{x',\xi'}(y',\eta')$ equals the identity on $\Gamma^0_+\cup\Gamma^0_-$; moreover, on $\Gamma^0_+$ we have $\partial_{x,\xi'}\eta_n=0$ and $\partial_{x_n}(y',\eta')=0$ and on $\Gamma^0_-$, we have $\partial_{x',\xi}y_n=0$ and $\partial_{\xi_n}(y',\eta')=0$. It follows that on $\Gamma^0_+\cup\Gamma^0_-$, $\det \partial_\xi\eta=\partial_{\xi_n}\eta_n$ and since $\varkappa$ is a diffeomorphism, $0\neq \det\partial_{(x,\xi)}(y,\eta)=\partial_{x_n}y_n\cdot\partial_{\xi_n}\eta_n$, yielding $\det \partial_\xi\eta\neq 0$. It remains to note that $\psi$ is one-to-one on $\Gamma^0_+\cup\Gamma^0_-$, which follows immediately from the identities $\psi(x,\xi',0)=(x,\xi',0)$ and $\psi(x',0,\xi)=\varkappa(x',0,\xi)$. General case {#s:general} ------------ We now consider the case of general $\Gamma_\pm^\circ,K^\circ,\Lambda^\circ$, satisfying the assumptions from the beginning of §\[s:calculus\]. We start by shrinking $\Gamma_\pm^\circ$ so that our setup can locally be conjugated to the model case of §\[s:model\]. (The set $\widehat K$ will be chosen in §\[s:construction-1\].) \[l:reduce-model\] Let $\widehat K\subset K^\circ$ be compact. Then there exist $\tilde\delta>0$ and - a finite collection of open sets $U_i\subset T^*X$, such that $$\widehat K\subset \widetilde K:=\bigcup_i K_i,\quad K_i:=K^\circ\cap U_i.$$ - symplectomorphisms $\varkappa_i$ defined in a neighborhood of $U_i$ and mapping $U_i$ onto $$\label{e:v-delta} V_{\tilde\delta}:=\{|(x',\xi')|<\tilde\delta,\ |x_n|<\tilde\delta,\ |\xi_n|<\tilde\delta\}\subset T^* \mathbb R^n,$$ such that, with $\Gamma_\pm^0$ defined in , $$\varkappa_i(U_i\cap\Gamma_\pm^\circ)=V_{\tilde\delta}\cap\Gamma_\pm^0;$$ - compactly microlocalized Fourier integral operators $$B_i:C^\infty(X)\to C_0^\infty(\mathbb R^n),\ B'_i:C^\infty(\mathbb R^n)\to C_0^\infty(X),$$ associated to $\varkappa_i$ and $\varkappa_i^{-1}$, respectively, such that $$\label{e:b-i-inv} B_iB'_i=1\quad\text{near }V_{\tilde\delta},\ B'_iB_i=1\quad\text{near }U_i.$$ It is enough to show that each point $\rho\in K^\circ$ has a neighborhood $U_\rho$ and a symplectomorphism $\varkappa_\rho:U_\rho\to V_\rho\subset T^*\mathbb R^n$ such that $\varkappa_\rho(U_\rho\cap\Gamma_\pm^\circ)=V_\rho\cap\Gamma_\pm^0$; see for example [@e-z Theorem 11.5] for how to construct the operators $B_i,B'_i$ locally quantizing the canonical transformations $\varkappa_\rho,\varkappa_\rho^{-1}$. By the Darboux theorem [@e-z Theorem 12.1] (giving a symplectomorphism mapping an arbitrarily chosen defining function of $\Gamma_-^\circ$ to $x_n$), we can reduce to the case $\rho=0\in T^*\mathbb R^n$ and $\Gamma_-^\circ=\{x_n=0\}$ near $0$. Since $\Gamma_+^\circ\cap\Gamma_-^\circ=K^\circ$ is symplectic, the Poisson bracket of the defining function $x_n$ of $\Gamma_-^\circ$ and any defining function $\varphi_+$ of $\Gamma_+^\circ$ is nonzero at $0$; thus, $\partial_{\xi_n}\varphi_+(0)\neq 0$ and we can write $\Gamma_+^\circ$ locally as the graph of some function: $$\Gamma_+^\circ=\{\xi_n=F(x,\xi')\}.$$ Put $\varphi_+'(x,\xi)=\xi_n-F(x,\xi')$, then $\{\varphi_+',x_n\}=1$. It remains to apply the Darboux theorem once again, obtaining a symplectomorphism preserving $x_n$ and mapping $\varphi'_+$ to $\xi_n$. We now consider the sets $$\label{e:smaller-stuff} \begin{gathered} \widetilde \Gamma_\pm:=\bigcup_i\Gamma_\pm^i,\quad \Gamma_\pm^i:=\Gamma_\pm^\circ\cap U_i,\\ \widetilde\Lambda:=\bigcup_i\Lambda_i,\quad \Lambda_i:=\{(\rho_-,\rho_+)\in\Lambda^\circ\mid \rho_\pm\in\Gamma_\pm^i\}. \end{gathered}$$ Let $\widehat\Gamma_\pm\subset\widetilde\Gamma_\pm$ be compact, with $\pi_\pm(\widehat\Gamma_\pm)=\widehat K$ and for each $\rho\in \widehat K$, the set $\pi_\pm^{-1}(\rho)\cap\widehat\Gamma_\pm$ is a flow line of $\mathcal V_\pm$ containing $\rho$. Define the compact set $$\label{e:hat-lambda} \widehat\Lambda:=\{(\rho_-,\rho_+)\in\Lambda^\circ\mid \rho_\pm\in\widehat\Gamma_\pm\}$$ and assume that $\widehat\Gamma_\pm$ are chosen so that $\widehat\Lambda\subset\widetilde\Lambda$. The goal of this subsection is to obtain an invariant notion of the principal symbol of Fourier integral operators in $\II$, microlocally near $\widehat\Lambda$. Define the diffeomorphisms $\widehat\varkappa_i:\Lambda_i\to V_{\tilde\delta}$ by the formula $$(\varkappa_i(\rho_-),\varkappa_i(\rho_+))=\phi(\widehat\varkappa_i(\rho_-,\rho_+)),\quad (\rho_-,\rho_+)\in\Lambda_i;$$ here $\phi$ is defined in . Consider some $A\in \II$, then $B_iAB'_i$ is a Fourier integral operator associated to the model canonical relation $\Lambda^0$ from §\[s:model\] (with the antiderivatives on $\Lambda^\circ$ and $\Lambda^0$ chosen in the beginning of §\[s:calculus\]). Therefore, there exists a compactly supported classical symbol $\tilde a^i(x,\xi;h)$ on $\mathbb R^{2n}$ such that, with $\operatorname{Op}_h^\Lambda$ defined in , $$\label{e:model-reduction} B_iAB'_i=\operatorname{Op}_h^\Lambda(\tilde a^i)+\mathcal O(h^\infty)_{\mathscr S'\to \mathscr S}.$$ By , we find $$A=B'_i\operatorname{Op}_h^\Lambda(\tilde a^i)B_i+\mathcal O(h^\infty)\quad\text{microlocally near }\Lambda_i.$$ Define the function $a^i\in C^\infty(\Lambda_i)$ using the principal symbol $\tilde a^i_0$ by $$a^i=\tilde a^i_0\circ\widehat\varkappa_i.$$ By Proposition \[l:change-of-variables\], applied to the Fourier integral operators $B_jB'_i$ and $B_iB'_j$ quantizing $\varkappa=\varkappa_j\circ\varkappa_i^{-1}$ and $\varkappa^{-1}$, respectively, with $U'=\varkappa_i(U_i\cap U_j)$, $V'=\varkappa_j(U_i\cap U_j)$ we see that whenever $\Lambda_i\cap\Lambda_j\neq\emptyset$, we have $$\label{e:transition} a^i|_{\Lambda_i\cap\Lambda_j}=(\gamma_{ij}^-\otimes \gamma_{ij}^+)a^j|_{\Lambda_i\cap\Lambda_j},$$ where $\gamma_{ij}^\pm$ are smooth functions on $\Gamma_\pm^i\cap \Gamma_\pm^j$ and $\gamma^\pm_{ij}|_K=1$. Moreover, $\gamma_{ji}^\pm=(\gamma_{ij}^\pm)^{-1}$ and $\gamma_{ij}^\pm\gamma_{jk}^\pm=\gamma_{ik}^\pm$ on $\Gamma_\pm^i\cap\Gamma_\pm^j\cap\Gamma_\pm^k$ (this can be seen either from the fact that the formulas  for different $i,j$ have to be compatible with each other, or directly from ). Therefore, we can consider smooth line bundles $\mathcal E_\pm$ over $\widetilde\Gamma_\pm$ with smooth sections $e^i_\pm$ of $\mathcal E_\pm|_{\Gamma_\pm^i}$ such that $e^j_\pm=\gamma_{ij}^\pm e^i_\pm$ on $\Gamma_\pm^i\cap\Gamma_\pm^j$ – see for example [@ho1 §6.4]. Define the line bundle $\mathcal E$ over $\widetilde\Lambda$ using the projection maps from : $$\mathcal E=(\tilde\pi_-^*\mathcal E^-)\otimes(\tilde \pi_+^*\mathcal E^+)$$ and for $A\in \II$, the symbol $\sigma_\Lambda(A)\in C^\infty(\widetilde\Lambda;\mathcal E)$ by the formula $$\label{e:symbol} \sigma_\Lambda(A)|_{\Lambda_i}=a^i(\tilde\pi_-^*e^i_-\otimes \tilde\pi_+^*e^i_+).$$ Note that the bundle $\mathcal E$ can be studied in detail using the global theory of Fourier integral operators (see for instance [@ho4 §25.1]). However, the situation in our special case is considerably simplified, since the Maslov bundle does not appear. We have $\sigma_\Lambda(A)=0$ near $\widehat\Lambda$ if and only if $A\in h\II$ microlocally near $\widehat\Lambda$. Moreover, for all $a\in C^\infty(\widetilde\Lambda;\mathcal E)$, there exists $A\in \II$ such that $\sigma_\Lambda(A)=a$ near $\widehat\Lambda$. The restrictions $\mathcal E_\pm|_{\widetilde K}$ are canonically trivial; that is, for $a_\pm\in C^\infty(\widetilde\Gamma_\pm;\mathcal E_\pm)$, we can view $a_\pm|_{\widetilde K}$ as a function on $\widetilde K$, by taking $e_\pm^i|_{K_i}=1$. The bundles $\mathcal E_\pm$ are trivial: \[l:trivial\] There exist sections $a_\pm\in C^\infty(\widetilde\Gamma_\pm;\mathcal E_\pm)$, nonvanishing near $\widehat\Gamma_\pm$ and such that $a_\pm|_{\widetilde K}=1$ near $\widehat K$. Since $\gamma_{ij}^\pm$ is a nonvanishing smooth function on $\Gamma_\pm^i\cap\Gamma_\pm^j$ such that $\gamma_{ij}^\pm|_{K_i\cap K_j}=1$, we can write $$\gamma_{ij}^\pm=\exp(f_{ij}^\pm),$$ where $f_{ij}^\pm$ is a uniquely defined function on $\Gamma_\pm^i\cap\Gamma_\pm^j$, such that $f_{ij}^\pm|_{K_i\cap K_j}=0$. We now put near $\widehat\Gamma_\pm$, $$a_\pm|_{\Gamma_\pm^i}=\exp(b_\pm^i)e_\pm^i,$$ where $b_\pm^i\in C^\infty(\Gamma_\pm^i)$ are such that near $\widehat\Gamma_\pm$ and $\widehat K$ respectively, $$(b_\pm^i-b_\pm^j)|_{\Gamma_\pm^i\cap\Gamma_\pm^j}=f_{ij}^\pm,\quad b_\pm^i|_{K_i}=0.$$ Such functions exist since $f_{ij}^\pm$ is a cocycle: $$f_{ii}^\pm=f_{ij}^\pm+f_{ji}^\pm=0;\quad f_{ij}^\pm+f_{jk}^\pm=f_{ik}^\pm\quad\text{on }\Gamma_\pm^i\cap\Gamma_\pm^j\cap\Gamma_\pm^k$$ and since the sheaf of smooth functions is fine; more precisely, if $1=\sum_i \chi_i$ is a partition of unity on $\widehat\Gamma_\pm$, with $\operatorname{supp}\chi_i\subset\Gamma_\pm^i$, we put $$b_\pm^i=\sum_k \chi_k f_{ik}^\pm. \qedhere$$ We now state the properties of the calculus, following directly from –, the general theory of Fourier integral operators, and Egorov’s Theorem [@e-z Theorem 11.1] (see the beginning of §\[s:calculus\] for multiplying two elements of $\II$): \[l:calculus\] Assume that $A_1,A_2\in \II,P\in \Psi^k(X)$. Then $A_1A_2,A_1P,PA_1$ lie in $\II$, and $$\begin{gathered} \label{e:good-calc-1} \sigma_\Lambda(A_1A_2)(\rho_-,\rho_+)=\sigma_\Lambda(A_2)(\rho_-,\pi_-(\rho_-)) \otimes\sigma_\Lambda(A_1)(\pi_+(\rho_+),\rho_+),\\ \label{e:good-calc-2} \sigma_\Lambda(A_1P)(\rho_-,\rho_+)=\sigma(P)(\rho_-)\cdot\sigma_\Lambda(A_1)(\rho_-,\rho_+),\\ \label{e:good-calc-3} \sigma_\Lambda(PA_1)(\rho_-,\rho_+)=\sigma(P)(\rho_+)\cdot\sigma_\Lambda(A_1)(\rho_-,\rho_+).\end{gathered}$$ Here in , $\sigma_\Lambda(A_2)(\rho_-,\pi_-(\rho_-))$ and $\sigma_\Lambda(A_1)(\pi_+(\rho_+),\rho_+)$ are considered as sections of $\mathcal E_-$ and $\mathcal E_+$, respectively. We next give a parametrix construction for operators of the form $1-A$, with $A\in \II$, needed in §\[s:grushin\]: \[l:funny-parametrix\] Let $A\in \II$ and assume that $$\WFh(A)\subset\widehat\Lambda;\quad \sigma_\Lambda(A)|_{\widetilde K}\neq 1\text{ everywhere}.$$ Then there exists $B\in \II$ with $\WFh(B)\subset\widehat\Lambda$, and such that $$(1-A)(1-B)=1+\mathcal O(h^\infty),\quad (1-B)(1-A)=1+\mathcal O(h^\infty).$$ Moreover, $B$ is uniquely defined modulo $\mathcal O(h^\infty)$ and $$\label{e:funny-symbol} \sigma_\Lambda(B)(\rho_-,\rho_+)= {\sigma_\Lambda(A)(\rho_-,\pi_-(\rho_-))\otimes \sigma_\Lambda(A)(\pi_+(\rho_+),\rho_+) \over \sigma_\Lambda(A)(\pi_-(\rho_-),\pi_+(\rho_+))-1}-\sigma_\Lambda(A)(\rho_-,\rho_+).$$ Take any $B_1\in \II$ with $\WFh(B_1)\subset\widehat\Lambda$ and symbol given by . By , $(1-A)(1-B_1)=1-hR$, for some $R\in \II$ with $\WFh(R)\subset\widehat\Lambda$. Define $B_2\in \II$ by the asymptotic Neumann series $$-B_2\sim\sum_{j\geq 1}h^jR^j.$$ Define $B\in \II$ by the identity $1-B=(1-B_1)(1-B_2)$, then $(1-A)(1-B)=1+\mathcal O(h^\infty)$. Similarly, we construct $B'\in \II$ such that $(1-B')(1-A)=1+\mathcal O(h^\infty)$. A standard algebraic argument, see for example the proof of [@ho3 Theorem 18.1.9], shows that $B'=B+\mathcal O(h^\infty)$ and both are determined uniquely modulo $\mathcal O(h^\infty)$. We finish this subsection with a trace formula for operators in $\II$, used in §\[s:trace\]: \[l:trace-basic\] Assume that $A\in \II$ and $\WFh(A)\subset\widehat\Lambda$. Then, with $d\operatorname{Vol}_\sigma=\sigma_S^{n-1}/(n-1)!$ denoting the symplectic volume form and $j_K:K^\circ\to\Lambda^\circ$ defined in , $$(2\pi h)^{n-1}\operatorname{Tr}A=\int_{\widehat K} \sigma_\Lambda(A)\circ j_K \,d\operatorname{Vol}_\sigma+\mathcal O(h).$$ By a microlocal partition of unity, we reduce to the case when $\WFh(A)$ lies entirely in one of the sets $\Lambda_i$ defined in . If $\tilde a_i$ is defined by , then by the cyclicity of the trace, $\operatorname{Tr}A=\operatorname{Tr}\operatorname{Op}_h^\Lambda(\tilde a_i)+\mathcal O(h^\infty)$. It remains to note that for any $a(x,\xi)\in C_0^\infty(\mathbb R^{2n})$, $$(2\pi h)^{n-1}\operatorname{Tr}\operatorname{Op}_h^\Lambda(a)=\int_{\mathbb R^{2n-2}}a(x',0,\xi',0)\,dx'd\xi'+\mathcal O(h),$$ seen directly from  by the method of stationary phase in the $x_n,\xi_n$ variables. Microlocal idempotents {#s:idempotents} ---------------------- In this subsection, we establish properties of microlocal idempotents associated to the Lagrangian $\Lambda^\circ$ considered in §\[s:general\], microlocally on the compact set $\widehat\Lambda$ defined in . We use the principal symbol $\sigma_\Lambda$ constructed in . \[d:microlocal-idempotent\] We call $A\in \II$ a *microlocal idempotent* of order $k>0$ near $\widehat\Lambda$, if $A^2=A+\mathcal O(h^k)_{\II}$ microlocally near $\widehat\Lambda$ and $\sigma_\Lambda(A)$ does not vanish on $\widehat\Lambda$. In the following Proposition, part 1 is concerned with the principal part of the idempotent equation; part 2 establishes a normal form for microlocal idempotents, making it possible to conjugate them microlocally to the operator $\operatorname{Op}_h^\Lambda(1)$ from . Part 3 is used to construct a global idempotent of all orders in Proposition \[l:idempotent-exists\] below, while part 4 establishes properties of commutators used in the construction of §\[s:global-construction\]. \[l:idempotents\] 1. $A\in \II$ is a microlocal idempotent of order 1 near $\widehat\Lambda$ if and only if near $\widehat\Lambda$, $$\label{e:idempotent-principal} \sigma_\Lambda(A)(\rho_-,\rho_+)=a_0^-(\rho_-)\otimes a_0^+(\rho_+)$$ for some sections $a_0^\pm\in C^\infty(\widetilde\Gamma_\pm;\mathcal E_\pm)$ nonvanishing near $\widehat\Gamma_\pm$ and such that $a_0^\pm|_{\widetilde K}=1$ near $\widehat K$. Moreover, $a_0^\pm$ are uniquely determined by $A$ on $\widehat \Gamma_\pm$. 2\. If $A,B\in \II$ are two microlocal idemptotents of order $k>0$ near $\widehat\Lambda$, then there exists an operator $Q\in\Psi^{\operatorname{comp}}(X)$, elliptic on $\widehat\Gamma_+\cup\widehat\Gamma_-$ and such that $B=Q A Q^{-1}+\mathcal O(h^k)_{\II}$ microlocally near $\widehat\Lambda$. Here $Q^{-1}$ denotes an elliptic parametrix of $Q$ constructed in Proposition \[l:eparametrix\]. 3\. If $A\in \II$ is a microlocal idempotent of order $k>0$ near $\widehat\Lambda$, and $A^2-A=h^kR_k+\mathcal O(h^\infty)$ microlocally near $\widehat\Lambda$ for some $R_k\in \II$, then for $\rho_\pm\in\widetilde\Gamma_\pm$ near $\widehat\Gamma_\pm$, $$\label{e:r-eq} \begin{gathered} \sigma_\Lambda(R_k)(\pi_+(\rho_+),\rho_+)=\sigma_\Lambda(R_k)(\pi_+(\rho_+),\pi_+(\rho_+))\cdot a_0^+(\rho_+),\\ \sigma_\Lambda(R_k)(\rho_-,\pi_-(\rho_-))=\sigma_\Lambda(R_k)(\pi_-(\rho_-),\pi_-(\rho_-))\cdot a_0^-(\rho_-), \end{gathered}$$ with $a_0^\pm$ defined in . 4\. If $A\in \II$ is a microlocal idempotent of all orders near $\widehat\Lambda$, $P\in\Psi^{\operatorname{comp}}(X)$ is compactly supported, and $[P,A]=h^k S_k+\mathcal O(h^\infty)$ microlocally near $\widehat\Lambda$ for some $S_k\in \II$, then near $\widehat\Lambda$, $$\sigma_\Lambda(S_k)(\rho_-,\rho_+)=a_0^-(\rho_-)\otimes \sigma_\Lambda(S_k)(\pi_+(\rho_+),\rho_+) +\sigma_\Lambda(S_k)(\rho_-,\pi_-(\rho_-))\otimes a_0^+(\rho_+).$$ In particular, $\sigma_\Lambda(S_k)\circ j_K=0$ near $\widehat K$, with $j_K:K^\circ\to\Lambda^\circ$ defined in . In this proof, all the equalities of operators in $\II$ and the corresponding symbols are presumed to hold microlocally near $\widehat\Lambda$. 1\. By , we have $A^2=A+\mathcal O(h)$ if and only if $$\sigma_\Lambda(A)(\rho_-,\rho_+)=\sigma_\Lambda(A)(\rho_-,\pi_-(\rho_-))\otimes \sigma_\Lambda(A)(\pi_+(\rho_+),\rho_+).$$ In particular, restricting to $\widetilde K$, we obtain $\sigma_\Lambda(A)=\sigma_\Lambda(A)^2$ near $\widehat K$. Since $\sigma_\Lambda(A)$ is nonvanishing, we get $\sigma_\Lambda(A)|_{\widetilde K}=1$ near $\widehat K$. It then remains to put $a^-_0(\rho_-)=\sigma_\Lambda(A)(\rho_-,\pi_-(\rho_-))$ and $a^+_0(\rho_+)=\sigma_\Lambda(A)(\pi_+(\rho_+),\rho_+)$. 2\. We use induction on $k$. For $k=1$, we have by  and , $$\sigma_\Lambda(QAQ^{-1})(\rho_-,\rho_+)={\sigma(Q)(\rho_+)\over\sigma(Q)(\rho_-)} \sigma_\Lambda(A)(\rho_-,\rho_+).$$ If $a_0^\pm$ and $b_0^\pm$ are given by , then it is enough to take any $Q$ with $$\label{e:q-principal} \sigma(Q)|_{\widetilde\Gamma_-}=a_0^-/b_0^-,\ \sigma(Q)|_{\widetilde\Gamma_+}=b_0^+/a_0^+,$$ this is possible since the restrictions of $a_0^\pm$ and $b_0^\pm$ to $\widetilde K$ are equal to 1. Now, assuming the statement is true for $k\geq 1$, we prove it for $k+1$. We have $B=\widetilde Q A\widetilde Q^{-1}+\mathcal O(h^k)$ for some $\widetilde Q\in\Psi^{\operatorname{comp}}$ elliptic on $\widehat\Gamma_+\cup\widehat\Gamma_-$; replacing $A$ by $\widetilde Q A\widetilde Q^{-1}$, we may assume that $B=A+\mathcal O(h^k)$. Then $B-A=h^k R_k$ for some $R_k\in \II$; since both $A$ and $B$ are microlocal idempotents of order $k+1$, we find $R_k=AR_k+R_kA+\mathcal O(h)$ and thus by , $$\label{e:r-eqn} \sigma_\Lambda(R_k)(\rho_-,\rho_+)=a_0^-(\rho_-)\otimes \sigma_\Lambda(R_k)(\pi_+(\rho_+),\rho_+) +\sigma_\Lambda(R_k)(\rho_-,\pi_-(\rho_-))\otimes a_0^+(\rho_+).$$ Take $Q=1+h^kQ_k$ for some $Q_k\in\Psi^{\operatorname{comp}}$, then $Q^{-1}=1-h^kQ_k+\mathcal O(h^{k+1})$ and $$QAQ^{-1}=A+h^k[Q_k,A]+\mathcal O(h^{k+1}).$$ Now, $B=QAQ^{-1}+\mathcal O(h^{k+1})$ if and only if $$(\sigma(Q_k)(\rho_+)-\sigma(Q_k)(\rho_-))a_0^-(\rho_-)\otimes a_0^+(\rho_+)=\sigma_\Lambda(R_k)(\rho_-,\rho_+).$$ By , it is enough to choose $Q_k$ such that for $\rho_\pm\in\widetilde \Gamma_\pm$, $$\sigma(Q_k)(\rho_-)=-{\sigma_\Lambda(R_k)(\rho_-,\pi_-(\rho_-))\over a_0^-(\rho_-)},\quad \sigma(Q_k)(\rho_+)={\sigma_\Lambda(R_k)(\pi_+(\rho_+),\rho_+)\over a_0^+(\rho_+)},$$ this is possible since $\sigma_\Lambda(R_k)\circ j_K=0$ (with $j_K$ defined in ) as follows from . 3\. Since this is a local statement, we can use  to reduce to the model case of §\[s:model\]. Using part 2 and the fact that the operator $\operatorname{Op}_h^\Lambda(1)$ considered in  is a microlocal idemptotent of all orders, we can write $$A=Q\operatorname{Op}_h^\Lambda(1)Q^{-1}+h^kA_k,$$ for some elliptic $Q\in\Psi^{\operatorname{comp}}$ and $A_k\in I_{\operatorname{comp}}(\Lambda^0)$. Then $$R_k=Q\operatorname{Op}_h^\Lambda(1)Q^{-1}A_k+A_kQ\operatorname{Op}_h^\Lambda(1)Q^{-1}-A_k+\mathcal O(h);$$ follows by  since $\sigma_\Lambda(Q\operatorname{Op}_h^\Lambda(1)Q^{-1})=\sigma_\Lambda(A)$ is given by . 4\. As in part 3, we reduce to the model case of §\[s:model\] and use part 2 to write $A=Q\operatorname{Op}_h^\Lambda(1)Q^{-1}+\mathcal O(h^\infty)$; then $$[P,A]=Q[Q^{-1}PQ,\operatorname{Op}_h^\Lambda(1)]Q^{-1}+\mathcal O(h^\infty).$$ Put $\widetilde P=Q^{-1}PQ$; by  and  we have $[\widetilde P,\operatorname{Op}_h^\Lambda(1)]=\operatorname{Op}_h^\Lambda(s\circ\phi)$, where $\phi$ is given by  and $$s(\rho_-,\rho_+;h)=\tilde p(\rho_+;h)-\tilde p(\rho_-;h),$$ where $\widetilde P=\operatorname{Op}_h(\tilde p)$; thus $$s(\rho_-,\rho_+;h)=s(\pi_+(\rho_+),\rho_+;h)+s(\rho_-,\pi_-(\rho_-);h).$$ It remains to conjugate by $Q$, keeping in mind . We can use part 3 of Proposition \[l:idempotents\], together with the triviality of the bundles $\mathcal E_\pm$, to show existence of a global idempotent, which is the starting point of the construction in §\[s:global-construction\]. \[l:idempotent-exists\] There exists a microlocal idempotent $\widetilde\Pi\in \II$ of all orders near $\widehat\Lambda$. We argue inductively, constructing microlocal idempotents $\widetilde\Pi_k$ of order $k$ for each $k$ and taking the asymptotic limit. To construct $\widetilde\Pi_1$, we use part 1 of Proposition \[l:idempotents\]; the existence of symbols $a_0^\pm$ was shown in Proposition \[l:trivial\]. Now, assume that $\widetilde\Pi_k$ is a microlocal idempotent of order $k>0$. By part 3 of Proposition \[l:idempotents\], we have $\widetilde\Pi_k^2-\widetilde\Pi_k=h^kR_k+\mathcal O(h^\infty)$ microlocally near $\widehat\Lambda$, where $R_k\in \II$ and $r_k=\sigma_\Lambda(R_k)$ satisfies . Put $\widetilde\Pi_{k+1}=\widetilde \Pi_k+h^k B_k$, for some $B_k\in \II$. We need to choose $B_k$ so that microlocally near $\widehat\Lambda$, $$R_k+\widetilde\Pi_k B_k+ B_k\widetilde \Pi_k- B_k=\mathcal O(h).$$ Taking $b_k=\sigma_\Lambda(B_k)$, by  this translates to $$b_k(\rho_-,\rho_+)=a_0^-(\rho_-)\otimes b_k(\pi_+(\rho_+),\rho_+)+ b_k(\rho_-,\pi_-(\rho_-))\otimes a_0^+(\rho_+) +r_k(\rho_-,\rho_+).$$ By , it is enough to take any $b_k^\pm\in C^\infty(\widetilde \Gamma_\pm;\mathcal E_\pm)$ such that near $\widehat K$, $b_k^\pm|_{\widetilde K}=-r_k\circ j_K$, with $j_K$ defined in  (for example, $b_k^\pm=-(r_k\circ j_K\circ \pi_\pm)a_0^\pm$) and put $$b_k(\rho_-,\rho_+):=a_0^-(\rho_-)\otimes b_k^+(\rho_+)+b_k^-(\rho_-)\otimes a_0^+(\rho_+) +r_k(\rho_-,\rho_+).\qedhere$$ Annihilating ideals {#s:ideals} ------------------- Assume that $\Pi\in \II$ is a microlocal idempotent of all orders near the set $\widehat\Lambda$ introduced in , see Definition \[d:microlocal-idempotent\]. We are interested in the following equations: $$\begin{gathered} \label{e:right-ideal} \Pi\Theta_-=\mathcal O(h^\infty)\quad\text{microlocally near }\widehat\Lambda,\\ \label{e:left-ideal} \Theta_+\Pi=\mathcal O(h^\infty)\quad\text{microlocally near }\widehat\Lambda,\end{gathered}$$ where $\Theta_\pm$ are pseudodifferential operators. The solutions to  form a right ideal and the solutions to  form a left ideal in the algebra of pseudodifferential operators. Moreover, by , , each solution $\Theta_\pm$ to the equations , satisfies $\sigma(\Theta_\pm)|_{\Gamma_\pm}=0$ near $\widehat\Gamma_\pm$ and each $\Theta_\pm$ such that $\WFh(\Theta_\pm)\cap\widehat\Gamma_\pm=\emptyset$ solves these equations. Note that in the model case of §\[s:model\], with $\Pi$ equaling the operator $\operatorname{Op}_h^\Lambda(1)$ from , and with the quantization procedure $\operatorname{Op}_h$ defined in , the set of solutions to  is the set of operators $\operatorname{Op}_h(\theta_-)$ with $\theta_-|_{x_n=0}=0$; that is, the right ideal generated by the operator $x_n$. The set of solutions to  is the set of operators $\operatorname{Op}_h(\theta_+)$ with $\theta_+|_{\xi_n=0}=0$; that is, the left ideal generated by the operator $hD_{x_n}$. This follows from the multiplication formulas  and , together with the multiplication formulas for the standard quantization [@e-z (4.3.16)]. We start by showing that our ideals are principal in the general setting: \[l:principal-ideal\] 1. For each defining functions $\varphi_\pm$ of $\Gamma_\pm^\circ$ near $\widehat\Gamma_\pm$, there exist operators $\Theta_\pm$ solving , , such that $\sigma(\Theta_\pm)=\varphi_\pm$ near $\widehat\Gamma_\pm$. Such operators are called *basic* solutions of the corresponding equations. 2\. If $\Theta_\pm,\Theta'_\pm$ are solutions to , , and moreover $\Theta_\pm$ are basic solutions, then there exist $Z_\pm\in\Psi^{\operatorname{comp}}$ such that $\Theta'_-=\Theta_-Z_-+\mathcal O(h^\infty)$ microlocally near $\widehat\Gamma_-$ and $\Theta'_+=Z_+\Theta_++\mathcal O(h^\infty)$ microlocally near $\widehat\Gamma_+$. We concentrate on the equation ; is handled similarly. Since the equations  and $\Theta'=\Theta_-Z_-$ are linear in $\Theta_-$ and $\Theta',Z_-$, respectively, we can use  and a pseudodifferential partition of unity to reduce to the model case of §\[s:model\]. Using part 2 of Proposition \[l:idempotents\], we can furthermore assume that $\Pi=\operatorname{Op}_h^\Lambda(1)$. To show part 1, in the model case, we can take $\Theta_-=\operatorname{Op}_h(\varphi_-)$, where $\varphi_-(x,\xi)$ is the given defining function of $\{x_n=0\}$. For part 2, if $\Theta_-=\operatorname{Op}_h(\varphi_-)$ and $\Theta'_-=\operatorname{Op}_h(\theta'_-)$, then we can write microlocally near $\widehat\Gamma_-$, $\Theta_-=x_n Y_-+\mathcal O(h^\infty)$, where $Y_-\in\Psi^{\operatorname{comp}}$ is elliptic on $\widehat\Gamma_-$; in fact, $Y_-=\operatorname{Op}_h(\varphi_-/x_n)$. Similarly we can write $\Theta'_-=x_nY'_-+\mathcal O(h^\infty)$ microlocally near $\widehat\Gamma_-$, for some $Y'_-\in\Psi^{\operatorname{comp}}$; it remains to put $Z_-=Y_-^{-1}Y'_-$ microlocally near $\widehat\Gamma_-$. For the microlocal estimate on the kernel of $\Pi$ in §\[s:estimate-kernel\], we need an analog of the following fact: $$\label{e:division} f\in C^\infty(\mathbb R^n)\ \Longrightarrow\ f(x)-f(x',0)=x_ng(x),\quad g\in C^\infty(\mathbb R^n),$$ where $f(x',0)$ is replaced by $\Pi f$ and multiplication by $x_n$ is replaced by a basic solution to . We start with a technical lemma for the model case: \[l:Xi-0\] Consider the operator $\Xi_0:C^\infty(\mathbb R^n)\to C^\infty(\mathbb R^n)$ defined by $$\Xi_0f(x',x_n)={f(x',x_n)-f(x',0)\over x_n}=\int_0^1 (\partial_{x_n}f)(x',tx_n)\,dt.$$ Then: 1\. $\Xi_0$ is bounded $H^1(\mathbb R^n)\to L^2(\mathbb R^n)$ and thus $\|\Xi_0\|_{H^1_h\to L^2}=\mathcal O(h^{-1})$. 2\. The wavefront set $\WFh(\Xi_0)$ defined in §\[s:prelim-basics\] satisfies[^5] $$\begin{gathered} \WFh(\Xi_0)\cap T^*(\mathbb R^n\times\mathbb R^n)\,\subset\,\Delta(T^*\mathbb R^n)\cup\Lambda^0\\ \cup \{(x',0,\xi,x',0,\xi',t\xi_n)\mid (x',\xi)\in\mathbb R^{2n-1},t\in [0,1]\}, \end{gathered}$$ where $\Delta(T^*\mathbb R^n)\subset T^*\mathbb R^n\times T^*\mathbb R^n$ is the diagonal and $\Lambda^0$ is defined using . 1\. Put $\lambda_t f(x',x_n)=(\partial_{x_n}f)(x',tx_n)$; then $$\|\Xi_0 f\|_{L^2}\leq \int_0^1\|\lambda_t f\|_{L^2}\,dt \leq \int_0^1 t^{-1/2}\|f\|_{H^1}\,dt \leq 2\|f\|_{H^1}.$$ 2\. Denote elements of $T^*(\mathbb R^n\times\mathbb R^n)$ by $(x,\xi,y,\eta)$. If $\chi\in C_0^\infty(\mathbb R)$ is supported away from zero, then, with $\operatorname{Op}_h^\Lambda(1)$ defined in , $$\chi(x_n) \Xi_0={\chi(x_n)\over x_n} (1-\operatorname{Op}_h^\Lambda(1)).$$ Since $\chi(x_n)/x_n$ is a smooth function, the identity operator has wavefront set on the diagonal, and $\WFh(\operatorname{Op}_h^\Lambda(1))\cap T^*(\mathbb R^n\times \mathbb R^n)\subset\Lambda^0$, we find $$\begin{gathered} \WFh(\Xi_0)\cap T^*(\mathbb R^n\times\mathbb R^n)\cap\{y_n\neq 0\}\,\subset\, \Delta(T^*\mathbb R^n) \cup \Lambda^0. \end{gathered}$$ Similarly, one has $\Xi_0\chi(x_n)=\chi(x_n)/x_n$; therefore, $$\WFh(\Xi_0)\cap\{x_n\neq 0\}\,\subset\, \Delta(T^*\mathbb R^n).$$ To handle the remaining part of the wavefront set, take $a,b\in C_0^\infty(T^*\mathbb R^n)$ such that $$(x',tx_n,\xi)\in\operatorname{supp}a,\ t\in [0,1]\Longrightarrow (x,\xi',t\xi_n)\not\in\operatorname{supp}b.$$ We claim that for any $\psi\in C_0^\infty(\mathbb R^n)$, $$\label{e:wf-internal} \operatorname{Op}_h(b)\psi\,\Xi_0\operatorname{Op}_h(a)\psi=\mathcal O(h^\infty);$$ indeed, the Schwartz kernel of this operator is $$\begin{gathered} \mathcal K(y,x)=(2\pi h)^{-2n}\int_{\mathbb R^{3n}\times [0,1]} e^{{i\over h}((y-z)\cdot\eta+(z'-x')\cdot\xi'+(tz_n-x_n)\xi_n)}\\ b(y,\eta)\psi(z)(ih^{-1}\xi_n a(z',tz_n,\xi)+(\partial_{z_n}a)(z',tz_n,\xi))\psi(x)\,d\xi d\eta dz dt. \end{gathered}$$ The stationary points of the phase in the $(\xi,\eta,z)$ variables are given by $$z=y,\ x'=y',\ x_n=ty_n,\ \eta'=\xi',\ \eta_n=t\xi_n$$ and lie outside of the support of the amplitude; by the method of nonstationary phase in the $(\xi,\eta,z)$ variables, the integral is $\mathcal O(h^\infty)_{C^\infty}$. Now, implies that $$\begin{gathered} \WFh(\Xi_0)\cap T^*(\mathbb R^n\times\mathbb R^n)\cap\{x_n=y_n=0\}\\\subset\, \{(x',0,\xi,x',0,\xi',t\xi_n)\mid (x',\xi)\in\mathbb R^{2n-1},\ t\in [0,1]\}, \end{gathered}$$ which finishes the proof. The microlocal analog of  in the general case is now given by \[l:Xi\] Let $\Pi\in \II$ be a microlocal idempotent of all orders near $\widehat\Lambda$ and $\Theta_-$ be a basic solution to , see Proposition \[l:principal-ideal\]. Then there exists an operator $\Xi:C^\infty(X)\to C_0^\infty(X)$ such that: 1\. $\WFh(\Xi)$ is a compact subset of $T^*(M\times M)$ and $\|\Xi\|_{L^2\to L^2}=\mathcal O(h^{-1})$; 2\. $\WFh(\Xi)\subset \Delta(T^*M)\cup \Lambda^\circ\cup \Upsilon$, where $\Delta(T^*M)\subset T^*M\times T^*M$ is the diagonal and $\Upsilon$ consists of all $(\rho_-,\rho'_-)$ such that $\rho_-,\rho'_-\in\Gamma_-^\circ$ and $\rho'_-$ lies on the segment of the flow line of $\mathcal V_-$ between $\rho_-$ and $\pi_-(\rho_-)$; 3\. $1-\Pi=\Theta_-\Xi+\mathcal O(h^\infty)$ microlocally near $\widehat K\times\widehat K$. By  and a microlocal partition of unity, we can reduce to the model case of §\[s:model\]. Moreover, by part 2 of Proposition \[l:idempotents\], we may conjugate by a pseudodifferential operator to make $\Pi=\operatorname{Op}_h^\Lambda(1)$. Finally, by part 2 of Proposition \[l:principal-ideal\] we can multiply $\Theta_-$ on the right by an elliptic pseudodifferential operator to make $\Theta_-=\operatorname{Op}_h(x_n)$. Then we can take $\Xi=A\Xi_0 A$, with $\Xi_0$ defined in Lemma \[l:Xi-0\] and $A\in\Psi^{\operatorname{comp}}(\mathbb R^n)$ compactly supported, with $A=1+\mathcal O(h^\infty)$ microlocally near $\widehat K$. The projector Pi {#s:global-construction} ================ In this section, we construct the microlocal projector $\Pi$ near a neighborhood $\widehat W$ of $K\cap p^{-1}([\alpha_0,\alpha_1])$ discussed in the introduction (Theorem \[t:our-Pi\] in §\[s:construction-1\]). In §\[s:construction-2\], we study the annihilating ideals for $\Pi$ in $\widehat W$ using §\[s:ideals\]. Construction of Pi {#s:construction-1} ------------------ Assume that the conditions of §§\[s:framework-assumptions\] and \[s:dynamics\] hold. Consider the sets $\Gamma_\pm^\circ$ and $K^\circ=\Gamma_+^\circ\cap\Gamma_-^\circ$ defined in  and let $\Lambda^\circ$ be given by . Put $$\widehat K := K\cap p^{-1}([\alpha_0-\delta_1/2,\alpha_1+\delta_1/2])\subset K^\circ,$$ here $\delta_1$ is defined in §\[s:projections\]. The sets $\Gamma_\pm^\circ$ satisfy the assumptions listed in the beginning of §\[s:calculus\], as follows from §§\[s:dynamics\] and \[s:projections\]. We choose $\delta>0$ small enough so that Lemma \[l:phi-pm\] holds (we will impose more conditions on $\delta$ in §\[s:construction-2\]) and consider the sets $$\label{e:w-hat} \begin{gathered} \widehat W:=U_\delta\cap p^{-1}([\alpha_0-\delta_1/2,\alpha_1+\delta_1/2]),\\ \widehat\Gamma_\pm:=\Gamma_\pm^\circ\cap\widehat W,\quad \widehat\Lambda:=\Lambda^\circ\cap (\widehat W\times\widehat W). \end{gathered}$$ Here $U_\delta$ is defined in . We now apply Proposition \[l:reduce-model\]; for $\delta$ small enough, $\widehat W,\widehat\Gamma_\pm$ are compact and $\widehat\Gamma_\pm,\widehat \Lambda$ satisfy the conditions listed after . Then  defines the principal symbol $\sigma_\Lambda(A)$ on a neighborhood of $\widehat\Lambda$ in $\Lambda^\circ$ for each $A\in\II$. \[t:our-Pi\] Let the assumptions of §§\[s:framework-assumptions\] and \[s:dynamics\] hold for all $r$, let $\Lambda^\circ$ be defined in  and $\widehat\Lambda\subset\Lambda^\circ$ be given by . Then there exists $\Pi\in \II$, uniquely defined modulo $\mathcal O(h^\infty)$ microlocally near $\widehat\Lambda$, such that the principal symbol of $\Pi$ is nonvanishing on $\widehat\Lambda$ and, with $P\in\Psi^{\operatorname{comp}}(X)$ defined in Lemma \[l:resolution\], $$\begin{aligned} \label{e:th-idempotent} \Pi^2-\Pi&=\mathcal O(h^\infty)\quad\text{microlocally near }\widehat\Lambda,\\ \label{e:th-commute} [P,\Pi]&=\mathcal O(h^\infty)\quad\text{microlocally near }\widehat\Lambda.\end{aligned}$$ Same can be said if we replace $\mathcal O(h^\infty)$ above by $\mathcal O(h^N)$, require that the full symbol of $\Pi$ lies in $C^{3N}$ for some large $N$ (rather than being smooth), and the assumptions of §\[s:dynamics\] hold for $r$ large enough depending on $N$. We argue by induction, finding a family $\Pi_k$, $k\geq 1$, of microlocal idempotents of all orders near $\widehat\Lambda$ (see Definition \[d:microlocal-idempotent\]) such that $[P,\Pi_k]=\mathcal O(h^{k+1})$ microlocally near $\widehat\Lambda$, and taking their asymptotic limit to obtain $\Pi$. We first construct $\Pi_1$. Take the microlocal idempotent of all orders $\widetilde\Pi\in \II$ near $\widehat\Lambda$ constructed in Proposition \[l:idempotent-exists\]. Since the Hamilton field of $p=\sigma(P)$ is tangent to $\Gamma_\pm$, $dp$ is annihilated by the subbundles $\mathcal V_\pm$ from §\[s:projections\]; therefore, $$p(\rho_\pm)=p(\pi_\pm(\rho_\pm)),\quad \rho_\pm\in\Gamma_\pm^\circ;$$ by  and , $[P,\widetilde\Pi]=\mathcal O(h)$ microlocally near $\widehat\Lambda$. We write $[P,\widetilde\Pi]=hS_0$ microlocally near $\widehat\Lambda$, where $S_0\in \II$ and by part 4 of Proposition \[l:idempotents\], $$\label{e:sigma-s-0} \sigma_\Lambda(S_0)(\rho_-,\rho_+)=\tilde a_0^-(\rho_-)\otimes s_0^+(\rho_+)+s_0^-(\rho_-)\otimes\tilde a_0^+(\rho_+),$$ with $s_0^\pm\in C^\infty(\widetilde\Gamma_\pm;\mathcal E_\pm)$ vanishing on $K$ near $\widehat K$ and $\tilde a_0^\pm\in C^\infty(\widetilde\Gamma_\pm;\mathcal E_\pm)$ giving the principal symbol of $\widetilde\Pi$ by . Here $\widetilde\Gamma_\pm$ are the neighborhoods of $\widehat\Gamma_\pm$ in $\Gamma_\pm^\circ$ defined in . We look for $\Pi_1$ in the form $$\label{e:pi-1} \Pi_1=e^{Q_0}\widetilde \Pi e^{-Q_0},$$ where $Q_0\in\Psi^{\operatorname{comp}}(X)$ is compactly supported and thus $e^{\pm Q_0}$ are pseudodifferential (see for example [@zeeman Proposition 2.7]). We calculate microlocally near $\widehat\Lambda$, $$e^{-Q_0}[P,\Pi_1]e^{Q_0}=[e^{-Q_0}Pe^{Q_0},\widetilde\Pi]=hS_0+[[P,Q_0],\widetilde\Pi]+\mathcal O(h^2).$$ Here we use that $e^{-Q_0}Pe^{Q_0}=P+[P,Q_0]+\mathcal O(h^2)$. By , , , $$\begin{gathered} \sigma_\Lambda(S_0+h^{-1}[[P,Q_0],\widetilde \Pi])(\rho_-,\rho_+)\\ =\tilde a_0^-(\rho_-)\otimes(s_0^+(\rho_+)-iH_p\sigma(Q_0)(\rho_+)\tilde a_0^+(\rho_+))\\ +(s_0^-(\rho_-)+iH_p\sigma(Q_0)(\rho_-)\tilde a_0^-(\rho_-))\otimes \tilde a_0^+(\rho_+). \end{gathered}$$ It is thus enough to take any $Q_0$ such that for the restrictions $q_0^\pm=\sigma(Q_0)|_{\widetilde\Gamma_\pm}$, the following transport equations hold near $\widehat\Gamma_\pm$: $$\label{e:transport-0} H_pq_0^\pm=\mp is_0^\pm/\tilde a_0^\pm,\quad q_0^\pm|_{\widetilde K}=0.$$ Such $q_0^\pm$ exist and are unique and smooth enough by Lemma \[l:ode\], giving $\Pi_1$. Note that Lemma \[l:ode\] can be applied near $\widehat\Gamma_\pm$, instead of the whole $\Gamma_\pm^\circ$, since $e^{tH_p}(\widehat\Gamma_\pm)\subset \widehat\Gamma_\pm$ for $\mp t\geq 0$ by part  of Lemma \[l:phi-pm\]. Now, assume that we have constructed $\Pi_k$ for some $k>0$. Let $a_0^\pm$ be the components of the principal symbol of $\Pi_k$ given by . Then microlocally near $\widehat\Lambda$, $[P,\Pi_k]=h^{k+1}S_k$, where $S_k\in \II$ and by part 4 of Proposition \[l:idempotents\], $$\sigma_\Lambda(S_k)(\rho_-,\rho_+)=a_0^-(\rho_-)\otimes s_k^+(\rho_+) +s_k^-(\rho_-)\otimes a_0^+(\rho_+),$$ where $s_k^\pm\in C^\infty(\widetilde\Gamma_\pm;\mathcal E_\pm)$ vanish on $K$ near $\widehat K$. We then take $$\label{e:pi-k} \Pi_{k+1}=(1+h^kQ_k)\Pi_k(1+h^kQ_k)^{-1}$$ where $Q_k$ is a compactly supported pseudodifferential operator. Microlocally near $\widehat\Lambda$, $$[P,\Pi_{k+1}]=h^{k+1}S_k+h^k[[P,Q_k],\Pi_k]+\mathcal O(h^{k+2}).$$ Therefore, $q_k^\pm=\sigma(Q_k)|_{\widetilde\Gamma_\pm}$ need to satisfy the transport equations near $\widehat\Gamma_\pm$ $$\label{e:transport-k} H_p q_k^\pm=\mp is_k^\pm/a_0^\pm,\quad q_k^\pm|_{\widetilde K}=0.$$ Such $q_k^\pm$ exist and are unique and smooth enough again by Lemma \[l:ode\], giving $\Pi_{k+1}$. To show that the operator $\Pi$ satisfying  and  is unique microlocally near $\widehat \Lambda$, we show by induction that each such $\Pi$ satisfies $\Pi=\Pi_k+\mathcal O(h^k)$ microlocally near $\widehat\Lambda$. First of all, $\Pi$ has the form  for some operator $Q_0$ microlocally near $\widehat\Lambda$, by part 2 of Proposition \[l:idempotents\]; moreover, by the proof of this fact, we can take $\sigma(Q_0)|_K=0$ near $\widehat K$. Now, $\sigma(Q_0)|_{\widehat\Gamma_\pm}$ are determined uniquely by the transport equations , and this gives $\Pi=\Pi_1+\mathcal O(h)$ microlocally near $\widehat\Lambda$. Next, if $\Pi=\Pi_k+\mathcal O(h^k)$ for some $k>0$, then, as follows from the proof of Part 2 of Proposition \[l:idempotents\], $\Pi$ has the form  for some operator $Q_k$ microlocally near $\widehat\Lambda$, such that $\sigma(Q_k)|_K=0$ near $\widehat K$. Then $\sigma(Q_k)|_{\widehat\Gamma_\pm}$ are determined uniquely by the transport equations , and this gives $\Pi=\Pi_{k+1}+\mathcal O(h^{k+1})$ microlocally near $\widehat\Lambda$. Annihilating ideals {#s:construction-2} ------------------- Let $\Pi\in \II$ be the operator constructed in Theorem \[t:our-Pi\]. In this section, we construct pseudodifferential operators $\Theta_\pm$ annihilating $\Pi$ microlocally near $\widehat\Lambda$; they are key for the microlocal estimates in §\[s:resolvent-bounds\]. More precisely, we obtain \[l:ideals\] If $\delta>0$ in the definition  of $\widehat W$ is small enough, then there exist compactly supported $\Theta_\pm\in\Psi^{\operatorname{comp}}(X)$ such that: 1. \[i:ideals-1\] $\Pi\Theta_-=\mathcal O(h^\infty)$ and $\Theta_+\Pi=\mathcal O(h^\infty)$ microlocally near $\widehat\Lambda$; 2. \[i:ideals-2\] $\sigma(\Theta_\pm)=\varphi_\pm$ near $\widehat W$, with $\varphi_\pm$ defined in Lemma \[l:phi-pm\]; 3. \[i:ideals-3\] if $P$ is the operator constructed in Lemma \[l:resolution\], then $$\label{e:ideals-3} [P,\Theta_-]=-ih\Theta_-Z_-+\mathcal O(h^\infty),\quad [P,\Theta_+]=ih Z_+\Theta_++\mathcal O(h^\infty)$$ microlocally near $\widehat W$, where $Z_\pm\in\Psi^{\operatorname{comp}}(X)$ are compactly supported and $\sigma(Z_\pm)=c_\pm$ near $\widehat W$, with $c_\pm$ defined in Lemma \[l:phi-pm\]; 4. \[i:ideals-4\] if $\Im\Theta_+={1\over 2i}(\Theta_+-\Theta_+^*)$ and $\zeta=\sigma(h^{-1}\Im\Theta_+)$, then $$\label{e:zeta} H_p\zeta=- c_+\zeta- {1\over 2}\{\varphi_+,c_+\}\quad \text{on }\Gamma_+\text{ near }\widehat W;$$ 5. \[i:ideals-5\] there exists an operator $\Xi:C^\infty(X)\to C_0^\infty(X)$, satisfying parts 1 and 2 of Proposition \[l:Xi\] and such that $$\label{e:Xi} 1-\Pi=\Theta_-\Xi+\mathcal O(h^\infty)\quad\text{microlocally near }\widehat W\times\widehat W.$$ The operators $\Theta_\pm$ satisfying conditions  and  exist by part 1 of Proposition \[l:principal-ideal\]. Next, since $[P,\Pi]=\mathcal O(h^\infty)$ microlocally near $\widehat\Lambda$, we find $$\Pi[P,\Theta_-]=\mathcal O(h^\infty),\quad [P,\Theta_+]\Pi=\mathcal O(h^\infty)$$ microlocally near $\widehat\Lambda$; condition  now follows from part 2 of Proposition \[l:principal-ideal\]. The symbols $\sigma(Z_\pm)$ can be computed using the identity $H_p\varphi_\pm=\mp c_\pm \varphi_\pm$ from part  of Lemma \[l:phi-pm\]. Condition  follows immediately from Proposition \[l:Xi\], keeping in mind that by making $\delta$ small we can make $\widehat W$ contained in an arbitrary neighborhood of $\widehat K$. Finally, we verify condition . Taking the adjoint of the identity $[P,\Theta_+]=ihZ_+\Theta_++\mathcal O(h^\infty)$ and using that $P$ is self-adjoint, we get microlocally near $\widehat W$, $$[P,\Theta_+^*]=ih\Theta_+^*Z_+^*.$$ Therefore, microlocally near $\widehat W$ $$2[P,h^{-1}\Im\Theta_+]=Z_+\Theta_+-\Theta_+^*Z_+^* =[Z_+,\Theta_+]+2i((\Im\Theta_+)Z_+^*+\Theta_+\Im Z_+).$$ By comparing the principal symbols, we get . Resolvent estimates {#s:resolvent-bounds} =================== In this section we give various estimates on the resolvent $\mathcal R(\omega)$, in particular proving Theorem \[t:gaps\]. In §\[s:estimate-full\], we reduce Theorem \[t:gaps\] to a microlocal estimate in a neighborhood of the trapped set, which is further split into two estimates: on the kernel of the projector $\Pi$ given by Theorem \[t:our-Pi\], proved in §\[s:estimate-kernel\], and on the image of $\Pi$, proved in §\[s:estimate-image\]. In §\[s:estimate-spectral\] we obtain a restriction on the wavefront set $\mathcal R(\omega)$ in $\omega$ on the image of $\Pi$, needed in §\[s:trace\]. Finally, in §\[s:resonant-states\], we discuss the consequences of our methods for microlocal concentration of resonant states and the corresponding semiclassical measures. Reduction to the trapped set {#s:estimate-full} ---------------------------- We take $\delta>0$ small enough so that the results of §\[s:construction-1\],\[s:construction-2\] hold, and define following  (with $\delta_1$ chosen in §\[s:projections\]), $$\label{e:w'} \widehat W:=U_\delta\cap p^{-1}([\alpha_0-\delta_1/2,\alpha_1+\delta_1/2]),\quad W':=U_{\delta/2}\cap p^{-1}([\alpha_0-\delta_1/4,\alpha_1+\delta_1/4]),$$ so that $W'$ is a neighborhood of $K\cap p^{-1}([\alpha_0,\alpha_1])$ compactly contained in $\widehat W$. Here $U_\delta$ is defined in . For the reductions of this subsection, it is enough to assume that $\omega$ satisfies . The region  will arise as the intersection of the regions  and  where the two components of the estimate will hold. To prove Theorem \[t:gaps\], it is enough to show the estimate $$\label{e:gaps2} \|\tilde u\|_{\mathcal H_1}\leq Ch^{-2}\|\tilde f\|_{\mathcal H_2}+\mathcal O(h^\infty)$$ for each $\tilde u=\tilde u(h)\in\mathcal H_1$ with $\|\tilde u\|_{\mathcal H_1}$ bounded polynomially in $h$ and for $\tilde f=\mathcal P(\omega)\tilde u$, where $\omega=\omega(h)$ satisfies . Subtracting from $\tilde u$ the function $v$ constructed in Lemma \[l:smart-parametrix\], we may assume that $$\WFh(\tilde f)\subset W'.$$ Let $\mathcal S(\omega)$ be the operator constructed in Lemma \[l:resolution\], $\mathcal S'(\omega)$ be its elliptic parametrix near $\mathcal U\supset\widehat W$ constructed in Lemma \[l:eparametrix\], and put $$u:=\mathcal S(\omega)\tilde u,\quad f:=\mathcal S'(\omega)\tilde f,$$ so by , for the operator $P$ constructed in Lemma \[l:resolution\], $$\label{e:assumption-general-1} (P-\omega)u=f\quad\text{microlocally near }\widehat W,\quad \WFh(f)\subset \widehat W.$$ By ellipticity (Proposition \[l:elliptic\]) and since $\WFh(f)\subset W'$, $$\label{e:assumption-general-1.5} \WFh(u)\cap\widehat W\subset p^{-1}([\alpha_0-\delta_1/4,\alpha_1+\delta_1/4]).$$ Let $\varphi_\pm$ be the functions constructed in Lemma \[l:phi-pm\]. By Lemma \[l:propagate-outgoing\], $u$ satisfies the conditions (see Figure \[f:reduction\]) $$\begin{gathered} \label{e:assumption-general-2} \WFh(u)\cap \widehat W\subset \{|\varphi_+|\leq\delta/2\},\\ \label{e:assumption-general-2.5} \WFh(u)\cap\Gamma_-^\circ\subset W'.\end{gathered}$$ Indeed, if $\rho\in\WFh(u)\cap\mathcal U$, then either $\rho\in\Gamma_+$ (in which case  and  follow immediately) or there exists $T\geq 0$ such that for $\gamma(t)=e^{tH_p}(\rho)$, $\gamma([-T,0])\subset\overline{\mathcal U}$ and $\gamma(-T)\in\WFh(\tilde f)\subset W'$. In the second case, if $\rho\in\widehat W$, then by convexity of $U_{\delta}$ (part  of Lemma \[l:phi-pm\]) we have $\gamma([-T,0])\subset\widehat W$. To show , it remains to use that $H_p\varphi_+^2\leq 0$ on $\widehat W$, following from part  of Lemma \[l:phi-pm\]. For , note that if $\rho\in\Gamma_-$, then $\gamma(-T)\in\Gamma_-\cap W'$; however, $e^{tH_p}(\Gamma_-\cap W')\subset \Gamma_-\cap W'$ for all $t\geq 0$ and thus $\rho\in W'$. ![A phase space picture of the geodesic flow near $\widehat W$. The shaded region corresponds to  and .[]{data-label="f:reduction"}](nhp.5) By Lemma \[l:smart-bound\], we reduce  to $$\label{e:gaps3} \|A_1u\|_{L^2}\leq Ch^{-2}\|f\|_{L^2}+\mathcal O(h^\infty),$$ where $A_1\in\Psi^{\operatorname{comp}}(X)$ is any compactly supported operator elliptic on $W'$. Now, let $\Pi\in \II$ be the operator constructed in Theorem \[t:our-Pi\] in §\[s:construction-1\]. Note that $$\label{e:commutation-yay} (P-\omega)\Pi u=\Pi f+\mathcal O(h^\infty)\quad\text{microlocally near } \widehat W,$$ since $[P,\Pi]=\mathcal O(h^\infty)$ microlocally near $\widehat W\times\widehat W$, $\WFh(\Pi)\subset\Lambda^\circ\subset\Gamma_-^\circ\times\Gamma_+^\circ$, and by . We finally reduce  to the following two estimates, which are proved in the following subsections: \[l:estimate-kernel\] Assume that $u,f$ are $h$-tempered families satisfying – and $$\label{e:assumption-kernel} \Re\omega\in [\alpha_0,\alpha_1],\quad \Im\omega\in [-(\nu_{\min}-\varepsilon)h,C_0h].$$ Then there exists compactly supported $A_1\in\Psi^{\operatorname{comp}}(X)$ elliptic on $W'$ such that $$\label{e:estimate-kernel} \|A_1(1-\Pi)u\|_{L^2}\leq Ch^{-1}\|\Xi f\|_{L^2}+\mathcal O(h^\infty),$$ where $\Xi$ is the operator from part  of Proposition \[l:ideals\]; note that by part 1 of Proposition \[l:Xi\], $\|\Xi\|_{L^2\to L^2}=\mathcal O(h^{-1})$. \[l:estimate-image\] Assume that $u,f$ are $h$-tempered families satisfying – and $$\label{e:assumption-image} \Re\omega\in[\alpha_0,\alpha_1],\quad \Im\omega\in [-C_0h,C_0h]\setminus\Big(-{\nu_{\max}+\varepsilon\over 2}h,-{\nu_{\min}-\varepsilon\over 2}h\Big),$$ Then there exists compactly supported $A_1\in\Psi^{\operatorname{comp}}(X)$ elliptic on $W'$ such that $$\label{e:estimate-image} \|A_1\Pi u\|_{L^2}\leq Ch^{-1}\|\Pi f\|_{L^2}+\mathcal O(h^\infty).$$ Note that by Proposition \[l:bund\] and the reduction to the model case of §\[s:general\], we have $\|\Pi\|_{L^2\to L^2}=\mathcal O(h^{-1/2})$. Estimate on the kernel of Pi {#s:estimate-kernel} ---------------------------- In this section, we prove Proposition \[l:estimate-kernel\], which is a microlocal estimate on the kernel of $\Pi$ (or equivalently, on the image of $1-\Pi$). We will use the identity  together with the commutator formula  to effectively shift the spectral parameter to the upper half-plane, where a standard positive commutator argument gives us the estimate. By , we have microlocally near $\widehat W$, $$\label{e:kernel-int-1} (P-\omega)(1-\Pi)u=(1-\Pi)f+\mathcal O(h^\infty)$$ Let $\Theta_-\in\Psi^{\operatorname{comp}}(X)$ and $\Xi$ be the operators constructed in Proposition \[l:ideals\], and denote $$v:=\Xi u,\quad g:=\Xi f.$$ Then microlocally near $\widehat W$, $$\label{e:kernel-int-1.5} (1-\Pi)u=\Theta_- v,\quad (1-\Pi)f=\Theta_- g.$$ Indeed, by part 2 of Proposition \[l:Xi\], , and the fact that $\WFh(\Pi)\subset\Lambda^\circ\subset\Gamma_-^\circ\times\Gamma_+^\circ$, we see that $1-\Pi=\Theta_-\Xi+\mathcal O(h^\infty)$ microlocally near $(\WFh(u)\setminus\widehat W)\times\widehat W$, since each of the featured operators is microlocalized away from this region. Combining this with , we see that $1-\Pi=\Theta_-\Xi+\mathcal O(h^\infty)$ microlocally near $\WFh(u)\times\widehat W$, and thus also near $\WFh(f)\times\widehat W$, yielding . By part 2 of Proposition \[l:Xi\] together with – and part  of Lemma \[l:phi-pm\], $$\begin{gathered} \label{e:wf-v-0} \WFh(v)\cup\WFh(g)\subset p^{-1}([\alpha_0-\delta_1/4,\alpha_1+\delta_1/4]),\\ \label{e:wf-v} (\WFh(v)\cup\WFh(g))\cap\widehat W\subset \{|\varphi_+|\leq\delta/2\}.\end{gathered}$$ We now obtain a differential equation on $v$; the favorable imaginary part of the operator in this equation, coming from commuting $\Theta_-$ with $P$, is the key component of the proof. \[l:estimate-kernel-int\] Let $Z_-$ be the operator from . Then microlocally near $\widehat W$, $$\label{e:kernel-int-2} (P-ihZ_--\omega)v=g+\mathcal O(h^\infty).$$ Given , the equation  becomes $(P-\omega)\Theta_-v=\Theta_- g+\mathcal O(h^\infty)$ microlocally near $\widehat W$. Using , we get microlocally near $\widehat W$, $$\Theta_-(P-ihZ_--\omega)v=\Theta_-g+\mathcal O(h^\infty).$$ To show , it remains to apply propagation of singularities (part 2 of Proposition \[l:microhyperbolic\]), for the operator $\Theta_-$. Indeed, by part  of Lemma \[l:phi-pm\], for each $\rho\in\widehat W$, there exists $t\geq 0$ such that the Hamiltonian trajectory $\{e^{sH_{\varphi_-}}(\rho)\mid 0\leq s\leq t\}$ lies entirely inside $\widehat W$ and $e^{tH_{\varphi_-}}(\rho)$ lies in $\{\varphi_+=-\delta\}$ and by  does not lie in $\WFh((P-ihZ_--\omega)v-\Theta_-g)$. We now use a positive commutator argument. Take a self-adjoint compactly supported $\mathcal X_-\in\Psi^{\operatorname{comp}}(X)$ such that $\WFh(\mathcal X_-)$ is compactly contained in $\widehat W$ and $\sigma(\mathcal X_-)=\chi(\varphi_-)$ near $\widehat W\cap\WFh(v)$, where $\varphi_-$ is defined in Lemma \[l:phi-pm\], $\chi\in C_0^\infty(-\delta,\delta)$, $s\chi'(s)\leq 0$ everywhere, and $\chi=1$ near $[-\delta/2,\delta/2]$. This is possible by  and . Put $\Im\omega=h\nu$; by  and since $P$ is self-adjoint, $$\label{e:pos-comm-1} \begin{gathered} \Im\langle\mathcal X_- v,g\rangle={h\over 2}\langle (Z_-^*\mathcal X_-+ \mathcal X_-Z_-+2\nu\mathcal X_-)v,v\rangle\\ +{1\over 2i}\langle[P,\mathcal X_-]v,v\rangle+\mathcal O(h^\infty) =h\langle\mathcal Y_- v,v\rangle+\mathcal O(h^\infty), \end{gathered}$$ where $\mathcal Y_-\in\Psi^{\operatorname{comp}}(X)$ is compactly supported, $\WFh(\mathcal Y_-)\subset\WFh(\mathcal X_-)\subset\widehat W$ and, using the function $c_-$ from part  of Lemma \[l:phi-pm\] together with part  of Proposition \[l:ideals\], we write near $\widehat W\cap \WFh(v)$, $$\begin{gathered} \sigma(\mathcal Y_-)=(c_-+\nu)\chi(\varphi_-)-{1\over 2}H_p\chi(\varphi_-) =(c_-+\nu)\chi(\varphi_-)-{1\over 2}c_-\varphi_-\chi'(\varphi_-). \end{gathered}$$ However, $\nu\geq-(\nu_{\min}-\varepsilon)$ by  and by , $c_->(\nu_{\min}-\varepsilon)$ on $\widehat W$; therefore $$\label{e:y-minus-positive} \sigma(\mathcal Y_-)\geq 0\quad\text{near }\WFh(v),\qquad \sigma(\mathcal Y_-)>0\quad\text{near }\WFh(v)\cap W'.$$ To take advantage of , we use the following combination of sharp Garding inequality with propagation of singularities: \[l:kernel-garding\] Assume that $Z,Q\in\Psi^{\operatorname{comp}}(X)$ are compactly supported, $\WFh(Z),\WFh(Q)$ are compactly contained in $\widehat W$, $Z^*=Z$, and $$\sigma(Z)\geq 0\quad\text{near }\WFh(v),\quad \sigma(Z)>0\quad\text{near }\WFh(v)\cap W'.$$ Then $$\label{e:kernel-garding} \|Qv\|_{L^2}^2\leq C\langle Zv,v\rangle +Ch^{-2}\|g\|^2_{L^2}+\mathcal O(h^\infty).$$ Without loss of generality, we may assume that $Q$ is elliptic on $\WFh(Z)\cup W'$. There exists compactly supported $Q_1\in\Psi^{\operatorname{comp}}(X)$, elliptic on $W'$, such that $\sigma(Z-Q_1^*Q_1)\geq 0$ near $\WFh(v)$ and $Q$ is elliptic on $\WFh(Q_1)$. Applying sharp Garding inequality (Proposition \[l:garding\]) to $Z-Q_1^*Q_1$, we get $$\label{e:kernel-garding-1} \|Q_1 v\|_{L^2}^2\leq C\langle Z v,v\rangle+Ch\|Q v\|_{L^2}^2+\mathcal O(h^\infty).$$ Now, by , , and since $H_p\varphi_-^2>0$ on $\widehat W\setminus \Gamma_-$ by part  of Lemma \[l:phi-pm\], each backwards flow line of $H_p$ starting on $\WFh(Q)$ reaches either $\WFh(Q_1)$ or the complement of $\WFh(v)$, while staying in $\widehat W$; by propagation of singularities (Proposition \[l:microhyperbolic\]) applied to , $$\label{e:kernel-garding-2} \|Qv\|_{L^2}\leq C\|Q_1 v\|_{L^2}+Ch^{-1}\|g\|_{L^2}+\mathcal O(h^\infty).$$ Combining  and , we get . Now, there exists $A_1\in\Psi^{\operatorname{comp}}(X)$ compactly supported, elliptic on $W'$ and with $\WFh(A_1)$ compactly contained in $\widehat W$, such that the estimate $$\label{e:product-estimate} |\langle \mathcal X_- v, g\rangle|\leq \tilde \varepsilon h \|A_1 v\|_{L^2}^2 +C_{\tilde\varepsilon}h^{-1}\|g\|_{L^2}^2+\mathcal O(h^\infty)$$ holds for each $\tilde\varepsilon>0$ and constant $C_{\tilde\varepsilon}$ dependent on $\tilde\varepsilon$. Taking $\tilde\varepsilon$ small enough and combining , (for $Z=\mathcal Y_-$ and $Q=A_1$), and , we arrive to $$\|A_1v\|_{L^2}\leq Ch^{-1}\|g\|_{L^2}+\mathcal O(h^\infty).$$ Since $(1-\Pi)u=\Theta_-v$ microlocally near $\widehat W$, we get . Estimate on the image of Pi {#s:estimate-image} --------------------------- In this section, we prove Proposition \[l:estimate-image\], which is a microlocal estimate on the image of $\Pi$. We will use the pseudodifferential operator $\Theta_+$ microlocally solving $\Theta_+\Pi=\mathcal O(h^\infty)$ to obtain an additional pseudodifferential equation satisfied by elements of the image of $\Pi$. This will imply that for a pseudodifferential operator $A$ microlocalized near $K$, the principal part of the expression $\langle A\Pi u(h),\Pi u(h)\rangle$ depends only on the integral of $\sigma(A)$ over the flow lines of $\mathcal V_+$, with respect to an appropriately chosen measure. A positive commutator estimate finishes the proof. By , we have microlocally near $\widehat W$, $$\label{e:eeq} (P-\omega)\Pi u=\Pi f+\mathcal O(h^\infty).$$ Let $\Theta_+\in\Psi^{\operatorname{comp}}(X)$ be the operator constructed in Proposition \[l:ideals\], then by , $$\label{e:theta-yay} \Theta_+\Pi u=\mathcal O(h^\infty)\quad\text{microlocally near }\widehat W.$$ We start with \[l:estimate-image-int\] Let $\zeta:=\sigma(h^{-1}\Im\Theta_+)$. Take the function $\psi$ on $\Gamma_+\cap \widehat W$ such that $$\label{e:psi} \{\varphi_+,\psi\}=2\zeta,\quad \psi|_K=0.$$ Assume that $A\in\Psi^{\operatorname{comp}}(X)$ satisfies $\WFh(A)\Subset\widehat W$ and $$\label{e:condiition} \int(e^\psi\sigma(A))\circ e^{sH_{\varphi_+}}\,ds=0\quad\text{on }K.$$ The integral in , and all similar integrals in this subsection, is taken over the interval corresponding to a maximally extended flow line of $H_{\varphi_+}$ in $\Gamma_+\cap\widehat W$. Then there exists compactly supported $A_0\in\Psi^{\operatorname{comp}}(X)$ with $\WFh(A_0)\Subset\widehat W$ such that $$|\langle A\Pi u,\Pi u\rangle|\leq Ch\|A_0\Pi u\|_{L^2}^2+\mathcal O(h^\infty).$$ By , there exists $q\in C_0^\infty(\widehat W)$ such that $ \{\varphi_+,e^\psi q\}=e^\psi\sigma(A)$ on $\Gamma_+$. We can rewrite this as $$\label{e:yay-int} \{\varphi_+,q\}+2\zeta q=\sigma(A)\quad\text{on }\Gamma_+.$$ Take $Q,Y\in\Psi^{\operatorname{comp}}(X)$ microlocalized inside $\widehat W$ and such that $\sigma(Q)=q$ and $$\sigma(A)=\{\varphi_+,q\}+2\zeta q+\sigma(Y)\varphi_+.$$ Then $A=(ih)^{-1}(Q\Theta_+-\Theta_+^*Q)+Y\Theta_++\mathcal O(h)_{\Psi^{\operatorname{comp}}}$ and thus for some $A_0$, $$\begin{gathered} \langle A\Pi u,\Pi u\rangle={\langle Q\Theta_+\Pi u,\Pi u\rangle -\langle Q\Pi u,\Theta_+\Pi u\rangle\over ih}+\langle Y\Theta_+\Pi u,\Pi u\rangle\\ +\mathcal O(h)\|A_0\Pi u\|_{L^2}^2+\mathcal O(h^\infty). \end{gathered}$$ The first three terms on the right-hand side are $\mathcal O(h^\infty)$ by . Now, take compactly supported self-adjoint $\mathcal X_+\in\Psi^{\operatorname{comp}}(X)$ such that $\WFh(\mathcal X_+)$ is compactly contained in $\widehat W$ and the symbol $\chi_+:=\sigma(\mathcal X_+)$ satisfies $\chi_+\geq 0$ everywhere, $\chi_+>0$ on $W'$, and $$\label{e:normal} \int (e^\psi\chi_+)\circ e^{sH_{\varphi_+}}\,ds=1\quad\text{on }K\cap p^{-1}([\alpha_0-\delta_1/4, \alpha_1+\delta_1/4]).$$ Putting $\Im\omega=h\nu$, we have by  $$\label{e:iimage} \begin{gathered} \Im\langle \mathcal X_+\Pi u,\Pi f\rangle=h\nu\langle\mathcal X_+\Pi u,\Pi u\rangle +{1\over 2i}\langle[P,\mathcal X_+]\Pi u,\Pi u\rangle+\mathcal O(h^\infty)\\ =h\langle \mathcal Y_+\Pi u,\Pi u\rangle+\mathcal O(h^\infty), \end{gathered}$$ where $\mathcal Y_+\in\Psi^{\operatorname{comp}}(X)$ is compactly supported, $\WFh(\mathcal Y_+)\subset \WFh(\mathcal X_+)\subset\widehat W$, and $$\sigma(\mathcal Y_+)=\nu\chi_+-H_p\chi_+/2.$$ We now want to use Lemma \[l:estimate-image-int\] together with Garding inequality to show that $\langle\mathcal Y_+\Pi u,\Pi u\rangle$ has fixed sign, positive for $\nu\geq -(\nu_{\min}-\varepsilon)/2$ and negative for $\nu\leq -(\nu_{\max}+\varepsilon)/2$. For that, we need to integrate $\sigma(\mathcal Y_+)$ over the Hamiltonian flow lines of $\varphi_+$ on $\Gamma_+$, with respect to the measure from . This relies on \[l:yay-int-2\] If $c_+$ is defined in Lemma \[l:phi-pm\], then $$\label{e:yay-int-2} \int (e^\psi H_p\chi_+)\circ e^{sH_{\varphi_+}}\,ds =-c_+\quad\text{on }K\cap p^{-1}([\alpha_0-\delta_1/4,\alpha_1+\delta_1/4]).$$ By part  of Lemma \[l:phi-pm\], we have on $\Gamma_+\cap \widehat W$ $$(e^{sH_{\varphi_+}})_*\partial_s (e^{-sH_{\varphi_+}})_*H_p =-[H_p,H_{\varphi_+}]=c_+H_{\varphi_+}.$$ Therefore, we can write (at $\rho\in K$ and $s$ such that $e^{sH_{\varphi_+}}(\rho)\in \widehat W$) $$(e^{-sH_{\varphi_+}})_* H_p=H_p+w(s)H_{\varphi_+}\quad\text{on }K$$ where $w(s)$ is the smooth function on $K\times \mathbb R$ given by $$\partial_s w(s)=c_+\circ e^{sH_{\varphi_+}},\ w(0)=0.$$ Now, differentiating  along $H_p$ and integrating by parts, we have on $K\cap p^{-1}([\alpha_0-\delta_1/4,\alpha_1+\delta_1/4])$ $$\begin{gathered} \int \big(H_p(e^\psi\chi_+)\big)\circ e^{sH_{\varphi_+}}\,ds =\int (H_p+w(s)\partial_s)\big((e^\psi\chi_+)\circ e^{sH_{\varphi_+}}\big)\,ds\\ =-\int (e^\psi c_+\chi_+)\circ e^{sH_{\varphi_+}}\,ds; \end{gathered}$$ therefore, $$\label{e:yayay} \int (e^\psi H_p\chi_+)\circ e^{sH_{\varphi_+}}\,ds =-\int (e^\psi (c_++H_p\psi)\chi_+)\circ e^{sH_{\varphi_+}}\,ds.$$ Now, we find on $\Gamma_+\cap \widehat W$ by  and , $$H_{\varphi_+}H_p \psi=(H_p+c_+)H_{\varphi_+}\psi=2(H_p+c_+)\zeta =-H_{\varphi_+}c_+.$$ We have on $K\cap \widehat W$, $H_p\psi=0$; thus $$c_++H_p\psi=c_+\circ\pi_+\quad\text{on }\Gamma_+\cap \widehat W$$ and by  and , on $K\cap p^{-1}([\alpha_0-\delta_1/4,\alpha_1+\delta_1/4])$, $$\int (e^\psi H_p\chi_+)\circ e^{sH_{\varphi_+}}\,ds =-c_+\int (e^\psi\chi_+)\circ e^{sH_{\varphi_+}}\,ds=-c_+.$$ This finishes the proof of . Using , , and Lemma \[l:estimate-image-int\] (taking into account ), we find for some compactly supported $A_1\in\Psi^{\operatorname{comp}}$ with $\WFh(A_1)\subset\widehat W$ and $A_1$ elliptic on $W'\cup \WFh(\mathcal X_+)$, $$\langle \mathcal Y_+\Pi u,\Pi u\rangle=\langle\mathcal Z_+\Pi u,\Pi u\rangle+\mathcal O(h)\|A_1\Pi u\|_{L^2}^2 +\mathcal O(h^\infty)$$ where $\mathcal Z_+\in\Psi^{\operatorname{comp}}(X)$ is any self-adjoint compactly supported operator with $\WFh(\mathcal Z_+)\subset\widehat W$ and $$\sigma(\mathcal Z_+)=(\nu+(c_+\circ\pi_+)/2)\chi_+\quad\text{on }\Gamma_+.$$ Then by , $$\label{e:image-ultimate} \Im\langle\mathcal X_+\Pi u,\Pi f\rangle=h\langle\mathcal Z_+\Pi u,\Pi u\rangle+\mathcal O(h^2)\|A_1\Pi u\|_{L^2}^2 +\mathcal O(h^\infty).$$ Now, by , $\nu_{\min}-\varepsilon<c_+< \nu_{\max}+\varepsilon$ on $K$, therefore, keeping in mind that $\WFh(\Pi u)\subset\Gamma_+^\circ$, we find $$\begin{gathered} \label{e:image-ineq-1} \sigma(\mathcal Z_+)\geq 0\quad\text{near }\WFh(\Pi u)\quad\text{for }\nu\geq-(\nu_{\min}-\varepsilon)/2, \\ \label{e:image-ineq-2} \sigma(\mathcal Z_+)\leq 0\quad\text{near }\WFh(\Pi u)\quad\text{for }\nu\leq-(\nu_{\max}+\varepsilon)/2.\end{gathered}$$ Moreover, in both cases $\sigma(\mathcal Z_+)\neq 0$ on $\WFh(\Pi u)\cap W'$. We now combine sharp Garding inequality and propagation of singularities for the operator $\Theta_+$: \[l:image-garding\] Assume that $Z,Q\in\Psi^{\operatorname{comp}}(X)$ are compactly supported, $\WFh(Z),\WFh(Q)$ are compactly contained in $\widehat W$, $Z^*=Z$, and $$\sigma(Z)\geq 0\quad\text{near }\WFh(\Pi u),\quad \sigma(Z)>0\quad\text{near }\WFh(\Pi u)\cap W'.$$ Then $$\label{e:image-garding} \|Q\Pi u\|_{L^2}^2\leq C\langle Z\Pi u,\Pi u\rangle +\mathcal O(h^\infty).$$ We argue similarly to the proof of Lemma \[l:kernel-garding\], with  replaced by $$\label{e:image-garding-2} \|Q\Pi u\|_{L^2}\leq C\|Q_1\Pi u\|_{L^2}+\mathcal O(h^\infty).$$ The estimate  follows from propagation of singularities (Proposition \[l:microhyperbolic\]) applied to . Indeed, by part  of Lemma \[l:phi-pm\] together with , for each $\rho\in\widehat W\cap\WFh(\Pi u)\subset\Gamma_+$, there exists $t\in \mathbb R$ such that $e^{tH_{\varphi_+}}(\rho)\in W'$ and $e^{sH_{\varphi_+}}(\rho)\in \widehat W$ for each $s$ between $0$ and $t$. Using  (for $Z=\pm \mathcal Z_+,Q=A_1$), , and an analog of , we complete the proof of . Microlocalization in the spectral parameter {#s:estimate-spectral} ------------------------------------------- In this section, we provide a restriction on the wavefront set of solutions to the equation $(P-\omega)u=f$ in the spectral parameter $\omega$, needed in §\[s:trace\]. We use the method of §\[s:estimate-image\], however since $\Re\omega$ is now a variable, we will get an extra term coming from commutation with the multiplication operator by $\omega$. Because of the technical difficulties of studying operators on product spaces (namely, a pseudodifferential operator on $X$ does not give rise to a pseudodifferential operator on $X\times (\alpha_0,\alpha_1)$ since the corresponding symbol does not decay under differentiation in $\xi$ and thus does not lie in the class $S^k$ of §\[s:prelim-basics\]), we use the Fourier transform in the $\omega$ variable. \[l:res-mic\] Fix $\nu\in [-C_0,C_0]$ and put $\omega=\alpha+ih\nu$, where $\alpha\in (\alpha_0,\alpha_1)$ is regarded as a variable. Assume that $u(x,\alpha;h)\in C([\alpha_0,\alpha_1];\mathcal H_1)$, $f(x,\alpha;h)\in C([\alpha_0,\alpha_1];\mathcal H_2)$ have norms bounded polynomially in $h$, satisfying – uniformly in $\alpha$. Define the semiclassical Fourier transform $$\label{e:fourier-transform} \hat u(x,s;h)=\int_{\alpha_0}^{\alpha_1} e^{-{is\alpha\over h}}u(x,\alpha;h)\,d\alpha,$$ and $\hat f(x,s;h)$ accordingly. Then there exists $A_1\in\Psi^{\operatorname{comp}}(X)$ elliptic on $W'$ such that: 1\. If $\nu\geq-(\nu_{\min}-\varepsilon)/2$, then for any fixed $s_0\in\mathbb R$, $$\label{e:res-mic-1} \qquad\|\Pi\hat f\|_{L^2_s((-\infty,s_0])L^2_{x}(X)}=\mathcal O(h^\infty)\ \Longrightarrow\ \|A_1\Pi\hat u(s_0)\|_{L^2_x(X)}=\mathcal O(h^\infty).$$ 2\. If $\nu\leq-(\nu_{\max}+\varepsilon)/2$, then for any fixed $s_0\in\mathbb R$, $$\label{e:res-mic-2} \|\Pi\hat f\|_{L^2_s([s_0,\infty))L^2_{x}(X)}=\mathcal O(h^\infty)\ \Longrightarrow\ \|A_1\Pi\hat u(s_0)\|_{L^2_x(X)}=\mathcal O(h^\infty).$$ We consider case 1; case 2 is handled similarly using  instead of . Since $u(\alpha),f(\alpha)$ are $h$-tempered uniformly in $\alpha$, their Fourier transforms $\hat u(s),\hat f(s)$ are $h$-tempered and satisfy – in the $L^2$ sense in $s$; therefore, the corresponding $\mathcal O(h^\infty)$ errors will be bounded in $L^2_s$ for expressions linear in $\hat u,\hat f$ and in $L^1_s$ for expressions quadratic in $\hat u,\hat f$. We also note that for each $j$, the derivatives $\partial_s^j\hat u(s)$, $\partial_s^j\hat f(s)$ are $h$-tempered uniformly in $s\in\mathbb R$ and also in the $L^2$ sense in $s$. Taking the Fourier transform of , we get $$\label{e:commutation-yayf} (hD_s+P-ih\nu)\Pi\hat u(s)=\Pi \hat f(s)+\mathcal O(h^\infty)_{L^2_s(\mathbb R)}\quad\text{microlocally near }\widehat W.$$ We use the operators $\mathcal X_+, \mathcal Z_+,A_1$ from §\[s:estimate-image\]. Similarly to , we find $$\begin{gathered} \Im\langle \mathcal X_+\Pi \hat u(s),\Pi\hat f(s)\rangle ={h\over 2}\partial_s\langle\mathcal X_+\Pi\hat u(s),\Pi\hat u(s)\rangle\\ +h\langle \mathcal Z_+\Pi\hat u(s),\Pi\hat u(s)\rangle +\mathcal O(h^2)\|A_1\Pi\hat u(s)\|_{L^2_x}^2 +\mathcal O(h^\infty)_{L^1_s(\mathbb R)}. \end{gathered}$$ Integrating this over $s\in (-\infty,s_0]$, by the assumption of , we find $$\label{e:bananas} \begin{gathered} \langle\mathcal X_+\Pi\hat u(s_0),\Pi\hat u(s_0)\rangle+2\int_{-\infty}^{s_0} \langle\mathcal Z_+\Pi\hat u(s),\Pi\hat u(s)\rangle\,ds\\ \leq Ch\|A_1\Pi \hat u(s)\|_{L^2_s((-\infty,s_0])L^2_x}^2+\mathcal O(h^\infty). \end{gathered}$$ Applying Lemma \[l:image-garding\] to $Q=A_1$ and $Z=\mathcal Z_+,\mathcal X_+$, and using , we get $$\begin{aligned} \label{e:zebra-garding} \|A_1\Pi\hat u(s)\|_{L^2_x}^2&\leq C\langle \mathcal Z_+\Pi\hat u(s),\Pi\hat u(s)\rangle+\mathcal O(h^\infty)_{L^1_s(\mathbb R)},\\ \label{e:zebra-garding2} \|A_1\Pi\hat u(s_0)\|_{L^2_x}^2&\leq C\langle \mathcal X_+\Pi\hat u(s_0),\Pi\hat u(s_0)\rangle+\mathcal O(h^\infty).\end{aligned}$$ Combining  with , integrated over $s\in (-\infty,s_0]$, and , we get the conclusion of . Localization of resonant states {#s:resonant-states} ------------------------------- In this section, we study an application of the estimates of the preceding subsections to microlocal behavior of resonant states, namely elements of the kernel of $\mathcal P(\omega)$ for a resonance $\omega$. Assume that we are given a sequence $h_j\to 0$, and $\omega(h)\in\mathbb C,\tilde u(h)\in\mathcal H_1$, defined for $h$ in this sequence, such that $$\label{e:resonant-state} \begin{gathered} \mathcal P(\omega)\tilde u=0,\quad \|\tilde u\|_{\mathcal H_1}=1;\\ \Re\omega\in [\alpha_0,\alpha_1],\quad \Im\omega\in [-(\nu_{\min}-\varepsilon)h,C_0h]; \end{gathered}$$ the condition on $\omega$ is just . We also use the operators $\mathcal S(\omega)$ and $P$ from Lemma \[l:resolution\] and put $$\label{e:resonant-state-u} u:=\mathcal S(\omega)\tilde u,$$ so that $$\label{e:resonant-state-p} (P-\omega)u=\mathcal O(h^\infty)\quad\text{microlocally near } \mathcal U.$$ We say that the sequence $u(h_j)$ converges to some Radon measure $\mu$ on $T^*X$, and we call $\mu$ the semiclassical defect measure of $u$ (see [@e-z Chapter 5]) if for each compactly supported $A\in\Psi^0(M)$, we have $$\label{e:measures} \langle Au,u\rangle\to \int_{T^*M}\sigma(A)\,d\mu\quad\text{as }h_j\to 0.$$ Such $\mu$ is necessarily a nonnegative measure, see [@e-z Theorem 5.2]. \[t:resonant-states\] Let $\tilde u(h)$ be a sequence of resonant states corresponding to some resonances $\omega(h)$, as in , and $u$ defined in . Take the neighborhood $\widehat W$ of $K\cap p^{-1}([\alpha_0,\alpha_1])$ defined in . Then: 1. \[rs:1\] $\WFh(\tilde u)\cap \mathcal U\,\subset\, \Gamma_+\cap p^{-1}([\alpha_0,\alpha_1])$; 2. \[rs:2\] for each $A_1\in \Psi^{\operatorname{comp}}(X)$ elliptic on $K\cap p^{-1}([\alpha_0,\alpha_1])$, there exists a constant $c>0$ independent of $h$ such that $\|A_1u\|_{L^2}\geq c$; 3. \[rs:3\] $u=\Pi u+\mathcal O(h^\infty)$ and $\Theta_+u=\mathcal O(h^\infty)$ microlocally near $\widehat W$, where $\Pi$ is constructed in Theorem \[t:our-Pi\] in §\[s:construction-1\] and $\Theta_+$ is the pseudodifferential operator from Proposition \[l:ideals\]; 4. \[rs:5\] there exists a smooth family of smooth measures $\mu_\rho$, $\rho\in K\cap p^{-1}([\alpha_0,\alpha_1])$, on the flow line segments $\pi_+^{-1}(\rho)\cap\widehat W\subset\Gamma_+$ of $\mathcal V_+$, independent of the choice of $u$, such that if $u$ converges to some measure $\mu$ on $T^*M$ in the sense of , and $\Re\omega(h_j)\to\omega_\infty$, $h^{-1}\Im\omega(h_j)\to\nu$ as $h_j\to 0$, then $\mu|_{\widehat W}$ has the form $$\label{e:measure-form} \mu|_{\widehat W}=\int_{K\cap p^{-1}(\omega_\infty)} \mu_\rho\,d\hat\mu(\rho),$$ for some nontrivial measure $\hat\mu$ on $K\cap p^{-1}(\omega_\infty)$, such that for each $b\in C^\infty(K)$, $$\label{e:measure-eqn} \int_{K\cap p^{-1}(\omega_\infty)} H_pb -(2\nu+c_+)b\,d\hat\mu=0,$$ with the function $c_+$ defined in Lemma \[l:phi-pm\]. **Remark**. The equation  is similar to the equation satisfied by semiclassical defect measures for eigenstates for the damped wave equation, see [@e-z (5.3.21)]. Part  follows immediately from Lemma \[l:propagate-outgoing\], part  follows from Lemma \[l:smart-bound\] and implies that $\mu|_{\widehat W}$ is a nontrivial measure in part . By the discussion in §\[s:estimate-full\], $u$ satisfies –, with $f=0$. The first statement of part  then follows from Proposition \[l:estimate-kernel\]. Indeed, we have $(1-\Pi)u=\mathcal O(h^\infty)$ microlocally near the set $W'$ introduced in ; it remains to apply propagation of singularities (Proposition \[l:microhyperbolic\]) to , using Lemma \[l:the-flow\]. The second statement of part  now follows from . Finally, we prove part . First of all, $\mu|_{\mathcal U}$ is supported on $\Gamma_+$ by part , and on $p^{-1}(\omega_\infty)$ by  and the elliptic estimate (Proposition \[l:elliptic\]; see also [@e-z Theorem 5.3]). Next, note that by Lemma \[l:estimate-image-int\] and since $u=\Pi u+\mathcal O(h^\infty)$ microlocally near $\widehat W$, we have for each $a\in C_0^\infty(\widehat W)$ and the function $\psi$ given by , $$\int(e^{\psi}a)(e^{s H_{\varphi_+}}(\rho))\,ds=0\quad\text{for all }\rho\in K\cap p^{-1}(\omega_\infty)\ \Longrightarrow\ \int a\,d\mu=0.$$ This implies , with $$\int a\,d\mu_\rho:=\int(e^{\psi}a)(e^{sH_{\varphi_+}}(\rho))\,ds,\quad a\in C_0^\infty(\widehat W),\ \rho\in K\cap p^{-1}(\omega_\infty).$$ To see , we note that by , for each $a\in C_0^\infty(\widehat W)$ we have (see the derivation of [@e-z (5.3.21)]) $$\label{e:texas} \int H_pa-2\nu a\,d\mu=0.$$ Put $b(\rho)=\int a\,d\mu_\rho$ for $\rho\in K\cap p^{-1}(\omega_\infty)$. Similarly to Lemma \[l:yay-int-2\] (replacing 1 by $b(\rho)$ on the right-hand side of ), we compute $$\int H_pa\,d\mu_\rho=H_p b(\rho)-c_+(\rho)b(\rho),\quad \rho\in K\cap p^{-1}(\omega_\infty)$$ and  follows by . Grushin problem {#s:grushin} =============== In this section, we construct a well-posed Grushin problem for the scattering resolvent, representing resonances in the region  as zeroes of a certain determinant $F(\omega)$ defined in  below. Together with the trace formulas of §\[s:trace\], this makes possible the proof of the Weyl law in §\[s:weyl-law\]. We assume that the conditions of §§\[s:framework-assumptions\] and \[s:dynamics\] hold, fix $\varepsilon>0$ (to be chosen in Theorem \[t:weyl-law\]), and use the neighborhoods $W'\subset \widehat W$ of $K\cap p^{-1}([\alpha_0,\alpha_1])$ defined in ; let $\delta,\delta_1>0$ be the constants used to define these neighborhoods. Take compactly supported $Q_1,Q_2\in\Psi^{\operatorname{comp}}(X)$ such that (with $U_\delta$ defined in Lemma \[l:phi-pm\]) $$\label{e:qz} \begin{gathered} Q_1=1+\mathcal O(h^\infty)\quad\text{microlocally near }U_{\delta/4}\cap p^{-1}([\alpha_0-\delta_1/6,\alpha_1+\delta_1/6]),\\ Q_2=1+\mathcal O(h^\infty)\quad\text{microlocally near }U_{\delta/3}\cap p^{-1}([\alpha_0-\delta_1/5,\alpha_1+\delta_1/5]),\\ \WFh(Q_1)\Subset U_{\delta/3}\cap p^{-1}([\alpha_0-\delta_1/5,\alpha_1+\delta_1/5]),\quad \WFh(Q_2)\Subset W'. \end{gathered}$$ We will impose more restrictions on $Q_1$ later in Lemma \[l:grushin-l\]. Using the operator $\mathcal P(\omega):\mathcal H_1\to\mathcal H_2$ from §\[s:framework-assumptions\] and the operator $\mathcal S(\omega)$ constructed in Lemma \[l:resolution\], define the holomorphic family of operators $$\mathcal G(\omega):=\begin{pmatrix} \mathcal P(\omega) & \mathcal S(\omega)Q_1\Pi Q_2 \\ Q_1 \Pi Q_2\mathcal S(\omega) & 1-Q_1\Pi Q_2 \end{pmatrix} :\mathcal H_1\oplus L^2(X)\to \mathcal H_2\oplus L^2(X).$$ Here $\Pi\in \II$ is the operator constructed in Theorem \[t:our-Pi\] in §\[s:construction-1\]; it is a microlocal idempotent commuting with the operator $P$ from Lemma \[l:resolution\] microlocally near the set $\widehat\Lambda=\Lambda^\circ\cap (\widehat W\cap\widehat W)$. Note that, since $Q_1,Q_2$ are microlocalized away from fiber infinity, $\mathcal G(\omega)$ is a compact perturbation of $\mathcal P(\omega)\oplus 1$, and therefore Fredholm of index zero. In this section, we will prove \[l:grushin\] There exists a global constant[^6] $N$ such that for $\omega$ satisfying , $$\label{e:grushin-bound} \|\mathcal G(\omega)^{-1}\|_{\mathcal H_2\oplus L^2\to \mathcal H_1\oplus L^2}=\mathcal O(h^{-N}).$$ Moreover, if $$\label{e:grushin-inverse} \mathcal G(\omega)^{-1}=\begin{pmatrix} \mathcal R_{11}(\omega) & \mathcal R_{12}(\omega) \\ \mathcal R_{21}(\omega) & \mathcal R_{22}(\omega) \end{pmatrix},$$ then $\mathcal R_{22}(\omega)=1-L_{22}(\omega)+\mathcal O(h^\infty)_{\mathcal D'\to C_0^\infty}$, where $L_{22}(\omega)\in \II$ is microlocalized inside $\widehat\Lambda$ and the symbol $\sigma_\Lambda(L_{22})$ defined in  satisfies $$\label{e:l-22-symbol} \sigma_\Lambda(L_{22}(\omega))(\rho,\rho)={\sigma(Q_1)(\rho)^2+(p(\rho)-\omega)\sigma(Q_1)(\rho)\over \sigma(Q_1)(\rho)^2+(p(\rho)-\omega)(\sigma(Q_1)(\rho)-1)},\quad \rho\in \widehat K.$$ To prove , we consider families of distributions $ u(h)\in\mathcal H_1$, $f(h)\in \mathcal H_2$, $ v(h), g(h)\in L^2(X)$, bounded polynomially in $h$ in the indicated spaces and satisfying $\mathcal G(u, v)=(f, g)$, namely $$\begin{aligned} \label{e:grushin-eq-1} \mathcal P(\omega) u+\mathcal S(\omega) Q_1\Pi Q_2 v&= f,\\ \label{e:grushin-eq-2} Q_1\Pi Q_2 \mathcal S(\omega) u+(1-Q_1\Pi Q_2) v&= g.\end{aligned}$$ Note that by , implies $$\label{e:grushin-eq-1'} (P-\omega)\mathcal S(\omega) u+ Q_1\Pi Q_2 v=\mathcal S'(\omega) f+\mathcal O(h^\infty)\quad \text{microlocally near }\mathcal U.$$ Here $\mathcal S'(\omega)$ is an elliptic parametrix of $\mathcal S(\omega)$ near $\mathcal U$ constructed in Proposition \[l:eparametrix\]. To show , it is enough to establish the bound $$\label{e:grushin-bound-2} \| u\|_{\mathcal H_1}+\| v\|_{L^2}\leq Ch^{-N}(\| f\|_{\mathcal H_2}+\| g\|_{L^2})+\mathcal O(h^\infty).$$ We start with a technical lemma: \[l:grushin-l\] There exists $Q_1\in\Psi^{\operatorname{comp}}(X)$ satisfying  and such that $$\label{e:grushin-l} \sigma(Q_1)^2+(p-\omega)(\sigma(Q_1)-1)\neq 0\quad\text{on }K\text{ for all } \omega\in [\alpha_0,\alpha_1].$$ It suffices to take $Q_1$ such that $\sigma(Q_1)|_K=\psi(p)$, where $\psi\in C_0^\infty(\alpha_0-\delta_1/5,\alpha_1+\delta_1/5)$ is equal to 1 near $[\alpha_0-\delta_1/6,\alpha_1+\delta_1/6]$ and $$\label{e:egg-2} \psi(\lambda)^2+(\lambda-\omega)(\psi(\lambda)-1)\neq 0,\quad \lambda\in\mathbb R,\ \omega\in [\alpha_0,\alpha_1].$$ We now show that such $\psi$ exists. The equation  holds automatically for $\lambda\not\in (\alpha_0-\delta_1/5,\alpha_1+\delta_1/5)$, as $\psi=0$ there and the left-hand side of  equals $\omega-\lambda\neq 0$. This however also shows that a real-valued $\psi$ with the desired properties does not exist. We take $\Re\psi\in C_0^\infty(\alpha_0-\delta_1/5,\alpha_1+\delta_1/5)$ equal to 1 near $[\alpha_0-\delta_1/6,\alpha_1+\delta_1/6]$ and take values in $[0,1]$ and $\Im\psi\in C_0^\infty(\alpha_1+\delta_1/6,\alpha_1+\delta_1/5)$ a nonnegative function to be chosen later. Then the left-hand side of  is equal to $1$ for $\lambda\in [\alpha_0-\delta_1/6,\alpha_1+\delta_1/6]$ and is positive for $\lambda\in [\alpha_0-\delta_1/5,\alpha_0-\delta_1/6]$. Next, the imaginary part of  is $$\Im\psi(\lambda)(2\Re\psi(\lambda)+\lambda-\omega).$$ Since $2\Re\psi(\lambda)+\lambda-\omega>0$ for $\lambda\in [\alpha_1+\delta_1/6,\alpha_1+\delta_1/5]$, it remains to take $\Im\psi(\lambda)>0$ on a large compact subinterval of $(\alpha_1+\delta_1/6,\alpha_1+\delta_1/5)$; then $\psi$ satisfies . Using Lemma \[l:grushin-l\], we determine $v$ microlocally outside of the elliptic region: \[l:grushin-1\] Let $Q_1$ be chosen in Lemma \[l:grushin-l\]. Then there exist $L_{21}^e(\omega),L_{22}^e(\omega)\in \II$ holomorphic in $\omega$, microlocalized inside $\widehat\Lambda$, and such that for all $ u, v, f, g$ satisfying , , $$\label{e:grushin-1-eq} v=L_{21}^e f+(1-L_{22}^e) g$$ microlocally outside of $\Gamma_+\cap\widehat W\cap p^{-1}([\alpha_0-\delta_1/8,\alpha_1+\delta_1/8])$. Moreover, $\sigma_\Lambda(L_{22}^e)$ satisfies  for $\rho\not\in p^{-1}([\alpha_0-\delta_1/8,\alpha_1+\delta_1/8])$. Using Proposition \[l:eparametrix\], construct compactly supported $R^e(\omega)\in\Psi^{\operatorname{comp}}(X)$ such that $R^e(\omega)(P-\omega)=1+\mathcal O(h^\infty)$ microlocally near $\widehat W\setminus p^{-1}(\alpha_0-\delta_1/8,\alpha_1+\delta_1/8)$. By , we get $$\mathcal S(\omega) u=R^e(\omega)(\mathcal S'(\omega) f-Q_1\Pi Q_2 v)+\mathcal O(h^\infty)$$ microlocally near $\widehat W\setminus p^{-1}(\alpha_0-\delta_1/8,\alpha_1+\delta_1/8)$. Substituting this into , we get $$\label{e:egg-0} (1-L') v= g-Q_1\Pi Q_2 R^e(\omega)\mathcal S'(\omega) f+\mathcal O(h^\infty)$$ microlocally outside of $\Gamma_+\cap \widehat W\cap p^{-1}([\alpha_0-\delta_1/8,\alpha_1+\delta_1/8])$, where $L'=Q_1\Pi Q_2(1+R^e(\omega)Q_1\Pi Q_2)\in \II$ and $\WFh(L')\subset\widehat\Lambda$. Let $\sigma_\Lambda(L')$ be the symbol of $L'$, defined in . By –, and since $\sigma_\Lambda(\Pi)|_K=1$ near $\widehat W$ (see part 1 of Proposition \[l:idempotents\] or §\[s:construction-1\]), we find for $\rho\in K\setminus p^{-1}(\alpha_0-\delta_1/8,\alpha_1+\delta_1/8)$, $$\sigma_\Lambda(L')(\rho,\rho)=\sigma(Q_1)(\rho)(1+\sigma(Q_1)(\rho)/(p(\rho)-\omega));$$ it follows from  that $$\label{e:egg-1} \sigma_\Lambda(L')|_K\neq 1\quad\text{outside of } p^{-1}(\alpha_0-\delta_1/8,\alpha_1+\delta_1/8).$$ By Proposition \[l:funny-parametrix\], there exists $L_{22}^e(\omega)\in \II$, with $\WFh(L_{22}^e)\subset \widehat\Lambda$, such that $(1-L_{22}^e)(1-L')=1+\mathcal O(h^\infty)$ microlocally outside of $p^{-1}([\alpha_0-\delta_1/8,\alpha_1+\delta_1/8])$, and note that the symbol $\sigma_\Lambda(L_{22}^e)$ satisfies  for $\rho\not\in p^{-1}([\alpha_0-\delta_1/8,\alpha_1+\delta_1/8])$ by . By , we get  with $L_{12}^e(\omega)=-(1-L_{22}^e(\omega))Q_1\Pi Q_2 R^e(\omega)\mathcal S'(\omega)$. By Proposition \[l:grushin-1\], replacing $ v$ by $A_e v$, where $A_e\in\Psi^{\operatorname{comp}}(X)$ is compactly supported, $\WFh(A_e)\subset \mathcal U\cap p^{-1}(\alpha_0-\delta_1/7,\alpha_1+\delta_1/7)$, and $A_e=1+\mathcal O(h^\infty)$ near $\widehat W\cap p^{-1}([\alpha_0-\delta_1/8,\alpha_1+\delta_1/8])$, we see that it is enough to prove  in the case $$\label{e:gcase-1} \WFh(v)\subset \mathcal U\cap p^{-1}([\alpha_0-\delta_1/7,\alpha_1+\delta_1/7]).$$ Using Lemma \[l:smart-parametrix\], consider $u'\in\mathcal H_1$ such that $\|u'\|_{\mathcal H_1}\leq Ch^{-1}\| f\|_{\mathcal H_2}$ and $\WFh(\mathcal P(\omega)u'- f)\subset \WFh(Q_1)\cap p^{-1}([\alpha_0-\delta_1/7,\alpha_1+\delta_1/7])$. Subtracting $u'$ from $ u$, we see that is suffices to prove  for the case $$\label{e:gcase-2} \WFh(f)\subset \WFh(Q_1)\cap p^{-1}([\alpha_0-\delta_1/7,\alpha_1+\delta_1/7]).$$ By , the wavefront set of $\mathcal P(\omega) u= f-\mathcal S(\omega)Q_1\Pi Q_2 v$ satisfies . Arguing as in §\[s:estimate-full\], and keeping in mind , we see that $ u$ satisfies –; in fact, can be strengthened to $$\label{e:assumption-general-1.75} \WFh(u)\cap \widehat W\subset p^{-1}([\alpha_0-\delta_1/7,\alpha_1+\delta_1/7]).$$ and  can be strengthened to $$\label{e:assumption-general-2.75} \WFh(u)\cap\Gamma_-^\circ\subset U_{\delta/3}\cap p^{-1}([\alpha_0-\delta_1/7,\alpha_1+\delta_1/7]).$$ We can now solve for $v$: \[l:horseradish\] Assume that $u,v,f,g$ satisfy , , , . Then $$\label{e:horseradish} v=Q_1\Pi\mathcal S'(\omega)f+(1-Q_1(P-\omega+1)\Pi Q_2)g+\mathcal O(h^\infty)_{C_0^\infty}.$$ Since $\Pi^2=\Pi+\mathcal O(h^\infty)$ microlocally near $\widehat W\times\widehat W$ and $Q_1=1+\mathcal O(h^\infty)$ microlocally near $K\cap p^{-1}([\alpha_0-\delta_1/6,\alpha_1+\delta_1/6]$, we have $$\label{e:kinda-i} \Pi Q_1 \Pi=\Pi+\mathcal O(h^\infty)\quad \text{microlocally near }(\widehat W\cap p^{-1}([\alpha_0-\delta_1/6,\alpha_1+\delta_1/6]))\times\widehat W.$$ We rewrite  as $$\label{e:grushin-eq-2.0} Q_1\Pi Q_2(\mathcal S(\omega) u- g)+(1-Q_1\Pi Q_2)( v- g)=0.$$ It follows immediately that $\WFh( v- g)\subset\WFh(Q_1)$ and thus $Q_2(v-g)=v-g+\mathcal O(h^\infty)_{C_0^\infty}$. Also, by , , and , $\WFh(g)\subset\mathcal U\cap p^{-1}([\alpha_0-\delta_1/7,\alpha_1+\delta_1/7])$. Applying $\Pi$ to  and using , , and , we get $\Pi Q_2\mathcal S(\omega) u-\Pi Q_2 g=\mathcal O(h^\infty)$ microlocally near $\widehat W$. By , we have $\Pi Q_2\mathcal S(\omega)u=\Pi\mathcal S(\omega)u+\mathcal O(h^\infty)_{C_0^\infty}$; therefore, $$\label{e:pi-u-bound} \Pi \mathcal S(\omega) u=\Pi Q_2 g+\mathcal O(h^\infty)\quad \text{microlocally near }\widehat W.$$ Then  becomes $$\label{e:grushin-eq-2.1} v=Q_1\Pi Q_2 v+(1-Q_1\Pi Q_2)g+\mathcal O(h^\infty)_{C_0^\infty}.$$ Applying $\Pi$ to , using that $[P,\Pi]=\mathcal O(h^\infty)$ microlocally near $\widehat W\times\widehat W$, and keeping in mind , we get $$\label{e:giraffe} (P-\omega)\Pi\mathcal S(\omega) u+\Pi Q_2 v=\Pi\mathcal S'(\omega) f+\mathcal O(h^\infty) \quad\text{microlocally near }\widehat W.$$ Together, and  give $$\Pi Q_2 v=\Pi \mathcal S'(\omega)f-(P-\omega)\Pi Q_2 g+\mathcal O(h^\infty) \quad\text{microlocally near }\widehat W.$$ By , we now get . By Proposition \[l:horseradish\], we see that $$\label{e:vbund} \|v\|_{L^2}\leq Ch^{-N}(\|f\|_{\mathcal H_2}+\|g\|_{L^2})+\mathcal O(h^\infty).$$ By Proposition \[l:estimate-kernel\] (using  instead of ), we get for some $A_1\in\Psi^{\operatorname{comp}}(X)$ elliptic near $W'$, $$\|A_1(1-\Pi) \mathcal S(\omega)u\|_{L^2}\leq Ch^{-N}(\|f\|_{\mathcal H_2}+\|g\|_{L^2})+\mathcal O(h^\infty).$$ Combining this with , we estimate $\|A_1u\|_{L^2}$ by the right-hand side of . Applying Lemma \[l:smart-bound\] to , we can estimate $\|u\|_{\mathcal H_1}$ by the same quantity, completing the proof of . It remains to describe the operator $\mathcal R_{22}$ from . We assume that $u,v,f,g$ satisfy , and $f=0$; then $\mathcal R_{22}g=v$. By Proposition \[l:grushin-1\], $v=(1-L_{22}^e)g+\mathcal O(h^\infty)$ microlocally outside of $\widehat W\cap p^{-1}([\alpha_0-\delta_1/8,\alpha_1+\delta_1/8])$; it then suffices to describe $v$ microlocally near $\widehat W\cap p^{-1}([\alpha_0-\delta_1/8,\alpha_1+\delta_1/8])$. Let $A_e$ be the operator introduced before  and $R^e(\omega)$ be an elliptic parametrix for $P-\omega$ constructed in the proof of Proposition \[l:grushin-1\]. Replacing $(u,v)$ by $(u+\mathcal S'(\omega)R^e(\omega)Q_1\Pi Q_2 (1-A_e)v,A_e v)$, we may assume that  and  hold, and in fact the resulting $f$ is $\mathcal O(h^\infty)_{C_0^\infty}$ and the resulting $g$ coincides with the original $g$ microlocally near $\widehat W\cap p^{-1}([\alpha_0-\delta_1/8,\alpha_1+\delta_1/8])$. By Proposition \[l:horseradish\], we now get for the original $v$ and $g$, $$v=(1-Q_1(P-\omega+1)\Pi Q_2)g+\mathcal O(h^\infty)\quad \text{microlocally near }\widehat W\cap p^{-1}([\alpha_0-\delta_1/8,\alpha_1+\delta_1/8]).$$ Note that $Q_1(P-\omega+1)\Pi Q_2\in \II$ and its principal symbol satisfies  in $p^{-1}([\alpha_0-\delta_1/8,\alpha_1+\delta_1/8])$, since $\sigma(Q_1)|_K=1$ in that region. This finishes the proof of Proposition \[l:grushin\]. By Proposition \[l:grushin\], $\mathcal R_{22}(\omega)-1$ is a compactly supported operator mapping $H^{-N}_h\to H^N_h$ for all $N$, therefore it is trace class. We can then define the determinant (see for instance [@tay1 (A.6.38)]) $$\label{e:S-omega} F(\omega):=\det \mathcal R_{22}(\omega),$$ which is holomorphic in the region  and $F(\omega)=0$ if and only if $\mathcal R_{22}(\omega)$ is not invertible (see [@tay1 Proposition A.6.16]). The key properties of $F$ needed in §\[s:weyl-law\] are established in \[l:grushin-ultimate\] 1. Resonances in the region  coincide (with the multiplicities defined in ) with zeroes of $F(\omega)$. 2\. For some constants $C$ and $N$, we have $|F(\omega)|\leq e^{Ch^{-N}}$ for $\omega$ in , and $|F(\omega)|\geq e^{-Ch^{-N}}$ for $\omega$ in the resonance free region . 3\. For $\omega$ in the resonance free region , we have $${\partial_\omega F(\omega)\over F(\omega)} =-\operatorname{Tr}((1-Q_1\Pi Q_2-Q_1\Pi \mathcal S(\omega)\mathcal R(\omega)\mathcal S(\omega) Q_1\Pi Q_2)\partial_\omega L_{22}(\omega))+\mathcal O(h^\infty).$$ Here $L_{22}(\omega)$ is defined in Proposition \[l:grushin\]. 1\. By Schur’s complement formula [@e-z (D.1.1)], and since $\mathcal G(\omega)$ is invertible by Proposition \[l:grushin\], we know that $\mathcal P(\omega)$ is invertible if and only if $\mathcal R_{22}(\omega)$ is, and in fact $$\label{e:schur} \mathcal P(\omega)^{-1}=\mathcal R_{11}(\omega)-\mathcal R_{12}(\omega)\mathcal R_{22}(\omega)^{-1} \mathcal R_{21}(\omega).$$ To see that the multiplicity of a resonance $\omega_0$ defined by  coincides with the multiplicity of $\omega_0$ as a zero of the function $F(\omega)$ (and in particular, to demonstrate that the multiplicity defined by  is a positive integer), it is enough to show that $$\label{e:multiplicity2} {1\over 2\pi i}\operatorname{Tr}\oint_{\omega_0}\mathcal P(\omega)^{-1}\partial_\omega \mathcal P(\omega)\,d\omega ={1\over 2\pi i}\operatorname{Tr}\oint_{\omega_0}\mathcal R_{22}(\omega)^{-1}\partial_\omega\mathcal R_{22}(\omega)\,d\omega;$$ indeed, since $\partial_\omega\mathcal R_{22}(\omega)$ is trace class, we can put the trace inside the integral on the right-hand side of , yielding $\partial_\omega F(\omega)/F(\omega)$; therefore, the right-hand side gives the mutliplicity of $\omega_0$ as a zero of $F(\omega)$ by the argument principle. Since $\partial_\omega (\mathcal G(\omega)^{-1})=-\mathcal G(\omega)^{-1}(\partial_\omega \mathcal G(\omega))\mathcal G(\omega)^{-1}$, we have $$\partial_\omega \mathcal R_{22}(\omega)=-\mathcal R_{21}(\omega)(\partial_\omega \mathcal P(\omega))\mathcal R_{12}(\omega) +\mathcal A(\omega)\mathcal R_{22}(\omega)+\mathcal R_{22}(\omega)\mathcal B(\omega),$$ where $\mathcal A(\omega),\mathcal B(\omega):L^2(X)\to L^2(X)$ are bounded operators holomorphic at $\omega_0$. By , follows from the two identities $$\begin{gathered} \operatorname{Tr}\oint_{\omega_0} \mathcal R_{12}(\omega)\mathcal R_{22}(\omega)^{-1}\mathcal R_{21}(\omega)\partial_\omega\mathcal P(\omega)\,d\omega =\operatorname{Tr}\oint_{\omega_0} \mathcal R_{22}(\omega)^{-1}\mathcal R_{21}(\omega) (\partial_\omega \mathcal P(\omega))\mathcal R_{12}(\omega)\,d\omega,\\ \operatorname{Tr}\oint_{\omega_0} \mathcal R_{22}(\omega)^{-1}(\mathcal A(\omega)\mathcal R_{22}(\omega)+\mathcal R_{22}(\omega)\mathcal B(\omega))\,d\omega=0. \end{gathered}$$ Both of them follow from the cyclicity of the trace, replacing $\mathcal R_{22}(\omega)^{-1}$ by its finite-dimensional principal part at $\omega_0$ and putting the trace inside the integral. 2\. By Proposition \[l:grushin\], the trace class norm $\|\mathcal R_{22}(\omega)-1\|_{\operatorname{Tr}}$ is bounded polynomially in $h$. Using the bound $|\det (1+T)|\leq e^{\|T\|_{\operatorname{Tr}}}$ (see for example [@tay1 (A.6.44)]), we get $|F(\omega)|\leq e^{Ch^{-N}}$. By Theorem \[t:gaps\], we have $\|\mathcal R(\omega)\|_{\mathcal H_2\to \mathcal H_1}\leq Ch^{-2}$ when $\omega$ satisfies . Using Schur’s complement formula again, we get $$\label{e:schur-strikes-back} \mathcal R_{22}(\omega)^{-1}=1-Q_1\Pi Q_2-Q_1\Pi Q_2 \mathcal S(\omega)\mathcal R(\omega)\mathcal S(\omega)Q_1\Pi Q_2.$$ Then $\|\mathcal R_{22}(\omega)^{-1}-1\|_{\operatorname{Tr}}\leq Ch^{-N}$ and thus $|F(\omega)|^{-1}=|\det (\mathcal R_{22}(\omega)^{-1})|\leq e^{Ch^{-N}}$. 3\. By Proposition \[l:grushin\], we have $\partial_\omega \mathcal R_{22}(\omega) =-\partial_\omega L_{22}(\omega)+\mathcal O(h^\infty)_{\mathcal D'\to \mathcal C_0^\infty}$, thus $${\partial_\omega F(\omega)\over F(\omega)}=-\operatorname{Tr}(\mathcal R_{22}(\omega)^{-1}\partial_\omega L_{22}(\omega))+\mathcal O(h^\infty).$$ By , it then suffices to prove that $$\operatorname{Tr}(Q_1\Pi (1-Q_2) \mathcal S(\omega) \mathcal R(\omega)\mathcal S(\omega)Q_1\Pi Q_2\partial_\omega L_{22}(\omega))= \mathcal O(h^\infty).$$ For that, it suffices to show that the intersection of the wavefront set of the operator on the left-hand side with the diagonal in $T^*X$ is empty. We assume the contrary, then there exists $\rho\in T^*X$ such that $$(\rho,\rho)\in\WFh(Q_1\Pi (1-Q_2) \mathcal S(\omega) \mathcal R(\omega)\mathcal S(\omega)Q_1\Pi Q_2\partial_\omega L_{22}(\omega)).$$ Since both $\Pi$ and $\partial_\omega L_{22}$ are microlocalized inside $\Lambda^\circ\subset\Gamma_-^\circ\cap\Gamma_+^\circ$, we see that $\rho\in K^\circ=\Gamma_+^\circ\cap\Gamma_-^\circ$. There exists $\rho'\in T^*X$ such that $$(\rho,\rho')\in\WFh(\mathcal S(\omega)\mathcal R(\omega) \mathcal S(\omega)Q_1\Pi Q_2\partial_\omega L_{22}(\omega)),\quad (\rho',\rho)\in \WFh(Q_1\Pi (1-Q_2)).$$ For any $h$-tempered $f\in L^2(X)$, we have $\WFh(\mathcal S(\omega)Q_1\Pi Q_2 \partial_\omega L_{22}(\omega)f)\subset \Gamma_+^\circ\cap \widehat W$, therefore by Lemma \[l:propagate-outgoing\] we have $\WFh(\mathcal R(\omega)\mathcal S(\omega)Q_1\Pi Q_2 \partial_\omega L_{22}(\omega)f)\cap \mathcal U\subset\Gamma_+$. It follows that $\rho'\in\Gamma_+$. Since $(\rho',\rho)\in\WFh(Q_1\Pi (1-Q_2))$, we see that $\rho'=\rho\in K^\circ$. However, then $\rho\in\WFh(Q_1)\cap \WFh(1-Q_2)$, which is impossible since $Q_2=1+\mathcal O(h^\infty)$ microlocally near $\WFh(Q_1)$. Trace formula {#s:trace} ============= In this section, we establish an asymptotic expansion for contour integrals of the logarithmic derivative of the determinant $F(\omega)$ of the effective Hamiltonian of the Grushin problem of §\[s:grushin\], defined in . By Proposition \[l:grushin-ultimate\], this reduces to computing contour integrals of operators of the form $\Pi\mathcal R(\omega)$, where $\Pi$ is the projector constructed in Theorem \[t:our-Pi\] in §\[s:construction-1\]. This in turn is done by approximating $\mathcal R(\omega)$ microlocally on the image of $\Pi$ by pseudodifferential operators, using Schrödinger propagators and microlocalization in the spectral parameter established in §\[s:estimate-spectral\]. We operate under the pinching condition  of Theorem \[t:weyl-law\], namely $\nu_{\max}<2\nu_{\min}$, and choose $\varepsilon>0$ such that $\nu_{\max}+\varepsilon<2(\nu_{\min}-\varepsilon)$. Take $\chi\in C_0^\infty(\alpha_0,\alpha_1)$ with $\alpha_0,\alpha_1$ from . Consider an almost analytic extension $\tilde\chi(\omega)$ of $\chi$, that is $\tilde\chi\in C^\infty(\mathbb C)$ such that $\tilde\chi|_{\mathbb R}=\chi$ and $\partial_{\bar\omega} \tilde\chi(\omega)=\mathcal O(|\Im\omega|^{\infty})$. We may take $\tilde\chi$ such that $\operatorname{supp}(\tilde\chi)\subset \{\Re\omega\in (\alpha_0,\alpha_1)\}$. The main result of this section is \[l:trace\] Take $$\label{e:nu-trace} \nu_-\in \Big[-(\nu_{\min}-\varepsilon),-{\nu_{\max}+\varepsilon\over 2}\Big],\quad \nu_+\in \Big[-{\nu_{\min}-\varepsilon\over 2},C_0\Big].$$ Let $F(\omega)$ be defined in  and put $$\label{e:trace-stuff} \mathcal I^\pm_\chi:=(2\pi h)^{n-1}\int_{\Im\omega=h\nu_\pm} \tilde\chi(\omega){\partial_\omega F(\omega)\over F(\omega)}\,d\omega.$$ Then, with $d\operatorname{Vol}_\sigma=\sigma_S^{n-1}/(n-1)!$ the symplectic volume form, $$\label{e:trace} \mathcal I_\chi^--\mathcal I_\chi^+=2\pi i\int_K \chi(p)\,d\operatorname{Vol}_\sigma + \mathcal O(h).$$ **Remark**. More precise trace formulas are possible; in particular, one can get a full asymptotic expansion in $h$ of each of $\mathcal I^\pm_\chi$. For simplicity, we prove here a less general version which suffices for the analysis of §\[s:weyl-law\]. The key feature of the expansions for the integrals , which produces a nontrivial asymptotics for resonances in Theorem \[t:weyl-law\], is that the principal part of $\mathcal I_\chi^\pm$ depends on the sign of $\pm$. The reason for this dependence is the difference of directions for propagation in the resolvent approximation $\mathcal R_\psi^\pm$ of Proposition \[l:trace-approx\] for the two cases; this in turn is explained by the difference between  and , which is due to the difference of the signs of the ‘commutator’ $\mathcal Z_+$ between  and . We start the proof by using Proposition \[l:res-mic\] to replace $\mathcal R(\omega)$ in the formula for $\partial_\omega F(\omega)/F(\omega)$ from Proposition \[l:grushin-ultimate\] by an operator $\mathcal R^\pm_\psi(\omega)$ obtained by integrating the Schrödinger propagator $e^{- it(P-\omega)/h}$ over a bounded range of times $t$. \[l:trace-approx\] Fix $\psi\in C_0^\infty(\mathbb R)$ such that $\psi=1$ near zero. For $\omega\in\mathbb C$, define the operators $\mathcal R^\pm_\psi(\omega):L^2(X)\to L^2(X)$ by $$\begin{gathered} \label{e:rr-1} \mathcal R^+_\psi(\omega):={i\over h}\int_{-\infty}^0 e^{is(P-\omega)/h}\psi(s)\,ds;\\ \label{e:rr-2} \mathcal R^-_\psi(\omega):=-{i\over h}\int_0^\infty e^{is(P-\omega)/h}\psi(s)\,ds;\end{gathered}$$ Then, if $\operatorname{supp}\psi$ is contained in a small enough neighborhood of zero, $$\label{e:trace-approx} \begin{gathered} \mathcal I_\chi^\pm=-(2\pi h)^{n-1}\operatorname{Tr}\int_{\Im\omega=h\nu_\pm}\tilde\chi(\omega) (1-Q_1\Pi Q_2\\-Q_1\mathcal R_\psi^\pm(\omega)\Pi Q_1\Pi Q_2)\partial_\omega L_{22}(\omega)\,d\omega +\mathcal O(h^\infty). \end{gathered}$$ We concentrate on the case of $\mathcal I_\chi^+$, the case of $\mathcal I_\chi^-$ is handled similarly, using  in place of . We denote $\omega=\alpha+ih\nu_+$, where $\alpha\in (\alpha_0,\alpha_1)$. By part 3 of Proposition \[l:grushin-ultimate\], it suffices to prove the trace norm bound $$\bigg\|\int_{\Im\omega=h\nu_+}\tilde\chi(\omega) Q_1(\Pi\mathcal S(\omega)\mathcal R(\omega)\mathcal S(\omega)-\mathcal R_\psi^+(\omega)\Pi)Q_1\Pi Q_2\partial_\omega L_{22}(\omega)\,d\omega\bigg\|_{\operatorname{Tr}}=\mathcal O(h^\infty).$$ Since the operator on the left-hand side is compactly supported and microlocalized away from the fiber infinity, it is enough to prove an estimate of the $L^2\to L^2$ operator norm instead of the trace class norm. Take arbitrary $h$-independent family $\tilde f=\tilde f(h)\in L^2(X)$ with $\|\tilde f\|_{L^2}\leq 1$ and put $$f(\alpha):=\tilde\chi(\omega) Q_1\Pi Q_2\partial_\omega L_{22}(\omega)\tilde f,\quad u(\alpha):=\mathcal S(\omega)\mathcal R(\omega)\mathcal S(\omega)f(\alpha).$$ Then $f(x,\alpha)$ is compactly supported in both $x\in X$ and $\alpha\in (\alpha_0,\alpha_1)$, $\|f\|_{L^\infty_\alpha L^2_x}$ is polynomially bounded in $h$, and $\WFh(f(\alpha))\subset \Gamma_+\cap W'$. Since $\mathcal R(\omega)_{\mathcal H_2\to \mathcal H_1}=\mathcal O(h^{-2})$ by Theorem \[t:gaps\], we see that $u(\alpha)\in\mathcal H_2$ is compactly supported in $\alpha\in (\alpha_0,\alpha_1)$ and the norm $\|u\|_{L^\infty_\alpha L^2_x}$ is bounded polynomially in $h$. Using Lemma \[l:propagate-outgoing\] similarly to §\[s:estimate-full\], we see that $u,f$ satisfy –, uniformly in $\alpha$. It now suffices to prove that for each choice of $\tilde f$, independent of $\alpha$, we have $$\label{e:trace-int} \int_{\alpha_0}^{\alpha_1}Q_1(\Pi u(\alpha)-R^+_\psi(\omega)\Pi f(\alpha))\,d\alpha=\mathcal O(h^\infty)_{L^2}.$$ Define the semiclassical Fourier transforms $\hat u(s),\hat f(s)$ by . Then  becomes $$\label{e:trace-int-2} Q_1\bigg(\Pi\hat u(0)-{i\over h}\int_{-\infty}^0e^{is(P-ih\nu_+)/h}\psi(s)\Pi\hat f(s)\,ds\bigg)=\mathcal O(h^\infty)_{L^2}.$$ By  and Proposition \[l:schrodinger\], we find microlocally near $W'$, $$\label{e:bazooka} \Pi\hat u(0)={i\over h}\int_{-\infty}^0e^{is(P-ih\nu_+)/h}(\psi(s)\Pi\hat f(s)-ih\psi'(s)\Pi\hat u(s))\,ds +\mathcal O(h^\infty).$$ Take $\tilde\varepsilon>0$ such that $\psi=1$ near $[-\tilde \varepsilon,\tilde\varepsilon]$, so that $\psi'(s)$ is compactly supported in $\{|s|>\tilde \varepsilon\}$. Since $\chi(\omega)$ and $\partial_\omega L_{22}(\omega)$ depend smoothly on $\alpha$, we see that $\|\partial^j_\alpha f(\alpha)\|_{L^\infty_\alpha L^2_x}=\mathcal O(h^{-1/2})$ for all $j$. By repeated integration by parts, we get $$\|\hat f(s)\|_{L^2_s((-\infty,-\tilde\varepsilon])L^2_x}=\mathcal O(h^\infty).$$ Then by , $\Pi\hat u(s)=\mathcal O(h^\infty)$ microlocally near $W'$ locally uniformly in $s\in (-\infty,-\tilde\varepsilon]$, and thus $Q_1e^{is(P-ih\nu_+)/h}\Pi\hat u(s)=\mathcal O(h^\infty)_{L^2}$ uniformly in $s\in (-\infty,0]\cap \operatorname{supp}\psi'$. By , we now get . Now, note that, since the expression under the integral in  is almost analytic in $\omega$, we can replace the integral over $\Im\omega=h\nu_\pm$ by the integral over the real line, with an $\mathcal O(h^\infty)$ error. Then $$\begin{gathered} \mathcal I^-_\chi-\mathcal I^+_\chi=(2\pi h)^{n-1}\operatorname{Tr}\mathcal A_\chi +\mathcal O(h^\infty),\\ \mathcal A_\chi:=\int_{\mathbb R}\chi(\alpha)\partial_\alpha L_{22}(\alpha) Q_1(\mathcal R^-_\psi(\alpha)-\mathcal R^+_\psi(\alpha))\Pi Q_1\Pi Q_2\,d\alpha. \end{gathered}$$ Proposition \[l:trace\] now follows from Proposition \[l:trace-basic\], the fact that $\WFh(\mathcal A_\chi)\subset \widehat W\times\widehat W$, and the following \[l:trace-kiwi\] The operator $\mathcal A_\chi$ lies in $\II$ and its principal symbol, as defined by , satisfies $\sigma_\Lambda(\mathcal A_\chi)\circ j_K=2\pi i \chi(p)$, with $j_K:K^\circ\to\Lambda^\circ$ defined in . Given the multiplication formula , the fact that $\sigma(Q_1)=\sigma(Q_2)=1$ and $\sigma_\Lambda(\Pi)\circ j_K=1$ on $K\cap p^{-1}([\alpha_0,\alpha_1])$ and $\operatorname{supp}\chi\subset (\alpha_0,\alpha_1)$, it is enough to prove the proposition with $\mathcal A_\chi$ replaced by $$\mathcal A'_\chi:=-{i\over h}\int_{\mathbb R^2}e^{-is\alpha/h}\chi(\alpha)\partial_\alpha L_{22}(\alpha)Q_1 e^{isP/h}\psi(s)\,dsd\alpha.$$ Denote $\mathcal L(\alpha)=\chi(\alpha)\partial_\alpha L_{22}(\alpha)Q_1$; it is an operator in $\II$. By applying a microlocal partition of unity to $\mathcal L(\alpha)$, we may reduce to the case when both $\mathcal L(\alpha)$ and $e^{isP/h}$ have local parametrizations (see  for the first one and for example [@e-z Theorem 10.4] for the second one) $$\begin{gathered} \mathcal L(\alpha)u(x)=(2\pi h)^{-(N+n)/2}\int_{\mathbb R^{N+n}} e^{{i\over h}\Phi(x,y,\theta)}a(x,y,\theta,\alpha;h)u(y)\,dyd\theta,\\ e^{isP/h}u(y)=(2\pi h)^{-n} \int_{\mathbb R^{2n}} e^{{i\over h}(S(y,\zeta,s)-z\cdot\zeta)}b(y,\zeta,s;h)u(z)\,dzd\zeta. \end{gathered}$$ Here $S(y,\zeta,s)=y\cdot\zeta+sp(y,\zeta)+\mathcal O(s^2)$ and $b(y,\zeta,0;0)=1$. Then $\mathcal A'_\chi$ takes the form $$\begin{gathered} \mathcal A'_\chi u(x)=-ih^{-1}(2\pi h)^{-(N+3n)/2}\int_{\mathbb R^{N+3n}} e^{{i\over h}(\Phi(x,y,\theta)+S(y,\zeta,s)-z\cdot\zeta-s\alpha)}\\ a(x,y,\theta,\alpha;h) b(y,\zeta,s;h)\psi(s) u(z)\,dy d\theta dzd\zeta dsd\alpha. \end{gathered}$$ We now apply the method of stationary phase in the $y,\zeta,s,\alpha$ variables. The stationary points are given by $s=0$, $\alpha=p(z,\zeta)$, $y=z$, $\zeta=-\partial_z \Phi(x,z,\theta)$. We get $$\mathcal A'_\chi u(x)=-2\pi i (2\pi h)^{-(N+n)/2}\int_{\mathbb R^{N+n}} e^{{i\over h}\Phi(x,z,\theta)}c(x,z,\theta;h)u(z)\,d\theta dz,$$ where $c$ is a classical symbol and $c(x,z,\theta;0)=a(x,z,\theta,p(z,-\partial_z\Phi(x,z,\zeta));0)$. It follows that $\mathcal A'_\chi\in \II$ and $\sigma_\Lambda(\mathcal A'_\chi)(\rho_-,\rho_+)=-2\pi i\sigma_\Lambda(\mathcal L(p(\rho_-)))(\rho_-,\rho_+)$. By , $\sigma_\Lambda(L_{22}(\alpha))(\rho,\rho)=p(\rho)-\alpha+1$ when $\rho\in K\cap p^{-1}([\alpha_0,\alpha_1])$, and thus $\sigma_\Lambda(\partial_\alpha L_{22}(\alpha))(\rho,\rho)=-1$. Therefore, we find $\sigma_\Lambda(\mathcal A'_\chi)(\rho,\rho)=2\pi i\chi(p(\rho))$ for $\rho\in K$. Weyl law for resonances {#s:weyl-law} ======================= In this section, we prove Theorem \[t:weyl-law\], using the Grushin problem from §\[s:grushin\], the trace formula of §\[s:trace\], and several tools from complex analysis. The argument below is quite standard, see for instance [@markus; @sj; @sj-bottles], and is simplified by the fact that we do not aim for the optimal $\mathcal O(h)$ remainder in the Weyl law, instead carrying out the argument in a rectangle of width $\sim 1$ and height $\sim h$. For more sophisticated techniques needed to obtain the optimal remainder, see [@sj-dwe]. First of all, by Proposition \[l:grushin-ultimate\], resonances in the region of interest are (with multiplicities) the zeroes of the holomorphic function $F(\omega)$ defined in . Take $\alpha''_0\in (\alpha_0,\alpha'_0)$, $\alpha''_1\in (\alpha'_1,\alpha_1)$. Fix $\nu_\pm$ satisfying  and let $\{\omega_j\}_{j=1}^{M(h)}$ denote the set of zeroes (counted with multiplicities) of $F(\omega)$ in the region (see Figure \[f:complex\]) ![The contour $\partial\Omega(h)$ (in blue). The horizontal shaded region is $\{\Im\omega\in (-(\nu_{\max}+\varepsilon)h/2,-(\nu_{\min}-\varepsilon)h/2)\}$, where Theorem \[t:gaps\] does not provide polynomial resolvent bounds; the vertical shaded region is the support of $\tilde\chi$.[]{data-label="f:complex"}](nhp.10) $$\Omega(h):=\{\Re\omega\in [\alpha''_0,\alpha''_1],\ \Im\omega\in [\nu_-h,\nu_+h]\}$$ By part 2 of Proposition \[l:grushin-ultimate\] and Jensen’s inequality, see for example [@fwl §2], we have the polynomial bound, for some $N,C$, $$\label{e:rough-upper-bound} M(h)\leq Ch^{-N}.$$ By a standard argument approximating the indicator function of $[\alpha'_0,\alpha'_1]$ by smooth functions from above and below, it is enough to prove that for each $\chi\in C_0^\infty(\alpha_0,\alpha_1)$, $$\label{e:weyl-law-int} (2\pi h)^{n-1}\sum_{j=1}^{M(h)} \chi(\Re\omega_j)=\int_K \chi(p)\,d\operatorname{Vol}_\sigma +\mathcal O(h).$$ Let $\tilde\chi(\omega)$ be an almost analytic continuation of $\chi$, as discussed in the beginning of §\[s:trace\]. We may assume that $\operatorname{supp}\tilde\chi\subset \{\Re\omega\in (\alpha''_0,\alpha''_1)\}$. By Proposition \[l:trace\], we have (with the integral over the vertical parts of $\partial\Omega(h)$ vanishing since $\tilde\chi=0$ there) $$\label{e:intint} {(2\pi h)^{n-1}\over 2\pi i}\oint_{\partial\Omega(h)}\tilde\chi(\omega){\partial_\omega F(\omega)\over F(\omega)} \,d\omega=\int_K\chi(p)\,d\operatorname{Vol}_\sigma+\mathcal O(h).$$ By Lemma $\alpha$ in [@tit §3.9] and the exponential estimates of part 2 of Proposition \[l:grushin-ultimate\] (splitting the region $\Omega(h)$ into boxes of size $h$ and applying Lemma $\alpha$ to each of these boxes, transformed into the unit disk by the Riemann mapping theorem), we have for some fixed $N$, $${\partial_\omega F(\omega)\over F(\omega)}=\sum_{j=1}^{M(h)}{1\over \omega-\omega_j}+G(\omega);\quad G(\omega)=\mathcal O(h^{-N}),\quad \omega\in\Omega(h)\cap\operatorname{supp}\tilde\chi.$$ Applying Stokes theorem to  (over the contour comprised of $\partial\Omega(h)$ minus the sum of circles of small radius $r$ centered at each $\omega_j$, and letting $r\to 0$) we get $$(2\pi h)^{n-1}\sum_{j=1}^{M(h)}\tilde\chi(\omega_j)=\int_K\chi(p)\,d\operatorname{Vol}_\sigma- {(2\pi h)^{n-1}\over 2\pi i}\int_{\Omega(h)}{\partial_\omega F(\omega)\over F(\omega)} \partial_{\bar\omega}\tilde\chi(\omega)\,d\bar\omega\wedge d\omega+\mathcal O(h).$$ Since $\tilde\chi$ is almost analytic and $\Omega(h)$ lies $\mathcal O(h)$ close to the real line, we have $\partial_{\bar\omega}\tilde\chi(\omega)=\mathcal O(h^\infty)$ for $\omega\in\Omega(h)$. Therefore, the second integral on the right-hand side is $\mathcal O(h^\infty)$ and we get $$(2\pi h)^{n-1}\sum_{j=1}^{M(h)}\tilde\chi(\omega_j)=\int_K \chi(p)\,d\operatorname{Vol}_\sigma+\mathcal O(h).$$ Since $\tilde\chi(\omega)=\chi(\Re\omega)+\mathcal O(h)$ for $\omega\in\Omega(h)$, we get $$\label{e:weyl-law-int-2} (2\pi h)^{n-1}\sum_{j=1}^{M(h)}\chi(\Re\omega_j)=\int_K\chi(p)\,d\operatorname{Vol}_\sigma+\mathcal O(h(1+h^{n-1}M(h))).$$ Since one can take $\chi$ to be any compactly supported function on $(\alpha_0,\alpha_1)$, and $M(h)=\mathcal O(h^{-N})$ for some fixed $N$ and any choice of $(\alpha''_0,\alpha''_1)$, by induction we see from  that $M(h)=\mathcal O(h^{1-n})$. Given this bound, implies , which finishes the proof. Example of a manifold with\ $r$-normally hyperbolic trapping {#s:example} ================================ In this Appendix, we provide a simple example of an even asymptotically hyperbolic manifold (as defined in §\[s:framework-ah\]) whose geodesic flow satisfies the dynamical assumptions of §\[s:dynamics\] and the pinching condition , therefore our Theorems \[t:gaps\]–\[t:resonant-states\] apply. This example is a higher dimensional generalization of the hyperbolic cylinder, considered for instance in [@fwl Appendix B]. The resonances for the provided example can be described explicitly via the eigenvalues of the Laplacian on the underlying compact manifold $N$, using separation of variables. However, our results apply to small perturbations of the metric (see §\[s:stability\]), as well as to subprincipal perturbations in the considered operator, when separation of variables no longer takes place. Let $(N,\tilde g)$ be a compact $n-1$ dimensional Riemannian manifold (at the end of this appendix, we will impose further conditions on $\tilde g$). We consider the manifold $M=\mathbb R_r\times N_\theta$ with the metric $$g=dr^2+\cosh^2 r\,\tilde g(\theta,d\theta).$$ Then $M$ has two infinite ends $\{r=\pm\infty\}$; near each of these ends, one can represent it as an even asymptotically hyperbolic manifold by taking the boundary defining function $\tilde x=e^{\mp r}$: $$g={d\tilde x^2\over \tilde x^2}+{(1+\tilde x^2)^2\over 4\tilde x^2}\tilde g(\theta,d\theta).$$ The resonances for the Laplace–Beltrami operator on $M$ therefore fit into the framework of §\[s:framework-assumptions\], as demonstrated in §\[s:framework-ah\]. The associated flow $e^{tH_p}$ is the geodesic flow on the unit cotangent $S^*M$, extended to a homogeneous flow of degree zero on the complement of the zero section in $T^*M$. We now verify the assumptions of §\[s:dynamics\]. If $\xi_r,\xi_\theta$ are the momenta dual to $r,\theta$, then $$p^2=\xi_r^2+\cosh^{-2}r\,\tilde g^{-1}(\theta,\xi_\theta),$$ where $\tilde g^{-1}$ is the dual metric to $g$, defined on the fibers of $T^*N$. We then have $$H_p r={\xi_r\over p},\quad H_p\xi_r={p^2-\xi_r^2\over p}\tanh r.$$ The trapped set $K$ and the incoming/outgoing tails $\Gamma_\pm$ are given by $$\Gamma_\pm=\{\xi_r=\pm p\tanh r\},\quad K=\{r=0,\ \xi_r=0\},$$ or strictly speaking, by the intersections of the sets above with the set $\overline{\mathcal U}$ from . Consider the following defining functions of $\Gamma_\pm$: $$\varphi_\pm=\xi_r\mp p\tanh r,$$ then $\{\varphi_+,\varphi_-\}|_K=2p$ and thus assumptions  and  of §\[s:dynamics\] are satisfied. Next, $$H_p\varphi_\pm=\mp c_\pm\varphi_\pm,\quad c_\pm=1\pm {\xi_r\over p}\tanh r.$$ In particular, $c_\pm|_K=1$ and, arguing as in the proof of Lemma \[l:phi-pm\], we get $$\nu_{\min}=\nu_{\max}=1.$$ In particular, the pinching condition  is satisfied. Finally, in order for the $r$-normal hyperbolicity assumption  of §\[s:dynamics\] to be satisfied, we need to make $\mu_{\max}\ll 1$, with $\mu_{\max}$ defined in . This is a condition on the underlying compact Riemannian manifold $(N,\tilde g)$, since $\mu_{\max}$ is the maximal expansion rate of the geodesic flow of $\tilde g$ on the unit cotangent bundle $S^*N$. To satisfy this condition, we can start with an arbitrary compact Riemannian manifold and multiply its metric by a large constant $C^2$; indeed, if $\varphi_t$ is the geodesic flow on the original manifold, then $\varphi_{C^{-1}t}$ is the geodesic flow on the rescaled manifold and the resulting $\mu_{\max}$ is divided by $C$. I would like to thank Maciej Zworski for plenty of helpful advice and constant encouragement throughout this project, and András Vasy, Stéphane Nonnenmacher, Charles Pugh, Colin Guillarmou, and Michael Hitrik for many very helpful discussions. I am very grateful to Frédéric Faure for a discussion of [@f-t; @f-t2; @f-t3]. Part of this work was completed while visiting Écolé Normale Supérieure in October 2012. This work was also partially supported by the NSF grant DMS-1201417. \#1[[arXiv:\#1](http://arxiv.org/abs/#1)]{} [0]{} Jacques Aguilar and Jean-Michel Combes, *A class of analytic perturbations for one-body Schrödinger Hamiltonians,* Comm. Math. Phys. **22**(1971), no. 4, 269–279. Jean–François Bony, Setsuro Fujiie, Thierry Ramond, and Maher Zerzeri, *Spectral projection, residue of the scattering amplitude, and Schrödinger group expansion for barrier-top resonances,* Ann. Inst. Fourier **61**(2011), no. 4, 1351–1406. Nicolas Burq and Maciej Zworski, *Control for Schrödinger operators on tori,* Math. Res. Lett. **19**(2012), no. 2, 309–324. Kiril Datchev and Semyon Dyatlov, *Fractal Weyl laws for asymptotically hyperbolic manifolds,* to appear in Geom. Funct. Anal., . Kiril Datchev, Semyon Dyatlov, and Maciej Zworski, *Sharp polynomial bounds on the number of Pollicott–Ruelle resonances,* to appear in Erg. Theory Dyn. Syst., . Mouez Dimassi and Johannes Sjöstrand, *Spectral asymptotics in the semi-classical limit,* London Mathematical Society Lecture Note Series **268**, Cambridge University Press, 1999. Dmitry Dolgopyat, *On decay of correlations in Anosov flows,* Ann. of Math. (2) **147**(1998), no. 2, 357–390. Johannes Duistermaat and Victor Guillemin, *The spectrum of positive elliptic operators and periodic bicharacteristics,* Invent. Math. **29**(1975), no. 1, 39–79. Semyon Dyatlov, *Quasi-normal modes and exponential energy decay for the Kerr–de Sitter black hole,* Comm. Math. Phys. **306**(2011), 119–163. Semyon Dyatlov, *Exponential energy decay for Kerr–de Sitter black holes beyond event horizons,* Math. Res. Lett. **18**(2011), 1023–1035. Semyon Dyatlov, *Asymptotic distribution of quasi-normal modes for Kerr–de Sitter black holes,* Annales Henri Poincaré **13**(2012), 1101–1166. Semyon Dyatlov, *Resonance expansions in general relativity,* Ph.D. thesis, in preparation. Semyon Dyatlov and Colin Guillarmou, *Microlocal limits of plane waves and Eisenstein functions,* preprint, . Frédéric Faure and Johannes Sjöstrand, *Upper bound on the density of Ruelle resonances for Anosov flows,* Comm. Math. Phys. **308**(2011), no. 2, 325–364. Frédéric Faure and Masato Tsujii, *Prequantum transfer operator for Anosov diffeomorphism (preliminary version),* preprint, . Frédéric Faure and Masato Tsujii, *Band structure of the Ruelle spectrum of contact Anosov flows,* preprint, . Frédéric Faure and Masato Tsujii, *Spectrum and zeta function of contact Anosov flows,* in preparation. Christian Gérard and Johannes Sjöstrand, *Semiclassical resonances generated by a closed trajectory of hyperbolic type,* Comm. Math. Phys. **108**(1987), no. 3, 391–421. Christian Gérard and Johannes Sjöstrand, *Resonances en limite semi-classique et exposants de Lyapunov,* Comm. Math. Phys. **116**(1988), no. 2, 193–213. Israel C. Gohberg and Efim I. Sigal, *An operator generalization of the logarithmic residue theorem and Rouché’s Theorem,* Mat. Sb. **84(126)**(1971), no. 4, 607–629. Arseni Goussev, Roman Schubert, Holger Waalkens, and Stephen Wiggins, *Quantum theory of reactive scattering in phase space,* Adv. Quant. Chem. **60**(2010), 269–332. Alain Grigis and Johannes Sjöstrand, *Microlocal analysis for differential operators: an introduction,* London Mathematical Society Lecture Note Series **196**, Cambridge University Press, 1994. Colin Guillarmou, *Meromorphic properties of the resolvent on asymptotically hyperbolic manifolds,* Duke Math. J. **129**(2005), no. 1, 1–37. Victor Guillemin and Shlomo Sternberg, *Geometric asymptotics,* AMS, 1990. Laurent Guillopé, Kevin K. Lin, and Maciej Zworski, *The Selberg zeta function for convex co-compact Schottky groups,* Comm. Math. Phys. **245**(2004), no. 1, 149–176. Morris W. Hirsch, Charles C. Pugh, and Michael Shub, *Invariant manifolds,* Lecture Notes in Mathematics **583**, Springer Verlag, 1977. Lars Hörmander, *The spectral function of an elliptic operator,* Acta Math. **121**(1968), 193–218. Lars Hörmander, *The Analysis of Linear Partial Differential Operators. I. Distribution Theory and Fourier Analysis,* Springer Verlag, 1990. Lars Hörmander, *The Analysis of Linear Partial Differential Operators. III. Pseudo-Differential Operators,* Springer Verlag, 1994. Lars Hörmander, *The Analysis of Linear Partial Differential Operators. IV. Fourier Integral Operators,* Springer Verlag, 1994. Alexei Iantchenko, Johannes Sjöstrand, and Maciej Zworski, *Birkhoff normal forms in semi-classical inverse problems,* Math. Res. Lett. **9**(2002), no. 2–3, 337–362. Dmitry Jakobson and Frédéric Naud, *Lower bounds for resonances of infinite-area Riemann surfaces,* Anal. PDE **3**(2010), no. 2, 207–225. Kostas D. Kokkotas and Bernd Schmidt, *Quasi-normal modes of stars and black holes,* Living Rev. Relativity **2**(1999), 2; <http://relativity.livingreviews.org/Articles/lrr-1999-2/> Carlangelo Liverani, *On contact Anosov flows,* Ann. of Math. (2) **159**(2004), no. 3, 1275–1312. Alexandr S. Markus, *Introduction to the spectral theory of polynomial operator pencils,* Translations of Mathematical Monographs **71**, AMS, 1998. Rafe R. Mazzeo and Richard B. Melrose, *Meromorphic extension of the resolvent on complete spaces with asymptotically constant negative curvature,* J. Func. Anal. **75**(1987), no. 2, 260–310. Richard B. Melrose and Maciej Zworski, *Scattering metrics and geodesic flow at infinity,* Invent. Math. **124**(1996), no. 1–3, 389–436. Werner Müller, *Spectral geometry and scattering theory for certain complete surfaces of finite volume,* Invent. Math. **109**(1992), no. 2, 265–305. Stéphane Nonnenmacher, *Spectral problems in open quantum chaos,* Nonlinearity **24**(2011), R123–R167. Stéphane Nonnenmacher, Johannes Sjöstrand, and Maciej Zworski, *From open quantum systems to open quantum maps,* Comm. Math. Phys. **304**(2011), no. 1, 1–48. Stéphane Nonnenmacher, Johannes Sjöstrand, and Maciej Zworski, *Fractal Weyl law for open quantum chaotic maps,* to appear in Ann. of Math. (2), . Stéphane Nonnenmacher and Maciej Zworski, *Quantum decay rates in chaotic scattering,* Acta Math. **203**(2009), no. 2, 149–233. Stéphane Nonnenmacher and Maciej Zworski, *Decay of correlations for normally hyperbolic trapping,* preprint. Tullio Regge, *Analytic properties of the scattering matrix,* Nuovo Cimento **8**(1958), 671–679. Antônio Sá Barreto and Maciej Zworski, *Distribution of resonances for spherical black holes,* Math. Res. Lett. **4**(1997), no. 1, 103–121. Johannes Sjöstrand, *Geometric bounds on the density of resonances for semiclassical problems,* Duke Math. J. **60**(1990), no. 1, 1–57. Johannes Sjöstrand, *A trace formula and review of some estimates for resonances,* in *Microlocal analysis and spectral theory* (Lucca, 1996), 377–437, NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., 490, Kluwer Acad. Publ., Dordrecht, 1997. Johannes Sjöstrand, *Asymptotic distribution of eigenfrequencies for damped wave equations,* Publ. Res. Inst. Math. Sci. **36**(2000), no. 5, 573–611. Johannes Sjöstrand, *Resonances for bottles and trace formulae,* Math. Nachr. **221**(2001), 95–149. Johannes Sjöstrand, *Weyl law for semi-classical resonances with randomly perturbed potentials,* preprint, . Johannes Sjöstrand and Georgi Vodev, *Asymptotics of the number of Rayleigh resonances,* with an appendix by Jean Lannes, Math. Ann. **309**(1997), no. 2, 287–306. Johannes Sjöstrand and Maciej Zworski, *Complex scaling and the distribution of scattering poles,* J. Amer. Math. Soc. **4**(1991), no. 4, 729–769. Johannes Sjöstrand and Maciej Zworski, *Asymptotic distribution of resonances for convex obstacles,* Acta Math. **183**(1999), no. 2, 191–253. Johannes Sjöstrand and Maciej Zworski, *Fractal upper bounds on the density of semiclassical resonances,* Duke Math. J. **137**(2007), no. 3, 381–459. Plamen Stefanov and Georgi Vodev, *Distribution of resonances for the Neumann problem in linear elasticity outside a strictly convex body,* Duke Math. J. **78**(1995), no. 3, 677–714. Siu-Hung Tang and Maciej Zworski, *From quasimodes to resonances,* Math. Res. Lett. **5**(1998), no. 3, 261–272. Michael E. Taylor, *Partial Differential Equations I. Basic theory,* Springer, 1996. Edward Charles Titchmarsh, *The Theory of the Riemann Zeta-function,* second edition, revised by David Rodney Heath-Brown, Oxford University Press, 1986. Masato Tsujii, *Contact Anosov flows and the FBI transform,* preprint, . András Vasy, *Microlocal analysis of asymptotically hyperbolic and Kerr–de Sitter spaces,* to appear in Invent. Math., . András Vasy, *Microlocal analysis of asymptotically hyperbolic spaces and high energy resolvent estimates,* Inverse Problems and Applications. Inside Out II, edited by Gunther Uhlmann, Cambridge University Press, MSRI publications **60**(2012), . San Vũ Ngoc, *Systèmes intégrables semi-classiques: du local au global,* Panoramas et Synthèses 22, 2006. Jared Wunsch and Maciej Zworski, *Resolvent estimates for normally hyperbolic trapped sets,* Ann. Henri Poincaré, **12**(2011), no. 7, 1349–1385. Maciej Zworski, *Distribution of poles for scattering on the real line,* J. Funct. Anal. **73**(1987), no. 2, 277–296. Maciej Zworski, *Semiclassical analysis,* Graduate Studies in Mathematics **138**, AMS, 2012. [^1]: The estimate  implies, in the case , cutoff resolvent bounds $\|\chi R_V(\omega) \chi\|_{L^2\to H^2_h}=\mathcal O(h^{-2})$ for any fixed $\chi\in C_0^\infty(\mathbb R^n)$. [^2]: [@e-z Theorem 10.4] is stated for self-adjoint $P$, rather than operators with real-valued principal symbols; however, the proof works similarly in the latter case, with the transport equation acquiring an additional zeroth order term due to the subprincipal part of $P$. [^3]: The parameter $\varepsilon$ is fixed in Theorem \[t:gaps\]; it is also taken small enough for the results of §\[s:transport\] to hold. [^4]: The choice of antiderivative (see §\[s:prelim-fio\]) is irrelevant here, since the phase factor in $B$ resulting from choosing another antiderivative will be cancelled by the phase factor in $B'$. [^5]: It would be interesting to understand the microlocal structure of $\Xi_0$, starting from the fact that its wavefront set lies in the union of three Lagrangian submanifolds. [^6]: A more careful analysis, as in §\[s:resolvent-bounds\], could give the optimal value of $N$; we do not pursue this here since the value of $N$ is irrelevant for our application in §\[s:weyl-law\].
{ "pile_set_name": "ArXiv" }
--- abstract: 'We reconsider chiral perturbation theory in a finite volume and develop a new computational scheme which smoothly interpolates the conventional $\epsilon$ and $p$ regimes. The counting rule is kept essentially the same as in the $p$ expansion. The zero-momentum modes of Nambu-Goldstone bosons are, however, treated separately and partly integrated out to all orders as in the $\epsilon$ expansion. In this new scheme, the theory remains infra-red finite even in the chiral limit, while the chiral-logarithmic effects are kept present. We calculate the two-point function in the pseudoscalar channel and show that the correlator has a constant contribution in addition to the conventional $\cosh$ function of time $t$. This constant term rapidly disappears in the $p$ regime but it is indispensable for a smooth convergence of the formula to the $\epsilon$ regime result. Our calculation is useful to precisely estimate the finite volume effects in lattice QCD simulations on the pion mass $M_\pi$ and kaon mass $M_K$, as well as their decay constants $F_\pi$ and $F_K$.' author: - Sinya Aoki - Hidenori Fukaya title: 'Interpolation between the $\epsilon$ and $p$ regimes' --- [99]{} Y. Nambu and G. Jona-Lasinio, Phys. Rev.  [**122**]{}, 345 (1961). S. Weinberg, Phys. Rev.  [**166**]{}, 1568 (1968). J. Gasser and H. Leutwyler, Annals Phys.  [**158**]{}, 142 (1984); Nucl. Phys.  B [**250**]{}, 465 (1985). J. Gasser and H. Leutwyler, Phys. Lett.  B [**184**]{}, 83 (1987). J. Gasser and H. Leutwyler, Nucl. Phys.  B [**307**]{}, 763 (1988). C. Bernard \[MILC Collaboration\], Phys. Rev.  D [**65**]{}, 054031 (2002) \[arXiv:hep-lat/0111051\]. G. Colangelo, S. Durr and C. Haefeli, Nucl. Phys.  B [**721**]{}, 136 (2005) \[arXiv:hep-lat/0503014\]. J. Gasser and H. Leutwyler, Phys. Lett.  B [**188**]{}, 477 (1987). H. Neuberger, Phys. Rev. Lett.  [**60**]{} (1988) 889. F. C. Hansen, Nucl. Phys.  B [**345**]{}, 685 (1990); F. C. Hansen and H. Leutwyler, Nucl. Phys.  B [**350**]{}, 201 (1991). P. Hasenfratz and H. Leutwyler, Nucl. Phys.  B [**343**]{}, 241 (1990). H. Leutwyler and A. Smilga, Phys. Rev.  D [**46**]{}, 5607 (1992). P. H. Damgaard, M. C. Diamantini, P. Hernandez and K. Jansen, Nucl. Phys.  B [**629**]{} (2002) 445 \[arXiv:hep-lat/0112016\]. P. H. Damgaard, P. Hernandez, K. Jansen, M. Laine and L. Lellouch, Nucl. Phys.  B [**656**]{}, 226 (2003) \[arXiv:hep-lat/0211020\]. P. Hernandez and M. Laine, JHEP [**0301**]{}, 063 (2003) \[arXiv:hep-lat/0212014\]. P. H. Damgaard and H. Fukaya, Nucl. Phys.  B [**793**]{}, 160 (2008) \[arXiv:0707.3740 \[hep-lat\]\]. F. Bernardoni and P. Hernandez, JHEP [**0710**]{}, 033 (2007) \[arXiv:0707.3887 \[hep-lat\]\]. F. Bernardoni, P. H. Damgaard, H. Fukaya and P. Hernandez, JHEP [**0810**]{}, 008 (2008) \[arXiv:0808.1986 \[hep-lat\]\]. G. Akemann, F. Basile and L. Lellouch, JHEP [**0812**]{}, 069 (2008) \[arXiv:0804.3809 \[hep-lat\]\]. A. Shindler, Phys. Lett.  B [**672**]{}, 82 (2009) \[arXiv:0812.2251 \[hep-lat\]\]. O. Bar, S. Necco and S. Schaefer, JHEP [**0903**]{}, 006 (2009) \[arXiv:0812.2403 \[hep-lat\]\]. F. Bernardoni, P. Hernandez and S. Necco, JHEP [**1001**]{}, 070 (2010) \[arXiv:0910.2537 \[hep-lat\]\]. O. Bar, S. Necco and A. Shindler, JHEP [**1004**]{}, 053 (2010) \[arXiv:1002.1582 \[hep-lat\]\]. F. Bernardoni, N. Garron, P. Hernandez, S. Necco and C. Pena, arXiv:1008.1870 \[hep-lat\]. C. Lehner, S. Hashimoto and T. Wettig, JHEP [**1006**]{}, 028 (2010) \[arXiv:1004.5584 \[hep-lat\]\]. T. DeGrand, Z. Liu and S. Schaefer, Phys. Rev.  D [**74**]{}, 094504 (2006) \[Erratum-ibid.  D [**74**]{}, 099904 (2006)\] \[arXiv:hep-lat/0608019\]. C. B. Lang, P. Majumdar and W. Ortner, Phys. Lett.  B [**649**]{}, 225 (2007) \[arXiv:hep-lat/0611010\]. P. Hasenfratz, D. Hierl, V. Maillart, F. Niedermayer, A. Schafer, C. Weiermann and M. Weingart, JHEP [**0911**]{}, 100 (2009) \[arXiv:0707.0071 \[hep-lat\]\]. H. Fukaya [*et al.*]{} \[JLQCD collaboration\], Phys. Rev.  D [**77**]{}, 074503 (2008) \[arXiv:0711.4965 \[hep-lat\]\]. A. Hasenfratz, R. Hoffmann and S. Schaefer, Phys. Rev.  D [**78**]{}, 014515 (2008) \[arXiv:0805.2369 \[hep-lat\]\]. A. Hasenfratz, R. Hoffmann and S. Schaefer, Phys. Rev.  D [**78**]{}, 054511 (2008) \[arXiv:0806.4586 \[hep-lat\]\]. O. Bar, S. Necco and S. Schaefer, PoS [**LAT2009**]{}, 078 (2009) \[arXiv:0910.2372 \[hep-lat\]\]. K. Jansen and A. Shindler, PoS [**LAT2009**]{}, 070 (2009) \[arXiv:0911.1931 \[hep-lat\]\]. P. H. Damgaard and H. Fukaya, JHEP [**0901**]{}, 052 (2009) \[arXiv:0812.2797 \[hep-lat\]\]. P. H. Damgaard and S. M. Nishigaki, Nucl. Phys.  B [**518**]{}, 495 (1998) \[arXiv:hep-th/9711023\]. T. Wilke, T. Guhr and T. Wettig, Phys. Rev.  D [**57**]{}, 6486 (1998) \[arXiv:hep-th/9711057\]. G. Akemann and P. H. Damgaard, Nucl. Phys.  B [**528**]{}, 411 (1998) \[arXiv:hep-th/9801133\]. A. V. Smilga and J. Stern, Phys. Lett.  B [**318**]{}, 531 (1993). J. C. Osborn, D. Toublan and J. J. M. Verbaarschot, Nucl. Phys.  B [**540**]{}, 317 (1999) \[arXiv:hep-th/9806110\]. H. Fukaya, S. Aoki, S. Hashimoto, T. Kaneko, J. Noaki, T. Onogi and N. Yamada \[JLQCD collaboration\], Phys. Rev. Lett.  [**104**]{}, 122002 (2010) \[Erratum-ibid.  [**105**]{}, 159901 (2010)\] \[arXiv:0911.5555 \[hep-lat\]\]. H. Fukaya [*et al.*]{} \[JLQCD and TWQCD collaborations\], arXiv:1012.4052 \[hep-lat\]. P. H. Damgaard and K. Splittorff, Nucl. Phys.  B [**572**]{}, 478 (2000) \[arXiv:hep-th/9912146\]. P. H. Damgaard and K. Splittorff, Phys. Rev.  D [**62**]{}, 054509 (2000) \[arXiv:hep-lat/0003017\]. P. H. Damgaard, Phys. Lett.  B [**476**]{}, 465 (2000) \[arXiv:hep-lat/0001002\]. J. Wess and B. Zumino, Phys. Lett.  B [**37**]{}, 95 (1971). E. Witten, Nucl. Phys.  B [**223**]{}, 422 (1983). K. Splittorff and J. J. M. Verbaarschot, Phys. Rev. Lett.  [**90**]{}, 041601 (2003) \[arXiv:cond-mat/0209594\]. Y. V. Fyodorov and G. Akemann, JETP Lett.  [**77**]{}, 438 (2003) \[Pisma Zh. Eksp. Teor. Fiz.  [**77**]{}, 513 (2003)\] \[arXiv:cond-mat/0210647\]. K. Splittorff and J. J. M. Verbaarschot, Nucl. Phys.  B [**683**]{}, 467 (2004) \[arXiv:hep-th/0310271\]. J. Lenaghan and T. Wilke, Nucl. Phys.  B [**624**]{}, 253 (2002) \[arXiv:hep-th/0108166\]. S. Aoki and H. Fukaya, Phys. Rev.  D [**81**]{}, 034022 (2010) \[arXiv:0906.4852 \[hep-lat\]\]. Y. Y. Mao and T. W. Chiu \[TWQCD Collaboration\], Phys. Rev.  D [**80**]{}, 034502 (2009) \[arXiv:0903.2146 \[hep-lat\]\]. S. R. Sharpe and N. Shoresh, Phys. Rev.  D [**62**]{}, 094503 (2000) \[arXiv:hep-lat/0006017\]. J. Bijnens, N. Danielsson and T. A. Lahde, Phys. Rev.  D [**73**]{}, 074509 (2006) \[arXiv:hep-lat/0602003\]. M. Luscher, Commun. Math. Phys.  [**104**]{}, 177 (1986). H. Leutwyler, Phys. Lett.  B [**189**]{}, 197 (1987). P. Hasenfratz and F. Niedermayer, Z. Phys. B [**92**]{} (1993) 91 \[arXiv:hep-lat/9212022\]. P. Hasenfratz, Nucl. Phys.  B [**828**]{}, 201 (2010) \[arXiv:0909.3419 \[hep-th\]\]. W. Bietenholz [*et al.*]{}, Phys. Lett.  B [**687**]{}, 410 (2010) \[arXiv:1002.1696 \[hep-lat\]\]. M. Weingart, arXiv:1006.5076 \[hep-lat\]. F. Niedermayer and C. Weiermann, Nucl. Phys.  B [**842**]{}, 248 (2011) \[arXiv:1006.5855 \[hep-lat\]\]. W. Bietenholz [*et al.*]{} \[QCDSF Collaboration\], J. Phys. Conf. Ser.  [**287**]{}, 012016 (2011) \[arXiv:1103.3311 \[hep-lat\]\].
{ "pile_set_name": "ArXiv" }
--- abstract: 'Current OFDMA systems group resource blocks into subband to form the basic feedback unit. Homogeneous feedback design with a common subband size is not aware of the heterogeneous channel statistics among users. Under a general correlated channel model, we demonstrate the gain of matching the subband size to the underlying channel statistics motivating heterogeneous feedback design with different subband sizes and feedback resources across clusters of users. Employing the best-M partial feedback strategy, users with smaller subband size would convey more partial feedback to match the frequency selectivity. In order to develop an analytical framework to investigate the impact of partial feedback and potential imperfections, we leverage the multi-cluster subband fading model. The perfect feedback scenario is thoroughly analyzed, and the closed form expression for the average sum rate is derived for the heterogeneous partial feedback system. We proceed to examine the effect of imperfections due to channel estimation error and feedback delay, which leads to additional consideration of system outage. Two transmission strategies: the fix rate and the variable rate, are considered for the outage analysis. We also investigate how to adapt to the imperfections in order to maximize the average goodput under heterogeneous partial feedback.' author: - 'Yichao Huang, , and Bhaskar D. Rao, [^1] [^2] [^3]' title: | Performance Analysis of Heterogeneous Feedback Design in an OFDMA Downlink\ with Partial and Imperfect Feedback --- [To Appear In IEEE Transactions on Signal Processing]{} Heterogeneous feedback, OFDMA, partial feedback, imperfect feedback, average goodput, multiuser diversity Introduction ============ Leveraging feedback to obtain the channel state information at the transmitter (CSIT) enables a wireless system to adapt its transmission strategy to the varying wireless environment. The growing number of wireless users, as well as their increasing demands for higher data rate services impose a significant burden on the feedback link. In particular for OFDMA systems, which have emerged as the core technology in 4G and future wireless systems, full CSIT feedback may become prohibitive because of the large number of resource blocks. This motivates more efficient feedback design approaches in order to achieve performance comparable to a full CSIT system with reduced feedback. In the recent years, considerable work and effort has been focused on limited or partial feedback design, e.g., see [@love08] and the references therein. To the best of our knowledge, most of the existing partial feedback designs are homogeneous, i.e., users’ feedback consumptions do not adapt to the underlying channel statistics. In this paper, we propose and analyze a heterogeneous feedback design, which aligns users’ feedback needs to the statistical properties of their wireless environments. Current homogeneous feedback design in OFDMA systems groups the resource blocks into subband [@zhu09] which forms the basic scheduling and feedback unit. Since the subband granularity is determined by the frequency selectivity, or the coherence bandwidth of the underlying channel, it would be beneficial to adjust the subband size of different users according to their channel statistics. Empirical measurements and analysis from the channel modeling field have shown that the root mean square (RMS) delay spread which is closely related to the coherence bandwidth, is both location and environment dependent [@asplund06; @huang12]. The typical RMS delay spread for an indoor environment in WLAN does not exceed hundreds of nanoseconds; whereas in the outdoor environment of a cellular system, it can be up to several microseconds. Intuitively, users with lower RMS delay spread could model their channel with a larger subband size and require less feedback resource than the users with higher RMS delay spread. Herein, we investigate this heterogeneous feedback design in a multiuser opportunistic scheduling framework where the system favors the user with the best channel condition to exploit multiuser diversity [@knopp95; @viswanath02]. There are two major existing partial feedback strategies for opportunistic scheduling, one is based on thresholding where each user provides one bit of feedback per subband to indicate whether or not the particular channel gain exceeds a predetermined or optimized threshold [@sanayei07; @hassel07; @chen08; @pugh10]. The other promising strategy currently considered in practical systems such as LTE [@sesia11] is the best-M strategy, where the receivers order and convey the M best channels [@jung07; @ko07; @choi07; @choi08; @pedersen09; @leinonen09; @donthi11; @hur11]. The best-M partial feedback strategy is embedded in the proposed heterogeneous feedback framework. Apart from the requirement of partial feedback to save feedback resource, the study of imperfections is also important to understand the effect of channel estimation error and feedback delay on the heterogeneous feedback framework. These imperfections are also considered in our work. Focus and Contributions of the Paper ------------------------------------ An important step towards heterogeneous feedback design is leveraging the “match" among coherence bandwidth, subband size and partial feedback. Under a given amount of partial feedback, if the subband size is much larger than the coherence bandwidth, then multiple independent channels could exist within a subband and the subband-based feedback could only be a coarse representative of the channels. On the other hand, if the subband size is much smaller than the coherence bandwidth, then channels in adjacent subbands are likely to be highly correlated and requiring feedback on adjacent subbands could be a waste of resource; or a small amount of subband-based partial feedback may not be enough to reflect the channel quality. In order to support this heterogeneous framework, we first consider the scenario of a general correlated channel model with one cluster of users with the same coherence bandwidth. The subband size is adjustable and each user employs the best-M partial feedback strategy to convey the M best channel quality information (CQI) which is defined to be the subband average rate. The simulation result shows that a suitable chosen subband size yields higher average sum rate under partial feedback conforming the aforementioned intuition. This motivates the design of heterogeneous feedback to “match" the subband size to the coherence bandwidth. The above-mentioned study, though closely reflects the relevant mechanism, is not analytically tractable due to two main reasons. Firstly, the general correlated channel model complicates the statistical analysis of the CQI. Secondly, the use of subband average rate as CQI makes it difficult to analyze the multi-cluster scenario. Therefore, a simplified generic channel model is needed that balances the competing needs of analytical tractability and practical relevance. In order to facilitate analysis, a subband fading channel model is developed that generalizes the widely used frequency domain block fading channel model. The subband fading model is suited for the multi-cluster analysis. According to the subband fading model, the channel frequency selectivity is flat within each subband, and independent across subbands. Since the subband sizes are different across different clusters, the number of independent channels are heterogeneous across clusters and this yields heterogeneous partial feedback design. Another benefit of the subband fading model is that the CQI becomes the channel gain and thus facilitate further statistical analysis. Under the multi-cluster subband fading model and the assumption of perfect feedback, we derive a closed form expression for the average sum rate. Additionally, we approximate the sum rate ratio for heterogeneous design, i.e., the ratio of the average sum rate obtained by a partial feedback scheme to that achieved by a full feedback scheme, in order to choose different best-M for users with different coherence bandwidth. We also compare and demonstrate the potential of the proposed heterogeneous feedback design against the homogeneous case under the same feedback constraint in our simulation study. The average sum rate helps in understanding the system performance with perfect feedback. In practical feedback systems, imperfections occur such as channel estimation error and feedback delay. These inevitable factors degrade the system performance by causing outage [@piantanida09; @isukapalli10]. Therefore, rather than using average sum rate as the performance metric, we employ the notion of average goodput [@lau08; @wu10; @akoum10] to incorporate outage probability. Under the multi-cluster subband fading model, we perform analysis on the average goodput and the average outage probability with heterogeneous partial feedback. In addition to examining the impact of imperfect feedback on multiuser diversity [@ma05; @kuhne08], we also investigate how to adapt and optimize the average goodput in the presence of these imperfections. We consider both the fixed rate and the variable rate scenarios, and utilize bounding technique and an efficient approximation to derive near-optimal strategies. To summarize, the contributions of this paper are threefold: a conceptual heterogeneous feedback design framework to adapt feedback amount to the underlying channel statistics, a thorough analysis of both perfect and imperfect feedback systems under the multi-cluster subband fading model, and the development of approximations and near-optimal approaches to adapt and optimize the system performance. The rest of the paper is organized as follows. The motivation under the general correlated channel model and the development of system model is presented in Section \[system\]. Section \[perfect\] deals with perfect feedback, and Section \[imperfect\] examines imperfect feedback due to channel estimation error and feedback delay. Numerical results are presented in Section \[numerical\]. Finally, Section \[conclusion\] concludes the paper. System Model {#system} ============ Motivation for Heterogeneous Partial Feedback {#motivation} --------------------------------------------- This part provides justification for the adaptation of subband size with one cluster of users under the general correlated channel model, and motivates the design of heterogeneous partial feedback for the multi-cluster scenario in Section \[multicluster\]. Consider a downlink multiuser OFDMA system with one base station and $K$ users. One cluster of user is assumed in this part and users in this cluster are assumed to experience the same frequency selectivity. The system consists of $N_\mathsf{c}$ subcarriers. $H_{k,n}$, the frequency domain channel transfer function between transmitter and user $k$ at subcarrier $n$, can be written as: $$\label{system:eq_1} H_{k,n}=\sum_{l=0}^{L-1}\sigma_lF_{k,l}\exp\left(-\frac{j2\pi ln}{N_\mathsf{c}}\right),$$ where $L$ is the number of channel taps, $\sigma_l$ for $l=0,\ldots,L-1$ represents the channel power delay profile and is normalized, i.e., $\sum_{l=0}^{L-1}\sigma_l^2=1$, $F_{k,l}$ denotes the discrete time channel impulse response, which is modeled as complex Gaussian distributed random processes with zero mean and unit variance $\mathcal{CN}(0,1)$ and is i.i.d. across $k$ and $l$. Only fast fading effect is considered in this paper, i.e., the effects of path loss and shadowing are assumed to be ideally compensated by power control. The received signal of user $k$ at subcarrier $n$ can be written as: $$\label{system:eq_2} u_{k,n}=\sqrt{P_{\mathsf{c}}}H_{k,n}s_{k,n}+v_{k,n},$$ where $P_{\mathsf{c}}$ is the average received power per subcarrier, $s_{k,n}$ is the transmitted symbol and $v_{k,n}$ is the additive white noise distributed as $\mathcal{CN}(0,\sigma_{n_{\mathsf{c}}}^2)$. From (\[system:eq\_1\]), it can be shown that $H_{k,n}$ is distributed as $\mathcal{CN}(0,1)$. The channels at different subcarriers are correlated, and the correlation coefficient between subcarriers $n_1$ and $n_2$ can be described as follows: $$\label{system:eq_3} \mathrm{cov}(H_{k,n_1},H_{k,n_2})=\sum_{l=0}^{L-1}\sigma_l^2\exp\left(-\frac{j2\pi l(n_2-n_1)}{N_\mathsf{c}}\right).$$ In general, adjacent subcarriers are highly correlated. In order to reduce feedback needs, $R_\mathsf{c}$ subcarriers are formed as one resource block, and $\eta$ resource blocks are grouped into one subband. Thus, there are $N=\frac{N_\mathsf{c}}{R_\mathsf{c}}$ resource blocks and $\frac{N}{\eta}$ subbands. In this manner, each user performs subband-based feedback to enable opportunistic scheduling at the transmitter. Since the channels are correlated and there is one CQI to represent a given subband, the CQI is a function of the all the individual channels within that subband. Herein, we employ the following subband (aggregate) average rate $S_{k,r}$ as the functional form [@forney98; @al96] of the CQI for user $k$ at subband $r$: $$\label{system:eq_4} S_{k,r}\triangleq\frac{1}{\eta R_{\mathsf{c}}}\sum_{n=(r-1)\eta R_{\mathsf{c}}+1}^{r\eta R_{\mathsf{c}}}\log_2\left(1+\frac{P_{\mathsf{c}}|H_{k,n}|^2}{\sigma_{n_{\mathsf{c}}}^2}\right).$$ Each user employs the best-M partial feedback strategy and conveys back the $M$ best CQI values selected from $S_{k,r}, 1\leq r\leq \frac{N}{\eta}$. A detailed description of the best-M strategy can be found in [@choi08; @leinonen09; @hur11]. After the base station receives feedback, it performs opportunistic scheduling and selects the user $k$ for transmission at subband $r$ if user $k$ has the largest CQI at subband $r$. Also, it is assumed that if no user reports CQI for a certain subband, scheduling outage happens and the transmitter does not utilize it for transmission. ![Comparison of average sum rate for different subband sizes ($\eta=1,2,4$) and partial feedback ($M=2,4$) with respect to the number of users. A general correlated channel model is assumed with an exponential power delay profile. ($N_{\mathsf{c}}=256$, $N=32$, $L=16$, $\delta=4$, $\frac{P_{\mathsf{c}}}{\sigma_{n_{\mathsf{c}}}^2}=10$ dB)[]{data-label="fig_1"}](Figure/fig1.eps){width="0.55\linewidth"} ![Illustration of the multi-cluster subband fading channel model for two different clusters with $16$ resource blocks. The subband sizes equal $2$ and $4$ for the two different clusters respectively. According to the subband fading model, the channel frequency selectivity is flat within each subband, and independent across subbands. The subband sizes can be heterogeneous across clusters, and this leads to heterogeneous channel frequency selectivity across clusters. The subband fading model approximates the general correlated channel model, and is useful for statistical analysis.[]{data-label="fig_2"}](Figure/fig2.eps){width="0.8\linewidth"} Now we demonstrate the need to adapt the subband size to achieve the potential “match" among coherence bandwidth, subband size and partial feedback through a simulation example. The channel is modeled according to the exponential power delay profile [@weinfurtner02; @mckay08; @eslami11]: $\sigma_l^2=\frac{1-\exp(-1/\delta)}{1-\exp(-L/\delta)}\exp\left(-\frac{l}{\delta}\right)$ for $0\leq l<L$, where the parameter $\delta$ is related to the RMS delay spread. The simulation parameters are: $N_{\mathsf{c}}=256$, $N=32$, $L=16$, $\delta=4$, $\frac{P_{\mathsf{c}}}{\sigma_{n_{\mathsf{c}}}^2}=10$ dB. The subband size $\eta$ can be adjusted and ranges from $1$ to $4$ resource blocks. We consider partial feedback with $M=2$ and $M=4$. The average sum rate of the system for different subband sizes and partial feedback with respect to the number of users is shown in Fig. \[fig\_1\]. Under the given coherence bandwidth, several observations can be made. Firstly, the curves with $\eta=4$ has the smallest increasing rate because a larger subband size gives a poor representation of the channel. Secondly, the curve with $\eta=1,M=2$ has the smallest average sum rate because a small amount of partial feedback is not enough to reflect the channel quality. Thirdly, the two curves $\eta=1,M=4$ and $\eta=2,M=2$ possess similar increasing rate. This is because the underlying channel is highly correlated within $2$ resource blocks and thus having $M$-best feedback with $\eta=2$ yields similar effect as having $2M$-best feedback with $\eta=1$. From the above observations, $\eta=2$ matches the frequency selectivity and there would be performance loss or waste of feedback resource when a subband size is blindly chosen. In a multi-cluster scenario where users in different clusters experience diverse coherence bandwidth, this advocates heterogeneous subband size and heterogeneous feedback. The general correlated channel model as well as the non-linearity of the CQI, though useful to demonstrate the need for heterogeneous feedback, does not lend itself to tractable statistical analysis. To develop a tractable analytical framework, an approximated channel model is needed. A widely used model is the block fading model in the frequency domain [@mceliece84; @medard02] due to its simplicity and capability to provide a good approximation to actual physical channels. According to the block fading model, the channel frequency selectivity is flat within each block, and independent across blocks [@chen08; @leinonen09; @hur11]. Herein, we generalize the block fading model to the subband fading model for the multi-cluster scenario. We assume that users possessing similar frequency selectivity are grouped into a cluster and the subband size is perfectly matched to the coherence bandwidth for a given cluster. According to the subband fading model, for a given cluster with a perfectly matched subband size, the channel frequency selectivity is flat within each subband, and independent across subbands. Fig. \[fig\_2\] demonstrates the subband fading model for two different clusters with different subband sizes under a given number of resource blocks. Multi-Cluster Subband Fading Model {#multicluster} ---------------------------------- We now present the multi-cluster subband fading model. Consider a downlink multiuser OFDMA system with one base station and $G$ clusters of users. The system consists of $N$ resource blocks and the total number of users equals $K$. Users in cluster $\mathcal{K}_g$ are indexed by the set $\mathcal{K}_g=\{1,\ldots,k,\ldots,K_g\}$ for $1\leq g\leq G$, with $|\mathcal{K}_g|=K_g$ and $\sum_{g=1}^G K_g=K$. In our framework, users in the same cluster group their resource blocks into subbands in the same manner while each cluster can potentially employ a different grouping which enables the subband size to be heterogeneous between clusters. Denote $\eta_g$ as the subband size for cluster $\mathcal{K}_g$, and $\eta_g \in \{2^0,2^1,\ldots,N\}$. The $\eta_g$’s are ordered such that $\eta_1<\cdots<\eta_G$. Based on the assumption for $\eta_g$, the number of subbands in cluster $\mathcal{K}_g$ equals $\frac{N}{\eta_g}$. Let $H_{k,r}^{(g)}$ be the frequency domain channel transfer function between transmitter and user $k$ in cluster $\mathcal{K}_g$ at subband $r$, where $1\leq k \leq K_g, 1\leq r \leq \frac{N}{\eta_g}$. $H_{k,r}^{(g)}$ is distributed as $\mathcal{CN}(0,1)$. According to the subband fading model, $H_{k,r}^{(g)}$ is assumed to be independent across users and subbands in cluster $\mathcal{K}_g$. The feedback for different clusters is at different granularity, and so to model the channel for the different clusters of users at the same basic resource block level, some additional notation is needed. Let $\tilde{H}_{k,n}^{(g)}=H_{k,\lceil \frac{n}{\eta_g}\rceil}^{(g)}$ with $1\leq n \leq N$ denote the resource block based channel transfer function. Then the received signals of user $k$ in cluster $\mathcal{K}_g$ at resource block $n$ can be represented by: $$\label{system:eq_5} u_{k,n}^{(g)}= \sqrt{P}\tilde{H}_{k,n}^{(g)}s_{k,n}^{(g)}+v_{k,n}^{(g)},$$ where $P$ is the average received power per resource block, $s_{k,n}^{(g)}$ is the transmitted symbol and $v_{k,n}^{(g)}$ is additive white noise distributed with $\mathcal{CN}(0,\sigma_n^2)$. Let $Z_{k,r}^{(g)}\triangleq |H_{k,r}^{(g)}|^2$ denote the CQI for user $k$ in cluster $\mathcal{K}_g$ at subband $r$. In order to reduce the feedback load, users employ the best-M strategy to feed back their CQI. In the basic best-M feedback policy, users measure CQI for each resource block at their receiver and feed back the CQI values of the $M$ best resource blocks chosen from the total $N$ values. For each resource block, the scheduling policy selects the user with the largest CQI among the users who fed back CQI to the transmitter for that resource block. However, in our heterogeneous partial feedback framework, since the number of independent CQI for cluster $\mathcal{K}_g$ is $\frac{N}{\eta_g}$, a fair and reasonable way to allocate the feedback resource is to linearly scale the feedback amount for users in cluster $\mathcal{K}_g$. To be specific, user $k$ in $\mathcal{K}_G$ (i.e., the cluster with the largest subband size) is assumed to feed back the $M$ best CQI selected from $\{Z_{k,r}^{(G)}\}, 1\leq r \leq \frac{N}{\eta_G}$, whereas user $k$ in $\mathcal{K}_g$ conveys the $\frac{\eta_G}{\eta_g}M$ best CQI selected from $\{Z_{k,r}^{(g)}\}, 1\leq r \leq \frac{N}{\eta_g}$. After receiving feedback from all the clusters, for each resource block the system schedules the user for transmission with the largest CQI. It is useful to note that the user feedback is based on the subband level, while the base station schedules transmission at the resource block level. Perfect Feedback {#perfect} ================ In this section, the CQI are assumed to be fed back without any errors and the average sum rate is employed as the performance metric for system evaluation. We derive a closed form expression for the average sum rate in Section \[spectral\_efficiency\] for the multi-cluster heterogeneous feedback system. In Section \[ratio\] we analyze the relationship between the sum rate ratio and the choice of the best-M. Derivation of Average Sum Rate {#spectral_efficiency} ------------------------------ According to the assumption, the CQI $Z_{k,r}^{(g)}$ is i.i.d. across subbands and users, and thus let $F_Z$ denote the CDF. Because only a subset of the ordered CQI are fed back, from the transmitter’s perspective, if it receives feedback on a certain resource block from a user, it is likely to be any one of the CQI from the ordered subset. We now aim to find the CDF of the CQI seen at the transmitter side as a consequence of partial feedback. Let $\tilde{Y}_{k,n}^{(g)}$ denote the reported CQI viewed at the transmitter for user $k$ in $\mathcal{K}_g$ at resource block $n$. Also, let $Y_{k,r}^{(g)}$ represent the subband-based CQI seen at the transmitter for user $k$ in $\mathcal{K}_g$ at subband $r$, then $\tilde{Y}_{k,n}^{(g)}=Y_{k,\lceil \frac{n}{\eta_g}\rceil}^{(g)}$. The following lemma describes the CDF of $\tilde{Y}_{k}^{(g)}$ (the index $n$ is dropped for notational simplicity), which is denoted by $F_{\tilde{Y}_{k}^{(g)}}$. \[lemma\_1\] The CDF of $\tilde{Y}_{k}^{(g)}$ is given by: $$\label{perfect:eq_1} F_{\tilde{Y}_{k}^{(g)}}(x)=\sum_{m=0}^{\frac{\eta_G}{\eta_g}M-1}\xi_g(N,M,\boldsymbol\eta,m)(F_Z(x))^{\frac{N}{\eta_g}-m},$$ where the vector $\boldsymbol\eta\triangleq(\eta_1,\cdots,\eta_g,\cdots,\eta_G)$ and $$\label{perfect:eq_2} \xi_g(N,M,\boldsymbol\eta,m)=\sum_{i=m}^{\frac{\eta_G}{\eta_g}M-1}\frac{\frac{\eta_G}{\eta_g}M-i}{\frac{\eta_G}{\eta_g}M}{\frac{N}{\eta_g}\choose i}{i\choose m}\left(-1\right)^{i-m}.$$ The proof is provided in Appendix \[appenA\]. Let $k_n^*$ demote the selected user at resource block $n$, then according to the scheduling policy: $$\label{perfect:eq_3} k_n^*=\arg\max_{k\in \mathcal{U}_n}\;\{\tilde{Y}_{k,n}^{(1)},\cdots,\tilde{Y}_{k,n}^{(g)},\cdots,\tilde{Y}_{k,n}^{(G)}\},$$ where $\mathcal{U}_n\triangleq\{\mathcal{U}_n^{(1)},\cdots,\mathcal{U}_n^{(g)},\cdots,\mathcal{U}_n^{(G)}\}$ is the set of users who convey feedback for resource block $n$, with $|\mathcal{U}_n^{(g)}|=\tau_g$ representing the number of users belonging to $\mathcal{U}_n$ in cluster $\mathcal{K}_g$. It can be easily seen that in the full feedback case, i.e., $M=M_F\triangleq\frac{N}{\eta_G}$, $|\mathcal{U}_n^{(g)}|=K_g$. For the general case when $1\leq M<M_F$, the probability mass function (PMF) of $\mathcal{U}_n$ is given by: $$\label{perfect:eq_4} \mathbb{P}(\mathcal{U}_n)=\left(\prod_{g=1}^G{K_g\choose \tau_g}\right)\left(\frac{\eta_G M}{N}\right) ^{\sum_{g=1}^G\tau_g}\left(1-\frac{\eta_G M}{N}\right)^{K-\sum_{g=1}^G\tau_g}, \quad 0\leq \tau_g\leq K_g.$$ *Remark:* Only the largest subband size $\eta_G$ appears in the expression of $\mathbb{P}(\mathcal{U}_n)$ instead of the vector $\boldsymbol\eta$. This is due to our heterogeneous partial feedback design to let users in cluster $\mathcal{K}_g$ convey back the $\frac{\eta_G}{\eta_g}M$ best CQI out of $\frac{N}{\eta_g}$ values. Now we turn to determine the conditional CDF of the CQI for the selected user at resource block $n$, conditioned on the set of users providing CQI for that resource block. Since users are equiprobable to be scheduled according to the fair scheduling policy, the condition on $k_n^*$ is not described explicitly, and so we denote the conditional CDF as $F_{X_n|\mathcal{U}_n}$, where $X_n|\mathcal{U}_n$ is the conditional CQI of the selected user at resource block $n$. Notice from Lemma \[lemma\_1\] that $\tilde{Y}_{k}^{(g)}$ possess a different distribution for different $g$ due to the heterogeneous feedback from different clusters. Using order statistics [@david03] yields $F_{X_n|\mathcal{U}_n}$ as: $$\label{perfect:eq_5} F_{X_n|\mathcal{U}_n}(x)=\prod_{g=1}^G(F_{\tilde{Y}_{k}^{(g)}}(x))^{\tau_g}.$$ Then the polynomial form of $F_{X_n|\mathcal{U}_n}$ can be obtained, which is stated in the following theorem. \[theorem\_1\] The CDF of $F_{X_n|\mathcal{U}_n}$ is given by: $$\label{perfect:eq_6} F_{X_n|\mathcal{U}_n}(x)=\sum_{m=0}^{\Phi(M,\boldsymbol\eta,\boldsymbol\tau)}\Theta_{G-1}(N,M,\boldsymbol\eta,\boldsymbol\tau,m)(F_Z(x))^{\sum_{g=1}^G\frac{N}{\eta_g}\tau_g-m},$$ where the vector $\boldsymbol\tau\triangleq(\tau_1,\cdots,\tau_g,\cdots,\tau_G)$, $\Phi(M,\boldsymbol\eta,\boldsymbol\tau)\triangleq\sum_{g=1}^G\tau_g\left(\frac{\eta_G}{\eta_g}M-1\right)$, $$\label{perfect:eq_7} \Theta_{g}(N,M,\boldsymbol\eta,\boldsymbol\tau,m)=\left\{ \begin{array}{l} \mathop{\sum}\limits_{i=0}^m\Lambda_1(N,M,\boldsymbol\eta,\boldsymbol\tau,i)\Lambda_2(N,M,\boldsymbol\eta,\boldsymbol\tau,m-i),\quad g=1\\ \mathop{\sum}\limits_{i=0}^m\Theta_{g-1}(N,M,\boldsymbol\eta,\boldsymbol\tau,i)\Lambda_{g+1}(N,M,\boldsymbol\eta,\boldsymbol\tau,m-i),\quad 2\leq g<G\\ \end{array} \right.$$ $$\label{perfect:eq_8} \Lambda_g(N,M,\boldsymbol\eta,\boldsymbol\tau,m)=\left\{ \begin{array}{l} (\xi_g(N,M,\boldsymbol\eta,0))^{\tau_g},\quad m=0\\ \frac{1}{m\xi_g(N,M,\boldsymbol\eta,0)}\sum_{\ell=1}^{\min\left(m,\frac{\eta_G}{\eta_g} M-1\right)}((\tau_g+1)\ell-m)\\ \times\xi_g(N,M,\boldsymbol\eta,\ell)\Lambda_g(N,M,\boldsymbol\eta,\boldsymbol\tau,m-\ell),\quad 1\leq m<\tau_g(\frac{\eta_G}{\eta_g}M-1)\\ (\xi_g(N,M,\boldsymbol\eta,\frac{\eta_G}{\eta_g} M-1))^{\tau_g},\quad m=\tau_g(\frac{\eta_G}{\eta_g} M-1).\\ \end{array} \right.$$ The proof is provided in Appendix \[appenA\]. After obtaining the conditional CDF $F_{X_n|\mathcal{U}_n}$, let $C_P(M)$ denote the average sum rate and it can be computed using the following procedure. $$\begin{aligned} C_P(M)&=\frac{1}{N}\sum_{n=1}^N\mathbb{E}[\log_2(1+X_n)]\notag\\ &\mathop{=}\limits^{(a)}\mathbb{E}_{\mathcal{U}}\left[\int_0^\infty \log_2(1+\rho x)d(F_{X|\mathcal{U}}(x))\right]\notag\\ &\mathop{=}\limits^{(b)}\mathbb{E}_{\mathcal{U}}\left[\sum_{m=0}^{\Phi(M,\boldsymbol\eta,\boldsymbol\tau)}\Theta_{G-1}(N,M,\boldsymbol\eta,\boldsymbol\tau,m)\int_0^\infty \log_2(1+\rho x)d(F_Z(x))^{\sum_{g=1}^G\frac{N}{\eta_g}\tau_g-m}\right]\notag\\ \label{perfect:eq_9}&\mathop{=}\limits^{(c)}\sum_{\boldsymbol\tau\neq\mathbf{0}}\mathbb{P}(\mathcal{U}) \sum_{m=0}^{\Phi(M,\boldsymbol\eta,\boldsymbol\tau)}\Theta_{G-1}(N,M,\boldsymbol\eta,\boldsymbol\tau,m)\mathcal{I}_1\left(\rho,\sum_{g=1}^G\frac{N}{\eta_g}\tau_g-m\right),\end{aligned}$$ where $\rho\triangleq \frac{P}{\sigma_n^2}$ and $\mathbb{P}(\mathcal{U})$ is given by (\[perfect:eq\_4\]). (a) follows from the conditional expectation of $X_n|\mathcal{U}_n$ and the identically distributed property (let $X$ and $\mathcal{U}$ represent $X_n$ and $\mathcal{U}_n$ respectively), (b) follows from (\[perfect:eq\_6\]) in Theorem \[theorem\_1\], (c) follows from (\[perfect:eq\_4\]), and define $\mathcal{I}_1(a,b)\triangleq \int_0^\infty \log_2(1+ax)d(F_Z(x))^b$. $\mathcal{I}_1(a,b)$ is computed in Appendix \[appenA\] to be: $$\label{perfect:eq_10} \mathcal{I}_1(a,b)=\frac{b}{\ln2}\sum_{\ell=0}^{b-1}\binom{b-1}{\ell}\frac{(-1)^{\ell}}{\ell+1}\exp\left(\frac{\ell+1}{a}\right)E_1\left(\frac{\ell+1}{a}\right),$$ where $E_1(x)=\int_x^\infty \exp(-t)t^{-1}dt$ is the exponential integral function [@abramowitz72]. The average sum rate for the full feedback is a special case and is given by: $$\label{perfect:eq_11} C_P(M_F)=\int_0^\infty \log_2(1+\rho x)d(F_Z(x))^K=\mathcal{I}_1(\rho,K).$$ *Remark:* It is noteworthy to mention that the functional form of $C_P(M)$ in (\[perfect:eq\_9\]) consists of two main parts. The first part, which involves $\mathbb{P}(\mathcal{U})$ and $\Theta_{G-1}(\cdot,\cdot,\cdot,\cdot,\cdot)$, accounts for the randomness of the set of users who convey feedback as well as the scheduling policy. This part is inherent to the heterogeneous partial feedback strategy, and is independent of the system metric for evaluation, such as the average sum rate employed in this paper. The second part $\mathcal{I}_1(\cdot,\cdot)$ depends on statistical assumption of the underlying channel and the system metric, and it is impacted by partial feedback as well. Sum Rate Ratio and Best-M {#ratio} ------------------------- We now examine how to determine the smallest $M$ that results in almost the same performance, in terms of average sum rate, as the full feedback case. Applying the same technique as in [@choi08; @hur11], define $\gamma_P$ as the spectral efficiency ratio and the problem can be formulated as: $$\label{perfect:eq_12} \mathrm{Find\; the\; minimum\;} M^{\ast},\quad s.t.\; \gamma_P=\frac{C_P(M^{\ast})}{C_P(M_F)}\geq \gamma.$$ The above problem can be numerically solved by substituting the expressions for $C_P(M)$ and $C_P(M_F)$. In order to obtain a simpler and tractable relationship between $M$ and $K$ given $\boldsymbol\eta$, i.e., the tradeoff between the amount of partial feedback and the number of users given existing heterogeneity of channel statistics in frequency domain, an approximation is utilized similar to that in [@hur11], by observing that $\mathcal{I}_1(a,b)$ in (\[perfect:eq\_10\]) is slowly increasing in $b$ with fixed $a$ (This phenomenon is due to the saturation of multiuser diversity [@sharif05]). Observing $\sum_{m=0}^{\Phi(M,\boldsymbol\eta,\boldsymbol\tau)}\Theta_{G-1}(N,M,\boldsymbol\eta,\boldsymbol\tau,m)=1$ and employing the binomial theorem yields the approximation for the spectral efficiency ratio as: $$\label{perfect:eq_13} \gamma_P\simeq 1-\left(1-\frac{\eta_G M^{\ast}}{N}\right)^K.$$ From (\[perfect:eq\_12\]) and (\[perfect:eq\_13\]), the minimum required $M^{\ast}$ can be obtained as follows: $$\label{perfect:eq_14} M^{\ast}\geq \frac{N}{\eta_G}\left(1-\left(1-\gamma\right)^{\frac{1}{K}}\right).$$ *Remark:* It can be seen that $M^*$ depends on the system parameters ($N,K,\gamma$) as well on the largest subband size $\eta_G$. It is also a consequence of our heterogeneous partial feedback assumption to let users in cluster $\mathcal{K}_g$ convey back the $\frac{\eta_G}{\eta_g}M$ best CQI out of $\frac{N}{\eta_g}$ values. This results in the fact that obtaining feedback information from users belonging to different clusters have almost the same statistical influence on scheduling performance. Imperfect Feedback {#imperfect} ================== After analyzing the heterogeneous partial feedback design with perfect feedback, we turn to examine the impact of feedback imperfections in this section. We develop the imperfect feedback model due to channel estimation error and feedback delay in Section \[imperfect\_model\], and investigate the influence of imperfections on two different transmission strategies in Section \[fix\_rate\] and \[vari\_rate\]. Then we propose how to optimize the system performance to adapt to the imperfections in Section \[optimize\]. Imperfect Feedback Model {#imperfect_model} ------------------------ The imperfect feedback model is built upon the subband fading model for the perfect feedback case. To differentiate from the notation for the perfect feedback case and focus on the imperfect feedback model, the resource block index is dropped. Let $h_k$ denote the frequency domain channel transfer function of user $k$ (users in different clusters are not temporally distinguished to avoid notational overload). Due to channel estimation error, the user only has its estimated version $\hat{h}_k$, and the relationship between $h_k$ and $\hat{h}_k$ can be modeled as: $$\label{imperfect:eq_1} h_k=\hat{h}_k+w_k,$$ where $w_k\sim \mathcal{CN}(0,\sigma_{w_k}^2)$ is the channel estimation error. The channel of each resource block is assumed to be estimated independently, which yields the channel estimation errors $w_k$ i.i.d. across users and resource blocks, i.e., $w_k\sim \mathcal{CN}(0,\sigma_{w}^2)$. It is clear that the base station makes decision on scheduling and adaptive transmission depending on CQI, a function of $\hat{h}_k$. Thus this information can be outdated due to delay between the instant CQI is measured and the actual instant of use for data transmission to the selected user. Let $\tilde{h}_k$ be the actual channel transfer function and we employ a first-order Gaussian-Markov model [@kuhne08; @isukapalli10; @akoum10] to describe the time evolution and to capture the relationship with the delayed version as follows: $$\label{imperfect:eq_2} \tilde{h}_k = \alpha_k(\hat{h}_k+w_k)+\sqrt{1-\alpha_k^2}\varepsilon_k,$$ where $\varepsilon_k$ accounts for the innovation noise and is distributed as $\mathcal{CN}(0,1)$. The delay time between $\tilde{h}_k$ and $\hat{h}_k$ is not explicitly written for notational simplicity, and $\alpha_k\in [0,1]$ is used to model the correlation coefficient. Since the feedback delay is mainly caused by the periodic feedback interval and processing complexity [@akoum10], the innovation noise $\varepsilon_k$ are i.i.d. across users and a common $\alpha$ is assumed. Moreover, $w_k$ and $\varepsilon_k$ are assumed independent. Therefore, for notational simplicity, the user index $k$ in the aforementioned parameters is dropped and $\hat{Z}\triangleq|\hat{h}|^2$ is denoted as CQI. Let $\tilde{\chi},\chi$ and $\hat{\chi}$ represent: the actual CQI of the selected user for transmission, its outdated version, and its outdated estimate respectively ($\hat{\chi}$ corresponds to $X$ for the perfect feedback case in Section \[spectral\_efficiency\]). Notice that the PDF of the outdated estimate $\hat{\chi}$ depends on the heterogeneous feedback design and the scheduling strategy, whereas the conditional PDF of $\tilde{\chi}|\hat{\chi}$ only depends on $\alpha$ and $\sigma_w^2$. Employing the same method in [@ma05; @kuhne08], the conditional PDF is obtained as follows: $$\label{imperfect:eq_3} f_{\tilde{\chi}|\hat{\chi}}(x|\hat{\chi})=\frac{\alpha_w^2}{2}\exp\left(-\frac{\alpha_w^2x+\alpha_w^2\alpha^2\hat{\chi}}{2}\right)I_0(\alpha_w^2\alpha\sqrt{\hat{\chi}x}),$$ where $\alpha_w=\sqrt{\frac{2}{\alpha^2\sigma_w^2+1-\alpha^2}}$, and $I_0(\cdot)$ is the zeroth-order modified Bessel function of the first kind [@abramowitz72]. Since the feedback is imperfect, there are two types of issues that arise. The first is the choice of the incorrect user to serve. However, because of the i.i.d nature of the errors this does not compromise the fairness and also does not complicate the determination of the CDF. The second problem is that of outage because the rate adaptation is made by the base station based on the erroneous CQI. Because of the error in the CQI, the rate chosen may exceed the rate that the channel can support and so the base station has to take steps to mitigate this effect of outage. A conservative strategy will result in less outage but under utilization of the channel while an aggressive strategy will result in good utilization of the channel but only for a small fraction of the time. We now present two transmission strategies to address the outage issue. Fix Rate Strategy {#fix_rate} ----------------- In the fix rate conservative scenario, a system parameter $\beta_0$ is chosen for rate adaptation, and outage results under the following condition: $$\label{imperfect:eq_4} \mathrm{Declare\; outage\; if:}\quad \{\tilde{\chi}\leq\beta_0|\hat{\chi}\}.$$ The system average goodput is defined as the total average bps/Hz successfully transmitted [@lau08]. We derive the average goodput and average outage probability for a given choice of system parameter $\beta_0$ in the following procedure. Firstly the conditional outage probability is expressed as: $$\label{imperfect:eq_5} \mathbb{P}(\tilde{\chi}\leq\beta_0|\hat{\chi})=1-\mathcal{Q}_1(\alpha_w\alpha\sqrt{\hat{\chi}},\alpha_w\sqrt{\beta_0}),$$ where $\mathcal{Q}_1(a,b)=\int_b^{\infty}t\exp(-\frac{t^2+a^2}{2})I_0(at)dt$ is the first-order Marcum-Q function [@nuttall72]. Denote $R_0(\beta_0,M)$ as the average goodput for the heterogeneous partial feedback system, which is written according to definition: $$\label{imperfect:eq_6} R_0(\beta_0,M)=\mathbb{E}_{\mathcal{U}}\left[\mathbb{E}_{\hat{\chi}|\mathcal{U}}\left[\mathbb{P}(\tilde{\chi}\geq\beta_0|\hat{\chi})\log_2(1+\rho\beta_0)\right]\right].$$ Then, from (\[perfect:eq\_4\]) and (\[perfect:eq\_9\]), $R_0(\beta_0,M)$ can be computed as: $$\begin{aligned} &R_0(\beta_0,M)\notag\\ &=\mathbb{E}_{\mathcal{U}}\left[\sum_{m=0}^{\Phi(M,\boldsymbol\eta,\boldsymbol\tau)}\Theta_{G-1}(N,M,\boldsymbol\eta,\boldsymbol\tau,m)\int_0^\infty \mathcal{Q}_1(\alpha_w\alpha\sqrt{x},\alpha_w\sqrt{\beta_0})\log_2(1+\rho\beta_0)d(F_{\hat{Z}}(x))^{\sum_{g=1}^G\frac{N}{\eta_g}\tau_g-m}\right]\notag\\ \label{imperfect:eq_7}&=\sum_{\boldsymbol\tau\neq\mathbf{0}}\mathbb{P}(\mathcal{U})\sum_{m=0}^{\Phi(M,\boldsymbol\eta,\boldsymbol\tau)}\Theta_{G-1}(N,M,\boldsymbol\eta,\boldsymbol\tau,m)\log_2(1+\rho\beta_0)\mathcal{I}_2\left(\beta_0,\sum_{g=1}^G\frac{N}{\eta_g}\tau_g-m\right),\end{aligned}$$ where $\mathcal{I}_2(a,b)\triangleq\int_0^\infty \mathcal{Q}_1(\alpha_w\alpha\sqrt{x},\alpha_w\sqrt{a})d(F_{\hat{Z}}(x))^b$. $\mathcal{I}_2(a,b)$ is computed in Appendix \[appenB\] to be: $$\label{imperfect:eq_8} \mathcal{I}_2(a,b)=\frac{2b}{(1-\sigma_w^2)\ln2}\sum_{\ell=0}^{b-1}\binom{b-1}{\ell}(-1)^{\ell}\frac{1}{\zeta_{\ell}}\left(\exp(-\frac{\vartheta^2}{2})+\exp(-\frac{\zeta_{\ell}\vartheta^2}{2(\varpi^2+\zeta_{\ell})})(1-\exp(-\frac{\varpi^2\vartheta^2}{2(\varpi^2+\zeta_{\ell})}))\right),$$ where $\varpi=\alpha_w\alpha$, $\vartheta=\alpha_w\sqrt{a}$, $\zeta_{\ell}=\frac{2(\ell+1)}{1-\sigma_w^2}$. The average outage probability $P_0(\beta_0,M)$ for the heterogeneous partial feedback design can be directly calculated from definition and (\[imperfect:eq\_7\]) as follows: $$\begin{aligned} P_0(\beta_0,M)&=\mathbb{E}_{\mathcal{U}}\left[\mathbb{E}_{\hat{\chi}|\mathcal{U}}\left[\mathbb{P}(\tilde{\chi}\leq\beta_0|\hat{\chi})\right]\right]\notag\\ \label{imperfect:eq_9}&=\sum_{\boldsymbol\tau\neq\mathbf{0}}\mathbb{P}(\mathcal{U})\sum_{m=0}^{\Phi(M,\boldsymbol\eta,\boldsymbol\tau)}\Theta_{G-1}(N,M,\boldsymbol\eta,\boldsymbol\tau,m)\left(1-\mathcal{I}_2\left(\beta_0,\sum_{g=1}^G\frac{N}{\eta_g}\tau_g-m\right)\right).\end{aligned}$$ The average goodput and average outage probability for the full feedback scenario is a special case and is given by: $$\begin{aligned} R_0(\beta_0,M_F)&=\log_2(1+\rho\beta_0)\mathcal{I}_2(\beta_0,K),\notag\\ \label{imperfect:eq_10}P_0(\beta_0,M_F)&=1-\mathcal{I}_2(\beta_0,K).\end{aligned}$$ Variable Rate Strategy {#vari_rate} ---------------------- Instead of choosing a conservative system parameter to account for the fix rate scenario as in the previous subsection, we consider an approach we refer to as the variable rate strategy. In the variable rate scenario, a system parameter $\beta_1$ is chosen and outage results under the following condition: $$\label{imperfect:eq_11} \mathrm{Declare\; outage\; if:}\quad \{\tilde{\chi}\leq\beta_1\hat{\chi}|\hat{\chi}\},$$ where $\beta_1$ can be regarded as the backoff factor. The system average goodput and average outage probability can be derived utilizing the following procedure. Now under the variable rate scenario, the conditional outage probability is expressed as: $$\label{imperfect:eq_12} \mathbb{P}(\tilde{\chi}\leq\beta_1\hat{\chi}|\hat{\chi})=1-\mathcal{Q}_1(\alpha_w\alpha\sqrt{\hat{\chi}},\alpha_w\sqrt{\beta_1\hat{\chi}}).$$ Using the same method as (\[imperfect:eq\_6\]) and (\[imperfect:eq\_7\]), let $R_1(\beta_1,M)$ denote the average goodput for the variable rate scenario whose expression can be written as follows: $$\begin{aligned} \label{imperfect:eq_13} R_1(\beta_1,M)&=\mathbb{E}_{\mathcal{U}}\left[\mathbb{E}_{\hat{\chi}|\mathcal{U}}\left[\mathbb{P}(\tilde{\chi}\geq\beta_1\hat{\chi}|\hat{\chi})\log_2(1+\rho\beta_1\hat{\chi})\right]\right]\notag\\ &=\sum_{\boldsymbol\tau\neq\mathbf{0}}\mathbb{P}(\mathcal{U})\sum_{m=0}^{\Phi(M,\boldsymbol\eta,\boldsymbol\tau)}\Theta_{G-1}(N,M,\boldsymbol\eta,\boldsymbol\tau,m)\mathcal{I}_3\left(\beta_1,\sum_{g=1}^G\frac{N}{\eta_g}\tau_g-m\right),\end{aligned}$$ where $\mathcal{I}_3(a,b)\triangleq\int_0^\infty \mathcal{Q}_1(\alpha_w\alpha\sqrt{x},\alpha_w\sqrt{ax})\log_2(1+\rho ax)d(F_{\hat{Z}}(x))^b$. For the full feedback case, the average goodput is given by: $$\label{imperfect:eq_14} R_1(\beta_1,M_F)=\mathcal{I}_3(\beta_1,K).$$ Note that unlike $\mathcal{I}_2(a,b)$, $\mathcal{I}_3(a,b)$ can not be written in closed form. Therefore, bounding technique and suitable approximation are attractive to find closed form alternatives for $\mathcal{I}_3(a,b)$. The following proposition presents a valid closed form upper bound for $\mathcal{I}_3(a,b)$ in the low $\mathsf{SNR}$ regime. \[proposition\_1\] In the low $\mathsf{SNR}$ regime, $\mathcal{I}_3(a,b)$ can be efficiently upper bounded by: $$\begin{aligned} \mathcal{I}_3^{\mathrm{UB}}(a,b)=&\frac{4\rho ab}{(1-\sigma_w^2)\ln2}\sum_{\ell=0}^{b-1}(-1)^{\ell}\frac{1}{\zeta_{\ell}^2}\Bigg(1+\frac{\vartheta^2}{\varphi_{\ell}}\Bigg(\frac{\varpi^2}{\varphi_{\ell}}{}_2F_1\left(1,\frac{3}{2};2;\frac{4\varpi^2\vartheta^2}{\varphi_{\ell}^2}\right)-{}_2F_1\left(\frac{1}{2},1;1;\frac{4\varpi^2\vartheta^2}{\varphi_{\ell}^2}\right)\notag\\ \label{imperfect:eq_15}&+\frac{2\zeta_{\ell}}{\varphi_{\ell}}\Bigg(\frac{\varpi^2}{\varphi_{\ell}}{}_2F_1\left(\frac{3}{2},2;2;\frac{4\varpi^2\vartheta^2}{\varphi_{\ell}^2}\right)-\frac{1}{2}{}_2F_1\left(1,\frac{3}{2};1;\frac{4\varpi^2\vartheta^2}{\varphi_{\ell}^2}\right)\Bigg)\Bigg)\Bigg),\end{aligned}$$ where $\varpi=\alpha_w\alpha$, $\vartheta=\alpha_w\sqrt{a}$, $\zeta_{\ell}=\frac{2(\ell+1)}{1-\sigma_w^2}$, $\varphi_{\ell}=\varpi^2+\vartheta^2+\zeta_{\ell}$, and ${}_2F_1(\cdot,\cdot;\cdot;\cdot)$ is the Gaussian hypergeometric function [@abramowitz72]. The proof is provided in Appendix \[appenB\]. $\mathcal{I}_3^{\mathrm{UB}}(a,b)$ is valid and tight especially for the low $\mathsf{SNR}$ regime. In order to track $\mathcal{I}_3(a,b)$ over the whole $\mathsf{SNR}$ regimes, we propose the following approximation method by leveraging Jensen’s inequality [@boyd04]. Recall the definition of $\mathcal{I}_3(a,b)=\mathbb{E}[\mathcal{Q}_1(\alpha_w\alpha\sqrt{\check{\chi}},\alpha_w\sqrt{a\check{\chi}})\log_2(1+\rho a\check{\chi})]$, where the random variable $\check{\chi}$ is defined to have CDF $(F_{\hat{Z}}(x))^b$. Firstly, $\mathbb{E}[\check{\chi}]$ can be computed and is given by: $$\begin{aligned} \mathbb{E}[\check{\chi}]&=\int_0^\infty x\frac{b}{1-\sigma_w^2}\sum_{\ell=0}^{b-1}\binom{b-1}{\ell}(-1)^{\ell}\exp\left({-\frac{(\ell+1)x}{1-\sigma_w^2}}\right)dx\notag\\ \label{imperfect:eq_16}&=\frac{b}{1-\sigma_w^2}\sum_{\ell=0}^{b-1}\binom{b-1}{\ell}(-1)^{\ell}\left(\frac{1-\sigma_w^2}{\ell+1}\right)^2.\end{aligned}$$ Then plugging (\[imperfect:eq\_16\]) into $\mathcal{Q}_1(\alpha_w\alpha\sqrt{x},\alpha_w\sqrt{ax})\log_2(1+\rho ax)$ yields: $$\label{imperfect:eq_17} \mathcal{I}_3^{A}(a,b)=\mathcal{Q}_1\left(\alpha_w\alpha\sqrt{\mathbb{E}[\check{\chi}]},\alpha_w\sqrt{a\mathbb{E}[\check{\chi}]}\right)\log_2(1+\rho a\mathbb{E}[\check{\chi}]).$$ Note that $\mathcal{I}_3^{A}(a,b)$ would serve as an upper bound from Jensen’s inequality if the function of interest $\mathcal{Q}_1(\alpha_w\alpha\sqrt{x},\alpha_w\sqrt{ax})\log_2(1+\rho ax)$ was concave in $x$. Properties of this function such as monotonicity and concavity are of interest and lead to rigorous arguments in support of this bound. If outage does not occur, extensive analysis can be carried out due to the well known properties of the $\log(\cdot)$ function. However, the concavity (or log-concavity) of $\mathcal{Q}_1(\alpha_w\alpha\sqrt{x},\alpha_w\sqrt{\beta_1x})$ in $x$ (notice that $x$ appears in both entries of $\mathcal{Q}_1(\cdot,\cdot)$) still remains an important open problem [@yu11]. Our numerical evidence suggests that $\mathcal{Q}_1(\alpha_w\alpha\sqrt{x},\alpha_w\sqrt{\beta_1x})\log_2(1+\rho \beta_1x)$ is concave and monotonically increasing in $x$ for practical choices of $\beta_1$. For any given $\beta_1$ preserving the aforementioned property, Jensen’s inequality yields an upper bound, whose tightness is of interest and discussed in the following proposition. The word practical is used to exclude the situation when $\beta_1$ approaches its maximum $1$ which in turn enables $\mathcal{Q}_1(\cdot,\cdot)$ to dominate the goodput to incur extreme outage. This makes intuitive sense according to the definition of average goodput. \[proposition\_2\] Let $\{\check{\chi}_b\}$ be the family of positive i.i.d. random variables. If $\mathcal{Q}_1(\alpha_w\alpha\sqrt{x},\alpha_w\sqrt{\beta_1x})\log_2(1+\rho \beta_1x)$ is concave and monotonically increasing in $x$ for any given $\beta_1$, then the Jensen bound is asymptotically tight, i.e., as $b\rightarrow\infty$, $$\label{imperfect:eq_18} \frac{\mathbb{E}[\mathcal{Q}_1(\alpha_w\alpha\sqrt{\check{\chi}_b},\alpha_w\sqrt{\beta_1\check{\chi}_b})\log_2(1+\rho \beta_1\check{\chi}_b)]}{\mathcal{Q}_1(\alpha_w\alpha\sqrt{\mathbb{E}[\check{\chi}_b]},\alpha_w\sqrt{\beta_1\mathbb{E}[\check{\chi}_b]})\log_2(1+\rho \beta_1\mathbb{E}[\check{\chi}_b])}\rightarrow 1.$$ The proof is provided in Appendix \[appenB\]. Nonetheless, when the aforementioned property is not preserved (e.g., $\beta_1$ approaches $1$), Jensen’s inequality does not hold but the expression has been experimentally found to be a good approximation and so can still be used. Therefore, (\[imperfect:eq\_17\]) is denoted as Jensen approximation. We conduct a numerical study and demonstrate the tightness of Jensen approximation in Fig. \[fig\_3\]. It is observed that the approximation method is very tight for moderate (even small) number of users and for all values of $\beta_1\in[0,1]$, which shows its potential in accurately tracking the performance of average goodput. ![Calculating the average goodput from numerical evaluation ($M=2,4,16$) and Jensen approximation ($M=16$) for the variable rate scenario under different $\rho$. ($N=64$, $\boldsymbol\eta=(1,4)$, $K_1=K_2=K/2=10$, $\alpha=0.98$, $\sigma_w^2=0.01$, $\rho=10$ dB, and $20$ dB)[]{data-label="fig_3"}](Figure/fig3.eps){width="0.68\linewidth"} Now we calculate the average outage probability. Since it does not involve the $\log(\cdot)$ function, it can be computed into closed form as follows: $$\begin{aligned} P_1(\beta_1,M)&=\mathbb{E}_{\mathcal{U}}\left[\mathbb{E}_{\hat{\chi}|\mathcal{U}}\left[\mathbb{P}(\tilde{\chi}\leq\beta_1\hat{\chi}|\hat{\chi})\right]\right]\notag\\ \label{imperfect:eq_19}&=\sum_{\boldsymbol\tau\neq\mathbf{0}}\mathbb{P}(\mathcal{U})\sum_{m=0}^{\Phi(M,\boldsymbol\eta,\boldsymbol\tau)}\Theta_{G-1}(N,M,\boldsymbol\eta,\boldsymbol\tau,m)\left(1-\mathcal{I}_4\left(\beta_1,\sum_{g=1}^G\frac{N}{\eta_g}\tau_g-m\right)\right),\end{aligned}$$ where $$\begin{aligned} \mathcal{I}_4(a,b)&\triangleq\int_0^\infty \mathcal{Q}_1(\alpha_w\alpha\sqrt{x},\alpha_w\sqrt{ax})d(F_{\hat{Z}}(x))^b\notag\\ &\mathop{=}\limits^{(a)}\frac{2b}{(1-\sigma_w^2)}\sum_{\ell=0}^{b-1}\binom{b-1}{\ell}(-1)^{\ell}\int_0^\infty\mathcal{Q}_1(\alpha_w\alpha x,\alpha_w\sqrt{a}x)\exp\left(-\frac{(\ell+1)x^2}{1-\sigma_w^2}\right)xdx\notag\\ \label{imperfect:eq_20}&\mathop{=}\limits^{(b)}\frac{b}{(1-\sigma_w^2)}\sum_{\ell=0}^{b-1}\binom{b-1}{\ell}(-1)^{\ell}\frac{1}{\zeta_{\ell}}\left(1+\frac{\psi_{\ell}}{\varsigma_{\ell}}\right),\end{aligned}$$ $\varpi=\alpha_w\alpha$, $\vartheta=\alpha_w\sqrt{a}$, $\zeta_{\ell}=\frac{2(\ell+1)}{1-\sigma_w^2}$, $\varphi_{\ell}=\varpi^2+\vartheta^2+\zeta_{\ell}$, $\psi_{\ell}=\varpi^2-\vartheta^2+\zeta_{\ell}$, $\varsigma_{\ell}=\sqrt{\varphi_{\ell}^2-4\varpi^2\vartheta^2}$. (a) follows from change of variables; (b) follows from applying [@simon02 B.48]. In the case of full feedback, the average outage probability $P_1(\beta_1,M_F)$ becomes: $$\label{imperfect:eq_21} P_1(\beta_1,M_F)=1-\mathcal{I}_4(\beta_1,K).$$ Optimization and Adaptation to Imperfections {#optimize} -------------------------------------------- We have obtained the relationship between the system parameter ($\beta_0$ or $\beta_1$) and the system average goodput, and we now aim to maximize the average goodput by adapting the system parameters. Consider the optimization of $R_1(\beta_1,M)$ to obtain the optimal backoff factor $\beta_1^{\ast}$. It is observed from (\[imperfect:eq\_13\]) that directly optimizing $R_1(\beta_1,M)$ is tedious, and a near-optimal method is now proposed to obtain $\beta_1^{\ast}$. This method is inspired by the results in Section \[ratio\], which show that the minimum required $M^{\ast}$ can be chosen to achieve almost the same performance as a system with full feedback. Thus an optimal $\beta_1^{\ast}$ for the full feedback scenario can be optimized first, and then $M^{\ast}$ is obtained to “match" the system performance. Looking again at Fig. \[fig\_3\] with emphasis on different number $M$ of partial feedback, as $M$ gets larger, the optimal $\beta_1$ converges to the full feedback case. In this example, $M^{\ast}=4$ is adequate to match the system performance. It is noteworthy to mention that this adaptation philosophy can be applied to partial feedback systems wherein system parameters are optimized according to full feedback assumption first and minimum required partial feedback is chosen subsequently. Note that a closed form approximation has been obtained to track $R_1(\beta_1,M_F)$ in Section \[vari\_rate\], which is denoted as $R_1^A(\beta_1,M_F)\triangleq\mathcal{I}_3^{\mathrm{A}}(\beta_1,K)$. The following proposition demonstrates the optimal property of $\beta_1$ when optimizing $R_1^A(\beta_1,M_F)$. \[proposition\_3\] There exists a unique global optimal $\beta_1$ that maximizes $R_1^A(\beta_1,M_F)$. The proof is provided in Appendix \[appenB\]. From the above analysis, the optimization strategy can be described as: $$\label{imperfect:eq_22} \beta_1^{\ast}=\arg\mathop{\max}\limits_{0\leq\beta_1\leq1} R_1^A(\beta_1,M_F)\simeq \arg\mathop{\max}\limits_{0\leq\beta_1\leq1} R_1(\beta_1,M_F).$$ Since it is proved in Proposition \[proposition\_3\] that $R_1^A(\beta_1,M_F)$ is quasiconcave [@boyd04] in $\beta_1$, numerical approach such as Newton-Raphson method can be applied to obtain $\beta_1^{\ast}$. As discussed before, once $\beta_1^*$ is found, the minimum required $M^*$ can be obtained by solving (\[perfect:eq\_12\]) or relying on (\[perfect:eq\_14\]). The same strategy can be carried over to the optimization of $\beta_0$, which is presented as follows: $$\label{imperfect:eq_23} \beta_0^{\ast}=\arg\mathop{\max}\limits_{\beta_0} R_0(\beta_0,M_F).$$ The impact of imperfections on system parameter adaptation, and the comparison between the fixed rate and variable rate strategies will be examined through simulations in Section \[numerical\_imperfect\]. Numerical Results {#numerical} ================= In this section, we conduct a numerical study to verify the results developed and to draw some insight. ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![The required minimum $M$ for heterogeneous perfect feedback design: (a) Comparison of the required minimum $M$ between numerically solving (\[perfect:eq\_12\]) and using approximation (\[perfect:eq\_14\]) under different $\gamma$ with respect to the number of users; ($N=64$, $\boldsymbol\eta=(1,4)$, $K_1=K_2=K/2$, $\rho=10$ dB) (b) Computing the required minimum $M$ with respect to different number of users when varying the ratio of the number of users in cluster $\mathcal{K}_1$. ($N=64$, $\boldsymbol\eta=(1,4)$, $\rho=10$ dB, $\gamma=0.99$)[]{data-label="fig_4"}](Figure/fig4_a.eps "fig:"){width="0.5\linewidth"} ![The required minimum $M$ for heterogeneous perfect feedback design: (a) Comparison of the required minimum $M$ between numerically solving (\[perfect:eq\_12\]) and using approximation (\[perfect:eq\_14\]) under different $\gamma$ with respect to the number of users; ($N=64$, $\boldsymbol\eta=(1,4)$, $K_1=K_2=K/2$, $\rho=10$ dB) (b) Computing the required minimum $M$ with respect to different number of users when varying the ratio of the number of users in cluster $\mathcal{K}_1$. ($N=64$, $\boldsymbol\eta=(1,4)$, $\rho=10$ dB, $\gamma=0.99$)[]{data-label="fig_4"}](Figure/fig4_b.eps "fig:"){width="0.5\linewidth"} ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Perfect Feedback Scenario {#numerical_perfect} ------------------------- The number of resource blocks $N$ is assumed to be $64$ for simulations throughout this section. We first consider a $2$-cluster system. Fig. \[fig\_4\] (a) plots the minimum required $M$ obtained by directly solving (\[perfect:eq\_12\]) and alternatively by the approximation (\[perfect:eq\_14\]) for two thresholds: $\gamma=0.99$ and $0.9$. Note that the result from (\[perfect:eq\_14\]) is rounded with the ceiling function since the required $M$ is an integer. The other simulation parameters are $\boldsymbol\eta=(1,4)$ (i.e., $M_F=16$), and $\rho=10$ dB. It is observed that the results from the approximate expression matches quite well with the exact computation. The question of whether the required $M$ is sensitive to the partition of users in the system is examined in Fig. \[fig\_4\] (b) wherein the ratio of the number of users in cluster $\mathcal{K}_1$ is changed and the minimum required $M$ is depicted for different total number of users with threshold $\gamma=0.99$. Interestingly, the result turns out to be “uniform". As discussed in Section \[perfect\], it is due to the heterogenous feedback design assumption to let users in cluster $\mathcal{K}_1$ consume $\frac{\eta_G}{\eta_1} M$ ($4M$ in this simulation) feedback which results in the fact that obtaining feedback information from users belonging to different clusters have almost the same influence on scheduling performance. Therefore, the representative simulation setup $K_g=K/G$ can be employed when the system performance metric is investigated with respect to the total number of users. ![Comparison of the average sum rate for a $4$-cluster system under different feedback strategies with respect to the number of users ($N=64$, $\boldsymbol\eta=(1,2,4,8)$, $K_1=K_2=K_3=K_4=K/4$, $M=2,4$, $\rho=10$ dB)[]{data-label="fig_5"}](Figure/fig5.eps){width="0.7\linewidth"} We now consider a $4$-cluster system with subband size vector $\boldsymbol\eta=(1,2,4,8)$ (i.e., $M_F=8$). Fig. \[fig\_5\] demonstrates the benefit of using heterogeneous feedback design. One of the competing strategies is also heterogeneous, but treats users from each cluster separately. In particular, the system firstly clusters the users based on their channel statistics, and then serves the clusters one by one requiring feedback only from the served cluster of users. In this way, the feedback amount is varying over time depending on the partition of users. This strategy is denoted as *separate* heterogeneous feedback compared to our *joint* heterogeneous feedback design. The other competing strategies are homogeneous without taking advantage of the channel statistics of different users. To maintain at least the same feedback amount for fair comparison, each user in the homogeneous case is assumed to feed back $\lceil\sum_{g=1}^G\frac{\eta_G}{\eta_g}\frac{M}{G}\rceil$ CQI values. Two subband sizes are assumed for the homogeneous feedback. It is clear that for the homogeneous case, users in cluster $\mathcal{K}_1$ have more independent feedback while users in cluster $\mathcal{K}_4$ suffer from redundant feedback. The average sum rate for two different values of $M$ are shown in Fig. \[fig\_5\]. The separate heterogenous feedback is observed to have the worst performance from a sum rate perspective because it does not fully exploit multiuser diversity, but it consumes the least feedback. Our joint heterogenous feedback design is shown to perform much better than the two homogeneous strategies for the $4$-cluster system. It is due to the fact that by considering the existing heterogeneity among users, the proposed heterogeneous design can make the best use of the degrees of freedom in the frequency domain in order to enhance the system performance as well as reduce feedback needs. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ![Comparison of fixed rate and variable rate strategies under normalized parameter $\beta$ ($\beta=\beta_1=\beta_0/10$) for different number of users $K$ ($N=64$, $\alpha=0.98$, $\sigma_w^2=0.01$, $\rho=10$ dB): (a) Comparison of average goodput; (b) Comparison of average outage probability.[]{data-label="fig_6"}](Figure/fig6_a.eps "fig:"){width="0.5\linewidth"} ![Comparison of fixed rate and variable rate strategies under normalized parameter $\beta$ ($\beta=\beta_1=\beta_0/10$) for different number of users $K$ ($N=64$, $\alpha=0.98$, $\sigma_w^2=0.01$, $\rho=10$ dB): (a) Comparison of average goodput; (b) Comparison of average outage probability.[]{data-label="fig_6"}](Figure/fig6_b.eps "fig:"){width="0.5\linewidth"} ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![The effect of channel estimation error ($\sigma_w^2$) and feedback delay ($\alpha$) on the optimal value of $\beta_0$ and $\beta_1$ ($\rho=10$ dB, $K=10$, $\sigma_w^2=0:0.005:0.1$, $\alpha=0.9:0.005:0.99$): (a) The optimal fix rate parameter $\beta_0$ with respect to $\sigma_w^2$ and $\alpha$; (b) The optimal variable rate parameter $\beta_1$ with respect to $\sigma_w^2$ and $\alpha$.[]{data-label="fig_7"}](Figure/fig7_a.eps "fig:"){width="0.5\linewidth"} ![The effect of channel estimation error ($\sigma_w^2$) and feedback delay ($\alpha$) on the optimal value of $\beta_0$ and $\beta_1$ ($\rho=10$ dB, $K=10$, $\sigma_w^2=0:0.005:0.1$, $\alpha=0.9:0.005:0.99$): (a) The optimal fix rate parameter $\beta_0$ with respect to $\sigma_w^2$ and $\alpha$; (b) The optimal variable rate parameter $\beta_1$ with respect to $\sigma_w^2$ and $\alpha$.[]{data-label="fig_7"}](Figure/fig7_b.eps "fig:"){width="0.5\linewidth"} ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Imperfect Feedback Scenario {#numerical_imperfect} --------------------------- Fig. \[fig\_6\] exhibits the comparison between the fix rate and variable rate outage scenarios as well as the effect of the number of users on the optimization of $\beta_0$ and $\beta_1$. In order to show the system performance of the two scenarios in one figure, a normalized parameter $\beta$ is defined. While examining the variable rate plots $\beta=\beta_1,$ and when considering the fixed rate plots $\beta=\beta_0/10$. The system parameters are: $\alpha=0.98$, $\sigma_w^2=0.01$, and $\rho=10$ dB. It can be seen that for both scenarios, larger number of users $K$ yields better system performance, i.e., higher average goodput and lower average outage probability. This is a consequence of increased multiuser diversity gain to combat the imperfections in the feedback system. ![Comparison of the average goodput for a $4$-cluster system with fix rate and variable rate strategies using optimized $\beta_0^*$ and $\beta_1^*$. The average goodput is calculated using the best-M partial feedback scheme when the minimum required $M^*$ is computed after obtaining $\beta_0^*$ or $\beta_1^*$. ($N=64$, $\boldsymbol\eta=(1,2,4,8)$, $K_1=K_2=K_3=K_4=K/4$, $\alpha=0.98$, $\sigma_w^2=0.01$, $\rho=10$ dB)[]{data-label="fig_8"}](Figure/fig8.eps){width="0.6\linewidth"} Fig. \[fig\_7\] illustrates the effect of channel estimation error $\sigma_w^2$ and feedback delay $\alpha$ on the optimal value of $\beta_0$ and $\beta_1$. Here $\sigma_w^2$ is varied from $0$ to $0.1$, and $\alpha$ is varied from $0.9$ to $0.99$ in steps of $0.005$. It can be observed from the changing profiles that both the optimal values of $\beta_0$ and $\beta_1$ get smaller as the imperfections become worse. Therefore, the system should adjust the system parameters to adapt to the encountered imperfections. Now we consider the adaptation of system parameters ($\beta_0$ or $\beta_1$) and partial feedback in a $4$-cluster heterogeneous feedback system. The system parameters are: $\boldsymbol\eta=(1,2,4,8)$, $\alpha=0.98$, $\sigma_w^2=0.01$, and $\rho=10$ dB. For both transmission strategies and for a given number of users $K$, the optimal value of $\beta_0^*$ or $\beta_1^*$ is first optimized according to the full feedback case discussed in Section \[optimize\]. Then, a minimum required $M^*$ is obtained by matching the system performance to that in the full feedback case. Fig. \[fig\_8\] demonstrates the average goodput for both transmission strategies with $M^*$ and $\beta_0^*$ (or $\beta_1^*$). We observe that there is almost a constant performance gain for the variable rate strategy compared with the fix rate one. This is due to the fact that for the variable rate scenario, the system is adapting the transmission parameters conditioned on the past memory even if it is the outdated one. If the channel estimation error and feedback delay are not severe, the imperfections can be compensated by multiplying with the backoff factor and relying on the past feedback. Conclusion ========== In this paper, we propose and analyze a heterogeneous feedback design adapting the feedback resource according to users’ frequency domain channel statistics. Under the general correlated channel model, we demonstrate the gain by achieving the potential match among coherence bandwidth, subband size and partial feedback. To facilitate statistical analysis, we employ the subband fading model for the multi-cluster heterogeneous feedback system. We derive a closed form expression of the average sum rate under perfect partial feedback assumption, and provide a method to obtain the minimum heterogeneous partial feedback required to obtain performance comparable to a scheme using full feedback. We also analyze the effect of imperfections on the heterogeneous partial feedback system. We obtain a closed form expression for the average goodput of the fix rate scenario, and utilize a bounding technique and tight approximation to track the performance of the variable rate scenario. Methods adapting the system parameters to maximize the average system goodput are proposed. The heterogeneous feedback design is shown to outperform the homogeneous one with the same feedback resource. With imperfections, the system adjusting the transmission strategy and the amount of partial feedback is shown to yield better performance. The developed analysis provides a theoretical reference to understand the approximate behavior of the proposed heterogeneous feedback system and its interplay with practical imperfections. Dealing with the general channel correlation and the corresponding nonlinear nature of the CQI are interesting directions for the heterogeneous feedback system. {#appenA} *Proof (sketch) of Lemma \[lemma\_1\]:* The methodology is an extension of the work in [@hur11] which deals with the homogeneous feedback case with one cluster of users and one specific subband size. Let $F_{Y_{k}^{(g)}}$ denote the CDF of $Y_{k}^{(g)}\triangleq Y_{k,r}^{(g)}$. Substituting the subband size $\frac{N}{\eta_g}$ and the number of reported CQI $\frac{\eta_G}{\eta_g} M$ for user $k$ in cluster $\mathcal{K}_g$ makes $F_{Y_{k}^{(g)}}$ satisfy (\[perfect:eq\_1\]). It can be shown that $F_{\tilde{Y}_{k}^{(g)}}(x)=\mathbb{P}(\tilde{Y}_{k,n}^{(g)}\leq x)=\mathbb{P}(Y_{k,\lceil\frac{n}{\eta_g}\rceil}^{(g)}\leq x)=F_{Y_{k}^{(g)}}(x)$, which concludes the proof. *Proof of Theorem \[theorem\_1\]:* Substituting the expressions of $F_{\tilde{Y}_{k}^{(g)}}$ from Lemma \[lemma\_1\] and combining (\[perfect:eq\_5\]) yield: $$\label{appendix:eq_1} F_{X_n|\mathcal{U}_n}(x)=(F_Z(x))^{\sum_{g=1}^G\frac{N}{\eta_g}\tau_g}\prod_{g=1}^G\left(\sum_{m=0}^{\frac{\eta_G}{\eta_g}M-1}\frac{\xi_g(N,M,\boldsymbol\eta,0)}{(F_Z(x))^m}\right)^{\tau_g}.$$ Applying [@gradshteyn07 0.314] to a finite-order power series in (\[appendix:eq\_1\]), $\left(\sum_{m=0}^{\frac{\eta_G}{\eta_g}M-1}\frac{\xi_g(N,M,\boldsymbol\eta,0)}{(F_Z(x))^m}\right)^{\tau_g}$ can be expressed as $\sum_{m=0}^{\tau_g(\frac{\eta_G}{\eta_g}M-1)}\frac{\Lambda_g(N,M,\boldsymbol\eta,\boldsymbol\tau,m)}{(F_Z(x))^m}$, where the expression for $\Lambda_g(N,M,\boldsymbol\eta,\boldsymbol\tau,m)$ is described in Theorem \[theorem\_1\]. Note that the coefficients of $\frac{1}{(F_Z(x))^m}$ can be computed in a recursive manner. Then we employ [@gradshteyn07 0.316] for the multiplication of power series. For $g=1$, $\Theta_1(N,M,\boldsymbol\eta,\boldsymbol\tau,m)$ can be calculated from $\Lambda_1(N,M,\boldsymbol\eta,\boldsymbol\tau,m)$ and $\Lambda_2(N,M,\boldsymbol\eta,\boldsymbol\tau,m)$ as: $$\Theta_1(N,M,\boldsymbol\eta,\boldsymbol\tau,m)=\mathop{\sum}\limits_{i=0}^m\Lambda_1(N,M,\boldsymbol\eta,\boldsymbol\tau,i)\Lambda_2(N,M,\boldsymbol\eta,\boldsymbol\tau,m-i).$$ For $2\leq g<G$, $\Theta_{g}(N,M,\boldsymbol\eta,\boldsymbol\tau,m)$ can be computed from $\Theta_{g-1}(N,M,\boldsymbol\eta,\boldsymbol\tau,m)$ and $\Lambda_{g+1}(N,M,\boldsymbol\eta,\boldsymbol\tau,m)$ in the following manner: $$\Theta_{g}(N,M,\boldsymbol\eta,\boldsymbol\tau,m)=\mathop{\sum}\limits_{i=0}^m\Theta_{g-1}(N,M,\boldsymbol\eta,\boldsymbol\tau,i)\Lambda_{g+1}(N,M,\boldsymbol\eta,\boldsymbol\tau,m-i).$$ This concludes the proof. *Derivation of $\mathcal{I}_1(a,b)$:* From the definition of $Z$, $F_Z(x)=1-\exp(-x)$ and $f_Z(x)=\exp(-x)$. Thus $d(F_Z(x))^b=b(F_Z(x))^{b-1}f_Z(x)dx=b\sum_{\ell=0}^{b-1}\binom{b-1}{\ell}(-1)^{\ell}\exp(-(\ell+1)x)dx$, where the last equality follows from the binomial theorem. Therefore, $\int_0^\infty \log_2(1+ax)d(F_Z(x))^b=\frac{b}{\ln2}\sum_{\ell=0}^{b-1}\binom{b-1}{\ell}(-1)^{\ell}\int_0^\infty\ln(1+ax)\exp(-(\ell+1)x)dx$. Applying [@gradshteyn07 4.337.2] yields (\[perfect:eq\_10\]). {#appenB} *Derivation of $\mathcal{I}_2(a,b)$:* From the definition of $\hat{Z}$, it can be shown that $F_{\hat{Z}}(x)=1-\exp(-\frac{1}{1-\sigma_w^2}x)$ and $f_{\hat{Z}}(x)=\frac{1}{1-\sigma_w^2}\exp(-\frac{1}{1-\sigma_w^2}x)$. Then $\mathcal{I}_2(a,b)$ can be calculated as: $$\begin{aligned} \mathcal{I}_2(a,b)&\mathop{=}\limits^{(a)}\frac{2b}{(1-\sigma_w^2)\ln2}\sum_{\ell=0}^{b-1}\binom{b-1}{\ell}(-1)^{\ell}\int_0^\infty\mathcal{Q}_1(\alpha_w\alpha x,\alpha_w\sqrt{a})\exp\left(-\frac{(\ell+1)x^2}{1-\sigma_w^2}\right)xdx\notag\\ &\mathop{=}\limits^{(b)}\frac{2b}{(1-\sigma_w^2)\ln2}\sum_{\ell=0}^{b-1}\binom{b-1}{\ell}(-1)^{\ell}\frac{1}{\zeta_{\ell}}\left(\mathcal{Q}_1(0,\vartheta)+\exp(-\frac{\zeta_{\ell}\vartheta^2}{2(\varpi^2+\zeta_{\ell})})(1-\mathcal{Q}_1(0,\frac{\varpi\vartheta}{\sqrt{\varpi^2+\zeta_{\ell}}}))\right)\notag\\ \label{appendix:eq_2}&\mathop{=}\limits^{(c)}\frac{2b}{(1-\sigma_w^2)\ln2}\sum_{\ell=0}^{b-1}\binom{b-1}{\ell}(-1)^{\ell}\frac{1}{\zeta_{\ell}}\left(\exp(-\frac{\vartheta^2}{2})+\exp(-\frac{\zeta_{\ell}\vartheta^2}{2(\varpi^2+\zeta_{\ell})})(1-\exp(-\frac{\varpi^2\vartheta^2}{2(\varpi^2+\zeta_{\ell})}))\right),\end{aligned}$$ where $\varpi=\alpha_w\alpha$, $\vartheta=\alpha_w\sqrt{a}$, $\zeta_{\ell}=\frac{2(\ell+1)}{1-\sigma_w^2}$. (a) is obtained by substituting the expression of $d(F_{\hat{Z}}(x))^b$ and using change of variables; (b) follows from applying [@simon02 B.18]; (c) follows from using the fact that $\mathcal{Q}_1(0,\vartheta)=\exp(-\frac{\vartheta^2}{2})$. *Proof of Proposition \[proposition\_1\]:* $$\begin{aligned} \mathcal{I}_3(a,b)&\mathop{=}\limits^{(a)}\frac{b}{(1-\sigma_w^2)\ln2}\sum_{\ell=0}^{b-1}\binom{b-1}{\ell}(-1)^{\ell}\int_0^\infty\mathcal{Q}_1(\alpha_w\alpha \sqrt{x},\alpha_w\sqrt{ax})\ln(1+\rho ax)\exp\left(-\frac{(\ell+1)x}{1-\sigma_w^2}\right)dx\notag\\ &\mathop{<}\limits^{(b)}\frac{2\rho ab}{(1-\sigma_w^2)\ln2}\sum_{\ell=0}^{b-1}\binom{b-1}{\ell}(-1)^{\ell}\int_0^\infty\mathcal{Q}_1(\alpha_w\alpha x,\alpha_w\sqrt{a}x)\exp\left(-\frac{(\ell+1)x^2}{1-\sigma_w^2}\right)x^3dx\notag\\ &\mathop{=}\limits^{(c)}\frac{4\rho ab}{(1-\sigma_w^2)\ln2}\sum_{\ell=0}^{b-1}(-1)^{\ell}\frac{1}{\zeta_{\ell}^2}\Bigg(1+\frac{\vartheta^2}{\varphi_{\ell}}\Bigg(\frac{\varpi^2}{\varphi_{\ell}}{}_2F_1\left(1,\frac{3}{2};2;\frac{4\varpi^2\vartheta^2}{\varphi_{\ell}^2}\right)-{}_2F_1\left(\frac{1}{2},1;1;\frac{4\varpi^2\vartheta^2}{\varphi_{\ell}^2}\right)\notag\\ \label{appendix:eq_3}&\quad+\frac{2\zeta_{\ell}}{\varphi_{\ell}}\Bigg(\frac{\varpi^2}{\varphi_{\ell}}{}_2F_1\left(\frac{3}{2},2;2;\frac{4\varpi^2\vartheta^2}{\varphi_{\ell}^2}\right)-\frac{1}{2}{}_2F_1\left(1,\frac{3}{2};1;\frac{4\varpi^2\vartheta^2}{\varphi_{\ell}^2}\right)\Bigg)\Bigg)\Bigg),\end{aligned}$$ where $\varpi=\alpha_w\alpha$, $\vartheta=\alpha_w\sqrt{a}$, $\zeta_{\ell}=\frac{2(\ell+1)}{1-\sigma_w^2}$, $\varphi_{\ell}=\varpi^2+\vartheta^2+\zeta_{\ell}$, and ${}_2F_1(\cdot,\cdot;\cdot;\cdot)$ is the Gaussian hypergeometric function [@abramowitz72]. (a) is obtained by substituting the expression of $d(F_{\hat{Z}}(x))^b$; (b) follows from the fact that when $\rho\ll1$, $\rho ax$ is a tight upper bound for $\ln(1+\rho ax)$; note that change of variables are used; (c) follows from applying [@simon02 B.60]. *Proof (sketch) of Proposition \[proposition\_2\]:* Define $s(\check{\chi}_b)\triangleq\mathcal{Q}_1(\alpha_w\alpha\sqrt{\check{\chi}_b},\alpha_w\sqrt{\beta_1\check{\chi}_b})\log_2(1+\rho \beta_1\check{\chi}_b)]$. Firstly it must be shown that $\frac{s(\check{\chi}_b)}{s(\mathbb{E}[\check{\chi}_b])}$ converges to $1$ in probability. For $\forall\epsilon>0$, it is now shown that: $$\begin{aligned} \mathbb{P}\left(\left|\frac{s(\check{\chi}_b)}{s(\mathbb{E}[\check{\chi}_b])}-1\right|\geq\epsilon\right)=&\mathbb{P}\left(\left|\frac{s(\check{\chi}_b)-s(\mathbb{E}[\check{\chi}_b])}{s(\mathbb{E}[\check{\chi}_b])}\right|\geq\epsilon\right)\notag\\ \label{appendix:eq_6}&\mathop{\leq}\limits^{(a)}\mathbb{P}\left(\frac{s(|\check{\chi}_b-\mathbb{E}[\check{\chi}_b]|)}{s(\mathbb{E}[\check{\chi}_b])}\geq\epsilon\right)\mathop{\rightarrow}\limits^{(b)}0,\end{aligned}$$ where (a) follows from the concave and monotonically increasing property of $s(\cdot)$: $|s(x)-s(y)|<s(|x-y|)$; (b) follows from the asymptotic scaling rate of $\mathbb{E}[\check{\chi}_b]$ and $|\check{\chi}_b-\mathbb{E}[\check{\chi}_b]|$, and the utilization of the Chebyshev’s inequality. From extreme value theory and asymptotic analysis of order statistics [@david03; @sharif05], it is known that the tail behavior of $\check{\chi}_b$ converges to type $3$ Gumbel distribution, which enables $\mathbb{E}[\check{\chi}_b]$ to scale as $\log b$ and $|\check{\chi}_b-\mathbb{E}[\check{\chi}_b]|$ to scale as $\log\log b$. Then a method similar to that in [@sanayei07] can be employed to prove the uniformly integrable property [@billingsley95] of $\frac{s(\check{\chi}_b)}{s(\mathbb{E}[\check{\chi}_b])}$. By combining the above property along with the convergence in probability leads to convergence in the mean [@billingsley95], which concludes the proof. *Proof of Proposition \[proposition\_3\]:* It must be shown that $\mathcal{I}_3^{\mathrm{A}}(\beta_1,K)=\mathcal{Q}_1\left(\alpha_w\alpha\sqrt{\mathbb{E}[\check{\chi}]},\alpha_w\sqrt{\beta_1\mathbb{E}[\check{\chi}]}\right)\log_2(1+\rho \beta_1\mathbb{E}[\check{\chi}])$ is strictly quasiconcave in $\beta_1$. This property can be proved by log-concavity [@boyd04]. It is shown in [@finner97; @yu11] that $\mathcal{Q}_1(\sqrt{a},\sqrt{b})$ is log-concave in $b\in[0,\infty)$ for $a\geq 0$. Also, $\log(1+b)$ is concave thus log-concave in $b\in[0,\infty)$. Since log-concavity is maintained in multiplication, $\mathcal{Q}_1(\sqrt{a},\sqrt{b})\log(1+b)$ is log-concave in $b\in[0,\infty)$. From the definition of $\mathcal{I}_3^{\mathrm{A}}(\beta_1,K)$, it is now proved to be log-concave in $\beta_1\in[0,\infty)$ since $\mathbb{E}[\check{\chi}]$ is irrelevant to $\beta_1$. Therefore, it is quasiconcave in $\beta_1\in[0,\infty)$ because log-concave functions are also quasiconcave. In addition, it is clear that $\mathop{\lim}\limits_{\beta_1\rightarrow 0}\mathcal{I}_3^{\mathrm{A}}(\beta_1,K)=0$. Also, it is now shown that: $$\begin{aligned} 0&\leq \mathop{\lim}\limits_{\beta_1\rightarrow\infty}\mathcal{I}_3^{\mathrm{A}}(\beta_1,K)\notag\\ &\mathop{\leq}\limits^{(a)}\mathop{\lim}\limits_{\beta_1\rightarrow\infty}\exp\left(-\frac{(\alpha_w\sqrt{\beta_1\mathbb{E}[\check{\chi}]}-\alpha_w\alpha\sqrt{\mathbb{E}[\check{\chi}]})^2}{2}\right)\log_2(1+\rho\beta_1\mathbb{E}[\check{\chi}])\notag\\ \label{appendix:eq_5}&\mathop{=}\limits^{(b)}\mathop{\lim}\limits_{\beta_1\rightarrow\infty}\frac{\rho}{2\alpha_w^2\ln2}\frac{1}{\left(1+\rho\beta_1\mathbb{E}[\check{\chi}]\right)\left(1-\frac{\alpha}{\sqrt{\beta_1}}\right)}\exp\left(-\frac{(\alpha_w\sqrt{\beta_1\mathbb{E}[\check{\chi}]}-\alpha_w\alpha\sqrt{\mathbb{E}[\check{\chi}]})^2}{2}\right)=0,\end{aligned}$$ where (a) follows from the upper bound $\mathcal{Q}_1(a,b)\leq \exp\left(-\frac{(b-a)^2}{2}\right)$ for $b>a\geq0$ [@simon02]; (b) follows from applying L’Hospital’s rule. Therefore, there exists a unique global optimal $\beta_1$ which maximizes $\mathcal{I}_3^{\mathrm{A}}(\beta_1,K)$. Acknowledgment {#acknowledgment .unnumbered} ============== The authors want to express their deep appreciation to the anonymous reviewers and the Associated Editor for their many valuable comments and suggestions, which have greatly helped to improve this paper. [10]{} D. J. Love, R. W. Heath, V. K. N. Lau, D. Gesbert, B. D. Rao, and M. Andrews, “[An overview of limited feedback in wireless communication systems]{},” *[IEEE]{} J. Sel. Areas Commun.*, vol. 26, no. 8, pp. 1341–1365, Oct. 2008. H. Zhu and J. Wang, “[Chunk-based resource allocation in OFDMA systems-part I: chunk allocation]{},” *[IEEE]{} Trans. Commun.*, vol. 57, no. 9, pp. 2734–2744, Sept. 2009. H. Asplund, A. A. Glazunov, A. F. Molisch, K. I. Pedersen, and M. Steinbauer, “[The COST 259 directional channel model-part II: macrocells]{},” *[IEEE]{} Trans. Wireless Commun.*, vol. 5, no. 12, pp. 3434–3450, Dec. 2006. Y. Huang and B. D. Rao, “[Awareness of channel statistics for slow cyclic prefix adaptation in an OFDMA system]{},” *[IEEE]{} Wireless Commun. Lett.*, vol. 1, no. 4, pp. 332–335, Aug. 2012. R. Knopp and P. A. Humblet, “Information capacity and power control in single-cell multiuser communications,” in *Proc. IEEE International Conference on Communications (ICC)*, Jun. 1995, pp. 331–335. P. Viswanath, D. N. C. Tse, and R. Laroia, “[Opportunistic beamforming using dumb antennas]{},” *[IEEE]{} Trans. Inf. Theory*, vol. 48, no. 6, pp. 1277–1294, Jun. 2002. S. Sanayei and A. Nosratinia, “[Opportunistic downlink transmission with limited feedback]{},” *[IEEE]{} Trans. Inf. Theory*, vol. 53, no. 11, pp. 4363–4372, Nov. 2007. V. Hassel, D. Gesbert, M. S. Alouini, and G. E. Oien, “[A threshold-based channel state feedback algorithm for modern cellular systems]{},” *[IEEE]{} Trans. Wireless Commun.*, vol. 6, no. 7, pp. 2422–2426, Jul. 2007. J. Chen, R. Berry, and M. Honig, “[Limited feedback schemes for downlink OFDMA based on sub-channel groups]{},” *[IEEE]{} J. Sel. Areas Commun.*, vol. 26, no. 8, pp. 1451–1461, Oct. 2008. M. Pugh and B. D. Rao, “[Reduced feedback schemes using random beamforming in MIMO broadcast channels]{},” *[IEEE]{} Trans. Signal Process.*, vol. 58, no. 3, pp. 1821–1832, Mar. 2010. S. Sesia, I. Toufik, and M. Baker, *[LTE–The UMTS Long Term Evolution]{}*, 2nd ed.1em plus 0.5em minus 0.4emWiley, 2011. B. C. Jung, T. W. Ban, W. Choi, and D. K. Sung, “[Capacity analysis of simple and opportunistic feedback schemes in OFDMA systems]{},” in *Proc. International Symposium on Communications and Information Technologies (ISCIT)*, Oct. 2007, pp. 203–208. J. Y. Ko and Y. H. Lee, “[Opportunistic transmission with partial channel information in multi-user OFDM wireless systems]{},” in *Proc. IEEE Wireless Communications and Networking Conference (WCNC)*, Mar. 2007, pp. 1318–1322. J. G. Choi and S. Bahk, “[Cell-throughput analysis of the proportional fair scheduler in the single-cell environment]{},” *[IEEE]{} Trans. Veh. Technol.*, vol. 56, no. 2, pp. 766–778, Mar. 2007. Y. J. Choi and S. Bahk, “[Partial channel feedback schemes maximizing overall efficiency in wireless networks]{},” *[IEEE]{} Trans. Wireless Commun.*, vol. 7, no. 4, pp. 1306–1314, Apr. 2008. K. Pedersen, T. Kolding, I. Kovacs, G. Monghal, F. Frederiksen, and P. Mogensen, “[Performance analysis of simple channel feedback schemes for a practical OFDMA system]{},” *[IEEE]{} Trans. Veh. Technol.*, vol. 58, no. 9, pp. 5309–5314, Nov. 2009. J. Leinonen, J. Hamalainen, and M. Juntti, “[Performance analysis of downlink OFDMA resource allocation with limited feedback]{},” *[IEEE]{} Trans. Wireless Commun.*, vol. 8, no. 6, pp. 2927–2937, Jun. 2009. S. Donthi and N. Mehta, “[Joint performance analysis of channel quality indicator feedback schemes and frequency-domain scheduling for LTE]{},” *[IEEE]{} Trans. Veh. Technol.*, vol. 60, no. 7, pp. 3096–3109, Sept. 2011. S. Hur and B. D. Rao, “[Sum rate analysis of a reduced feedback OFDMA downlink system employing joint scheduling and diversity]{},” *[IEEE]{} Trans. Signal Process.*, vol. 60, no. 2, pp. 862–876, Feb. 2012. Y. Huang and B. D. Rao, “[Environmental-aware heterogeneous partial feedback design in a multiuser OFDMA system]{},” in *Proc. Asilomar Conference on Signals, Systems, and Computers*, Nov. 2011, pp. 970–974. P. Piantanida, G. Matz, and P. Duhamel, “[Outage behavior of discrete memoryless channels under channel estimation errors]{},” *[IEEE]{} Trans. Inf. Theory*, vol. 55, no. 9, pp. 4221–4239, Sept. 2009. Y. Isukapalli and B. D. Rao, “[Packet error probability of a transmit beamforming system with imperfect feedback]{},” *[IEEE]{} Trans. Signal Process.*, vol. 58, no. 4, pp. 2298–2314, Apr. 2010. V. Lau, W. K. Ng, and D. S. W. Hui, “[Asymptotic tradeoff between cross-layer goodput gain and outage diversity in OFDMA systems with slow fading and delayed CSIT]{},” *[IEEE]{} Trans. Wireless Commun.*, vol. 7, no. 7, pp. 2732–2739, Jul. 2008. T. Wu and V. Lau, “[Design and analysis of multi-user SDMA systems with noisy limited CSIT feedback]{},” *[IEEE]{} Trans. Wireless Commun.*, vol. 9, no. 4, pp. 1446 –1450, Apr. 2010. S. Akoum and R. W. Heath, “[Limited feedback for temporally correlated MIMO channels with other cell interference]{},” *[IEEE]{} Trans. Signal Process.*, vol. 58, no. 10, pp. 5219–5232, Oct. 2010. Q. Ma and C. Tepedelenlioglu, “[Practical multiuser diversity with outdated channel feedback]{},” *[IEEE]{} Trans. Veh. Technol.*, vol. 54, no. 4, pp. 1334–1345, Jul. 2005. A. Kuhne and A. Klein, “[Throughput analysis of multi-user OFDMA-systems using imperfect CQI feedback and diversity techniques]{},” *[IEEE]{} J. Sel. Areas Commun.*, vol. 26, no. 8, pp. 1440–1450, Oct. 2008. E. Dahlman, S. Parkvall, and J. Skold, *[4G LTE/LTE-Advanced for Mobile Broadband]{}*.1em plus 0.5em minus 0.4emAcademic Press, 2011. Ericsson, “[System-level evaluation of OFDM - further consideration]{},” 3GPP, TSG-RAN WG1R1-031303, Tech. Rep., Nov. 2003. H. Song, R. Kwan, and J. Zhang, “[Approximations of EESM effective SNR distribution]{},” *[IEEE]{} Trans. Commun.*, vol. 59, no. 2, pp. 603–612, Feb. 2011. S. N. Donthi and N. B. Mehta, “[An accurate model for EESM and its application to analysis of CQI feedback schemes and scheduling in LTE]{},” *[IEEE]{} Trans. Wireless Commun.*, vol. 10, no. 10, pp. 3436–3448, Oct. 2011. L. Wan, S. Tsai, and M. Almgren, “[A fading-insensitive performance metric for a unified link quality model]{},” in *Proc. IEEE Wireless Communications and Networking Conference (WCNC)*, Apr. 2006, pp. 2110–2114. J. Fan, Q. Yin, G. Y. Li, B. Peng, and X. Zhu, “[Adaptive block-level resource allocation in OFDMA networks]{},” *[IEEE]{} Trans. Wireless Commun.*, vol. 10, no. 11, pp. 3966–3972, Nov. 2011. G. D. Forney Jr and G. Ungerboeck, “[Modulation and coding for linear Gaussian channels]{},” *[IEEE]{} Trans. Inf. Theory*, vol. 44, no. 6, pp. 2384–2415, Oct. 1998. N. Al-Dhahir and J. M. Cioffi, “[Optimum finite-length equalization for multicarrier transceivers]{},” *[IEEE]{} Trans. Commun.*, vol. 44, no. 1, pp. 56–64, Jan. 1996. S. H. Muller-Weinfurtner, “[Coding approaches for multiple antenna transmission in fast fading and OFDM]{},” *[IEEE]{} Trans. Signal Process.*, vol. 50, no. 10, pp. 2442–2450, Oct. 2002. M. R. McKay, P. J. Smith, H. A. Suraweera, and I. B. Collings, “[On the mutual information distribution of OFDM-based spatial multiplexing: exact variance and outage approximation]{},” *[IEEE]{} Trans. Inf. Theory*, vol. 54, no. 7, pp. 3260–3278, Jul. 2008. M. Eslami and W. A. Krzymien, “[Net throughput maximization of per-chunk user scheduling for MIMO-OFDM downlink]{},” *[IEEE]{} Trans. Veh. Technol.*, vol. 60, no. 9, pp. 4338–4348, Nov. 2011. R. McEliece and W. E. Stark, “[Channels with block interference]{},” *[IEEE]{} Trans. Inf. Theory*, vol. 30, no. 1, pp. 44–53, Jan. 1984. M. M[é]{}dard and R. G. Gallager, “[Bandwidth scaling for fading multipath channels]{},” *[IEEE]{} Trans. Inf. Theory*, vol. 48, no. 4, pp. 840–852, Apr. 2002. H. A. David and H. N. Nagaraja, *[Order Statistics]{}*, 3rd ed.1em plus 0.5em minus 0.4emWiley-Interscience, 2003. M. Abramowitz and I. A. Stegun, *[Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables]{}*.1em plus 0.5em minus 0.4emDover, 1972. M. Sharif and B. Hassibi, “[On the capacity of MIMO broadcast channels with partial side information]{},” *[IEEE]{} Trans. Inf. Theory*, vol. 51, no. 2, pp. 506–522, Feb. 2005. A. H. Nuttall, “[Some integrals involving the Q-function]{},” Naval Underwater Systems Center, Tech. Rep., Apr. 1972. S. P. Boyd and L. Vandenberghe, *[Convex Optimization]{}*.1em plus 0.5em minus 0.4emCambridge Univ Pr, 2004. Y. Yu, “[On log-concavity of the generalized Marcum Q function]{},” *Arxiv Preprint*, 2011. \[Online\]. Available: <arXiv:1105.5762> M. K. Simon, *[Probability Distributions Involving Gaussian Random Variables: A Handbook For Engineers and Scientists]{}*.1em plus 0.5em minus 0.4emSpringer Netherlands, 2002. I. S. Gradshteyn and I. M. Ryzhik, *[Tables of Integrals, Series and Products]{}*, 7th ed., D. Zwillinger and A. Jeffrey, Eds.1em plus 0.5em minus 0.4emAcademic Press, 2007. P. Billingsley, *[Probability and Measure]{}*, 3rd ed.1em plus 0.5em minus 0.4emJohn Wiley & Sons, 1995. H. Finner and M. Roters, “[Log-concavity and inequalities for Chi-square, F and Beta distributions with applications in multiple comparisons]{},” *Statistica Sinica*, vol. 7, pp. 771–788, 1997. [Yichao Huang]{} (S’10–M’12) received the B.Eng. degree in information engineering with highest honors from the Southeast University, Nanjing, China, in 2008, and the M.S. and Ph.D. degrees in electrical engineering from the University of California, San Diego, La Jolla, in 2010 and 2012, respectively. He then join Qualcomm, Corporate R&D, San Diego, CA. He interned with Qualcomm, Corporate R&D, San Diego, CA, during summer 2011 and summer 2012. He was with California Institute for Telecommunications and Information Technology (Calit2), San Diego, CA, during summer 2010. He was a visiting student at the Princeton University, Princeton, NJ, during spring 2012. Mr. Huang received the Microsoft Young Fellow Award in 2007 from Microsoft Research Asia. He received the ECE Department Fellowship from the University of California, San Diego in 2008, and was a finalist of Qualcomm Innovation Fellowship in 2010. His research interests include communication theory, optimization theory, wireless networks, and signal processing for communication systems. [Bhaskar D. Rao]{} (S’80–M’83–SM’91–F’00) received the B.Tech. degree in electronics and electrical communication engineering from the Indian Institute of Technology, Kharagpur, India, in 1979, and the M.Sc. and Ph.D. degrees from the University of Southern California, Los Angeles, in 1981 and 1983, respectively. Since 1983, he has been with the University of California at San Diego, La Jolla, where he is currently a Professor with the Electrical and Computer Engineering Department. He is the holder of the Ericsson endowed chair in Wireless Access Networks and was the Director of the Center for Wireless Communications (2008–2011). His research interests include digital signal processing, estimation theory, and optimization theory, with applications to digital communications, speech signal processing, and human–computer interactions. Dr. Rao’s research group has received several paper awards. His paper received the Best Paper Award at the 2000 Speech Coding Workshop and his students have received student paper awards at both the 2005 and 2006 International Conference on Acoustics, Speech, and Signal Processing, as well as the Best Student Paper Award at NIPS 2006. A paper he coauthored with B. Song and R. Cruz received the 2008 Stephen O. Rice Prize Paper Award in the Field of Communications Systems. He was elected to the Fellow grade in 2000 for his contributions in high resolution spectral estimation. He has been a Member of the Statistical Signal and Array Processing technical committee, the Signal Processing Theory and Methods technical committee, and the Communications technical committee of the IEEE Signal Processing Society. He has also served on the editorial board of the EURASIP Signal Processing Journal. [^1]: Copyright (c) 2012 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]. [^2]: This research was supported by Ericsson endowed chair funds, the Center for Wireless Communications, UC Discovery grant com09R-156561 and NSF grant CCF-1115645. The material in this paper was presented in part at the 45th Asilomar Conference on Signals, Systems, and Computers, Pacific Grove, CA, November 2011. [^3]: The authors are with Department of Electrical and Computer Engineering, University of California, San Diego, La Jolla, CA 92093-0407, USA (e-mail: [email protected]; [email protected]).
{ "pile_set_name": "ArXiv" }
--- author: - 'Julien Grain,' - Vincent Vennin bibliography: - 'StochaAnisotropy.bib' date: today title: Unavoidable shear from quantum fluctuations in contracting cosmologies --- Introduction {#sec:intro} ============ Although inflation is the leading paradigm to describe the early universe, it is sometimes assumed to be preceded by a phase of slow contraction, followed by a bounce. This is indeed naturally expected in most theories of quantum gravity, it avoids the initial singularity that is otherwise still present in inflation when considered alone, and can also solve the trans-Planckian problem (see and references therein). In a homogeneous cosmology, the energy density associated with the anisotropic stress however grows as $a^{-6}$, where $a$ is the scale factor. Since the background energy density grows as $a^{-3(1+w)}$, where $w$ is the equation-of-state parameter of the background fluid, any initial anisotropy grows unstable as the universe contracts, hence driving a new kind of singularity [@Belinsky:1970ew], unless $w>1$. Several solutions to this problem have been proposed. One can for instance assume that the universe contracts with a stiff equation of state, $w>1$, preventing the energy density associated with the anisotropic stress to overcome the isotropic contribution. This corresponds for instance to the ekpyrotic model [@Khoury:2001wf]. Another solution more rooted in quantum gravity consists in regularising the space-time curvature. A concrete example of this is Loop Quantum Cosmology of anisotropic models of the universe [@Ashtekar_2009]. Here the quantisation of the gravitational degrees of freedom makes all energy densities (including the one associated with the anisotropic stress) effectively bounded from above, then guaranteeing a regular bounce for any equation of state. Obviously, a third solution is to fine tune the initial anisotropy to be sufficiently small such that it remains negligible until the bounce occurs. The amount of required fine tuning may be large for a long phase of contraction, but, at the classical level, it is always possible to set the initial anisotropy to arbitrarily small values. However, in the presence of vacuum quantum fluctuations, this is not true anymore and there is a minimum, inevitable amount of anisotropic stress sourced by quantum fluctuations. The goal of this work is to study this unavoidable shear production sourced by quantum fluctuations. It is organised as follows. In , we formulate the problem, and show why quantum fluctuations of a test scalar field provide a non-vanishing source of anisotropic stress (more precisely, we explain why the one-point function of the anisotropic stress produced by scalar-field fluctuations vanish, but not its two-point function). In , we derive a stochastic formalism, analogous to the stochastic-inflation formalism, to describe these quantum fluctuations in an inflating or a contracting cosmology, and to compute the expectation value of the shear they generate. In , we apply these tools to the case of a contracting phase with a constant equation-of-state parameter $w$. We find that, when $w>-1/9$, the shear produced out of massless scalar field fluctuations is always negligible until the energy density of the universe reaches the Planck scale, regardless of initial conditions; while for $w<-1/9$, it becomes sizeable much before then, at a point that depends on the time at which the contraction is initiated. In , we compare the shear backreaction with the direct contribution of the test scalar field fluctuations to the energy density of the universe, and conclude that the later is always more important than the former. We also discuss the case of inflating backgrounds, where it is found that the shear backreaction is always negligible. We present our conclusions in . Formulating the problem {#sec:FormulationProblem} ======================= We consider the case where deviations from the isotropic Friedmann-Lemaître-Robertson-Walker (FLRW) space-time can be parametrised by a Bianchi I metric [@Ellis:1968vb; @Pereira:2007yy], \[eq:metric:Bianchi\] s\^2=-t\^2+a\^2(t)\_[ij]{}x\^ix\^j, with $a(t)$ the isotropised scale factor and $\gamma_{ij}(t)=\exp[2\beta_i(t)]\delta_{ij}$. The $\beta_i$ functions are constrained to satisfy $\sum_i\beta_i=0$, such that $\det \gamma_{ij}=1$. The shear tensor, $\sigma_{ij}$, is defined as \_[ij]{}= , where a dot denotes a derivative with respect to cosmic time $t$. We further assume that the universe is filled with a homogeneous but anisotropic matter fluid, described by its energy density $\rho$, its pressure $P$ and its anisotropic stress $\Pi_{\mu\nu}$, such that its stress-energy tensor reads \[eq:Tmunu:fluid\] T\_ = (+P) u\_u\_ + P g\_ + \_ . In this expression, $u^\mu$ is the four velocity of a comoving observer, normalised such that $u_\mu u^\mu=-1$. From the metric , a comoving observer has $u^i=0$, and the normalisation condition imposes that $u^\mu=-\delta^\mu_0$. The anisotropic stress is a symmetric tensor (since the stress-energy tensor is symmetric) that parametrises deviations from a perfect fluid. Given that its “$00$” component can always be reabsorbed in the first term of , hence in the definition of $\rho$, it is set to zero without loss of generality. The “$0i$” component of the Einstein equation then imposes that $\Pi_{0i}=0$, which implies that the anisotropic stress is orthogonal to the comoving velocity, $\Pi_{\mu\nu} u^\nu=0$. If one decomposes $\Pi_{ij}$ as $\Pi_{ij}=\Pi_{ij}- g^{mn}\Pi_{mn} g_{ij}/3 + g^{mn}\Pi_{mn}g_{ij} /3$, the component $ g^{mn}\Pi_{mn}g_{ij} /3$ can always be reabsorbed in the second term of , hence in the definition of $P$, and the remaining part $\Pi_{ij}- g^{mn}\Pi_{mn} g_{ij}/3$ is traceless, so we assume that $\Pi_{ij}$ is traceless without loss of generality, $\Pi_{ij} g^{ij}=0$. Under these conditions, the Einstein equations can be written as \[eq:Friedmann\] H\^2& = & + \_[ij]{} \^[ij]{} ,\ \[eq:Raychaudhuri\] & = & -- \_[ij]{} \^[ij]{} ,\ \[eq:sigmadot:Pi\] (\^i\_j)+3 H \^i\_j &=& , which correspond respectively to the “$00$” component, the trace and the traceless part of the “$ij$” component, and where $H\equiv\dot{a}/a$ is the Hubble parameter. By combining these three equations, one obtains a conservation relation, \[eq:conservation:equation\] +3H(+P)=-\_[ij]{}\^[ij]{} . At the background level, one can see that the only effect of the shear is to add a contribution $\rho_\sigma = \Mp^2 \sigma_{ij} \sigma^{ij}/2$ to the energy density, and $p_\sigma = \rho_\sigma$ to the pressure, in the Friedmann equation  and in the Raychaudhuri equation . This additional contribution is sourced by the anisotropic stress through , which can be formally solved as \[eq:sigma:sol\] \^i\_j (t) = \^i\_j (t\_)\^3 + \_[t\_]{}\^t \^3 \^[i]{}\_[j]{}(t\^)t\^ . From this expression, one can see that, in the absence of anisotropic stress, any initial amount of shear tensor grows as $\sigma^i_{j} \propto a^{-3}$, which gives rise to $\rho_\sigma\propto a^{-6}$. Denoting by $w$ the equation-of-state parameter of the background fluid, such that $P=w\rho$, gives rise to $\rho\propto a^{-3(1+w)}$ in the absence of anisotropic stress, so the shear contribution eventually dominates unless $w>1$, as mentioned above. The time at which $\rho_\sigma$ takes over $\rho$ obviously depends on its initial value, hence on $\sigma^i_j(t_\uin)$. Since the contracting phase ends at the time when the universe experiences a bounce, and given that the bounce is followed by an expanding phase, during which $\rho_\sigma$ decreases faster than $\rho$ (unless $w>1$), it is always possible to make the shear irrelevant by requiring its initial value to be sufficiently small. The shear problem thus becomes an initial fine-tuning problem. There is, however, no natural choice for neither the values of the shear tensor nor the probabilities associated to them, and this could have important consequences for the state of the universe after the bounce (see for a concrete example in Loop Quantum Cosmology). Alternatively one could imagine the existence of mechanisms in the early stage of, or prior to, the contracting phase, that would set the shear to the required small value, as could be the case in cyclic models of the universe. These considerations are however valid only under the assumption that the anisotropic stress vanishes. While this is true in principle for a perfect fluid, at the high energy scales that operate in the early universe, field theory is the relevant framework to describe matter, and as we shall see, the unavoidable vacuum fluctuations present in quantum fields are enough to generate an anisotropic stress. For explicitness, let us consider the case of a test scalar field $\phi$, minimally coupled to gravity, described by the action \[eq:action:phi\] S\^[()]{} = -\^4 , where $V(\phi)$ is the potential associated with $\phi$. The stress-energy tensor can be obtained from $T_{\mu\nu}^{(\phi)} = -(2/\sqrt{-g}) \delta S^{(\phi)}/\delta g^{\mu\nu} $, which leads to \[eq:StressEnergyTensor:ScalarField\] T\_\^[()]{} = \_\_ - g\_ . As explained above, the anisotropic stress corresponds to the transverse and traceless part of the stress-energy tensor. Transverseness is obtained by projecting $T_{\mu\nu}^{(\phi)}$ in the hyperplane orthogonal to $u^\mu$, by using the projector $\perp_{\mu\nu}=g_{\mu\nu}+u_\mu u_\nu$.[^1] In the Bianchi I space time , its only non-zero components are $\perp_{ij}=g_{ij}$, which is no more than the induced metric. Tracelessness is obtained by removing the trace of the projected tensor, and one obtains $\Pi_{ij}^{(\phi)} = T^{(\phi)}_{ij }-\frac{1}{3}g_{ij}g^{mn} T_{mn}^{(\phi)}$, hence \[eq:AnisotropicStress:ScalarField\] \_[ij]{}\^[()]{} = \_i \_j -(g\^[mn]{}\_m\_n) g\_[ij]{} . Let us consider the case where the initial shear vanishes, so space time is initially isotropic. If the quantum state of the scalar field $\phi$ shares the same symmetries as the background, for instance if it is in its vacuum state $\vert 0 \rangle$, then there is no preferred direction and $\langle 0 \vert \Pi_{ij}^{(\phi)} \vert 0 \rangle =0 $ (this will be shown explicitly below). This implies that, when taking the expectation value of , the anisotropic stress disappears, so $\langle \sigma^i_{j} \rangle \propto a^{-3}$ and the above considerations still apply. However, the two-point function of the anisotropic stress does not necessarily vanish, and since the self contracted-shear appears in $\rho_\sigma$, this means that quantum fluctuations of a scalar field can source an effective shear. In other words, by self contracting , in the absence of initial shear, one has \[eq:rhoSigma:AnisotropicStress\] \_= \_[t\_]{}\^t t\_1\_[t\_]{}\^t t\_2 \^3 \_[i]{}\^[j]{} (t\_1) \^[i]{}\_[j]{}(t\_2) . The goal of the following is thus to compute the two-point function of the anisotropic stress for a test scalar field, in order to evaluate . Let us note that, since $\phi$ must be homogeneous at the background level, is of second order in the field perturbations, hence $\rho_\sigma$ is of fourth order in those perturbations. This small contribution is however enhanced by $a^{-6}$, so the (ir)relevance of the effect is a priori not obvious. Let us also note that other types of fields, such as vector fields or gauge fields, can give a non-vanishing mean anisotropic stress, hence source the shear at second rather than fourth order in perturbations. In this work, we choose to study the case of a scalar field, in order to extract the minimum amount of shear one cannot avoid, simply by postulating the existence of light scalar degrees of freedom during the contraction. Stochastic shear from a test scalar field {#sec:Stochastic:Formalism} ========================================= In order to compute the statistics of the anisotropic stress generated by a scalar field in a contracting phase, we employ a stochastic formalism reminiscent of the “stochastic inflation” formalism [@Starobinsky:1982ee; @Starobinsky:1986fx; @Starobinsky:1994bd]. While this makes practical calculations easier, a comparison with a full quantum field theoretic treatment is performed in , where we check that the “stochastic contraction” approach developed in this work is indeed valid (a result that is well-established in the context of inflation, see ). Another reason for using a stochastic formalism is that, although only the case of a test field is considered hereafter, it may help to generalise our results to describe the backreaction of this field on the background dynamics, see the discussion in . Langevin equations for a test scalar field {#sec:Langevin:Free:Scalar} ------------------------------------------ During inflation, because of the presence of a phase-space attractor, the so-called slow-roll solution [@Salopek:1990jq; @Liddle:1994dx], the stochastic formalism reduces to a single Langevin equation once this attractor is reached. In a contracting cosmology, there is no such attractor and the stochastic formalism has to be extended to the full phase space, as was done in , which we follow below. Starting from , and working with cosmic time,[^2] the Hamilton equations read \[eq:eom:phi\] &=& ,\ &=&a(t)e\^[-2\_i(t)]{}\^[ij]{}\_i\_j-a\^3 m\^2 . \[eq:eom:pi\] Hereafter, for simplicity, we consider the case of a free field with potential $V(\phi)=m^2\phi^2/2$. The coarse-graining procedure then consists in splitting the scalar field into a large-scale (classical) part and small-scale (quantum) fluctuations as $\phi=\bar\phi+\phi_Q$ and $\pi=\bar\pi+\pi_Q$, with \[eq:phiQ\] \_Q&=&W(,t),\ \_Q&=&W(,t). \[eq:piQ\] Here, $W$ is a window function that selects out small-wavelength modes, it is such that $W \simeq 1$ for $k\gg k_\epsilon(\vec{k}/k,t)$ and $W\simeq 0$ otherwise. In an FLRW background, the scale of coarse-graining is chosen as $k_\epsilon=\epsilon a \vert H\vert$ with $\epsilon\ll1$, for two reasons. First, this ensures that the coarse-grained part is indeed comprised of large-scale modes only, well above the Hubble radius, where gradients can be neglected (see below and the discussion in ). Second, this guarantees that the inflow of modes into the large-scale sector is made of quantum fluctuations that are sufficiently squeezed and amplified to be considered as a classical stochastic noise. Here, since the background is anisotropic, the coarse-graining scale not only depends on time but also on the direction of $\vec{k}$, which is why the window function depends both on the norm and direction of $\vec{k}$. In each spatial direction, from one can define a scale factor $a_i=a e^{\beta_i}$, hence a Hubble rate $H_i = H + \dot{\beta}_i$. If one defines the coarse-grained field as being made of wavelengths that are much larger than the Hubble radius in *all* directions, k\_i&lt;a\_i H\_i= a e\^[\_i]{} H+\_i i, where $\epsilon\ll 1$, this gives a prescription for $ k_\epsilon(\vec{k}/k,t)$. In and , $\hat{a}^\dagger_{\vec{k}}$ and $\hat{a}_{\vec{k}}$ are creation and annihilation operators, that satisfy the canonical commutation relations $[a_{\vec{k}},a^\dag_{\vec{q}}]=\delta^3(\vec{k}-\vec{q})$. The mode functions $\phi_{\vec{k}}$ and $\pi_{\vec{k}}$, which satisfy the equations of motion  and  when written in Fourier space, have to be normalised according to the Klein-Gordon product $i\int_{\Sigma_t}\dd^3x(\Phi_{\vec{k}}\Pi^*_{\vec{q}}-\Pi_{\vec{k}}\Phi^*_{\vec{q}})=\delta^3(\vec{k}-\vec{q})$, where $\Phi_{\vec{k}}=\phi_{\vec{k}}(t)e^{-i\vec{k}\cdot\vec{x}}$ and $\Pi_{\vec{k}}=\pi_{\vec{k}}(t)e^{-i\vec{k}\cdot\vec{x}}$, and where $\Sigma_t$ is a space-like hypersurface at fixed time $t$. Inserting the field decomposition $\phi=\bar\phi+\phi_Q$ and $\pi=\bar\pi+\pi_Q$ into and , one obtains Langevin equations for the coarse-grained fields, \[eq:Langevin:phi\] &=&+\_ ,\ &=&-a\^3 m\^2 | + \_ , \[eq:Langevin:pi\] where gradient terms are neglected since they are suppressed by $\epsilon^2$. The two noise terms $\xi_\phi$ and $\xi_\pi$ are random Gaussian noises, that share the same statistics as their quantum counterparts \[eq:xi:phi\] \_&=&-,\ \_&=&- . \[eq:xi:pi\] At leading order in perturbation theory, this statistics is Gaussian, hence fully characterised by its two-point correlation functions, \[eq:2pt:xi:gen\] 0 \_f (,t)\_g (’,t’)0&=& f\_(t) g\_\^\*(t’) e\^[-i (-)]{}, where $f$ and $g$ can either be $\phi$ or $\pi$, and where we have used the fact that the only non-zero contribution comes from $\left<0\right|\hat{a}_{\vec{k}}\hat{a}^\dag_{\vec{q}}\left|0\right>=\delta^3(\vec{k}-\vec{q})$. In an anisotropic universe, the resulting integral is however not straightforward to compute, since both the mode functions $\phi_{\vec{k}}$ and $\pi_{\vec{k}}$, and the window function $W(\vec{k},t)$, depend on the direction of $\vec{k}$. For simplicity, we will assume that the universe is initially close to being isotropic, and that the anisotropy is only sourced by the test scalar field $\phi$. Since we want to determine the point at which the backreaction of the shear becomes problematic from the point of view of the isotropy assumption, the effect of the anisotropy of the mode and window functions can indeed be neglected, since it only gives rise to corrections of higher order in the shear. In this case, the coarse-graining scale becomes independent of $\vec{k}$, $k_\epsilon=\epsilon a \left\vert H \right\vert $. If the window function is set to a Heaviside function, \[eq:Window\] W(,t)=, one has $\partial W(\vec{k},t)/\partial t = -\dot{k}_\epsilon \delta(k-k_\epsilon)$. If the mode functions are isotropic, $\phi_{\vec{k}} = \phi_k$ and $\pi_{\vec{k}}=\pi_k$, the integral over the angles defining the direction of $\vec{k}$ can be performed in , and the resulting integral over the norm of $\vec{k}$ can be carried out with the delta function appearing in $\partial W(\vec{k},t)/\partial t$. This gives rise to \[eq:noise:correlator\] 0 \_f (,t)\_g (’,t’)0&=& \_[(\_t)\_[f,g]{}(-’)]{} (t-t’) , which shows that noises are white (uncorrelated in time) in this case. This also defines the covariance matrix $\boldmathsymbol{\Sigma}$ of the stochastic noises $\xi_\phi$ and $\xi_\pi$ in the Langevin equations  and , which can be formally integrated as \[eq:Phi:Solution:Green\] |(t) = (t,t\_) |(t\_) + \_[t\_]{}\^t s  (t,s) (s) . In this expression, we have arranged the coarse-grained phase-space coordinates into the vector $\bar{\boldmathsymbol{\Phi}} = \left(\bar{\phi}, \bar{\pi} \right)^\mathrm{T}$, and, similarly $\boldmathsymbol{\xi} = \left( \xi_\phi, \xi_\pi\right)^\mathrm{T}$. The Green’s matrix $\boldmathsymbol{G}$ is associated with the homogeneous dynamics, it is the solution of \[eq:Green:def\] (t,t’)= ( [cc]{} 0 & a\^[-3]{}\ -m\^2 a\^3 & 0 ) (t,t’) + (t-t’) (see Appendix A of for more details). From these relations, all statistical moments of the stochastic phase-space variables $\bar{\phi}$ and $\bar{\pi}$ can be computed. For instance, the two-point functions are given by \[eq:TwoPoint:phi\] \_[t,t’]{}(’-) 0 |(,t) |(’,t’)\^0 = \_[t\_]{}\^[(t,t’)]{}s (t,s) \_s(-’) \^(t’,s)\ where we have defined $\delta \bar{f} = \bar{f}-\langle 0 \vert \bar{f} \vert 0\rangle$. Stochastic anisotropic stress ----------------------------- Let us now derive the expectation value of the anisotropic stress. As announced in , since the scalar field $\phi$ is placed in a quantum state that is statistically isotropic, we will find that the mean anisotropic stress vanishes. The following calculation is however more than a simple consistency check, since it provides the building blocks for the calculation of the mean shear, which will be performed in . As can be seen in , the calculation of the mean anisotropic stress requires to evaluate quantities of the type $\langle 0 \vert \partial_i \bar{\phi} \partial_j \bar{\phi} \vert 0 \rangle$. This gradient correlator can be obtained from the field correlator , by expanding $\partial_i \bar{\phi} =\lim_{\alpha\to 0} [\bar{\phi}(\vec{x}+\alpha \vec{e}_i,t)-\bar{\phi}(\vec{x},t)]/\alpha$, where $\vec{e}_i$ is a unit vector pointing towards the $i^{\mathrm{th}}$ spatial direction. This allows one to write the gradient correlator as the sum of four field correlators, and in the coincident configuration where the gradients are evaluated at the same location $\vec{x}=\vec{x}'$, one finds[^3] \_[t,t’;(i,j)]{}&& 0 \_i|(,t) \_j|\^(,t’) 0\ &=& \_[0,0]{}\ & = & = - \_[t,t’]{}”(0) \_[ij]{} . \[eq:Xi\_ij:Xi”\] In the first equality, we have used that $\boldmathsymbol{\Xi}_{t,t'}$ only depends on $\vert \vec{x}'-\vec{x}\vert$, see , and the primes in our final expression denote derivation with respect to that quantity. In the second equality, we have used that $\boldmathsymbol{\Xi}_{t,t'}$ is an even function of $\vert \vec{x}'-\vec{x}\vert$ (since the cardinal sine function $\mathrm{sinc}(x)\equiv\sin (x)/x$ appearing in is even), hence $\boldmathsymbol{\Xi}_{t,t'}'(0)=0$; together with the fact that $\vert \alpha \vec{e}_i-\beta\vec{e}_j \vert^2 = \alpha^2+\beta^2-2\alpha\beta\delta_{ij}$. This expression can be made further explicit by making use of and , and since $\mathrm{sinc}''(0)=-1/3$, one finds \[eq:gradient:two:point:function\] \_[t,t’;(i,j)]{}= \_[t\_]{}\^[(t,t’)]{}s k\_\^2(s) (t,s) \_s(0 ) \^(t’,s) . It is then easy to see from that, in an isotropic background, the fact that $\boldmathsymbol{\Xi}^{(i,j)} _{t,t'}$ is proportional to $\delta_{ij}$ implies that \[eq:anisotropic:stress:one:point:function\] 0 \_[ij]{} 0 = 0 . This confirms that the one-point function of the anisotropic stress vanishes in an isotropic background. Stochastic shear {#sec:Stochastic:Shear} ---------------- From , the two-point function of the anisotropic stress can be expressed as combinations of four-point functions of field gradients. Since the statistics of $\phi$ is Gaussian \[given that $\boldmathsymbol{\xi}$ is a Gaussian noise and that $\bar{\boldmathsymbol{\Phi}}$ is linear in the noise, see \], these four-point functions can be evaluated using Wick theorem, namely $\langle 0 \vert \partial_i \bar{\phi} \partial_j \bar{\phi} \partial_m \bar{\phi}\partial_n \bar{\phi} \vert 0 \rangle = \langle 0 \vert \partial_i \bar{\phi} \partial_j \bar{\phi} \vert 0\rangle \langle 0 \vert \partial_m \bar{\phi}\partial_n \bar{\phi} \vert 0 \rangle +\langle 0 \vert \partial_i \bar{\phi} \partial_m \bar{\phi} \vert 0 \rangle \langle 0 \vert \partial_j \bar{\phi}\partial_n \bar{\phi} \vert 0 \rangle + \langle 0 \vert \partial_i \bar{\phi} \partial_n \bar{\phi} \vert 0 \rangle \langle 0 \vert \partial_j \bar{\phi} \partial_m \bar{\phi} \vert 0 \rangle$. This gives rise to 0 \_[i]{}\^[j]{} (t) \^[i]{}\_[j]{} (t’) 0 &=& g\^[ij]{}(t)g\^[mn]{}(t’) . In an isotropic universe, making use of , the terms involving products of the form ${\Xi}_{t,t;(i,j)}^{\bar{\phi},\bar{\phi}} {\Xi}_{t,'t';(k,\ell)}^{\bar{\phi},\bar{\phi}} $ cancel out. Such terms would indeed correspond to the squared expectation value of the anisotropic stress, which was already shown to vanish, see . Thus only the cross terms remain, those of the form ${\Xi}_{t,t';(i,j)}^{\bar{\phi},\bar{\phi}} {\Xi}_{t,t';(k,\ell)}^{\bar{\phi},\bar{\phi}} $, leading to[^4] 0 \_[i]{}\^[j]{} (t) \^[i]{}\_[j]{} (t’) 0 = \^2 . \[eq:Piji:Piij:Xi”\] This can be cast in a form similar to , and this shows that, even in an isotropic background, scalar quantum fluctuations make the two-point function of the anisotropic stress not vanish. Shear backreaction in a contracting cosmology {#sec:Contraction} ============================================= Let us now apply the calculational program sketched above to the case of a contracting cosmology (expanding cosmologies, in particular inflating cosmologies, are discussed in ). In an isotropic universe, defining $u=a \phi$, - give rise to \[eq:mode:MS:u\] [u]{}\_k”+(a\^2m\^2+k\^2-\^2-’)u\_k=0 , where a prime denotes derivation with respect to conformal time $\eta$, related to cosmic time through $\dd t = a \dd \eta$, and where $\mathcal{H}=a'/a$. In order to set initial conditions in the Bunch-Davies vacuum state [@Bunch:1978yq], the pulsation $\omega_k^2$, defined as the term between the braces in , must be dominated by $k^2$ at early time. In a contracting cosmology, this clearly imposes $m=0$, and for simplicity, we consider that this is the case hereafter. If the equation-of-state parameter of the background is constant, one has $\omega_k^2 = k^2-2(1-3w)/[(1+3w)^2\eta^2]$. The resulting mode equation can be solved by means of Hankel functions, and setting integration constants such that the mode functions are normalised according to the Klein-Gordon product mentioned above, one obtains \_k = e\^[i(2+1)]{} H\_\^[(1)]{}(-k) = in the vacuum state, where $H_{\nu}^{(1)}$ is the first Hankel function and where, hereafter, we assume that $-1/3<w<1$. On super-Hubble scales, the Hankel function can be expanded, and one finds $\phi_{k\ll a H} \simeq \ee^{i\pi (2\nu-1)/4}2^{\nu-1}\Gamma(\nu)(-k\eta)^{1/2-\nu}/(\sqrt{\pi k}a)$ and $\pi_k = a^2\phi_k'$. Plugging these expressions into , one obtains -2em \_t(-’) &=& (1+3w)\^ \^H\^3\ && ( [ccc]{} 1 && (w-1) a\^3 H\ (w-1) a\^3 H & & \^2 ) .\ \[eq:Sigma:MasslessField\] If the mass vanishes, the background equation of motion  can be readily solved, and one obtains for the Green function (t,t’) = ( [cc]{} 1& \_[t’]{}\^t\ 0 & 1 )(t-t’) . \[eq:Green:massless\] Recalling that $\mathrm{sinc}''(0)=-1/3$, by making use of and , then gives rise to -2em 0 \_[i]{}\^[j]{} (t\_1) \^[i]{}\_[j]{} (t\_2) 0 &= & \^H\_\^[8]{} ()\^[[3w-5]{}]{}\ & & 2em \^[-1-9w]{}-1\^2 , \[eq:Piji:Piij:Xi”:explicit\] where $a_1$ and $a_2$ denote $a(t_1)$ and $a(t_2)$ respectively. The contribution from the shear to the overall energy density can then be computed with . In order to specify the term $\mathrm{max}(a_1,a_2)$ appearing in , the rectangular integration domain of can be split into two triangles, one where $t_1<t_2$ and one where $t_1>t_2$. Since the integrand is in fact symmetric in $t_1$ and $t_2$, both these triangular regions give the same contribution and it is enough to compute the integral over one of the triangles and multiply it by two. This leads to 0 \_0 & =& \^4() ()\^[4]{} .\ \[eq:rho\_shear:contraction\] One can check that, when $t=t_\uin$, $\left\langle \rho_\sigma \right\rangle=0$. When $w>-1/9$, the terms have been ordered in decreasing power, while for $w<-1/9$, the dominant term is the third one (the first one on the second line), which increases as $\propto a^{2+18w}$. These two cases need therefore to be distinguished. ![Left panel: contribution from the shear to the overall energy density in a contracting cosmology, as a function of time labeled by the value of the Hubble scale in Planckian units. The result is computed with , using $\epsilon=0.1$ Three different equation-of-state parameters are displayed, namely $w=-2/9$ (blue lines), $w=0$ (green lines) and $w=1/3$ (red lines), for three different initial times, $H_\uin/\Mp = 10^{-10}$ (solid lines), $H_\uin/\Mp = 10^{-15}$ (dashed lines) and $H_\uin/\Mp = 10^{-10}$ (dotted lines). The grey shaded area corresponds to $\langle\rho_\sigma\rangle>\rho_{\mathrm{tot}}$, where the backreaction from the shear becomes important (and where our calculation does not apply). Right panel: direct contribution from the scalar field fluctuations to the overall energy density, in the same situations as those shown in the left panel.[]{data-label="fig:shear"}](figs/MeanShear "fig:"){width="49.60000%"} ![Left panel: contribution from the shear to the overall energy density in a contracting cosmology, as a function of time labeled by the value of the Hubble scale in Planckian units. The result is computed with , using $\epsilon=0.1$ Three different equation-of-state parameters are displayed, namely $w=-2/9$ (blue lines), $w=0$ (green lines) and $w=1/3$ (red lines), for three different initial times, $H_\uin/\Mp = 10^{-10}$ (solid lines), $H_\uin/\Mp = 10^{-15}$ (dashed lines) and $H_\uin/\Mp = 10^{-10}$ (dotted lines). The grey shaded area corresponds to $\langle\rho_\sigma\rangle>\rho_{\mathrm{tot}}$, where the backreaction from the shear becomes important (and where our calculation does not apply). Right panel: direct contribution from the scalar field fluctuations to the overall energy density, in the same situations as those shown in the left panel.[]{data-label="fig:shear"}](figs/Mean_rho_phi "fig:"){width="49.60000%"} When $w>-1/9$, since the dominant term, the one proportional to $1/(5+27 w)$, does not depend on the initial time (contrary to the other terms, it does not involve $a_\uin$), the contribution from the shear to the energy density quickly reaches an attractor, which scales as $\langle \rho_\sigma\rangle \propto H^6/\Mp^2$, hence $\langle \rho_\sigma\rangle /\rho_{\mathrm{tot}} \propto (H/\Mp)^4$. As a consequence, until the background energy density reaches the Planck scale, the contribution from the shear to that energy density is negligible. This can be clearly seen in the left panel of (see the cases $w=0$ and $w=1/3$). In fact, because of the prefactor in , the ratio $\langle \rho_\sigma\rangle /\rho_{\mathrm{tot}}$ is still negligible when $H$ reaches the Planck scale. In bouncing cosmologies, $|H|$ is bounded from above since the bounce demands for a smooth transition from contraction, $H<0$, to expansion, $H>0$. The upper-bound of $|H|$ can be safely taken to be of the order of the Planck scale at most[^5] and the maximal value that $\langle \rho_\sigma\rangle /\rho_{\mathrm{tot}}$ can achieve is thus of order $\sim\epsilon^{4\frac{1+9w}{1+3w}}$. Hence, one concludes that there is no shear backreaction problem in this case. When $w<-1/9$, since the dominant term, the first one of the second line in , explicitly depends on $a_\uin$, the late-time behaviour of $\langle\rho_\sigma\rangle$ strongly depends on the initial time of the contraction and there is no attractor. Besides, this dominant term increases in time (recall that $a$ decreases in a contracting cosmology), so $\langle \rho_\sigma\rangle /\rho_{\mathrm{tot}}$ does not scale as $(H/\Mp)^4$ anymore, but rather as $(H/\Mp)^{2(7-9w)/[3(1+w)]} [\Mp/(\epsilon H_\uin)]^{-4(1+9w)/[3(1+w)]}$. This implies that, if the contracting phase starts in the infinite past ($\epsilon H_\uin=0$), then there is an IR divergence in the shear. Indeed, if $w<-1/9$, then the power spectrum of the field fluctuations is too red for the large wavelengths to give a finite contribution to the shear. This problem can be solved by considering a finite contraction. However, if $H_\uin$ is sufficiently small, if the contracting phase starts sufficiently early on, the contribution from the shear to the overall energy density becomes sizeable before the energy density of the universe reaches the Planck scale. This is again well visible in the left panel of , see the case $w=-2/9$. Whether or not a backreaction problem occurs before $H$ reaches $\Mp$ thus depends on the value of $H_\uin$, on the total amount of contraction. This is why, in , we display the minimum value of $H_\uin$ such that the shear does not backreact on the background geometry until $H$ reaches the Planck scale. One can see that, when $w\lesssim -0.2$, this is impossible to achieve. Let us note that the case $w=-1/9$ is a priori singular. Indeed, in the prefactor of , one notices the presence of the term $(1+9w)^{-2}$, which diverges when $w$ approaches $-1/9$. However, in that limit, all three first terms in the square brackets become of the same order, and, when combined together, they precisely give a contribution proportional to $(1+9w)^{2}$ (with a logarithmic dependence on $a/a_\uin$), making this limit regular. In all cases, let us also stress that quantum diffusion makes the shear grow faster than classically. Indeed, if $w>-1/9$, one has $\langle\rho_\sigma\rangle \propto a^{-9(1+w)}$ and if $w<-1/9$, $\langle\rho_\sigma\rangle \propto a^{9w-7}$. So $\langle\rho_\sigma\rangle$ always increases faster than $a^{-8}$, while it scales as $a^{-6}$ classically. Discussion {#sec:Discussion} ========== ![Minimum value of the initial Hubble parameter (parametrising the maximum amount of contraction) such that the backreaction coming from the shear (blue curve) and from the spectator field fluctuations (green) remain negligible until the time when the overall energy density reaches the Planck scale.[]{data-label="fig:HinMin"}](figs/HinMin){width="49.60000%"} For comparison, it is interesting to compute the contribution to the overall energy density arising directly from the scalar field fluctuations themselves. Indeed, from , the “$00$” component of the stress-energy tensor reads $\rho_\phi=\dot{\phi}^2/2+V(\phi)+\delta^{ij}\partial_i\phi\partial_j\phi/(2a^2)$. In the absence of a potential, and after neglecting gradient terms, which are suppressed by higher powers of $\epsilon$ in the coarse-grained sector, this gives rise to[^6] $\rho_\phi=\dot{\phi}^2/2=\pi^2/(2 a^6)$, see . One thus has to compute $\langle \rho_\phi \rangle =\Xi_{t,t}^{\bar{\pi},\bar{\pi}}(0) /(2 a^6)$, where the two-point functions $\boldmathsymbol{\Xi}_{t,t'}(\vert \vec{x}'-\vec{x}\vert)$ are given in . Making use of for the noise correlator and of for the Green function, this gives rise to 0 \_0 = ()\^ H\^4 \[eq:rho:phi:contraction\] if $w\neq 0$, while $\langle \rho_\phi \rangle = 9 H^4/(256\pi^2)\ln(a_\uin/a)$ when $w=0$. One can see that two regimes need again to be distinguished. If $w>0$, the second term in the square brackets, $(a/a_\uin)^{6w}$, quickly becomes negligible as contraction proceeds, and one reaches an attractor where $\langle \rho_\phi\rangle \propto H^4$, hence $\langle \rho_\phi\rangle/\rho_{\mathrm{tot}}$ is suppressed by $(H/\Mp)^2$. In this case, the spectator field does not backreact until $H$ reaches the Planck scale, and the direct contribution from the scalar field fluctuations to the overall energy density is larger than that of the shear, which we recall is suppressed by $(H/\Mp)^4$. If $w<0$, the second term in the square brackets provides the dominant contribution. Since it depends explicitly on $a_\uin$, there is no attractor anymore, $\langle \rho_\phi\rangle/\rho_{\mathrm{tot}}$ scales as $(H/\Mp)^{2(1-w)/(1+w)} (\Mp/H_\uin)^{-4w/(1+w)}$, and there is an IR divergence (similar to the one discussed for the shear) if the contracting phase extends infinitely far in the past. The energy density contained in $\phi$ can therefore become large before $H$ reaches the Planck mass, provided $H_\uin$ is small enough, provided contraction lasts long enough. The direct contribution from the scalar field fluctuations to the overall energy density is displayed in the right panel of , and in , we show the minimum value of the initial Hubble rate such that no backreaction occurs, as a function of the equation-of-state parameter. Since this value is always larger than the one coming from the requirement of no shear backreaction, one concludes that the direct contribution of the scalar field fluctuations to the energy density becomes sizeable before its contribution to the shear does. As a consequence, the assumption of working with a test field breaks down before the shear becomes sizeable, and one would need to relax it in order to investigate the dynamics of the universe when it becomes truly anisotropic. Let us note that, in an inflating cosmology, a light ($m\ll H$) spectator scalar field also generates an anisotropic stress, although there is no classical shear instability in that case. In a de-Sitter space-time, once the field has reached the slow-roll attractor, at the background level, it evolves as $\phi(t) = e^{-m^2(t-t_\uin)/(3H)}\phi(t_\uin)$, and if its mode functions are normalised to the Bunch-Davies vacuum, on super-Hubble scales, they are given by $\phi_{k\ll aH}\simeq 2^{\nu-1}\Gamma(\nu)H^{\nu-1/2}a^{\nu-3/2}/(k^\nu\sqrt{\pi})$, where $\nu=3/2\sqrt{1-4m^2/(9 H^2)}$. Using the same procedure as above, this gives rise to 0 \_0& =& . One can see that, up to quickly decaying terms, $\rho_\sigma$ acquires a constant value, of order $H^6/\Mp^2$. Therefore, its relative contribution to the Friedmann equation  is suppressed by $(H/\Mp)^4$ and remains small at sub-Planckian scales. For instance, in the light-field limit, when $m\ll H$, the above expression gives ()\^4. In single-field slow-roll models of inflation, with current measurements [@Aghanim:2018eyx; @Akrami:2018odb] of the scalar power spectrum amplitude $\calP_\zeta = H^2/(8\pi^2\epsilon_1\Mp^2)\simeq 2\times 10^{-9}$ where $\epsilon_1$ is the first slow-roll parameter, and current constraints from the Planck data on the tensor-to-scalar ratio $r=16\epsilon_1<0.1$, one obtains $\rho_\sigma/(3\Mp^2 H^2) < 10^{-24}\epsilon^4$ at the time when observed scales are produced, a tiny value indeed. Let us note however that, compared to the Friedmann equation, the relative contribution of the shear to the Raychaudhuri equation  is enhanced by an additional factor $1/\epsilon_1$, since $\dot{H}=-\epsilon_1 H^2$, but using the above formulas the contribution is proportional to $\epsilon^4 (H/\Mp)^2 \calP_\zeta$ and is therefore still negligible. Let us finally point out that the energy density contained in the scalar field fluctuations reads $\langle \rho_\phi \rangle \simeq m^2 \langle\bar{\phi}^2\rangle/2 \simeq \epsilon^{3-2\nu} 2^{2\nu-5} 3 \Gamma^2(\nu) H^4/\pi^3$, where one has used to evaluate $\langle\bar{\phi}^2\rangle$. This is in fact larger than $\langle \rho_\sigma \rangle$, so $\phi$ would stop acting as a test field before it generates a substantial shear, as in the contracting case discussed in . There is therefore no shear backreaction problem in inflating cosmologies. This is in agreement with , where, instead of working with Bianchi I cosmologies, the effect of quantum anisotropies on quasi de-Sitter backgrounds was studied in the context of Taub cosmologies. In , the classical isotropisation of inflating space times in the presence of SU(2)-gauge fields coupled with an axion was also studied, and it was shown that some initial configurations lead to a premature termination of slow-roll inflation if the axion drives inflation. In , the case of inflating Bianchi I space times in the presence of a U(1) gauge field coupled to the inflaton was also investigated, and the stochastic formalism was employed to show that the quantum fluctuations in the gauge field can produce a large statistical anisotropy. Conclusion {#sec:Conclusion} ========== Isotropic contracting cosmologies with $w<1$ are unstable fixed points, since any amount of initial shear leads to an effective contribution to the energy density that grows as $a^{-6}$. Nonetheless, if one sets initial conditions in an exactly isotropic universe, classically, the universe remains isotropic. When quantum fluctuations are included however, they make the system diffuse away from the isotropic fixed point, and the classical instability then has the potential to drive a shear backreaction problem. In this work, we have quantified this effect if the anisotropic stress arises from quantum fluctuations of a test, massless scalar field. We have found that, if $w>-1/9$, the shear contribution to the overall energy density reaches an attractor that remains negligible until the energy density of the universe reaches the Planck scale, where a bounce is expected to take place (and close to which our treatment does not apply anyway). If $w<-1/9$ however, the shear backreaction can become substantial before one reaches the Planck scale, if the contracting phase lasts long enough (which signals the presence of an IR divergence for an infinite contraction). In either case, we have also found that the direct contribution from the scalar field quantum fluctuations to the energy density of the universe is always more important than that coming from the shear. As a consequence, in order to study the dynamics of the universe as it becomes truly anisotropic, one would first need to incorporate the backreaction of the scalar field on the overall energy density. Let us also note that the stochastic formalism was already employed in the context of contracting cosmologies in , where the case of a contracting phase driven by a scalar field with an exponential potential was studied. In such systems, phase space possesses classical unstable fixed points. Even if one sets initial conditions exactly on such fixed points, quantum fluctuations drive the system away from these configurations, and the outcome of this quantum instability was found to depend on the equation-of-state parameter during the contraction: if $w=1/3$, there is no substantial backreaction before one reaches the Planck scale, while if $w=0$, the quantum instability can become effective before the Planck scale if contraction lasts long enough. Those results are therefore very similar to ours. Although our calculation was performed for a massless spectator scalar field, it can be easily generalised to other setups. For instance, if the mass of the field is not vanishing, the dynamics remains linear and the methods presented in this work can still be directly applied. However, in that case, three physical scales play a role in the dynamics of a given Fourier mode: its physical wavenumber $k/a$, the Hubble scale $H$, and the mass of the field $m$. In the asymptotic past, both $k/a$ and $H$ become small, but $k/a$ takes over $H$ if $w>-1/3$ and the effect of the expansion becomes negligible. However, since $m$ is constant, it provides the dominant contribution to the frequency of the field fluctuations in the asymptotic past. Since they cannot be described as massless fluctuations on a flat space-time in that limit, the Bunch-Davies vacuum cannot be imposed, which makes the choice of the initial quantum state more involved. The geometry in which the anisotropic degrees of freedom are allowed to develop could finally be extended to more generic setups than Bianchi I metrics. The case of non-test fields could also be studied. When the contribution of the anisotropic stress to the background dynamics becomes sizeable, its backreaction could be modelled with the stochastic formalism (provided the separate-universe approach holds in that case), which may help to understand how the geometry of the universe behaves as one approaches the bounce. This would notably imply to compute the field perturbations on an anisotropic configuration, and to incorporate anisotropic noises in the Langevin equations. We leave these extensions for future work. It is a pleasure to thank Emmanuel Frion, Tays Miranda, Patrick Peter and David Wands for very interesting comments and discussions. Comparison between stochastic and quantum-field-theoretic calculations {#sec:app:QFT} ====================================================================== In this appendix, we explain how our results could be obtained by performing a full-quantum-field theoretic (QFT) calculation, and show that the stochastic procedure employed in this work provides a good approximation to the full QFT result. The stochastic procedure leading to the expectation value of $\rho_\sigma$ relies on two approximations. First, by identifying the quantum operators $\hat{\xi}_\phi$ and $\hat{\xi}_\phi$, defined in and , with their stochastic counterparts, with Gaussian stochastic noises sharing the same two-point function , non-vanishing commutators are neglected. However, in the present case, since $\langle \hat{\rho}_\sigma \rangle$ involves four-point functions of $\phi$ only (it does not involve $\pi$, the momentum conjugated to $\phi$), there is no effect from non-vanishing commutators in that quantity, and the stochastic procedure is exact with that respect (see ). Second, once a given Fourier mode has joined the coarse-grained sector, once it has crossed out the coarse-graining scale $k_\epsilon(t)$, it is evolved with the background equations of motion  and  while it should in principle be evolved with the full equations of motion  and , where the gradient terms are not neglected. The non-inclusion of gradient terms on super coarse-graining scales leads to an error that we now try to characterise. We consider a free spectator scalar field and employ the same vectorial notations as in : $\bar{\boldsymbol{\Phi}}:=\left(\bar\phi,\bar\pi\right)^\mathrm{T}$, $\boldsymbol{\xi}:=\left(\xi_\phi,\xi_\pi\right)^\mathrm{T}$ and $\boldsymbol{\Phi}_k:=\left(\phi_k,\pi_k\right)^\mathrm{T}$. The scale of coarse-graining, $k_\epsilon(t)$, is labelled by the small parameter $\epsilon$. It is a bijective function from $t$ to $k$ ($k=k_\epsilon(t)$), and we denote its inverse bijection from $k$ to $t$ as $k_\epsilon^{-1}$ ($t=k_\epsilon^{-1}(k)$). Following the prescription of , we let $k_\epsilon(t) = \epsilon a(t) \vert H(t) \vert$, and use a step function for the window function, see , which leads to $\partial W(\vec{k},t)/\partial t = -\dot{k}_\epsilon \delta(k-k_\epsilon)$. In vectorial notations, - can be recast as =-(t). The Langevin equations  and , and the equations of motion for the mode functions  and , read &=&\_[(0)]{}|+,\ \_k&=&\_[(k)]{}\_k, with \_[(k)]{}(t):=( [cc]{} 0 & a\^[-3]{}\ -a\^3(+m\^2) & 0 )=\_[\_[(0)]{}]{}+\_[\_[(k)]{}]{}. The Green function introduced in generates the solutions of the classical background dynamics, driven by $\boldsymbol{A}_{(0)}$. For this reason, we denote it $\boldsymbol{G}_{(0)}(t,s)$ in this appendix, while $\boldsymbol{G}_{(k)}(t,s)$ stands for the solution of the quantum mode functions, is the solution of where $\boldsymbol{A}_{(0)}$ is replaced with $\boldsymbol{A}_{(k)}$. With these notations, on the one hand, and  can be recast as \[eq:Xi:stochastic\] \_[t,t’]{}(’-) &=& \_[t\_]{}\^[(t,t’)]{} s\ & & \_[(0)]{}(t,s)\_[k\_(s)]{}(s)\^\_[k\_(s)]{}(s)\^\_[(0)]{}(t’,s) . On the other hand, a full QFT calculation would instead lead to the following covariance matrix \_[t,t’]{}\^(’-) &=& \^[k\_\[(t,t’)\]]{}\_[k\_(t\_)]{}k\^3 (k’-) \_[k]{}(t)\^\_[k]{}(t’) . To make the comparison between these two expressions explicit, let us first note that \_[k]{}(t)=\_[(k)]{}(t,s)\_[k]{}(s). In the above, the time $s$ can be arbitrarily chosen, with the only constraint that is has to be anterior to $t$. Let us thus choose $s=t_k=k^{-1}_\epsilon(k)$, which is possible thanks to the bijective nature of $k_\epsilon(t)$. This leads to \_[t,t’]{}\^(’-) &=& \^[k\_\[(t,t’)\]]{}\_[k\_(t\_)]{}k\^3 (k’-) \_[(k)]{}(t,t\_k)\_[k]{}(t\_k)\^\_[k]{}(t\_k)\^\_[(k)]{}(t’,t\_k).\ Let us now perform the change of integration variable $k\to s$ defined by $k=k_\epsilon(s)$. Then $\dd k^3=\dd s \left(\frac{\dd k^3_\epsilon}{\dd s}\right)$ and $t_k$ becomes $t_{k_\epsilon(s)}=k^{-1}_\epsilon[k_\epsilon(s)]=s$. One finally obtains \_[t,t’]{}\^(’-)&=&\^[(t,t’)]{}\_[t\_]{}s\ & & \_[(k\_(s))]{}(t,s)\_[k\_(s)]{}(s)\^\_[k\_(s)]{}(s)\^\_[(k\_(s))]{}(t’,s). \[eq:covqft\] Comparing this expression with , one clearly sees that the stochastic approach provides a good approximation of the full QFT result if the scale of coarse graining is chosen such that $\boldsymbol{G}_{(k_\epsilon(s))}(t,s)$ is sufficiently close to $\boldsymbol{G}_{(0)}(t,s)$. In other words, the stochastic calculation is valid if gradient terms can be neglected in the evolution of quantum modes at scales larger than the coarse-graining scale. In order to make this statement more quantitative, let us compare $\boldsymbol{G}_{(0)}(t,s)$ and $\boldsymbol{G}_{(k)}(t,s)$ in more details. As explained above, they satisfy the differential equations \[eq:ode:G0\] \_[(0)]{}(t,s) &=& \_[(0)]{} (t)\_[(0)]{}(t,s) + (t-s)\ \_[(k)]{}(t,s) &=& \_[(k)]{}(t,s) + (t-s), see . By taking the difference between these two equations, one obtains = \_[(0)]{} (t) + \_[(k)]{} (t) \_[(k)]{}(t,s) . Let us see this relation as a linear, first-order differential equation for $ \boldsymbol{G}_{(k)}(t,s) - \boldsymbol{G}_{(0)}(t,s)$, with a source term given by $\boldsymbol{V}_{(k)} (t) \boldsymbol{G}_{(k)}(t,s)$. By variation of constants, this equation can be formally solved as \_[(k)]{}(t,s) - \_[(0)]{}(t,s) = \_s\^[t]{}s’ \_[\_[(0)]{}(t,s’)]{} \_[(k)]{} (s’) \_[(k)]{}(s’,s), where $\mathcal{T}$ denotes the time-ordering operator, and where one recognises the formal solution to , $\boldsymbol{G}_{(0)}(t,s')$. This gives rise to \[eq:Dyson\] \_[(k)]{}(t,s)=\_[(0)]{}(t,s)+\^t\_ss’\_[(0)]{}(t,s’)\_[(k)]{}(s’)\_[(k)]{}(s’,s). This is only an implicit solution since $\boldsymbol{G}_{(k)}(t,s)$ appears on both sides of this relation. However, since both $\boldsymbol{V}_{(k)}$ and $\boldsymbol{G}_{(k)}-\boldsymbol{G}_{(0)}$ are suppressed by powers of $k^2$, this allows one to make a perturbative expansion in $k^2$ (this is the so-called Dyson series procedure) to relate $\boldsymbol{G}_{(k)}$ to $\boldsymbol{G}_{(0)}$. The first order of this expansion is obtained by replacing $\boldsymbol{G}_{(k)}(s',s)$ by $\boldsymbol{G}_{(0)}(s',s)$ in the right-hand side of , leading to \[eq:Dyson:1stOrder\] \_[(k)]{}(t,s) = \_[(0)]{}(t,s)+\^t\_ss’\_[(0)]{}(t,s’)\_[(k)]{}(s’)\_[(0)]{}(s’,s)+(k\^4). The integral on the right-hand side of this expression can be computed making use of . Denoting $\mathcal{I}(t_1,t_2)\equiv \int_{t_1}^{t_2}\dd t/a^3(t)$, one has \_[(0)]{}(t,s’)\_[(k)]{}(s’)\_[(0)]{}(s’,s) = -k\^2 a(s’) ( [ccc]{} (s’,t) & & (s’,t)(s,s’)\ 1 & & (s,s’) ). If the equation-of-state parameter of the universe, $w$, is constant, then $a(t)=a_\uin (t/t_\uin)^{\frac{2}{3(1+w)}}$, and the integral $\mathcal{I}$ as well as the integration over $s'$ in can be performed. This gives rise to \[eq:Dyson:def:M\] \_[(k)]{}(t,s) - \_[(0)]{}(t,s) = (t,s)+, where the matrix $\boldsymbol{M}(t,s)$ has entries \[eq:M00\] M\_[00]{}(t,s)&=& (3w+5)+3(w-1) \^[3w+1]{} .\ & & . -2(3w+1)\^[(w-1)]{}\ M\_[10]{}(t,s)&=&1-\^\ M\_[01]{}(t,s)&=& (3w+5)-(3w+5) \^ .\ & & . -(9w-1)\^[(w-1)]{}+(9w-1)\^[3w+1]{}\ \[eq:M11\] M\_[11]{}(t,s)&=& 3(w-1)+(3w+5) \^[3w+1]{} .\ & & . -2(3w+1)\^. With $k_\epsilon(s) = \epsilon a(s) H(s)$, gives rise to $\boldsymbol{G}_{(k_\epsilon(s))}(t,s) - \boldsymbol{G}_{(0)}(t,s) = \epsilon^2 \boldsymbol{M}(t,s)+\order{k^4}$, so taking the difference between and  leads to & & -2em \_[t,t’]{}\^(’-) - \_[t,t’]{}(’-) =\ & & \^2\^[(t,t’)]{}\_[t\_]{}s (t,s)\_[s]{}\^\_[(0)]{}(t’,s) + \_[(0)]{}(t,s)\_[s]{}\^(t’,s) + . In this expression, $\boldsymbol{G}_{(0)}$ is given by , $\boldsymbol{\Sigma}$ by and $\boldmathsymbol{M}$ by -. For explicitness, and in order to avoid displaying cumbersome formulas, let us evaluate this formula in the coincident configuration where $\vec{x}'=\vec{x}$ and $t=t'$. One obtains \[eq:Xi:QFT:Sto:Comp\] \_[t,t]{}\^(0) - \_[t,t]{}(0) = \^2 (t) + , where the matrix $\boldsymbol{N}(t)$ has entries \[eq:N00\] N\_[00]{}&=&\ N\_[01]{}&=&N\_[10]{}=\ N\_[11]{}&=& a\^6H\^4 . \[eq:N11\] This needs to be compared with the stochastic result we have been using, namely -2em \_[t,t]{}(0) &=& \_[t\_]{}\^t s \_[(0)]{}(t,s)\_[s]{}\^\_[(0)]{}(t,s)\ & = & ( [cc]{} & a\^3 H\^3\ a\^3 H\^3 & (w-1)\^2 a\^6 H\^4 ) . \[eq:Xi:sto:explicit\] Let us note that in the square bracketed terms of -, soon after the onset of the contracting phase, since $a/a_\uin \ll 1$, the dominant term is the one with the smallest power of $a/a_\uin$. Assuming that $w\in ]-1/3,1[$, this power is either $0$, when $w>0$, or $6w$ when $w<0$. Indeed, $3w+1>0$ for $w>-1/3$ so $3/2(3w+1)$ is never the smallest power index, and $9w+1>6w$ for $w>-1/3$ so $9w+1$ is never the smallest power index either. This means that, as in , the result either scales as $(a/a_\uin)^0$ or as $(a/a_\uin)^{6w}$. It is then easy to see that the prefactors appearing in and  are the same, up to overall constants of order one, and up to the (crucial) additional $\epsilon^2$ factor appearing in . One concludes that, provided $\epsilon\ll 1$, the stochastic result provides a good approximation to the full QFT result. This has been derived here for the coincident configuration of the field correlators, but this can be easily extended to the non-coincident configuration, and to the two-point function of the anisotropic stress since the two are simply related via . This validates the use of the stochastic formalism to perform the calculation presented in this work. [^1]: From the normalisation condition of $u^\mu$, $u_\mu u^\mu=-1$, one can indeed check that (i) $\perp_{\mu\nu} u^\nu =0$, (ii) if $v^\mu$ is a vector orthogonal to $u^\mu$, $u_\mu v^\mu=0$, then $\perp_{\mu\nu} v^\nu = v_\mu$, and (iii) $\perp_{\mu}^\rho \perp_{\rho\sigma} = \perp_{\mu\nu}$. [^2]: In general, the choice of the time coordinate with which the Langevin equations are derived determines the gauge in which the calculation is performed [@Pattison:2019hef] (namely the synchronous gauge for that time variable). When the field perturbations couple to metric perturbations, this sets the gauge in which the noise needs to be computed. However, in the present case of a test field, such a coupling is neglected and there is no subtlety associated with the choice of time coordinate. [^3]: In the non-coincident configuration, if one expands + = + ++, one finds \_[t,t’;(i,j)]{}(,’) &=& 0 \_i|(,t) \_j|\^(’,t’) 0\ &=& - - ,\ where $\boldmathsymbol{\Xi}^\prime_{t,t'}(\vert \vec{x}'-\vec{x}\vert)$ and $\boldmathsymbol{\Xi}^{\prime\prime}_{t,t'}(\vert \vec{x}'-\vec{x}\vert)$ can be made explicit making use of and . However, the coincident limit of the resulting expression is singular since $\lim_{\vert\Delta\vec{x}\vert\to 0} \Delta x_i \Delta x_j/\vert\Delta \vec{x}\vert^2$ is ill-defined (it depends on the direction of $\Delta\vec{x} \equiv \vec{x}'-\vec{x}$ along which this limit is taken), which explains why we treat the coincident configuration separately. [^4]: Here this result has been obtained by first coarse-graining the scalar field, and then calculating its two-point function and associated anisotropic stress. We have checked that by first deriving the full anisotropic stress, which includes all scales, and then coarse-graining the result, one obtains the same formula. [^5]: Either the bounce is classical and we can expect the maximal value of $|H|$ to be below the Planck scale; or the bounce is of quantum-gravitational origin, hence expected to occur at the Planck energy density, as can explicitly be proven in Loop Quantum Cosmology for example. [^6]: In this case, by computing the “$(ij)$” component of the stress-energy tensor in , one can see that the pressure associated with $\phi$ takes the same form, $p_\phi=\rho_\phi$. This implies that the direct fluctuations of the scalar field also have an effective equation-of-state parameter $w=1$ (provided the potential vanishes), as for the shear (for which $w=1$ even in the presence of a non-vanishing potential). At the classical level, these two therefore backreact in the same way, but at the quantum level, they behave differently, see and .
{ "pile_set_name": "ArXiv" }
--- abstract: 'Topological insulators (TIs) hold great promise for realizing zero-energy Majorana states in solid-state systems. Recently, several groups reported experimental data[@PhysRevLett.109.056803; @PhysRevLett.110.186807; @Appll; @sreport; @NatMat.11.417; @cihan] suggesting that signatures of Majorana modes in topological insulator Josephson junctions (TIJJs) have – indeed – been observed. To verify this claim, one needs to study the topological properties of low-energy Andreev-bound states (ABS) in TIs of which the Majorana modes are a special case. It has been shown theoretically that topologically non-trivial low-energy ABS are also present in TIJJs with doped topological insulators up to some critical level of doping at which the system undergoes a topological phase transition[@PhysRevLett.107.097001; @PhysRevLett.109.237009; @PhysRevB.87.035401]. Here, we present first experimental evidence for this topological transition in the bulk band of a doped TI. Our theoretical calculations, and numerical modeling link abrupt changes in the critical current of top-gated TIJJs to moving the chemical potential in the charge-accumulation region on the surface of the doped TI across a band-inversion point. We demonstrate that the critical-current changes originate from a shift of the spatial location of low-energy ABS from the surface to the boundary between topologically-trivial and band-inverted regions after the transition. The appearance of a decay channel for surface ABS is related to the vanishing of the band effective mass in the bulk and thus exemplifies the topological character of surface ABS as boundary modes. Importantly, the mechanism suggest a means of manipulating Majorana modes in future experiments.' author: - 'V. Orlyanchik$^{1}$, M. P. Stehno$^{1}$, C. D. Nugroho$^{1}$, P. Ghaemi$^{1}$, M. Brahlek$^{2}$, N. Koirala$^{2}$, S. Oh$^{2}$, and D. J. Van Harlingen$^{1}$' bibliography: - 'gatedTI.bib' title: Signature of a topological phase transition in the Josephson supercurrent through a topological insulator --- Department of Physics, University of Illinois at Urbana-Champaign, Urbana, IL 61801, USA Department of Physics and Astronomy, Rutgers, The State University of New Jersey, Piscataway, New Jersey 08854, USA In small-gap semiconductors strong spin-orbit interactions may cause an inversion between valence and conduction bands generating a new class of insulators which is called topological insulator (TI)[@NatPhys.5.378; @RevModPhys.82.3045; @RevModPhys.83.1057]. One intriguing feature of these systems is the emergence of gapless (i.e. metallic), spin-momentum-locked states on the interface with ordinary insulators, e.g. a region with topologically-trivial bandstructure or vacuum. There has been growing interest in the properties of ordered phases of these helical states. Although, in two spatial dimensions, fluctuations prohibit spontaneous symmetry breaking[@PhysRevLett.17.1133], new phases may be induced by the proximity effect when magnetic or superconducting materials are brought into contact with TIs. Coupling TIs and conventional superconductors (SC) introduces helical surface states with superconducting pair-correlations and is of particular interest in the present context. It was argued that in-gap vortex modes[@caroli], and Andreev states in $\pi$-Josephson junctions of conventional superconductors which are coupled by TI surface states may carry a zero-energy Majorana mode[@PhysRevLett.100.096407], a fermionic mode which is its own antiparticle[@NatPhys.5.614]. Early models considered an idealized scenario in which helical states existed only on a single, isolated surface of the TI, and the bulk of the material was perfectly insulating. As most of the initially discovered TIs were not true insulators in the bulk, it seemed difficult to realize the proposal in experiments. Later it was shown theoretically that topologically protected zero-energy modes exist also on the surface of doped, superconducting TIs up to a critical level of doping. At this doping level, a topological phase transition occurs in the superconducting TI[@PhysRevLett.107.097001; @PhysRevLett.109.237009; @PhysRevB.87.035401]. Several groups demonstrated supercurrent transport in SC-TI-SC hybrid structures[@NatMat.11.417; @ZhangInducedSCinBiSe; @MorpurgoBiSe; @PhysRevLett.109.056803] as well as the ability to control the chemical potential and the magnitude of supercurrent in a TI by electrostatic gating[@MorpurgoBiSe; @NadyaBiSe]. Furthermore, it was argued[@NatMat.11.417; @NadyaBiSe] that the majority of the supercurrent is carried by a set of low-energy Andreev bound states located on the surface of the TI. Recently, a new generation of high-quality, bismuth-selenide TI material became available for which it was shown that only a few quintuple layers close to the surface contribute to electrical transport[@ThicknessIndependentTRansport]. It was then concluded that the electric current is carried by the helical surface states and by carriers in a charge-accumulation region close to the surface of the TI where the conduction band bends down and crosses the chemical potential. As a result, a 2-dimensional electron gas (2DEG) layer is formed below the surface whereas the chemical potential in the bulk of the TI remains in the gap of the TI. We incorporate this material into a Josephson device which provides us with a unique opportunity to study the rôle of low-energy Andreev states and band topology in supercurrent transport as we are able to adjust the chemical potential of the 2DEG layer within the conduction band and in the gap. Here, we report measurements of the [D]{}[C]{} Josephson effect in SC-TI-SC devices and focus on the dependence of the critical current ($I_{C}$) on the voltage that is applied to a top-gate which is used for electrostatic gating. Unlike the normal state resistance, which is a smooth function of gate voltage throughout, the Josephson current exhibits an abrupt drop followed by a gradual decrease as the chemical potential is lowered. We attribute the sharp change in $I_{C}$ to a topological transition in the conduction band of the charge accumulation region which shifts the spatial location of low-energy Andreev-bound states pinned on the boundary between parts of the sample with topologically-trival and band-inverted bandstructure. The samples in this study were fabricated from high-quality Bi$_{2}$Se$_{3}$ thin-films grown by molecular-beam epitaxy on sapphire (Al$_{2}$O$_{3}$) (0001) substrates[@GrowBiSE1; @ThicknessIndependentTRansport]. The thickness of the films varied between 6 QL and 60 QL. To pattern planar Josephson junctions, we used standard electron-beam lithography techniques. The length of the junction was set by the separation between two sputtered Nb electrodes and varied between $90$ nm and $120$ nm. The junction width was defined by dry-etching and ranged between 0.1 $\mu$m and 1 $\mu$m. The top-gate was fabricated by evaporating 70 nm of Au on top of the 40 nm-thick dielectric layer of ALD-grown alumina (Al$_{2}$O$_{3}$) or hafnia (HfO$_{2}$). Most of the measurements were carried out in a dilution refrigerator equipped with a 10 Tesla superconducting magnet and a base temperature of 8 mK. The data, we present here, were obtained on a junction of 0.1 $\mu$m length, 0.4 $\mu$m width, and a thickness of 15 QL and are representative for all 15 measured samples. The current-voltage (I-V) characteristics of a Nb-Bi$_{2}$Se$_{3}$-Nb junction is shown in Fig. \[fig:I-V\](a) for different sample temperatures. Here, we define the critical current ($I_{C}$) of the device as the current at which a finite voltage drop develops between the two superconducting electrodes. A maximum Josephson supercurrent of $I_{C}^{max}=190$ nA was measured at temperature $T=45$ mK. Other devices showed values of $I_{C}^{max}$ between 10 nA and 300 nA scaling with the width of junction. The hysteresis in the low-temperature I-V characteristics is consistent with electron heating which develops after switching the junction into the resistive state[@HysteresisSNS]. As shown in Fig. \[fig:I-V\](b), a Fraunhofer-like pattern is generated by applying a perpendicular magnetic field which modulates the critical current in the device. While the general shape of the curve is typical for Josephson devices, the details specific to SC-TI-SC junctions have been discussed elsewhere[@cihan]. [![(a) The current-voltage characteristics of the Josephson junction at different temperatures. Inset: A schematic representation of the junction cross-section. (b) The dependence of the critical current on perpendicular magnetic field exhibits a Fraunhofer-like pattern.[]{data-label="fig:I-V"}](Fig1_1.pdf "fig:"){width="7.5cm"}]{} Having verified the presence of Josephson coupling in our devices, we turn to the doping dependence of the Josephson current which was studied by electrostatic depletion of charge carriers in the junction. The intrinsic n-doping that we observed in our devices is commonly attributed to selenium vacancies forming in MBE-grown Bi$_{2}$Se$_{3}$ films shortly after the growth is completed[@EviromentalDisorder1; @EviromentalDisorder2]. The excess electronic density raises the Fermi energy in the device and can be removed effectively by applying a negative voltage to the top-gate of the junction (cp. device schematics in the inset of Fig.\[fig:I-V\](a)). A change in charge-carrier density affects resistance and critical current of the junction in different ways. As shown in Fig. \[fig:Ic-Vg\](a), applying a negative gate voltage reduces the critical current in a non-monotonic manner. The initial, gradual reduction of $I_C$ is followed by a rather abrupt drop. This rapid change in critical current takes place in a narrow region of gate voltages, $\Delta V_{G}<1~$V, which we call transition region and mark it by its center value, the critical gate voltage $V_{G}^{C}\approx -12$ V. The transition region is trailed by another gradual and featureless decrease of $I_{C}$. This highly non-monotonic behavior of $I_{C}(V_{G})$ is reproducible. It was observed in multiple, consecutive gate voltage scans for each individual Josephson junction, and it was present in all measured devices (15 Josephson junctions). The value of $V_{G}^{C}$ varies accordingly with thickness and dielectric constant of the gate dielectric (Al$_{2}$O$_{3}$ or HfO$_{2}$). In contrast, the abrupt change in $I_{C}$ was *not* accompanied by a fast variation in the normal state resistance of the junction, which is a smooth function of $V_G$ (see inset in Fig. \[fig:Ic-Vg\](a)). Moreover, as shown in Fig. \[fig:Ic-Vg\](b), the product of critical current and normal state resistance, $I_{C}R_{N}$, which reflects the characteristic energy of the Josephson coupling, shows a distinct step at $V_{G} \approx V_{G}^{C}$. [![(a) The critical current as a function of the gate voltage measured at $T=50$ mK. Inset: Dependence of the normal state resistance of the junction on gate voltage. (b) The variation of the product of the critical current and normal state resistance ($I_{C}R_{N}$) measured as a function of gate voltage. The dashed lines emphasize two values of $I_{C}R_{N}$ above and below the transition region.[]{data-label="fig:Ic-Vg"}](Fig2_1.pdf "fig:"){width="9.5cm"}]{} We propose that the abrupt change in the critical current is a manifestation of a topological phase transition in the charge-accumulation region of the sample. Below, we argue that a considerable fraction of the supercurrent is carried by Andreev-bound states (ABS) which are pinned on the interface ($\it{boundary}$ ABS) between the topologically-trivial region and the band-inverted insulating bulk of the TI. Electrostatic gating changes the bending of the conduction band at the surface and, thus, moves the location where the chemical potential intersects with respect to the bottom of the band. This allows the 2DEG layer to transition between a topologically-trivial and a band-inverted phase. To demonstrate the mechanism behind the topological transition, we developed a model based on the concepts laid out in Refs.[@PhysRevLett.107.097001; @PhysRevLett.109.237009; @PhysRevB.87.035401]. It was shown that low-temperature transport in MBE-grown Bi$_{2}$Se$_{3}$ films is described fully by two conductance channels, the topological surface and a quantum-confined electron gas (2DEG) which is a few quintuple layers thick. The latter results from a downward bending of the conduction band due to charge-accumulation in a few nm-wide zone below the film surface[@ThicknessIndependentTRansport; @TranspotinBiSE1]. Schematically, we can picture the SC-TI-SC Josephson junction as a two-layered structure of doped TI material comprised of an insulating bulk (i.e. the chemical potential is in the bulk gap) and the charge-accumulation region. Due to the short charge-screening length in the TI, we assume that electrostatic gating acts only on the 2DEG layer. This allows us to adjust its chemical potential, which – for simplicity – is taken to be constant throughout the layer. Looking at the cross-section of the TIJJ, we identify two important boundaries: $L_{1}$ is the physical boundary of the TIJJ between the 2DEG and vacuum (or the gate dielectric in an actual device, see Fig.\[fig:Fig3\](a)), and $L_{2}$ is the boundary between the insulating bulk of the TI and the 2DEG (see Fig. \[fig:Fig3\](b)). Let us first assume that the chemical potential in the 2DEG layer is well within the insulating gap (i.e. a large negative voltage is applied to the top-gate), such that all 2DEG states are depopulated (see Fig.\[fig:Fig3\](c)). In this case, the 2DEG region is part of the insulating bulk. The proximity effect induces pair-correlations in the topologically-protected surface layer, $L_{1}$, and the supercurrent is carried exclusively by ABS close to the physical surface of the TIJJ. Since the occupation probability of ABS follows a thermal distribution, i.e. $I_C \propto - \sum_{E_n \geq 0} \frac{\partial E_n}{\partial \phi} \tanh\left({\frac{E_n}{ 2k_B T}}\right)$, at low temperatures a large fraction of the supercurrent is carried by the ABS which are lowest in energy ($E_n$ is the energy of the $n$-th ABS for a phase difference of $\phi$ across the junction, $k_B$ is the Boltzmann constant, and $T$ is the temperature). Typically, close to $\phi = \pi$, the slope $\frac{\partial E_n}{\partial \phi}$ is largest for the lowest ABS band in a TIJJ[@PhysRevB.86.214515], and under appropriate conditions this band may host a zero-energy Majorana state. By contrast, when the chemical potential enters the 2DEG conduction band, superconductivity is induced in this region as well, and a new set of ABS emerges, which extends throughout the 2DEG layer. The energy of 2DEG ABS exceeds that of the boundary ABS[@PhysRevLett.107.097001] located at $L_{1}$ (see supplementary materials) and, thus, they are decoupled from ABS on the physical surface. Notice that they do not provide a decay – or delocalization – channel for Majorana modes indicating that the band topology has not changed. Indeed, previous theoretical work[@PhysRevLett.107.097001; @PhysRevB.87.035401; @PhysRevLett.109.237009] suggests that the energy of ABS in a doped TI depends on the chemical potential in a nontrivial way relating to the topology of the band. In particular, the energy of the lowest ABS is minimal and equal to the energy of the boundary ABS when the chemical potential reaches the point of a topological transition in the conduction band (i.e. the band effective mass is zero). Below, we outline the energetics of ABS close to a topological phase transition and argue that they drive the step-like change in the critical current of our TIJJs. For a detailed theoretical discussion, we refer the reader to the supplementary material with this Letter. To study the low-energy ABS in the 2DEG, we use the low-energy effective Hamiltonian of a doped TI with $s$-wave superconducting pair-correlations: $$H_\mathrm{eff}=\left[v_F\left(k_x \sigma_x+k_y \sigma_y+k_z \sigma_z \right) \tau_x+m \left(|\textbf{k}| \right) \tau_z - \mu \right] \eta_z + \Delta\left(\cos(x) \eta_x+\sin(x) \eta_y \right)\ .$$ \[eq.Hamiltonian\] Here, $\eta_i$ are Pauli matrices that act on superconducting particle- and hole-states, and the representations $\tau_i$ and $\sigma_i$ are chosen for orbital- and (physical) spin-degrees of freedom[@PhysRevLett.107.097001]. The Hamiltonian is translational invariant along the $\hat{y}$-direction, which is in-plane and perpendicular to the current flow in the junction (see Fig.\[fig:Fig3\](a)), and $k_{y}$ denotes the wave vector along $\hat{y}$. The charge-carrier concentration is determined by the chemical potential in the 2DEG region which controls the accessible Fermi momenta $\textbf{k}$ in the band structure. These enter the equations in form of the momentum-dependent effective mass term for the conduction band, $m(|\textbf{k}|)$. A simplified form of the effective mass term, which captures the essential features of the bulk band structure of TIs, is $m \left(|\textbf{k}| \right)=M-\epsilon |\textbf{k}|^2 $ with $M\cdot \epsilon>0$. As mentioned above, at lower doping levels (i.e. closer to the Dirac point where $k = |\textbf{k}|<M/\epsilon$), the effective mass has the same sign as the mass of the bulk TI, which is given by $m \left(0\right)=M$. The low-energy ABS are localized at $L_1$, the physical surface of the TIJJ. Further increasing the chemical potential decreases the magnitude of the effective mass $|m(k)|$, which vanishes when $k=\sqrt{M/\epsilon}$. Beyond this point, the mass term reverses its sign with respect to the bulk mass $M$. As derived in the supplementary material section, the energy difference between 2DEG ABS and boundary ABS at $L_1$ is given approximately by: $$|\Delta E| \propto \alpha \frac{m\left(|\textbf{k}_f|\right)^2}{m\left(|\textbf{k}_f|\right)^2+v_F^2 |\textbf{k}_f|^2}\ .$$ \[eq.deltaE\] where $\alpha >0$ is of order of $|\Delta|^2/\mu$. Hence, for $m\left(|\textbf{k}_f|\right) = 0$, the energy of 2DEG ABS is equal to the surface bound-state energy, which opens a decay channel for the ABS at $L_1$. At the same time, the 2DEG layer becomes a region with topologically-trivial bandstructure similar to an ordinary superconducting metal and, thus, provides a de-localization path for surface ABS. As we increase the chemical potential further, we move the topological boundary to $L_2$, the interface between the insulating TI and the 2DEG region (see Fig.\[fig:Fig3\](b)). Here, a new set of low-energy boundary ABS, which carries a significant portion of the supercurrent, appears after the topological transition of the 2DEG layer. To confirm the conjecture regarding the distribution of the supercurrent within the sample, we performed a series of numerical calculations that map out the evolution of the spatial location of low-energy ABS as a function of chemical potential. We used a simple four-band model for strong topological insulators[@PhysRevLett.107.097001; @PhysRevB.87.035401] and considered $s$-wave superconducting pairing at the mean-field level. The energy of ABS is derived using exact diagonalization (cp. supplementary material section). Figures \[fig:Fig3\](f) and (e) show – respectively – the amplitudes of the low-energy ABS and the band structure used in the calculations. For the chemical potential situated within the Dirac cone ($\mu < 1.4$), supercurrent transport is localized close to the surface where the amplitude of the ABS wavefunction is large in Fig. \[fig:Fig3\](f). When the critical chemical potential ($\mu_{C}$) associated with the topological transition is reached, the low-energy ABS migrate abruptly to the interface between the 2DEG layer and the insulating bulk TI. Importantly, the value of $\mu_{C}=1.8$ is significantly above the bottom of the 2DEG conduction band. Increasing the chemical potential beyond $\mu_{C}=1.8$ does not move the position of low-energy ABS. [![(a) and (b) show a schematic representation of the location of topological Andreev-bound states corresponding the positions of the chemical potential in (c) and (d). (e) The band structure used in numerical calculations. (f) The evolution of the spatial location of low-energy Andreev-bound states as a function of the chemical potential assuming the band structure in (e).[]{data-label="fig:Fig3"}](Fig3_2.pdf "fig:"){width="13.5cm"}]{} Indeed, close inspection of the $R(V_{G})$-dependence in Fig.\[fig:Ic-Vg\](a) reveals that the inflection point, which is associated with the depletion of the 2DEG band, occurs at $V_{G}^\mathrm{2DEG}\sim -20$ V (marked by an arrow in the inset). This value is distinctly lower than the critical gate voltage, $V_{G}^{C}=-12$ V. Although the position of the chemical potential at which the transition takes place varies slightly between samples (likely reflecting different levels of disorder in the devices), the width of the transition is always much smaller than the distance to the band edge, i.e. $\Delta V_{G} \ll |V_{G}^\mathrm{2DEG}-V_{G}^{C}|$, which is consistent with our assumption that the transition is controlled by properties of the bulk band structure. The pronounced changes in the critical current across the topological transition reflect a difference in the effective transmission of individual ABS. Whereas we may assume that the SC-TI interface barrier is lower and structural disorder is less for ABS at $L_{2}$, the effective thickness of the boundary-zone is larger on this interface. All three aspects enhance supercurrent transport by ABS at $L_{2}$. This explains the observed increase in critical current as the chemical potential is shifted up in energy across the transition and exemplifies a direct consequence of band structure topology on Andreev transport. In turn, this effect suggests a control mechanism for the spatial location of low-energy ABS and, perhaps, Majorana fermions. When the junction is in the dissipative state, the current is carried by ordinary quasi-particle excitations. All conduction channels are expected to contribute, and a strong dependence on the position of the topological boundary is absent. Indeed, the normal-state resistance varies smoothly with gate voltage (i.e. charge-carrier density), cp. inset in Fig.\[fig:Ic-Vg\](a). As a result, a step in the $I_{C}R_{N}$-product marks the crossover between the two topologically-distinct bandstructure configurations of the 2DEG region (see Fig. \[fig:Ic-Vg\](b)) and illustrates the difference in the effective Josephson coupling for supercurrent transport in the two respective locations of the topological boundary. In conclusion, we presented the experimental observation of a topological phase transition in the bulk band of a doped, proximity-coupled 3-dimensional topological insulator. Hallmark of the transition is a shift in the spatial location of low-energy Andreev-bound states which follow the position of the topological boundary. In particular, we demonstrated that the charge-accumulation zone (2DEG channel) close to the surface of a doped 3-d TI can be driven through the transition by shifting the chemical potential electrostatically with a top-gate. This was registered as a jump in the magnitude of the critical current of the Josephson junction. The abrupt change occurred within the bulk band of the 2-DEG region and correlated with a sign-change of the effective mass in the TI band structure model. The transition in the bandstructure topology of the gapped, 2DEG proximity region resulted in an altered current-flow pattern due to a displacement of low-energy ABS. Consequently, the change in the effective Josephson coupling led to a sudden and unusual drop in the $I_CR_N$-product well above the band edge. Notice that in our model for the sudden drop in the critical current, it is not necessary for the bulk of the junction to be insulating. Our model is valid even when the bulk has the fermi level in the conduction band, as long as it is not above the critical chemical potential where the topological phase transition happens. Exploration of Majorana physics in 3-dimensional topological insulators requires control over the spatial location of zero-energy Andreev states. Our result directs toward an efficient way of manipulating low-energy Andreev-bound states in 3-d TI Josephson devices. acknowledgments {#acknowledgments .unnumbered} =============== We would like to thank Cihan Kurter, Aaron Finck, Ashvin Vishwanath, Taylor Hughes, Jeffrey Teo and Eduardo Fradkin for useful discussions. VO, MPS, CDN, and DJVH acknowledge funding by Microsoft Station-Q. For the device fabrication, we acknowledge use of the facilities of the Frederick Seitz Materials Research Laboratory at the University of Illinois at Urbana-Champaign. PG acknowledges support from NSF DMR-1064319. The work at Rutgers was supported by IAMDN of Rutgers University, NSF DMR-0845464 and ONR N000141210456.   **Low energy Andreev states in the bulk and on the surface of doped topological insulators** Gapless helical surface states are considered the characteristic feature of topological insulators. Topological properties of bulk bands are difficult to observe. However, close to a transition, the Andreev-bound state (ABS) spectrum in Josephson junctions depends sensitively on the bulk-band topology which allows us to detect a clear signature of a topological transition in the band. At wave vectors close to different time-reversal symmetric points (TRs), the Hamiltonian of bulk TI bands resembles the three-dimensional massive Dirac Hamiltonian. At the TRs, a relative change in the sign of the mass term in the Dirac Hamiltonian occurs[@PhysRevB.76.045302]. As a result, the minimal effective Hamiltonian for the wave vectors near the bottom of the conduction band of 3-D TIs is given by: $$\label{bh} \mathcal{H}=v_F \tau_x \ {\boldsymbol \sigma} \cdot \textbf{k} + \tau_z \ m(\textbf{k})$$ where ${\boldsymbol \sigma}=(\sigma_x,\sigma_y,\sigma_z)$ are the Pauli matrices acting on spin space, $\tau_x$, $\tau_y$ are the Pauli matrices acting on orbital space, and $\textbf{k}$ is the wave vector relative to the TR at the bottom of the conduction band. The momentum-dependent mass term $m(\textbf{k})=M-\epsilon |\textbf{k}|^2$, with $M\cdot \epsilon >0$, changes sign at $|\textbf{k}| =\sqrt{M/\epsilon}$. It was shown before that the topological properties of the superconducting phase of a doped topological insulator change when the chemical potential corresponds to the wave vectors at which the effective mass $m(\textbf{k})$ vanishes[@PhysRevLett.107.097001; @PhysRevLett.109.237009; @PhysRevB.87.035401]. One manifestation of this topological transition is the appearance of edge modes, e.g. the zero-energy Majorana states at the ends of a vortex passing through a doped topological insulator[@PhysRevLett.107.097001; @PhysRevLett.109.237009; @PhysRevB.87.035401; @PhysRevLett.100.096407]. In this case, the bound states that extend along the vortex[@caroli; @PhysRevLett.107.097001; @PhysRevLett.109.237009; @PhysRevB.87.035401] become gapless at the transition and provide a channel for coupling Majorana modes at both ends of the vortex thus allow them to annihilate. Indeed, zero-modes in the energy spectrum of the superconducting bulk were the first theoretical evidence for a topological phase transition as function of the doping level in doped TIs[@PhysRevLett.107.097001]. Later the result was confirmed by looking directly at the evolution of Majorana states at the end of the vortex as a function of the chemical potential[@PhysRevB.84.144507; @PhysRevB.87.035401]. An analogous scenario is realized in a topological insulator Josephson junction (TIJJ). When the chemical potential is in the bulk gap of the TI, ABS are found only at the TI surface. At a phase difference of $\pi$, two of the localized Andreev states in a TIJJ are zero-energy Majorana modes[@PhysRevLett.100.096407]. On the other hand, when the TI is doped and the chemical potential enters the conduction band, superconductivity is induced in the bulk, and additional ABS are formed throughout the TIJJ. Similar to the topological transition in the vortex, the energy of ABS evolves with the chemical potential in the TI. As we move it through the bulk band of the TI, the shape of the Fermi surface changes, i.e. the Fermi surfaces is composed of different sets of **k**-states. For particular **k**-vectors, the mass term $m(\textbf{k})$ vanishes, and the bandstructure undergoes a topological phase transition. An important signature is that the ABS spectrum becomes gapless thus decay channels for Majorana modes open, which – at this point – can no longer exist on the surface. The presence of Majorana modes on the surface and gapless modes in the bulk of TIJJs corresponds to a phase difference of $\pi$, precisely. Under simplifying assumptions, it can be shown that this condition is fulfilled in TIJJs at maximum critical current[@PhysRevB.86.214515]. We, however, are interested in the full evolution of low-energy Andreev states. Thus, we study the general case of TIJJs with arbitrary phase difference. In what follows, we first derive the energy of ABS on the surface of a narrow junction. Next, we find the energy of bulk ABS and show that it decreases as the magnitude of the effective mass, $|m(\textbf{k})|$, decreases. In particular, we see that – for vanishing effective mass $m(\textbf{k})$ – the (finite) energy of the lowest-lying ABS on the surface will become equal to the energy of the bulk Andreev states even if the phase of the junction is not $\pi$. When the chemical potential is in the bulk band-gap of the TI, the only gapless states are helical states localized at the boundary of the TI sample. They realize the gapless, 2-dimensional Dirac Hamiltonian: $$\label{sdh} \mathcal{H}_s= iv_F{\boldsymbol \sigma} \cdot \textbf{k}$$ where ${\boldsymbol \sigma}=(\sigma_x,\sigma_y)$ are the Pauli matrices in the bases $\left(\psi_\uparrow,\psi_\downarrow\right)$, and $\psi_\sigma$ is the electronic state with spin $\sigma$ localized on the surface of the TI. The low-energy effective Hamiltonian describing a Josephson junction on the surface of the TI with supercurrent along $\hat{x}$ (such that the superconducting $\phi$ varies in $\hat{x}$-direction) is given by[@PhysRevLett.107.097001] $$\label{sh} H=\left(-iv_F{\boldsymbol{\nabla}}\cdot\boldsymbol{\sigma}-\mu\right)\eta_z+\Delta\left[\cos(\phi(x))\eta_x+\sin(\phi(x))\eta_y\right]\ ,$$ with a convenient choice of bases $\left(\psi_\uparrow,\psi_\downarrow,\psi_\downarrow^\dagger,-\psi^\dagger_\uparrow\right)^T$. In this Hamiltonian, the Fermi velocity at chemical potential $\mu$ is denoted by $v_F$, and $\Delta$ is the superconducting gap. The matrices $\boldsymbol{\sigma}$ act on physical spin space (which is locked with the momentum) whereas the $\eta_i$ act on the superconducting particle-hole space. As the Hamiltonian, eqn. \[sh\], is invariant under translation along $\hat{y}$, the momentum $k_y$ in this direction is conserved. The lowest-energy Andreev states in the junction correspond to $k_y=0$ for which equation \[sh\] reduces to: $$\label{Hamil} \left[\left(-iv_F\partial_x\sigma_x-\mu\right)\eta_z+\Delta\left(\cos(\phi(x))\eta_x+\sin(\phi(x))\eta_y\right)\right]|v\rangle=E|v\rangle\ .$$ As an operator, $\sigma_x$ commutes with the above Hamiltonian. Therefore, it is possible to divide the eigenstates into two separate sets with eigenvalue $\langle\sigma_x\rangle=1$, or $-1$: $$\left[\left(\mp iv_F\partial_x-\mu\right)\eta_z+\Delta\left(\cos(\phi(x))\eta_x+\sin(\phi(x))\eta_y\right) \right]|v\rangle=E|v\rangle\ ,$$ where the negative (positive) sign of the first term corresponds to $\langle\sigma_x\rangle = 1$ (or $-1$), respectively. Similarly, at finite, positive chemical potential ($\mu>0$), the right and left Fermi points correspond to states with opposite $\sigma_x$ eigenvalues – $\langle\sigma_x \rangle > 0$ for the right, and $\langle\sigma_x \rangle < 0$ for the left Fermi point, respectively. The eigenvectors $|v\rangle$ have the form $|v\rangle=e^{i k_\pm x}|u\rangle$, where $k_\pm=\pm k_f= \pm \frac{\mu}{v_F}$ are the Fermi momenta at the two Fermi points, and the eigenvectors $|u \rangle$ satisfy: $$\label{Hamils} \left[\left(\mp iv_F\partial_x\right)\eta_z+\Delta\left(\cos\left(\phi(x)\right)\eta_x+\sin\left(\phi(x)\right)\eta_y\right) \right]|u\rangle=E|u\rangle\ .$$ A short Josephson junction is modeled by the following phase distribution: $$\left\{\begin{array}{cll} \phi(x)=& 0 \; &\mathrm{for} \ x<0,\ \mathrm{and} \\ \phi(x)=& \phi_0 \; &\mathrm{for} \ x\ge 0\ . \end{array}\right.$$ Since we are interested in in-gap Andreev states, we define $E=\Delta \cos(\beta)$ with $0 \le \beta \le \pi$. The eigenvector $|u\rangle$ has the form: $$|u\rangle=e^{\frac{\kappa \Delta}{v_F} x}\left(\begin{array}{ccc} a \\ b \end{array} \right)$$ where the vectors $\left(a,b\right)^T$ satisfy the following equation: $$\Delta\left(\begin{array}{ccc} \mp i\kappa-\cos(\beta) & e^{i\phi(x)} \\ e^{-i\phi(x)} & \pm i\kappa-\cos(\beta) \end{array}\right)\left(\begin{array}{ccc} a \\ b \end{array} \right)=0\ ,$$ which leads to $\kappa=\sin(\beta)$ for $x<0$, and $\kappa=-\sin(\beta)$ for $x>0$. Notice, since we have $0 \le \beta \le \pi$, it follows $\sin(\beta)\ge 0$. The wave functions are: $$x<0:\ \left(\begin{array}{ccc} a \\ b \end{array} \right)=\frac{1}{\sqrt{2}} \left(\begin{array}{ccc} 1 \\ e^{\pm i\beta} \end{array} \right)\ ,\ \mathrm{and}$$ $$x\ge 0:\ \left(\begin{array}{ccc} a \\ b \end{array} \right)=\frac{1}{\sqrt{2}} \left(\begin{array}{ccc} 1 \\ e^{i(\phi \mp \beta)} \end{array} \right)\ .$$ From the continuity condition at $x=0$, we obtain $\beta=\frac{\phi}{2}$ for the right Fermi point, and $\beta=\pi-\frac{\phi}{2}$ for the left Fermi point such that $0 \leq \phi,\ \beta \leq \pi$. Close to right and left Fermi points, the energies of the ABS are found to be $E=\Delta \cos(\phi/2)$ and $E=\Delta\cos(\pi-\phi/2)=-\Delta\cos(\phi/2)$, respectively. The two modes correspond to eigenvalues of $\sigma_x$ with opposite sign, thus, there is no mixing between them. Importantly, their wave functions stay localized on the surface of the TI because bulk ABS have larger energy. In what follows, we show that the energy of Andreev states in the bulk of a TIJJ is always larger than the energy of low-energy Andreev states on the surface, unless the chemical potential corresponds to the Fermi wave vectors $\textbf{k}_f$ where the effective mass $|m(\textbf{k}_f)|$ vanishes. The low-energy Hamiltonian in the bulk of a topological insulator is the massive Dirac Hamiltonian, eqn. \[bh\]. Adding terms for superconducting pairing and the finite chemical potential, the Hamiltonian reads as $$H_{b}=\left[v_F\left( {\boldsymbol{\nabla}}\cdot\boldsymbol{\sigma} \right) \tau_x+m \left(|\textbf{ k}| \right) \tau_z - \mu \right] \eta_z + \Delta\left[\left(\cos\left(\phi(x)\right)\eta_x+\sin\left(\phi(x)\right)\eta_y\right) \right]\ .$$ Here, we would like to point out that the minimal model for TI bulk states has two orbitals, and another set of Pauli matrices, $\tau_i$, for orbital space had to be introduced in the Hamiltonian. Therefore, the algebraic structure of the wave function is different for bulk and surface states. Similar to surface states, low-energy bulk ABS have momenta parallel to the $\hat{x}$-direction (i.e. $k_y=k_z=0$), and the effective Hamiltonian is $$H_{b}=\left[ v_F k_x \sigma_x \tau_x+m \left(|\textbf{k}| \right) \tau_z - \mu \right] \eta_z +\Delta\left[\left(\cos\left(\phi(x)\right)\eta_x+\sin\left(\phi(x)\right)\eta_y\right) \right]\ .$$ Again, the operator $\sigma_x$ commutes with the effective Hamiltonian thus we can divide the eigenstates into two separate sets with eigenvalue $\langle\sigma_x\rangle=1$, or $-1$: $$\label{Hamilb} H_{b}=\left[\pm v_F k_x \tau_x+m \left(|\textbf{ k}|\right) \tau_z - \mu \right] \eta_z + \Delta\left[\left(\cos\left(\phi(x)\right)\eta_x+\sin\left(\phi(x)\right)\eta_y\right) \right]\ .$$ In the following, we only present the solution for $\langle\sigma_x\rangle=1$, which corresponds to the $+$ sign for the first term of $H_{b}$. The case of $\langle\sigma_x\rangle=-1$ can be treated identically. When the chemical potential is in the conduction band, spectrum and orbital wave function are solutions of the kinetic Hamiltonian, $\mathcal{H}=v_F \tau_x k_x + \tau_z \ m(\textbf{k})$. Its eigenvalues and eigenfunctions are given by: $$\label{wf} \begin{split} \mathcal{E}(k_x)&=\sqrt{v_F^2 k_x^2+ m \left(|k_x|\right) ^2}\ , \\ |\Phi_\tau(k_x)\rangle &=\frac{1}{\mathcal{N}} \left(m(|k_x|),\sqrt{m(|k_x|)^2+v_F^2 k_x^2}-v_F k_x\right)^T\ , \end{split}$$ with the normalization factor $\mathcal{N}^2=2m(|k_x|)^2+2v_F^2 k_x^2-2v_F k_x\sqrt{m(|k_x|)^2+v_F^2 k_x^2}$. The low-energy ABS close to the Fermi points, $k^{\pm}_f=\pm \frac{\sqrt{\mu^2-m \left(|k_f|\right) ^2}}{v_F}$, are derived by setting $k_x=k^{\pm}_f\mp i\partial_x$ in equation \[Hamilb\] and by projecting onto the corresponding orbital wave functions $|\Phi_\tau(k^\pm_f)\rangle$. We obtain the equations $$\label{fp} \left[\left(\mp iv_F\partial_x\right)\eta_z+\Delta\left(\cos(\phi(x))\eta_x+\sin(\phi(x))\eta_y\right) \right]|w\rangle=E|w\rangle$$ The Hamiltonian in eqn. \[fp\] is formally identical to the previously discussed Hamiltonian, eqn. \[Hamils\]. For the energies at the right and left Fermi points, we have $E=\Delta \cos(\phi/2)$ and $E=\Delta\cos(\pi-\phi/2)=-\Delta\cos(\phi/2)$. Unlike before, the wave functions at the two Fermi points correspond to the same eigenvector of $\sigma_x$, and the orbital parts of the wave functions, eqn. \[wf\], are – in general – not orthogonal. Since the phase $\phi$ varies along the $\hat{x}$-direction, states at the two Fermi points are scattered into each other and repel. The energy shift due to scattering is calculated using second-order perturbation theory. It is proportional to the orbital overlap of the wave functions at the two Fermi points, i.e. $\langle \Phi_\tau(k^-_f)|\Phi_\tau(k^+_f)\rangle = \frac{| m \left(|k_f|\right)|}{\sqrt{m \left(|k_f|\right)^2+v_F^2 k_f^2}}$, and takes the form $$E=\pm\left( \Delta |\cos(\phi/2)|+\alpha \frac{m \left(|k_f|\right)^2}{m \left(|k_f|\right)^2+v_F^2 k_f^2}\right)$$ where $\alpha>0$. Assuming the magnitude of the superconducting gap is equal on the surface and in the bulk, it follows that the energies of bulk Andreev states are larger than those of surface Andreev states unless $m\left(\mathbf{k}_f\right)=0$.   **Numerical simulation** In order to further support our model, we study the structure of low-energy Andreev states in TIJJs numerically. A simple discrete model for TIs includes 4 orbitals on a cubic lattice with orbital-dependent nearest-neighbor hopping[@PhysRevB.81.045120; @PhysRevLett.107.097001]. We implement superconductivity at the mean-field level, i.e. we double the number of orbitals at each lattice site to represent superconducting particle- and hole-states, and we add a coupling term between the two sectors. Allowing the phase of this coupling to vary along the $\hat{x}$-direction, we can model a TIJJ. As we set the momentum along the $\hat{y}$-direction (i.e. “parallel” to the junction) equal to zero, it is sufficient to discretize the Hamiltonian on a square, real-space lattice with dimensions $L_x \times L_z$ for which the two site-labels $(x,z)$ are chosen along the directions of the superconducting phase variation, $\hat{x}$, and the chemical potential shift, $\hat{z}$. Band-bending on the surface of the TI is modeled by choosing the chemical potential for sites with index $0<z \le L_z/3$ in the conduction band whereas it is fixed at zero for sites with $z>L_z/3$. We then solve the discrete Hamiltonian exactly and plot the wave function amplitudes associated with the lowest-energy ABS against the chemical potential in the surface layer, see Fig. 3(f) in the main article. We observe that, as the chemical potential enters the conduction band, most of the weight of the ABS wave functions remains close to the boundary of the model (i.e. at $L_z = 0$). When we further increase the chemical potential, the wave function spreads gradually across the surface region. As the critical chemical potential ($\mu_C$) is reached, and the effective mass vanishes ($m\left(\textbf{k}_f\right)=0$), the wave function is fully delocalized and extends throughout the region (i.e. sites for which $0<z<L_z/3$) where band-bending occurs in our model. Past the critical value $\mu_C$, the ABS are strongly localized at $z=L_z/3$, which is the interface between the – now – topologically-trivial surface and the band-inverted bulk. In conclusion, we showed that in a doped topological insulator, where superconductivity is induced in the bulk (in addition to the surface states), the energies of low-energy bulk Andreev-bound states are related to the magnitude of the effective mass of the bulk band and have a minimum when the effective mass vanishes. At this point, the energies of boundary and bulk ABS are equal, which couples them and annihilates the boundary modes. Most prominently, this mechanism destroys surface Majorana modes when the bulk substrate undergoes a topological transition which makes it relevant in a wider context. Experimental signatures of the spatial displacement of boundary ABS due to a topological transition in the bulk band are covered in the main text.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We consider the quantization of scalar fields in spacetimes such that, by means of a suitable scaling of the field by a time dependent function, the field equation can be regarded as that of a field with a time dependent mass propagating in an auxiliary ultrastatic static background. For Klein-Gordon fields, it is well known that there exist an infinite number of nonequivalent Fock representations of the canonical commutation relations and, therefore, of inequivalent quantum theories. A context in which this kind of ambiguities arises and prevents the derivation of robust results is, e.g., in the quantum analysis of cosmological perturbations. In these situations, typically, a suitable scaling of the field by a time dependent function leads to a description in an auxiliary static background, though the nonstationarity still shows up in a time dependent mass. For such a field description, and assuming the compactness of the spatial sections, we recently proved in three or less spatial dimensions that the criteria of a natural implementation of the spatial symmetries and of a unitary time evolution are able to select a unique class of unitarily equivalent vacua, and hence of Fock representations. In this work, we succeed to extend our uniqueness result to the consideration of all possible field descriptions that can be reached by a time dependent canonical transformation which, in particular, involves a scaling of the field by a function of time. This kind of canonical transformations modify the dynamics of the system and introduce a further ambiguity in its quantum description, exceeding the choice of a Fock representation. Remarkably, for [*any*]{} compact spatial manifold in less than four dimensions, we show that our criteria eliminate any possible nontrivial scaling of the field other than that leading to the description in an auxiliary static background. Besides, we show that either no time dependent redefinition of the field momentum is allowed or, if this may happen –something which is typically the case only for one-dimensional spatial manifolds–, the redefinition does not introduce any Fock representation that cannot be obtained by a unitary transformation.' author: - Jerónimo Cortez - 'Guillermo A. Mena Marugán, Javier Olmedo' - 'José M. Velhinho' title: Criteria for the determination of time dependent scalings in the Fock quantization of scalar fields with a time dependent mass in ultrastatic spacetimes --- Introduction {#sec:intro} ============ It is well known that the relation between classical and quantum systems is not a one to one correspondence. In fact, the construction of a quantum theory that corresponds to a given classical system is generally plagued with ambiguities. Usually, one first selects a specific set of variables which provides an (over-)complete set of coordinates on phase space, assumed to be a symplectic manifold, and requires this set to be closed under Poisson brackets. In short, one considers then a suitable Poisson algebra of phase space functions, able to distinguish points, and looks for a representation of it as an algebra of linear operators on a Hilbert space [@quantprocess]. Even ignoring all the freedom existing in the choices that lead to a particular algebra of functions, so that one admits the identification of classical systems directly with these algebras, their representation as an algebra of operators introduces ambiguities which affect the physics derived with the resulting quantum theory. In the simplest cases studied in Quantum Mechanics, where the classical system has a finite number of degrees of freedom and the phase space possesses a linear structure, the ambiguities are surpassed in the following way. First, one passes to the exponentiated version of ($i$ times) the natural position and momentum variables, so that one concentrates the analysis just on bounded functions, and arrives to the so-called Weyl algebra as the characteristic algebra of the system. Next, one restricts all discussions exclusively to strongly continuous, unitary, and irreducible representations of this algebra. The Stone-von Neumann theorem [@simon] guaranties then that the allowed representations are all unitarily equivalent, so that the quantum physics is univocally determined. It is worth emphasizing that the uniqueness of the representation is achieved only when one imposes certain criteria, assumed for the validity of the Stone-von Neumman theorem. In particular, if one renounces to the requirement of strong continuity, one can obtain representations which are not unitarily equivalent to the standard one. For instance, this is the situation that is found in the polymer representation [@polymer; @bohr] adopted in Loop Quantum Cosmology [@LQC; @LQCap], namely, the quantization of simple cosmological spacetimes following the methods put forward in Loop Quantum Gravity [@LQG]. The picture gets more complicated when one analyzes systems which possess an infinite number of degrees of freedom. This is so even for the simplest fieldlike systems, with a phase space described by a field and its momentum, and a dynamics determined by linear field equations. If one considers the associated canonical commutation relations (CCR’s), or more precisely the field analogue of the Weyl algebra, one finds that there exist infinitely many possibilities of representing them which are not related by unitary transformations. This infinite ambiguity still arises if one restricts all considerations to Fock representations [@wald], where one describes the field in terms of creation and annihilation operators. Different representations can be interpreted as corresponding to different choices of vacuum, which in turn implies a different identification of the creation and annihilation parts. These alternatives can also be viewed as distinct choices of a basis of solutions for the dynamical equations, with a different characterization of the field in terms of the coefficients of the expansion in that basis. Hence, the possible choices of (suitable orthonormalized) bases are related among them by means of linear canonical transformations, often called Bogoliubov transformations, which change the sets of creation and annihilationlike variables. The essential difference with respect to Quantum Mechanics is that such linear canonical transformations cannot always be implemented as unitary transformations in the quantum theory. As a consequence, unless one includes additional criteria [@wald; @ash-magnon; @kay; @jackie] to select a vacuum state (or rather a unitarily equivalent class of them), one has to deal with an infinite number of nonequivalent Fock quantizations, each leading to different physical predictions. Furthermore, in nonstationary scenarios, like those arising in cosmology, there exists an additional ambiguity which is previous to the selection of a Fock representation, and which is related to the choice of a canonical pair to describe the field when one allows that part of its evolution be assigned to the time dependent spacetime in which the propagation takes place. In fact, in nonstationary settings, it is customary to scale the field configurations by time varying functions. This is so irrespective of whether the spacetime in which the propagation occurs is a true physical background [@mukhanov], an effective spacetime (e.g., a quantum corrected background in effective Loop Quantum Cosmology [@LQCap; @fmov; @fmov2]), or an auxiliary spacetime (like for dimensional reductions of systems with two commuting spacelike Killing vectors, as in Gowdy models [@Gowdy; @jeme; @unit-gt3]). A scaling of this type is found, for instance, in the study of Klein-Gordon (KG) fields in Friedmann-Robertson-Walker (FRW) spacetimes, in the treatment of scalar perturbations around FRW spacetime –like in the analysis of Mukhanov-Sasaki variables [@muk-sas]–, or in the consideration of Bardeen potentials [@bar]. As we will comment in more detail below, in such cases the field is typically changed by a function of the scale factor of the geometry, but the specific functional dependence depends on the problem under consideration. This scaling of the field configurations can always be completed into a linear and time dependent canonical transformation, which leads to a new canonical pair of field variables. Since the transformation varies in time, the new pair has a different (but still linear) dynamics. Hence, the freedom to perform a transformation of this type introduces a fundamental ambiguity in the description of the linear system and of its properties under quantum evolution. It is mainly on this kind of ambiguity that we will concentrate our discussion in this work, proposing criteria that remove it in situations of interest in cosmology and, besides, determine a unique representation of the CCR’s for the corresponding privileged scaling. Let us recall that, given a linear field phase space, the relevant information on the choice of creation and annihilationlike variables is encoded in a basic structure called the [*complex structure*]{} [@wald; @cocoque]. A complex structure $J$ is a real, linear map on the phase space which preserves the symplectic form, $\sigma$, and whose square is minus the identity. In addition, it is required that the composition of the complex structure (acting in one of the entries of $\sigma$) and the symplectic form provides a positive definite bilinear map on phase space. Every such complex structure defines a vacuum state which subsequently determines a Fock representation of the CCR’s [@wald] (or, strictly speaking, of the corresponding Weyl relations). A result due to Shale [@shale; @honegger] tells us that, if we have a Fock representation of the CCR’s determined by a complex structure $J$, a linear canonical transformation $T$ admits a unitary implementation in that representation if and only if the antilinear part of $T$, namely $(T+JTJ)/2$, is a Hilbert-Schmidt operator[^1]. Obviously, in infinite dimensions this requirement is not satisfied by all conceivable canonical transformations, so that not all of them lead to unitarily related quantum theories. It is worth commenting that the Hilbert-Schmidt requirement can be reinterpreted as the condition that the analyzed transformation maps the vacuum to a new state with a finite particle content (to the extent that a particle concept can be employed in the scenario under discussion). In practical situations, as we have mentioned, one looks for reasonable criteria which can remove the ambiguity in the representation and select a preferred vacuum, or equivalence class of vacua. For instance, one can require a natural quantum implementation of the classical symmetries of the system [@wald]. However, in general cases, and in particular in generic nonstationary settings, one simply has not sufficient symmetry to pick out a unique Fock representation. This is particularly important in cosmology. When considering fields that propagate in cosmological backgrounds, which are nonstationary, the lack of uniqueness criteria renders the predictions of the Fock quantization devoid of physical relevance, inasmuch as they depend on particular choices and, furthermore, there exist an infinite number of them. At least for cases in which the cosmological background still possesses some spatial symmetries, it is a standard procedure to keep the requirement that the quantization structures be invariant under those symmetry transformations, even if this does not totally fix the representation. Provided that these transformations are symplectomorphisms, this amounts to the requirement that the complex structure be invariant. We will call [*invariant*]{} the representations with this property. In addition, in the lack of a time symmetry, it sounds reasonable to demand at least that the dynamical evolution be implemented as a family of unitary transformations. Precisely this combined criteria of spatial symmetry invariance and unitary dynamics have been used to determine a unique Fock quantization for certain scalar fields describing gravitational waves [@jeme; @unit-gt3; @scho; @unique-gowdy-1; @BVV2; @CQG25], in the context of inhomogeneous cosmologies of the Gowdy type. The criteria have been proven to apply as well to scalar fields with a generic time dependent mass defined on $d$-spheres, with $d=1,2,3$ [@PRD79; @CMV8], including the commented (dimensionally reduced) description of the Gowdy fields as particular cases. More recently, it has been possible to extend the result of the uniqueness of the Fock quantization of scalar fields satisfying a KG equation with time varying mass to fields defined on [*any*]{} compact spatial manifold in three or less dimensions [@CMOV-FTC]. Actually, once one allows for a scaling of the field by a time dependent function (treated classically), as we have commented that frequently happens in cosmology, the description of the (scalar) field propagation in certain nonstationary spacetimes can be reformulated as that of a field with a time varying mass in a static background. This typically occurs in FRW spacetimes. The simplest example is that of a test KG field, which after a rescaling by the FRW scale factor (and in conformal time) obeys a field equation of the form $$\label{1new} \ddot \varphi - \Delta \varphi+ s(t) \varphi =0,$$ which precisely corresponds to the propagation of a free field with a time dependent mass. Besides, in source-free Einstein-Maxwell theory, using conformal time and adopting a suitable Lorentz gauge, the vector potential can be scaled in a similar way to arrive at a massless wave equation in a static spacetime [@jantzen]. A context in which the discussion encounters a natural application is in the quantization of cosmological perturbations [@bar; @mukhanov; @muk2]. In particular, for perturbations of the energy-momentum tensor that are isotropic and adiabatic, the gauge invariant energy density perturbation amplitude can be scaled by a suitable time function (other than the scale factor) so as to satisfy (in conformal time) a field equation of the above type (\[1new\]), in an effective static background [@bar]. One also finds this same kind of equation with varying mass in the asymptotic analysis of the dynamics of the perturbations of a massive scalar field in an FRW spacetime, after a suitable gauge fixing and a scaling of the field [@hh; @fmov][^2]. In addition, the tensor perturbations of an FRW cosmological background, describing its gravitational wave content, are subject as well to a field equation of this type after scaling them (and choosing again conformal time) [@bar]. Therefore, the result of uniqueness of the Fock representation for a KG field with time varying mass and in a static spacetime under the criteria of symmetry invariance and unitary dynamics finds immediate applications in cosmology, and in particular in the study of cosmological perturbations, if one contemplates the possibility of scaling the fields by time dependent functions, which partially absorb the evolution of the cosmological background. Recall that these results are valid in models with compact spatial topology. This includes the physically important case of flat models with compact sections of 3-torus topology [@threetorus]. Let us emphasize that different scalings lead to different field descriptions, each of them with a different dynamics. The Fock quantization of each of these descriptions does not necessarily provide unitarily equivalent quantum theories. Let us see this in more detail. We already mentioned that, on phase space, the scaling of the field by a time function can be regarded as part of a time dependent linear canonical transformation. The scaling of the field is then completed by a transformation of the momentum, in which the latter suffers just the inverse scaling, so as to maintain the canonical structure. Besides, in this transformation, the momentum may acquire a contribution linear in the field. In order to respect locality and the spatial dependence of the fields, the most general linear contribution to the momentum that we will consider consists of the field multiplied by a (conveniently densitized) function of time. The resulting family of canonical transformations, being time dependent, generally modify the dynamical evolution of the system. In this regard, it is important to contemplate the presence of a field contribution to the new momentum if one wants to maintain a dynamics dictated by a quadratic Hamiltonian with certain good properties, like e.g. the absence of crossed terms mixing the configuration and momentum fieldlike variables. But the fact that the dynamics changes implies that the criteria for uniqueness, which in particular include a unitary implementation of the time evolution, must be applied independently to each field description, at least in principle. Besides, since the descriptions are related by linear canonical transformations (varying in time, actually), and not all of these transformations can be implemented in terms of unitary operators in the quantum theory, it is not granted that the different formulations attained in this manner result to be unitarily equivalent. Hence, if we want to reach a privileged Fock quantization for our system, we need to fix this ambiguity in the field description. A quite remarkable result, proven first for the case in which the spatial manifold on which the field is defined is a circle [@CMVPRD75], and demonstrated recently for the 3-sphere and the sphere in two dimensions [@zejaguije], is that the proposed criteria of natural invariance under the spatial symmetries and of unitary dynamics happen to select also a unique field description among this class of time dependent canonical transformations. The description selected is precisely the one in which the field equations are of the type (\[1new\]), with time varying mass, in a static background[^3]. When the spatial manifold is a circle, it was shown that field descriptions differing just in the inclusion of a field contribution to the momentum are possible, but they are all unitarily equivalent, so that a representation of the new canonical pair can be directly constructed from the original one in such a way that the relation is unitary [@CMVPRD75]. The aim of the present work is to extend this result to any compact spatial manifold in three or less dimensions. Namely, we want to prove that our criteria of symmetry invariance [*and unitary time evolution*]{} select in fact a unique field description for our system on [*any*]{} compact spatial manifold in three or less spatial dimensions. This, together with the already obtained result about the uniqueness of the Fock representation for the specific field description in which the KG equation does not contain any dissipative term \[that is, the description in which the background appears to be static and the field equation takes the form (\[1new\])\], provides a considerable robustness to the quantization, choosing a unique Fock quantum theory up to unitary equivalence. In particular, this guaranties the reliability of the quantum predictions. The rest of the paper is organized as follows. We start by introducing the model in Sec. II. The uniqueness result about the choice of Fock representation for a scalar field with varying mass propagating in a static spacetime whose spatial sections are compact is reviewed in Sec. III. Although this result was proven in Ref. [@CMOV-FTC], we succinctly revisit the arguments of the demonstration for completeness in the presentation and because they provide the basis for the proof of the result of this work, namely, that our criteria select also a unique field description among all those related by a time dependent scaling. The proof that all nontrivial scalings are excluded is presented in Sec. IV. In addition, in Sec. V we show that either there is no freedom to include a time dependent linear contribution of the field in the momentum or, if the freedom exists (something that may typically happen only for one-dimensional spatial manifolds), the change does not introduce any Fock representation which is not attainable from the original one by a unitary transformation. The relation between the Fock quantization selected by our criteria and the choice of vacuum in terms of the Hadamard condition [@wald] is briefly discussed in Sec. VI. We present our conclusions in Sec. VII. Finally, two appendices are added. The model and its quantization {#sec:model} ============================== We begin by considering the Fock quantization of a real scalar field with a time varying mass function. The field $\varphi$ is defined on a general Riemannian compact space $\Sigma$ of three or less (spatial) dimensions, and propagates in a globally hyperbolic background of the form $\mathbb{I} \times \Sigma$, where $\mathbb{I}$ is a (not necessarily unbounded) time interval. We call $h_{ab}$ the metric on the spatial manifold $\Sigma$ ($a,b$ denoting spatial indices), and restrict the discussion here to the case of orthogonal foliations [and a time independent $h_{ab}$]{}. As we have already commented, under very mild assumptions (in particular on the mass function) it is then possible to show that a preferred Fock representation is selected by imposing the criteria that the dynamics be unitary and that one achieves a natural unitary implementation of the spatial symmetries of the field equations [@CMOV-FTC]. For our analysis, we choose an (arbitrarily) fixed time $t_0$ and, at that instant of time, we consider the field data $(\varphi,P_{\varphi})=(\varphi,\sqrt{h}{\dot\varphi})_{|t_0}$, where the dot denotes the time derivative and $h$ is the determinant of the spatial metric. By construction, we identify the canonical phase space of the system with the set of data pairs $\{(\varphi,P_{\varphi})\}$, equipped with the symplectic form $\sigma$ that is determined by the standard Poisson brackets $\{\varphi(t_0,x),P_{\varphi}(t_0,y)\}=\delta(x-y)$. These brackets are taken independent of the choice of $t_0$, so that the time independence of $\sigma$ is granted. Note also that the configuration variable $\varphi$ is defined as a scalar, and hence the momentum $P_{\varphi}$ is a scalar density. We call $\Delta$ the standard Laplace-Beltrami (LB) operator associated with the metric $h_{ab}$. Note that $-\Delta$ is a nonnegative operator, i.e., with the exception of possibly null eigenvalues (in this respect, see the comments below about zero modes), all eigenvalues of $\Delta$ are real and negative. Employing this operator, we introduce the complex structure $J_0$ determined by: $$\begin{aligned} \label{cano-cs} J_0(\varphi ) &=& -(-h \Delta)^{-1/2} P_{\varphi} , \nonumber \\J_0(P_{\varphi}) &=& (-h \Delta)^{1/2} \varphi .\end{aligned}$$ The Fock representation defined by $J_0$ is the analogue of the free massless field representation. In fact, $J_0$ is constructed from the LB operator ignoring the existence of a mass in the system. Nonetheless, rather than the massless case, we are going to consider the general case of the field equation $$\label{1} \ddot \varphi - \Delta \varphi+ s(t) \varphi =0,$$ which, given the expression of the field momentum, is equivalent to the canonical equations of motion: $$\label{fieldequations} \dot\varphi=\frac{1}{\sqrt{h}}P_{\varphi}, \quad {\dot P}_{\varphi}=\sqrt{h}\big[\Delta\varphi -s(t)\varphi\big].$$ The mass function $s(t)$ is allowed to be quite arbitrary, except for some weak conditions that were specified in Ref. [@CMV8]. Namely, we assume that it has a second derivative which is integrable in any closed subinterval of $\mathbb{I}$. In order to discuss whether the dynamics admits a unitary implementation with respect to the Fock representation determined by $J_0$, essential ingredients are the general properties of the LB operator in any compact space [@compact]. In particular, the eigenmodes of the LB operator allow us to decompose the field in a series expansion. In the considered general setting, the natural space of functions on $\Sigma$ is that of square integrable functions in the inner product provided by the metric volume element (constructed with $h_{ab}$). Let then $\{\Psi_{n,l}\}$ be a complete set of real orthonormal eigenmodes of the LB operator with respect to this inner product, with corresponding discrete set of eigenvalues given by $\{-\omega^2_n\}$, with $n \in \mathbb{N}$. Necessarily, these eigenvalues are such that $\omega^2_n$ tends to infinity when so does $n$. In general, the spectrum of the LB operator may be [degenerate]{}, so that two or more of the eigenmodes $\Psi_{n,l}$ have the same eigenvalue. The label $l$ takes this degeneracy into account. We call $g_n$ the dimension of the eigenspace with eigenvalue $-\omega^2_n$. This degeneracy number is always finite, $\Sigma$ being compact. For each $n$, the label $l$ runs from $1$ to $g_n$. In the following, all sums performed over the spectrum of the LB operator include this degeneracy. Using these eigenmodes, we can express the field $\varphi$ as a series $\varphi=\sum_{n,l}q_{n,l}\Psi_{n,l}$. With this expansion at hand, it is clear that the degrees of freedom of the field reside in the discrete set of real modes $\{q_{n,l}\}$, which vary only in time. Since the eigenmodes are orthonormal with respect to the inner product provided by the metric volume element, one gets that the canonical momentum conjugate to $q_{n,l}$ is $p_{n,l}={\dot q}_{n,l}$. Besides, recalling that $J_0$ is obtained from the LB operator, it is easy to realize that this complex structure is block diagonal by modes in the introduced field expansion and, furthermore, independent of the degeneracy labeled by $l$. Let us then define $$\label{basic-var} a_{n,l}=\sqrt{\frac{\omega_n}{2}}q_{n,l} + i \frac{p_{n,l}}{\sqrt{2\omega_n}} ,$$ which, together with their complex conjugates $a_{n,l}^*$ [form]{} a set of annihilation and creationlike variables[^4]. In these variables, the complex structure $J_0$ is totally diagonal, taking the standard form $J_0(a_{n,l})=i a_{n,l}$ and $J_0(a_{n,l}^{\ast})=-i a_{n,l}^{\ast}$. In other words, $a_{n,l}$ and $a_{n,l}^{\ast}$ can be regarded as the variables that are promoted to annihilation and creation operators in the Fock representation determined by $J_0$. Returning to the dynamics, one can check that the modes obey the equations of motion: $$\label{q-eq} \ddot q_{n,l}+\big[\omega_n^2+s(t) \big]q_{n,l}=0.$$ It is worth noticing that all the modes are decoupled, and that the evolution equations are the same for all modes in the same eigenspace (indicated by the label $n$). The evolution of the variables $(a_{n,l},a_{n,l}^*)$ from the fixed reference time $t_0$ to any other time $t$ is a linear transformation which is then block diagonal, owing to the decoupling of the modes, and insensitive to the degeneracy label $l$. Thus, the transformation adopts the general form $$\label{bogo-transf} a_{n,l}(t)=\alpha_n(t,t_0)a_{n,l}(t_0)+ \beta_n(t,t_0)a_{n,l}^*(t_0).$$ Since the evolution respects the symplectic structure, this transformation must be canonical. This implies that, for all values of $n$ and $t$ and independently of the value of $t_0$, one has $$\label{symp} |\alpha_n(t,t_0)|^2 =1+|\beta_n(t,t_0)|^2.$$ Actually, a canonical transformation of the type can be implemented in terms of a unitary operator in the Fock representation defined by the complex structure $J_0$ if and only if the sequence formed by its corresponding beta-functions $\beta_n(t,t_0)$ is square summable, namely, if $\sum_n g_n |\beta_n(t,t_0)|^2$ is finite [@honegger] (note that the degeneracy has been taken into account). To elucidate whether this sum is finite or not, we need to know the behavior of the beta-functions for large $n$, i.e., to know the asymptotic behavior of the dynamics for modes with large value of $\omega_n^2$. This asymptotic analysis was carried out in Ref. [@CMV8]. It was proven there that, for any possible mass function $s(t)$ and any values of $t$ and $t_0$, the leading term in the beta-function is proportional to $1/\omega_n^2$. It then turns out that the requirement that the sum of $|\beta_n(t,t_0)|^2$ be finite is equivalent to the finiteness of $\sum_n g_n/\omega_n^4$. Indeed, this condition is satisfied for all Riemannian compact manifolds in three or less dimensions. This fact follows from the asymptotic properties of the spectrum of the LB operator. In particular, the number of eigenstates whose eigenvalue does not exceed $\omega^2$ in norm is known to grow in $d$ dimensions at most like $\omega^{d}$ [@compact]. With this bound in the growth rate, one can prove that $g_n/\omega_n^4$ is summable. If the manifold $(\Sigma,h_{ab})$ possesses an isometry group, the LB operator is automatically invariant under it. Therefore, these symmetries are directly transmitted to the field equations (\[fieldequations\]). In the canonical formulation, the group translates into canonical transformations which commute with the dynamics. More generally, we will consider the subgroup of the unitary transformations \[in the Hilbert space of square integrable (configuration) functions with respect to the measure defined by the metric volume element associated with $h_{ab}$\] that commute with the LB operator, or a convenient subgroup of it determined by the isometries, provided that this latter subgroup satisfies certain conditions which we will explain later on. We will call this symmetry group $G$, which leaves the dynamics invariant. As part of our criteria for the uniqueness of the quantization, we demand that these symmetries find a natural unitary implementation in the quantum theory. In fact, this is ensured in the Fock representation determined by the complex structure $J_0$, since this structure depends exclusively on the LB operator (and the metric volume element), and hence inherits its invariance under the symmetry group $G$. Thus, the complex structure $J_0$ is invariant under $G$ and determines a Fock representation in which the quantum counterpart of Eq. (\[q-eq\]) is a unitary dynamics. In the next section, we will prove that, if there exists another Fock representation with the same properties, it has to be unitarily equivalent to the one defined by $J_0$. Uniqueness of the representation {#unique rep} ================================ In order to obtain a natural unitary implementation of the symmetry group $G$ in the Fock representation, we just have to concentrate our attention on complex structures $J$ that are invariant under its action. Therefore, the first step in our analysis is to characterize these $G$-invariant complex structures, something that is possible by means of a suitable application of Schur’s lemma [@unique-gowdy-1; @BVV2; @CMV8]. Let us analyze the action of the group $G$ on the canonical phase space. We start by studying its action on the configuration space, formed from the values of the field $\varphi$ at time $t_0$. We will call $\cal Q$ this configuration space. Recall that, by construction, the action of $G$ is naturally unitary on $\cal Q$ (with respect to the inner product obtained with the metric volume element) and commutes with the LB operator. Therefore, each of the eigenspaces of the LB operator corresponding to different eigenvalues provides an irreducible representation of $G$ or, otherwise, can be composed in a finite number of mutually orthogonal irreducible subspaces. In this way, we can decompose the configuration space $\cal Q$ in a convenient hierarchy of finite dimensional subspaces: first, as a direct sum of eigenspaces ${\cal Q}^{n}$ of the LB operator (the superscript $n$ labeling the associated eigenvalue), and then each of these eigenspaces as a direct sum of irreducible representations ${\cal Q}^{n}_m$ of the symmetry group $G$ (the label $m$ counting the different components for each $n$). Note that, if $G$ is taken as the maximal subgroup of the unitary group that commutes with the LB operator, all these irreducible representations are distinct. On the other hand if, starting with the spatial isometries, we rather identify $G$ with a subgroup of the former maximal subgroup, we now [*assume*]{} that all such representations differ (this is the case, e.g., with the isometry group of the $d$-sphere or the $d$-torus). Clearly, if we call $g_{n,m}$ the dimension of those representations, ${\cal Q}^{n}_m$, the sum of $g_{n,m}$ over $m$ must equal the degeneracy $g_n$ for each value of $n$. In particular, the integers $g_{n,m}$ can never exceed $g_n$. We can proceed similarly to get a decomposition in irreducible representations of the space $\cal P$ formed by the momentum fields $P_{\varphi}$ at the fixed time $t_0$. Since the momenta are scalar densities, the integral for the inner product is performed in this case with the inverse volume element. Altogether, we arrive at a decomposition of the phase space of the system, $\Gamma$, in the form ${\Gamma}=\oplus_n {\Gamma}^{n}= \oplus_{n,m} {\Gamma}^{n}_m$, where we have called ${\Gamma}^{n}_m= {\cal Q}^{n}_m\oplus {\cal P}^{n}_m$. Besides, given that $G$ acts in the same way on fields and on their momenta, the group action coincides on the subspace ${\cal Q}^{n}_m$ and on its counterpart ${\cal P}^{n}_m$. Via Schur’s lemma [@schur], a direct consequence of this decomposition in irreducible representations is that the $G$-invariant complex structures must be block diagonal, with a (possibly) different block $J_{n,m}$ for each ${\Gamma}^{n}_m$, since they commute with $G$ and cannot mix [*different*]{} irreducible representations[^5]. Therefore, the allowed complex structures $J$ must all admit the generic expression $J=\oplus {J}_{n,m}$. In each component ${\Gamma}^{n}_m$, one can always find a basis of configuration variables and corresponding momentum variables which arises from a suitable choice of orthonormal eigenmodes of the LB operator, like those that we introduced in the previous section to expand the field. For each given $n$, the complete set $\{q_{n,l}, p_{n,l}\}$ is obtained as the union of all such bases when the whole set of subspaces $\Gamma^{n}_m$ of $\Gamma^{n}$ are considered[^6]. Besides, on each ${\Gamma}^{n}_m$, the corresponding complex structure $J_{n,m}$ consists of four maps, $J^{qq}_{n,m}$, $J^{qp}_{n,m}$, $J^{pq}_{n,m}$, and $J^{pp}_{n,m}$, according to the four choices of initial and final space between ${\cal Q}^{n}_m$ and ${\cal P}^{n}_m$. Moreover, each of these four maps, established between the same irreducible representation of $G$, is itself invariant under the action of the group, and therefore must be proportional to the identity map ${\bf I}$ by Schur’s lemma (the proportionality constants being restricted by the imposition that the complex structure be a real map). In total, we conclude that the $G$-invariant complex structures adopt also a block diagonal form in each subspace $\Gamma^{n}_m$, the blocks being given by a 2-dimensional complex structure formed out of the four proportionality constants mentioned above. This 2-dimensional complex structure only mixes $q_{n,l}$ with $p_{n,l}$ for each value of $l$, and coincides for all the labels $l$ in the same subspace $\Gamma^{n}_m$. To compare a generic $G$-invariant complex structure $J$ with the reference one, $J_0$, it is convenient to change the basis on phase space to the complex variables $a_{n,l}$ and $a_{n,l}^{\ast}$. Since all invariant complex structures have the same block form, and they are symplectomorphisms, one can easily show that they are always related by a transformation of the type $J=K J_0 K^{-1}$, where $K$ is a symplectic map which admits the same decomposition in $2\times 2$ blocks that we have found for $J$ [@unique-gowdy-1]. Likewise, all the 2-dimensional blocks of $K$ are identical in each space $\Gamma^{n}_m$. Hence, each invariant complex structure is totally characterized by a discrete set of 2-dimensional symplectic maps $K_{n,m}$. We can view each of these (real) maps as $2\times 2$ matrices and express them in terms of two complex numbers, $\kappa_{n,m}$ and $\lambda_{n,m}$, which provide their diagonal and nondiagonal elements, respectively [@CMV8]. The condition that the map preserves the symplectic form translates into the relation $|\kappa_{n,m}|^2=1+|\lambda_{n,m}|^2$. Note that, then, the complex structures $J$ and $J_0$ will be unitarily equivalent if and only if the symplectic transformation between them, $K$, admits a unitary implementation with respect to (e.g.) $J_0$. We have already commented that this amounts to demand the square summability (including degeneracy) of the beta-functions (or rather beta-coefficients, in this case) corresponding to the map $K$, which are nothing but the complex numbers $\lambda_{n,m}$ [@CMOV-FTC]. Hence, the necessary and sufficient condition for $J$ and $J_0$ to be unitarily related is that $\sum_{n,m} g_{n,m}|\lambda_{n,m}|^2$ be finite. On the other hand, let us assume that the evolution map, $U$, admits a unitary implementation with respect to a $G$-invariant complex structure, $J$. This is equivalent to say (via a change of basis from the creation and annihilationlike variables that diagonalize $J$ to those for $J_0$) that $K^{-1}UK$ can be implemented as a unitary transformation with respect to $J_0$ or, alternatively, that the beta-functions of $K^{-1}UK$ are square summable. The effect of the transformation $K$ is to replace the functions $\alpha_n$ and $\beta_n$ for $J_0$ with new ones, adapted to the basis which diagonalizes $J$. We emphasize that these new functions depend no more just on $n$, but also on the index $m$. A direct calculation leads to the following expression for these new beta-functions: $$\hspace*{-2pc} \beta^J_{n,m}(t,t_0)=(\kappa_{n,m}^*)^2\beta_n (t,t_0)-\lambda_{n,m}^2\beta^*_n(t,t_0)+2 i \kappa_{n,m}^*\lambda_{n,m} {\Im}[\alpha_n(t,t_0)].\label{betaJ}$$ Here, the symbol $\Im$ denotes the imaginary part. Therefore, a $G$-invariant complex structure allows for a unitary implementation of the dynamics if and only if $\sum_{n,m} g_{n,m}|\beta^J_{n,m}(t,t_0)|^2$ is finite at all instants of times $t$. We can then easily adapt the discussion of Ref. [@CMV8] to show that the unitary implementation of the dynamics with respect to $J$ implies indeed that this complex structure is unitarily equivalent to $J_0$. A sketch of the proof goes as follows. Employing that $\sqrt{g_{n,m}}\beta^J_{n,m}(t,t_0)$ and $\sqrt{g_{n}}\beta_{n}(t,t_0)$ are square summable (because the dynamics is unitary with respect to $J$ –by hypothesis– and $J_0$), we conclude that the sequences formed by $\sqrt{g_{n,m}}\,{\Im}[\alpha_n(t,t_0)]\, \lambda_{n,m}/\kappa_{n,m}^*$ must also be square summable at all times. Then, making use then of the asymptotic behavior of ${\Im}[\alpha_n(t,t_0)]$, which was discussed in Ref. [@CMV8], we can easily deduce the square summability, at all instants of time, of $$\label{43} \left\{ \,\sqrt{g_{n,m}}\frac{\lambda_{n,m}}{\kappa_{n,m}^*}\,\sin{\left[ \omega_n(t-t_0)+\int_{t_0}^t d\bar{t}\frac{s(\bar{t})}{2\omega_n}\right]}\right\}.$$ We can now appeal to Luzin’s theorem and integrate the finite sums of the squared elements of this sequence (which are measurable functions) over a suitable set in the time interval $\mathbb{I}$ in order to show that, actually, the sum $\sum_{n,m} g_{n,m}|\lambda_{n,m}|^2$ has to be finite [@CMV8]. But this finiteness is precisely the necessary and sufficient condition for the unitary equivalence between the two complex structures $J$ and $J_0$. This proves that any complex structure that is invariant under the group $G$ and allows for a unitary implementation of the dynamics turns out to be related with $J_0$ by a unitary transformation, so that there exists one and only one equivalence class of complex structures satisfying our criteria. Uniqueness of the field description {#result1} =================================== In the previous sections, we have demonstrated the uniqueness of the Fock quantization adopting since the very beginning a specific field description for our system. However, in nonstationary backgrounds, as we have discussed in the Introduction, it seems most natural to allow for time dependent scalings of the fields, which may absorb part of the dynamical variation of the background. In this context, one must consider the possibility of performing linear canonical transformations that depend on time and that, as far as the field is concerned, amount to a scaling by a time function. This introduces a new ambiguity in our quantization, different in extent to the one considered so far, because this type of canonical transformations change the field dynamics. Hence, one may wonder whether it is still possible to use our criteria and select not just one privileged Fock representation for the KG field description with time dependent mass in an auxiliary static background, but also a unique field description for our system when scalings are contemplated. This is the subject that we will address in the following. The main aim of this work is to prove that our criteria eliminate in fact this apparent freedom in the choice of field description. Unitary implementability condition {#sub1} ---------------------------------- The most general linear canonical transformation depending (only) on time and which changes the field just by a scaling has the form $$\label{transform} \displaystyle \phi=f(t) \varphi, \qquad P_{\phi}=\frac{P_{\varphi}}{f(t)}+g(t)\sqrt{h} \,\varphi .$$ Note that we have allowed for a contribution of the field $\varphi$ in the new momentum, and that this contribution has been multiplied by $\sqrt{h}$ so as to obtain a scalar density. The function $f(t)$, which provides the scaling of the field, is assumed to be nonvanishing, to avoid the artificial introduction of singularities. In addition, the two functions $f(t)$ and $g(t)$ are real, and we suppose that they are at least twice differentiable, so that the transformation does not spoil the differential structure formulation of the field theory. Furthermore, there is no loss of generality in assuming that $f(t_0)=1$ and $g(t_0)=0$ at the reference time $t_0$. In fact, the values of these two functions at $t_0$ can be set equal to those data by means of a constant linear canonical transformation. But, given a Fock representation for the original fields with symmetry invariance and a unitary dynamics, we immediately obtain a Fock representation for any constant linear combination of the canonical fields which possesses the same properties [@CMVPRD75]. Therefore, in the following we restrict our discussion to functions $f(t)$ and $g(t)$ with the above initial data. The dynamics of the new canonical pair $(\phi,P_{\phi})$ admits a description in terms of a Bogoliubov transformation similar to that in Eq. (\[bogo-transf\]), but with different functions $\tilde \alpha_n(t,t_0)$ and $\tilde \beta_n(t,t_0)$. Adopting again creation and annihilationlike variables like those for the massless case, but now constructed from the new canonical pair, one can calculate the relation between the new alpha and beta-functions and the original ones. Ignoring the explicit reference to the dependence on $t_0$ in all functions, and defining $2 f_{\pm}(t)=f(t)\pm 1/f(t)$, one obtains: $$\begin{aligned} \label{26} \displaystyle \tilde \alpha_n(t) &=& f_+(t)\alpha_n(t) + f_-(t)\beta^*_n(t) + \frac {i}{2} \frac{g(t)}{\omega_n}[\alpha_n(t)+\beta_n^*(t)] \, , \\ \label{27} \tilde \beta_n(t) &=& f_+(t) \beta_n(t) + f_-(t)\alpha^*_n(t)+\frac{i}{2}\frac{g(t)}{\omega_n}[ \alpha^*_n(t)+\beta_n(t)] .\end{aligned}$$ In the following, we will demonstrate that, if one performs [*any*]{} canonical transformation of the above type with $f(t)$ other than the unit function, the dynamics becomes such that one cannot implement it as a unitary transformation with respect to [*any*]{} invariant Fock representation. The arguments of the proof are a suitable generalization of those presented in Refs. [@CMVPRD75; @zejaguije]. Let us first make fully explicit the condition for a unitary implementation. Suppose that we are given an invariant Fock representation of the CCR’s, determined by a sequence of pairs $(\kappa_{n,m},\lambda_{n,m})$ as explained in the previous section. The dynamics associated with the new canonical pair $(\phi,P_{\phi})$ can be implemented as a unitary transformation in the considered invariant Fock quantum theory if and only if the sequences with elements $\sqrt{g_{n,m}}\tilde\beta^J_{n,m}(t,t_0)$ are square summable for all possible values of $t$ [@CMV8; @CMVPRD75], where $$\hspace*{-2pc} \tilde\beta^J_{n,m}(t,t_0)=(\kappa_{n,m}^*)^2\tilde\beta_n (t,t_0)-\lambda_{n,m}^2\tilde\beta^*_n(t,t_0)+2 i \kappa_{n,m}^*\lambda_{n,m} {\Im}[\tilde\alpha_n(t,t_0)],\label{betaJtilde}$$ in complete parallelism with Eq. (\[betaJ\]). For simplicity, we obviate the reference to $t_0$ from now on. Thus, assuming a unitary evolution with respect to $J$ in the new field description is equivalent to [saying]{} that $$\sum_n^{\infty}\sum_m g_{n,m}|\tilde\beta^J_{n,m}(t)|^2<\infty$$ at all instants of time $t$. Since every term in the sum is positive, it follows that, if we select a particular value $M$ of $m$ for each $n$, the sequence $\{ g_{n,M}|\tilde\beta^J_{n,M}(t)|^2\}$ is also summable. We emphasize that this is so for any possible choice of $M$. In turn, this summability immediately implies that $\{\tilde\beta^J_{n,M}(t)/(\kappa_{n,M}^*)^2\}$ is square summable, because both $g_{n,M}$ and $|\kappa_{n,M}|$ are always greater than (or equal to) 1. In particular, it is then guaranteed that, for every $t$, the terms of these sequences tend to zero in the limit of infinite $n$. The next step in our line of reasoning is to introduce the asymptotic behavior of $\alpha_n(t)$ and $\beta_n(t)$ in the expression of $\tilde\beta^J_{n,M}(t)$, using relations (\[26\],\[27\]). According to the analysis performed in Ref. [@CMV8], one can take $\beta_n(t)=0$ and $\alpha_n(t)=e^{-i\omega_n\tau}$ up to order $1/\omega_n$ (at least), where $\tau=t-t_0$. As a consequence, we arrive at the result that the condition of a unitary implementation of the dynamics implies that the sequences with elements $$\label{lead-seq} \left[e^{i\omega_n\tau}-z_{n,M}^2e^{-i\omega_n\tau}\right] f_-(t) -2iz_{n,M}\sin(\omega_n\tau)f_+(t)$$ must tend to zero at all times in the limit of large $n$. We have called $z_{n,M}=\lambda_{n,M}/\kappa_{n,M}^*$. Splitting $z_{n,M}$ in its real and imaginary parts, $z_{n,M}=x_{n,M}+iy_{n,M}$, we introduce the definitions: $$\begin{aligned} A_{n,M}&=2 y_{n,M}(f_+-x_{n,M}f_-),\quad\quad\quad\quad\quad\quad & B_{n,M}=(1+ y^2_{n,M}-x^2_{n,M})f_-, \nonumber\\ C_{n,M}&=(1+ x^2_{n,M}-y^2_{n,M})f_{-}-2x_{n,M}f_+,\quad\quad & D_{n,M}=-2 x_{n,M}y_{n,M}f_-,\end{aligned}$$ where, to simplify the notation, we have ignored the explicit time dependence of the functions $f_{\pm}(t)$. Note that, since $|\lambda_{n,M}|\leq |\kappa_{n,M}|$, we have $$\label{bound} |z_{n,M}|^2=|x_{n,M}|^2+|y_{n,M}|^2\leq 1.$$ Taking the real and imaginary parts of the expression (\[lead-seq\]), we see that the sequences given by $$\label{re} A_{n,M}\sin(\omega_n\tau)+ B_{n,M}\cos(\omega_n\tau)$$ and $$\label{im} C_{n,M}\sin(\omega_n\tau)+ D_{n,M}\cos(\omega_n\tau)$$ have to vanish in the limit $n\to \infty$ at all instants of time $t\in\mathbb{I}$. These conditions can be employed to prove that, indeed, unitarity of the dynamics can be attained only if the function $f(t)$ in Eq. (\[transform\]) is the unit function. Proof of the non-unitarity of time dependent scalings {#sub2} ----------------------------------------------------- We notice first that all sequences $A_{n,M}$, $B_{n,M}$, $C_{n,M}$, and $D_{n,M}$ are bounded, owing to inequality (\[bound\]). Using this fact, one can form suitable linear combinations of the expressions (\[re\]) and (\[im\]) and conclude that the following sequences must have a vanishing limit as well: $$\begin{aligned} \label{sin} &&(A_{n,M}D_{n,M}-B_{n,M}C_{n,M})\sin(\omega_n\tau), \\ \label{cos} &&(A_{n,M}D_{n,M}-B_{n,M}C_{n,M})\cos(\omega_n\tau).\end{aligned}$$ Obviously, this is only possible if $A_{n,M}D_{n,M}-B_{n,M}C_{n,M}$ tends to zero. A simple calculation shows that $$\label{master} A_{n,M}D_{n,M}-B_{n,M}C_{n,M}= f_{-}\,(x^2_{n,M}+y^2_{n,M}-1)[(1+ x^2_{n,M}+y^2_{n,M})f_{-}-2x_{n,M}f_{+}].$$ We now prove that a further necessary condition for the unitary implementability of the dynamics is that the sequence [of]{} elements $(x^2_{n,M}+y^2_{n,M}-1)$ \[one of the factors in Eq. (\[master\])\] does not tend to zero. Let us suppose that it does and show that this leads to a contradiction. In this case, [while]{} expression (\[master\]) automatically has a vanishing limit, this is not sufficient to guaranty unitarity. In particular, we still have to check that both expressions (\[re\]) and (\[im\]) tend to zero for all values of $t$. By taking the sum of the squares of those expressions, and using our hypothesis that $x^2_{n,M}+y^2_{n,M}\to 1$, we obtain that $$\label{sproof1} (f_{+}-x_{n,M}f_{-})\sin(\omega_n\tau)+y_{n,M}f_{-}\cos(\omega_n\tau)$$ must tend to zero at all times, $t$. At this stage, two possibilities are available. We consider first the case in which $y_{n,M}\to 0$. Since we have already assumed that $x^2_{n,M}+y^2_{n,M}\to 1$, it follows that $x^2_{n,M}$ tends to 1. From expression (\[sproof1\]), we then conclude that there must exist a subsequence of values of $n$ such that one gets a zero limit either for $f\sin(\omega_n\tau)$ or $\sin(\omega_n\tau)/f$ (or both, if both types of subsequences exist). In either case, recalling the positivity of the function $f$, we have that $\sin(\omega_n\tau)$ must tend to zero, on some subsequence, for all times $t$. However, this is actually impossible, as we show in Appendix A (see also Ref. [@zejaguije]). We consider now the alternate case in which $y_{n,M}$ does not tend to zero. As explained in detail in Appendix B, this leads to the conclusion that $$\label{bigsinus} \sin[\omega_n\tau+\Theta_{n,M}(t)]$$ must have a vanishing limit on some subsequence of values of $n$, at all instants of time $t$, where $$\label{theta} \cot[\Theta_{n,M}]=\frac{1}{y_{n,M}}\frac{f_{+}}{f_{-}}-\frac{x_{n,M}}{y_{n,M}}.$$ Again, using the result proven in Appendix A, one concludes that the sequence given by expression (\[bigsinus\]) cannot tend to zero for all values of $t$ in any given interval $\mathbb{I}$. Therefore, the only possibility which is compatible with our hypothesis of a unitary implementation of the dynamics is that the sequence $\{x^2_{n,M}+y^2_{n,M}-1\}$ does not tend to zero in the limit of large $n$. The next step in our demonstration is to show that, in addition to the condition proven above, the unitary implementation is not achievable unless the function $f(t)$ is the unit function. Let us suppose that, on the contrary, this is not the case. Then, there exist values of $t$ such that $f(t)\not =1$. We will consider those values of $t$, and only those, and will show that the existence of those times leads in fact to a contradiction. Recall that the function $f$ is strictly positive and continuous ([actually]{}, we have assumed that it is twice differentiable). In particular, this implies that $f(t)\not =1$ if and only if $f_{-}(t)\not =0$. Therefore, we are considering points where $f_{-}(t)\not =0$, and we have assumed that such points exist. Going back to expression (\[master\]), a necessary condition for the unitary implementation of the dynamics is that the sequences with elements $$\label{smaster} (x^2_{n,M}+y^2_{n,M}-1)[(1+ x^2_{n,M}+y^2_{n,M})f_{-}-2x_{n,M}f_{+}]$$ tend to zero, at all the considered values of $t$. Moreover, we know that the sequence formed by $(x^2_{n,M}+y^2_{n,M}-1)$ cannot tend to zero at infinitely large $n$. Hence, there exists $\epsilon>0$ and a subsequence $S$ of positive integers $n$ such that $|x^2_{n,M}+y^2_{n,M}-1|>\epsilon$ in $S$. This in turn implies that the second factor in Eq. (\[smaster\]) must tend to zero on that subsequence, a result from which one easily concludes that $$\label{firstproof} f^2(t)[(1- x_{n,M})^2+y^2_{n,M}] - [(1+ x_{n,M})^2+y^2_{n,M}]$$ must have a vanishing limit [on]{} the subsequence $S$. It then immediately follows that the function $f(t)$ must coincide at all the considered values of $t$, simply because the time independent sequences $(1- x_{n,M})^2+y^2_{n,M}$ and $(1+ x_{n,M})^2+y^2_{n,M}$ cannot both tend to zero. Thus, we reach the conclusion that the function $f$ can attain at most two distinct values, one of them equal to 1 (e.g., at the reference time $t_0$) and the other assumed to be different from it. But this is forbidden by continuity. The contradiction shows that the only consistent possibility is that $f(t)$ [is indeed]{} the unit function, as we wanted to prove. Uniqueness of the field description: momentum redefinition {#result2} ========================================================== In the previous section, we have proven that a unitary implementation of the dynamics with respect to an invariant Fock representation requires the function $f$ in Eq. (\[transform\]) to be the unit function. There remains however the possibility of a nontrivial time dependent canonical transformation, coming from the redefinition of the momentum $P_{\phi}=P_{\varphi}+g(t)\sqrt{h}\varphi$. [We will now show that (in less than four spatial dimensions) two distinct scenarios may occur.]{} If the sequence [of]{} elements $g_n/\omega_n^2$ is not summable, then unitarity can only be achieved with $g(t)=0$. Alternatively, if $g_n/\omega_n^2$ gives in fact a summable sequence, then one can attain a unitary dynamics for any function $g(t)$, but this is possible only in the representation defined by $J_0$, or in representations that are unitarily equivalent to it, and therefore the physical predictions remain uniquely determined. Let us return to the summability condition that guaranties the unitary implementation of the dynamics with respect to the representation selected by the complex structure $J$, condition which in particular implies that the sequence $\{\sqrt{g_{n,m}}\tilde\beta^J_{n,m}(t)/(\kappa_{n,m}^*)^2\}$ is also square summable (over $n$ and $m$). We particularize the discussion to the only allowed case, $f(t)=1$, as we have seen. Then, a direct calculation shows that $$\begin{aligned} \label{g1} \frac{\tilde\beta^J_{n,m}(t)}{(\kappa_{n,m}^*)^2}&=& \beta_n\left(1+\frac{ig(t)}{2\omega_n}\right)-z_{n,m}^2\,\beta^*_n \left(1-\frac{ig(t)}{2\omega_n}\right)+ iz_{n,m}\frac{g(t)}{\omega_n}\left[\Re(\beta_n)+\Re(\alpha_n)\right]\nonumber\\ &&+ i \frac{g(t)}{2\omega_n}\alpha_n^*+ i\frac{g(t)}{2\omega_n}z_{n,m}^2\alpha_n +2iz_{n,m}\Im(\alpha_n).\end{aligned}$$ [The symbol $\Re$ stands for real part.]{} Note that the square summability of $\sqrt{g_{n,m}}\beta_n$ and the boundedness of $|z_{n,m}|$ imply that all the terms in $\beta_n$ lead to square summable contributions. Since the set of square summable objects is a linear space, we conclude that a necessary condition for the unitary implementation of the dynamics is that the sum $\sum_n\sum_m g_{n,m}|B_{n,m}|^2$ be finite, where $$\label{g2} B_{n,m}(t)= 2z_{n,m}\Im(\alpha_n)+ \frac{g(t)}{2\omega_n}\left[\alpha_n^* + z_{n,m}^2\alpha_n +2z_{n,m}\Re(\alpha_n)\right]$$ is the remaining part of $\tilde\beta^J_{n,m}(t)/(\kappa_{n,m}^*)^2$ (divided by $i$). We now make use of the analysis performed in Ref. [@CMV8], where it was demonstrated that, up to order $1/\omega_n$, one gets the asymptotic behavior $\alpha_n(t)\approx e^{-i\omega_n\tau}$ for large $n$. As a consequence, it is easy to see that a necessary condition for a unitary quantum dynamics is the finiteness of $\sum_n\sum_m g_{n,m}|A_{n,m}|^2$, where we have called $$\label{g4} A_{n,m}(t)= 2|z_{n,m}|\Im(\alpha_n)+ \frac{g(t)}{2\omega_n}\left[e^{i(\omega_n\tau-\delta)} + |z_{n,m}|^2e^{-i(\omega_n\tau-\delta)} +2|z_{n,m}|\cos(\omega_n\tau)\right].$$ Here, we have introduced the notation $z_{n,m}=|z_{n,m}|e^{i\delta}$. Since $\omega_n\to\infty$, it is clear that the sequence [of]{} elements $\sqrt{g_{n,m}}A_{n,m}/\omega_n$ must also be square summable (over $n$ and $m$). In addition, we know that the contribution to this sequence coming from the second term in Eq. (\[g4\]) is square summable, because so is $\sqrt{g_{n,m}}/\omega_n^2$ (as discussed in Sec. \[sec:model\]) and the multiplying factor is bounded in norm for each $t$, as one can easily check (recall that $|z_{n,m}|\leq 1$). Hence, the contribution of the first term, namely the sequence formed by $\sqrt{g_{n,m}}|z_{n,m}|\Im[\alpha_n(t)]/\omega_n$, must be square summable as well for all times $t$. But then, the kind of arguments presented at the end of Sec. \[unique rep\] (and discussed in more detail in Ref. [@CMV8]) lead us to conclude that $\{\sqrt{g_{n,m}}|z_{n,m}|/\omega_n\}$ must be square summable. Let us consider again the sequence given by $\sqrt{g_{n,m}}A_{n,m}$. The terms coming from the two last summands in Eq. (\[g4\]) are clearly square summable, since $|z_{n,m}|e^{-i(\omega_n\tau-\delta)} +2\cos(\omega_n\tau)$ is bounded in norm by 3 and we have already seen that $\sqrt{g_{n,m}}|z_{n,m}|/\omega_n$ has this summability property. Therefore, the rest of summands provide also a square summable sequence. In particular, the imaginary part is necessarily square summable by its own. In this way, we deduce that $$\label{mark} \frac{g(t)}{\omega_n}\sqrt{g_{n,m}}\sin(\omega_n\tau-\delta)$$ has to be square summable at all instants of time, $t$. Obviously, this condition is satisfied if the function $g(t)$ vanishes identically. On the contrary, let us suppose that this is not the case. Since the function $g(t)$ is continuous, if it is not the null function there must exist an interval of values of $t$ for which it differs from zero. In consequence, $\sqrt{g_{n,m}}\sin(\omega_n\tau-\delta)/{\omega_n}$ must provide a square summable sequence at all values of $t$ in that interval. Then, applying once more the type of arguments employed at the end of Sec. \[unique rep\] and detailed in Ref. [@CMV8] (actually, in this case one can appeal to simpler arguments like those published in Refs. [@unique-gowdy-1; @CQG25; @PRD79]), we conclude that the sequence formed by $\sqrt{g_{n,m}}/{\omega_n}$ must be square summable. We thus see that, in those cases where the sum of $g_{n,m}/\omega_n^2$ (over $n$ and $m$) diverges, we arrive at a contradiction, proving that unitarity can be reached exclusively if $g(t)$ vanishes. This happens, for instance, when the spatial manifold is the 2-sphere [@CQG25] or the 3-sphere [@CMV8]. On the other hand, in the case that $\{g_{n,m}/\omega_n^2\}$ has a finite sum (like, e.g., when the manifold is $S^1$ [@CMVPRD75]), we consider again the sequence [of]{} elements $\sqrt{g_{n,m}}A_{n,m}$ and analyze in further detail the condition that it be square summable. From our discussion in the paragraph above Eq. (\[mark\]) and the assumed summability of $g_{n,m}/\omega_n^2$, we get that the contribution coming from the first term in Eq. (\[g4\]), namely $\sqrt{g_{n,m}}|z_{n,m}|\Im[\alpha_n(t)]$ (up to an irrelevant multiplicative factor), is actually square summable for all the values of $t$ in the studied interval. Then, a straightforward generalization of the discussion presented in Ref. [@CMV8] (see Sec. IV.C) allows us to conclude that $\sqrt{g_{n,m}}|z_{n,m}|$ forms a square summable sequence and, moreover, that the same applies to $\sqrt{g_{n,m}}|\lambda_{n,m}|$. This last step follows from the fact that the convergence of the partial sums of $g_{n,m}|z_{n,m}|^2$ implies that $|\lambda_{n,m}|$ tends to zero when $n\to \infty$. Since $|\kappa_{n,m}|^2=1+|\lambda_{n,m}|^2$, we then have that $|\kappa_{n,m}|\to 1$ in that limit, and thus the value of $1/|\kappa_{n,m}|$ is bounded at large $n$. Summarizing, $g(t)$ is necessarily the zero function unless $\{g_{n,m}/\omega_n^2\}$ is summable, and in that case one must have that $\sum_{n,m} g_{n,m}|\lambda_{n,m}|^2$ is finite. Remarkably, this is precisely the condition that guaranties that the representation defined by the complex structure $J$ (with Bogoliubov coefficients of the “beta” type given by [$\lambda_{n,m}$]{}) is unitarily equivalent to the representation determined by the complex structure $J_0$. Therefore, $g(t)$ must vanish identically unless $g_{n,m}/\omega_n^2$ is summable. If this last property is satisfied, one may change the momentum by adding a time dependent, linear contribution of the field, while respecting the existence of invariant representations which implement the dynamics as a unitary transformation. However, all such representations belong to the same unitary class of equivalence, which is just the class containing the representation determined by the complex structure associated with the massless [situation]{}, $J_0$. In this sense, we can ensure the uniqueness of the field description and its corresponding Fock representation under our criteria of symmetry invariance and unitary evolution. This is the main result of the present paper. For the sake of completeness, the next section will be devoted to discuss how the selected [*unitary*]{} Fock quantization is related with the Fock quantization obtained by imposing the so-called Hadamard condition [@wald]. To make the discussion more accessible, we will start by briefly recalling the context in which the Hadamard approach arises, emphasizing the physical relevance of this formulation, and the uniqueness result that it provides in universes with compact spatial sections. Connection with the Hadamard quantization {#sec:hadamard} ========================================= As it is well known, in the theory of scalar fields there exist classical observables which have no counterpart within the Weyl algebra of quantum observables. This happens with the stress-energy tensor, which is excluded from the Weyl algebra owing to its quadratic dependence on the field, involving the (mathematically ill-defined) product of distributions. In order to incorporate this tensor in the quantum theory, a procedure was introduced in the seventies called point-splitting (see for instance Ref. [@point-splitting]). This method provides a consistent regularization scheme by extracting the spurious infinities associated with quadratic field terms. Roughly speaking, the point-splitting renormalization method assumes that the expectation value of the anticommutator function $G(x,y)=\langle\phi(x)\phi(y)+\phi(y)\phi(x)\rangle$, for the state of interest, possesses a Hadamard singularity structure [@Hadamard-book] in small normal neighborhoods. Since the expectation value of the stress-energy tensor can be obtained from $G(x,y)$ by differentiation, the regularization of $G(x,y)$ provides a renormalized value of it. The prescription consists then in subtracting a suitable Hadamard solution to $G(x,y)$ and declaring the coincidence limit of this difference as the regularized value of the two-point function. The limit $x\to y$ in the formal point separated expression of the expectation value of the stress-energy tensor will exist and define a finite value. The point-splitting prescription relies on the use of Hadamard states (i.e., states satisfying the Hadamard ansatz), which can be proven to exist in [*any*]{} globally hyperbolic spacetime. Therefore, given a free scalar field in an arbitrary (globally hyperbolic) spacetime, one can specify a Hadamard representation of the CCR’s by looking for a Fock vacuum state satisfying the Hadamard condition \[i.e., a state whose two-point function $G(x,y)$ has a short-distance behavior of the Hadamard type\]. This approach rules out infinitely many Fock representations. Since this Hadamard condition is sufficient to ensure that a well-defined quantum stress-energy tensor is obtained, it is reasonable from a physical point of view to impose it (i.e., implement the Hadamard approach) as a criterion to select the representation of the CCR’s, at least if the classical background in which the field propagates is given a physical significance. Unfortunately, the Hadamard criterion does not suffice to pick out a [*unique*]{} preferred quantization in general; indeed, generically there exist infinitely many non unitarily equivalent Hadamard vacuum states. Remarkably, for free scalar fields in spacetimes with compact Cauchy surfaces, it has been shown [@wald] that all Hadamard vacua belong to the same class of unitarily equivalent states. This result, together with the uniqueness discussed in the previous sections, imply that we have at our disposal two different criteria in order to select a unique preferred quantization of the linear KG field. Thus, for such systems, one may wonder whether the [*unitary*]{} and the [*Hadamard*]{} quantizations are in conflict or not. This is the question that we want to address in this section. For the sake of conciseness, let us consider the case of a KG field $\phi$ with mass $m$ on a closed FRW spacetime with the spatial topology of a 3-sphere ($k=+1$). It is a simple exercise to see, in conformal time, that under the time dependent scaling $\varphi=a\phi$, where $a$ is the scale factor, the dynamics of the scaled field $\varphi$ coincides with that of a scalar field with time varying mass $s(t)=m^{2}a^{2}-(\ddot{a}/a)$ propagating in a static background whose Cauchy surfaces are 3-spheres. Now, the first thing we must notice is that the Hadamard and the unitary quantizations are constructed from different phase space descriptions: on the one hand, the unitary quantization is based on a preferred representation for the scaled field $\varphi$, selected as the fundamental field by the criteria of unitarity and spatial symmetry invariance (see Sec. \[result1\]), which is determined by the complex structure $J_{0}$; on the other hand, the Hadamard quantization rests on a preferred representation of the field $\phi$ obtained by imposing the Hadamard condition. In short, the Hadamard and the unitarity (combined with spatial symmetry invariance) criteria select representations of the CCR’s for [*distinct*]{} fields, related by a time dependent canonical transformation. In order to properly compare these quantizations we have to: (i) choose (once and for all) a basic field variable, say $\varphi$ (ii) determine how the Hadamard quantization can be translated to the $\varphi$-description, and (iii) compare the result with the representation selected by unitary evolution and spatial symmetry invariance. As we will show below, the Hadamard quantization defines a representation of the CCR’s, when reformulated in the $\varphi$-description, which is related by means of a unitary transformation with the quantization picked out by our criteria. This result will be achieved by employing that, on closed FRW spacetimes and in the $\phi$-description, Hadamard states are indeed unitarily equivalent to adiabatic vacuum states [@erratajunker] [^7]. Translating the form of adiabatic states to the $\varphi$-description, we will establish the equivalence of the quantization with the one selected by $J_0$ by proving that the transformation that relates the corresponding vacuum states is unitary. Hence, in the framework of the $\varphi$-description, the Hadamard quantization defines a theory which allows for the same physical predictions than the quantum theory specified by the requirement of a unitary evolution, together with the invariance under the spatial symmetries. In this sense, we can assure that there is no tension between the unitary and the Hadamard quantizations. To demonstrate that the vacuum state defined by $J_0$ is unitarily equivalent to an adiabatic vacuum state in the $\varphi$-description, we will consider four steps. In the first one, we will extract the Cauchy data for an adiabatic state (in particular of zeroth order) for the field $\phi$. Next, we will find (via the time dependent canonical transformation) the corresponding Cauchy data in the $\varphi$-description. Then we will consider the Cauchy data that parametrize our $J_0$-state. And, finally, we will compare the two sets of Cauchy data parameterizing the different states, concluding that they are unitarily related. Let us start by recalling the definition of adiabatic states. In a closed FRW spacetime, with metric $g_{ab}=-d\tau_{a} d\tau_{b}+a^{2}(\tau)h_{ab}$, where $\tau$ denotes the cosmological time and $h_{ab}$ stands for the round metric of the 3-sphere, the dynamics of the field $\phi$ is dictated by the differential equation $$\label{phi-frw-cosmo-time} \phi^{\prime\prime}+3\frac{a^{\prime}}{a}\phi^{\prime}-\frac{1}{a^{2}}\Delta\phi+m^{2}\phi=0.$$ Here, the prime denotes the derivative with respect to $\tau$. One can perform a mode decomposition of the field: $$\phi(\tau,{\bf{x}})=\sum_{{\bf{n}}} \left[ a_{{\bf{n}}}\phi_{{\bf{n}}}(\tau,{\bf{x}})+ a^{*}_{{\bf{n}}}\phi^{*}_{{\bf{n}}}(\tau,{\bf{x}}) \right]; \qquad \phi_{{\bf{n}}}(\tau,{\bf{x}})=Q_{{\bf{n}}}({\bf{x}})u_{n}(\tau),$$ where $\{Q_{{\bf{n}}}({\bf{x}})\}$ is a complete set of eigenfunctions of the LB operator, $\Delta Q_{{\bf{n}}}=-n(n+2)Q_{{\bf{n}}}$, and $\bf{n}$ denotes the tuple formed by the eigenvalue integer label $n$ and the degeneration labels $l$ and $m$, standard for the harmonics on the 3-sphere (see, e.g., Ref. [@CMV8]). The time dependent part of the mode solutions, $u_{n}$, satisfies $$\label{timemodes-frw-cosmo-time} u_{n}^{\prime\prime}+3\frac{a^{\prime}}{a}u_{n}^{\prime}+w_{n}^{2}u_{n}=0; \quad w_{n}^{2}=\frac{n(n+2)}{a^{2}}+m^{2}.$$ In addition, the modes $u_n$ are subject to the normalization condition $u_{n}(u^{*}_{n})^{\prime}-u^{*}_{n}u_{n}^{\prime}=ia^{-3}$, coming from the requirement that the corresponding field solutions be normalized with respect to the KG inner product and the fact that the eigenfunctions $\{Q_{\bf{n}}\}$ are orthonormal on the 3-sphere. At cosmological time $\tau_0$, the Cauchy data of the field modes $u_{n}$ are $$\label{f-cauchy-data} q_{n}=u_{n}\vert_{\tau_0},\quad p_{n}=a^{3}u_{n}^{\prime}\vert_{\tau_0}.$$ In terms of the Cauchy data $q_{n}$ and $p_{n}$, the normalization condition reads $q_{n}p^{*}_{n}-q^{*}_{n}p_{n}=i$. Let us focus our attention on solutions of the form $$\label{adiab-sol} u_{n}(\tau )=\frac{1}{\sqrt{2a^{3}\Omega_{n}}} \exp\left(-i\int_{\bar{\tau}}^{\tau} \Omega_{n}(\tilde{\tau})d\tilde{\tau}\right).$$ Substituting this formula in Eq. (\[timemodes-frw-cosmo-time\]), we get that the positive functions $\Omega_{n}$ must satisfy $$\label{lambda-sol} \Omega^{2}_{n}=w_{n}^{2}-\frac{3}{4}\left(\frac{a^{\prime}}{a}\right)^{2} -\frac{3}{2}\frac{a^{\prime\prime}}{a}+\frac{3}{4}\left(\frac{\Omega_{n}^{\prime}}{\Omega_{n}}\right)^{2} -\frac{1}{2}\frac{\Omega^{\prime\prime}_{n}}{\Omega_{n}}.$$ We can try to solve this equation by an iterative process, in which one obtains the $r$-th (positive) function $\Omega^{(r)}_{n}$ from the preceding one $\Omega^{(r-1)}_{n}$; namely, $$\label{n-order} \left(\Omega^{(r+1)}_{n}\right)^{2}=w_{n}^{2}-\frac{3}{4}\left(\frac{a^{\prime}}{a}\right)^{2} -\frac{3}{2}\frac{a^{\prime\prime}}{a}+\frac{3}{4}\left(\frac{\Omega^{(r)\,\prime}_{n}}{\Omega^{(r)}_{n}}\right)^{2} -\frac{1}{2}\frac{\Omega^{(r)\,\prime\prime}_{n}}{\Omega^{(r)}_{n}},\quad r\in\mathbb{N};\quad \left(\Omega^{(0)}_{n}\right)^{2}=w_{n}^{2}.$$ In general, because of the arbitrariness of the scale factor $a$, one cannot ensure the positivity of the right-hand side in the first formula of Eq. (\[n-order\]), so that the iteration procedure may break down. However, it can be shown that, for a sufficiently large $n$, $\big(\Omega^{(r+1)}_{n}\big)^2$ is always strictly positive in a finite time interval [@luders]. Hence, the iteration procedure can be safely performed whenever a finite time interval and an ultraviolet regime are considered. An adiabatic vacuum state of $r$-th order is a Fock state constructed from a solution $u_{n}$ to Eq. (\[timemodes-frw-cosmo-time\]) with initial conditions at time ${\tau}_0$: $$\label{initial-cond} u_{n}(\tau_0)=W^{(r)}_{n}(\tau_0),\quad u^{\,\prime}_{n}(\tau_0)=W^{\,(r)\,\prime}_{n}(\tau_0),$$ where $W^{(r)}_{n}(\tau_0)$ is given by $$\label{ansatz-adiabatic} W^{(r)}_{n}(\tau_0)=\frac{1}{\sqrt{2a^{3}\Omega^{(r)}_{n}}} \exp\left(-i\int_{\bar{\tau}}^{\tau_0} \Omega^{(r)}_{n}(\tilde{\tau})d\tilde{\tau}\right).$$ In particular, using $\Omega^{(0)}_{n}=w_{n}=[n(n+2)+m^{2}a^{2}]^{1/2}/a$ one obtains the adiabatic solution of zeroth order, $W^{(0)}_{n}$. Then, from Eq. (\[f-cauchy-data\]), we get that the Cauchy data for the zeroth order adiabatic state at time $\tau_0$ are $$\label{cd-zoas} q_{n}=W^{(0)}_{n},\quad p_{n}=-a^{2}W^{(0)}_{n}\left[a^{\prime}\left(1+\frac{m^{2}}{2w^{2}_{n}}\right)+iaw_{n}\right].$$ By using the map $\varphi=a\phi$, as well as the relationship between conformal and cosmological times[[^8]]{}, the corresponding Cauchy data in the $\varphi$ description at $t_{0}$ are given by, $$\label{cd-jo} Q_{n}=a W^{(0)}_{n},\qquad P_{n}=-aW^{(0)}_{n}\left(\frac{\dot{a}m^{2}}{2aw^{2}_{n}}+iaw_{n}\right).$$ It is straightforward to check that $Q_{n}P^{*}_{n}-Q^{*}_{n}P_{n}=i$. Next, let us consider the mode solutions of the field $\varphi$ associated with the complex structure $J_0$. We will call $v_{n}(t)$ the time dependent part of these solutions. At the reference conformal time $t_{0}$, the Cauchy data of $v_n$ defining (and defined by) the field decomposition dictated by $J_{0}$ are $$\label{cd-scaledfield} \bar{Q}_{n}=v_{n}\vert_{t_{0}}=\frac{1}{[4n(n+2)]^{1/4}},\qquad \bar{P}_{n}=\dot{v}_{n}\vert_{t_{0}}=-i\left[{\frac{n(n+2)}{4}}\right]^{1/4}.$$ Clearly, this pair of data satisfies the normalization condition $\bar{Q}_{n}\bar{P}^{*}_{n}-\bar{Q}^{*}_{n}\bar{P}_{n}=i$. The zeroth order adiabatic state, parametrized by the Cauchy data (\[cd-jo\]) obtained by “dragging” the state to the $\varphi$-description, is related to the vacuum state characterized by the data (\[cd-scaledfield\]) via a Bogoliubov transformation of the form: $$\label{cd-relation} Q_{n}=\alpha_{n}\bar{Q}_{n}+\beta_{n}\bar{Q}^{*}_{n},\qquad P_{n}=\alpha_{n}\bar{P}_{n}+\beta_{n}\bar{P}^{*}_{n},$$ where $$\label{bogo-coeff} \alpha_{n}=i(P_{n}\bar{Q}^{*}_{n}-Q_{n}\bar{P}_{n}^{*}),\quad \beta_{n}=i(Q_{n}\bar{P}_{n}-\bar{Q}_{n}P_{n}).$$ The equivalence of the considered states depends on whether the antilinear part of the Bogoliubov transformation defines a square summable sequence; namely, $\sum_{{\bf{n}}}|\beta_{n}|^2<\infty$, where we have already taken into account that $\beta_{n}$ depends on $n$ only. Since each eigenspace of the LB operator on $S^{3}$ has dimension $g_n=(n+1)^2$, the square summability condition reads $\sum_{n}g_{n}|\beta_{n}|^2<\infty$. That is, the states will be unitarily related if and only if this sum is finite. To elucidate whether this is the case or not, we will analyze the asymptotic behavior of $\beta_{n}$ and prove that the answer is in the positive. Therefore, the unique (up to unitary equivalence) Hadamard vacuum state gives, in the $\varphi$-description, a state which is in fact unitarily equivalent to the vacuum determined by $J_0$. From Eqs. (\[cd-jo\]) and (\[cd-scaledfield\]), it is straightforward to see that $$\label{quasi-beta} Q_{n}\bar{P}_{n}-\bar{Q}_{n}P_{n}=\frac{aW^{(0)}_{n}}{[4n(n+2)]^{1/4}} \left[i\left(aw_{n}-\sqrt{n(n+2)}\right)+\frac{\dot{a}m^{2}}{2aw^{2}_{n}}\right].$$ Substituting in this equation the expression of $W^{(0)}_{n}$, and writing $n(n+2)=a^{2}w^{2}_{n}(1-x_{n}^{2})$, where $x_{n}=m/w_{n}$, we get $$Q_{n}\bar{P}_{n}-\bar{Q}_{n}P_{n}=\frac{1}{2(1-x_{n}^{2})^{1/4}}\left[i\left(1-\sqrt{1-x_{n}^{2}}\right)+\frac{\dot{a}x^{3}_{n}}{2ma^2} \right] e^{-i\int w_{n}}.$$ Thus, in the asymptotic limit $n>>1$ (i.e., when $x_{n}<<1$) the ultraviolet behavior of $\beta_{n}$ is $$\beta_{n}=i\left[\frac{im^{2}a^{2}}{4 n^{2}}+O\left(\frac{1}{n^{3}}\right)\right] e^{-i\int w_{n}}.$$ Therefore $\sqrt{g_{n}}\beta_{n}\sim O(1/n)$, a fact that implies that $\{\sqrt{g_{n}}\beta_{n}\}$ is square summable. So, the analyzed states are equivalent. In conclusion, the Fock quantization selected by the criterion of a unitary evolution (together with the invariance under the spatial symmetries) defines a representation of the CCR’s which is unitarily equivalent to the one determined by the Hadamard criterion when the latter is translated to the $\varphi$-description. On the one hand, the fact that the two approaches, namely the Hadamard criterion and the unitary one, select the same unitary equivalence class of representations –in the spatially compact case and using the $\varphi$-description– is probably not completely unexpected, since both approaches rely on related dynamical aspects. However, the two perspectives are, at least [*a priori*]{}, intrinsically different. In the unitary approach, what is imposed is only the existence of unitary transformations implementing the evolution between any two (regular) instants separated by a finite (not [*infinitesimal*]{}) interval of time, with no further requirement regarding continuity with respect to time, or any pre-established local form of the vacuum state. On the other hand, in the Hadamard approach a seemingly stronger condition, fixing the local singularity structure of the vacuum state, is imposed, which is strong enough to ensure the regularization of the stress-energy tensor. It seems far from obvious whether these two approaches should lead to equivalent quantizations. If one adopts the point of view, as we do, that preserving unitarity of the dynamics is a desirable aspect in quantum physics, the fact that the two perspectives actually lead to equivalent quantum theories appears by itself as an interesting and reassuring result. It is also worthwhile mentioning that the Hadamard condition essentially translates the information about the causal structure of the classical background into the local structure of the quantum states. This is of course what one wants when the classical background has a true physical meaning, but things are less clear when the background is only an effective or an auxiliary one. In particular, when part or all of the degrees of freedom are gravitational, the true causal structure is a dynamical entity with possibly little or no relation with the causal structure of the auxiliary background where the degrees of freedom are represented as scalar fields. This happens e.g., in the case of Gowdy models and in the treatment of cosmological perturbations [@unit-gt3; @fmov]. Similarly, when quantum corrections are partially incorporated in the spacetime where the scalar field propagates, its causal structure is only an effective concept. In such cases, we find it important that one can take advantage of criteria which do not make explicit use of the causal structure of the background as a fundamental entity. Finally, let us emphasize that the established relation between the Hadamard criterion and the unitarity criterion applies just to the $\varphi$-description, while it is exclusively the latter of these criteria (together with the invariance under the spatial symmetries) which picks out that description as a privileged one. Conclusions =========== As we have discussed, a major problem in the quantization of (scalar) fields in nonstationary scenarios is the ambiguity that generically appears in the selection of a Fock quantum description. On the one hand, the possibility of absorbing part of the field evolution in the time dependence of the spacetime where the propagation takes place affects the choice of a canonical pair for the field, as well as the dynamics of the system that we want to quantize. On the other hand, even if a specific pair is picked out, among all those related by time dependent linear canonical transformations, it is well known that there exists an infinite number of unitarily inequivalent representations for the corresponding CCR’s and, therefore, of physically different quantum theories, each of them leading to different results. In this situation, it is clear that the quantum predictions have doubtful significance, because if they are falsified one can always adhere to another inequivalent Fock quantization in the infinite collection at hand. This problem is especially relevant in cosmology, a context where the setting is naturally nonstationary, and is so both because the window for quantum effects seems to be narrow and because one cannot falsify the quantum physics by an unlimited number of repeated measurements, but rather by observing the Universe in which we live. In these circumstances, determining an unambiguous quantization whose predictions can be trusted is essential if one wants to develop a realistic program of quantum cosmology. We recently proved that, when the field dynamics can be put in the form of that of a KG field in a static spacetime but with a time varying mass, there exist some reasonable criteria which allow one to select a unique unitary class of equivalence of Fock representations, and hence one reaches uniqueness in the Fock quantization. These criteria are the invariance of the vacuum under the spatial symmetries of the field equations and the unitary implementation of the field dynamics. This uniqueness result is valid for fields defined on [*any*]{} compact spatial manifold in three or less dimensions [@zejaguije]. In other words, in less than five spacetime dimensions, the spatial topology is not relevant as far as compactness is guaranteed. In noncompact cases, the infrared divergences play an important role and generically prevent the extension of the result. Even so, in cosmology for instance, one can appeal to the physical irrelevance of large scales beyond a causal [radius]{} to justify that the results obtained with the assumption of compactness should still be applicable. In many practical situations, and in particular for fields in cosmological spacetimes, the above field description, for which our uniqueness theorem had been proven, is reached indeed after a suitable scaling of the field by a function of time. This scaling can be considered, as we have commented, part of a linear canonical transformation, obviously time dependent, in which the momentum suffers the inverse scaling. Besides, in this canonical transformation, it is extremely convenient to allow for a possible time dependent linear contribution of the field to the redefined momentum. In this work, we have analyzed the effect of this class of canonical transformations on the quantization. Since the transformations are time dependent, they actually modify the dynamics of the field, and hence affect the restrictions imposed by our uniqueness criteria, which include the unitarity of the evolution. In consequence, these canonical transformations introduce a new infinite ambiguity in the quantization of the system, previous to the choice of Fock representation once a particular field description is accepted. The main result of this work is to demonstrate that, again for [*any*]{} compact spatial manifold in three or less dimensions, there exists no ambiguity in the choice of field description if one insists in our criteria of vacuum invariance under the spatial symmetries and a unitary implementation of the dynamics. More specifically, we have proven that no scaling of the field is permitted with respect to the description in which the propagation occurs apparently in a static background, if one wants to reach a Fock representation in which the vacuum has the spatial symmetries of the field equations and the corresponding dynamics is implemented as a unitary transformation. This only leaves the [freedom]{} of changing the momentum by adding a time dependent contribution that is linear in the field. We have shown that there exist two possibilities. If the LB operator, excluding the subspace of zero modes, has an inverse that is not trace class (so that the sum of $g_{n,m}/\omega^2_n$ diverges), then the form of the momentum is totally fixed by our two requirements of vacuum invariance and unitary evolution. No freedom exists to add a linear contribution of the field. In this way, the field description of the system is completely determined by our criteria, and the studied time dependent canonical transformations are all precluded, except the trivial one. This is in fact the situation encountered, e.g., in the case of $T^3$ topology [@threetorus] or $S^3$ topology [@zejaguije]. The other possibility is that, on the opposite, the inverse of the LB operator, once its kernel is removed, is indeed trace class. Typically, this happens if the spatial manifold on which the field theory is defined is one dimensional. The number of eigenstates of the LB operator with eigenvalue smaller or equal than $\omega_{n}$ (i.e. $\sum_{\tilde{n}\leq n} \sum_m g_{\tilde{n},m})$ grows then at most like $\omega_{n}$, and the eigenvalue itself should grow like $n$. It is then not difficult to check that the sum of $g_{n,m}/\omega^2_n$ is finite. In this case, changes in the momentum that add a term which is linear in the field, multiplied by any function of time and properly densitized, are allowed while respecting the existence of a Fock representation which satisfies our criteria in the field description with the new momentum. However, all these field descriptions can be obtained then directly from the original one, by a straightforward implementation of the canonical transformation. None of these descriptions admit a Fock representation that, while fulfilling the criteria of vacuum invariance and unitary evolution, turns out to be inequivalent to the representation adopted in the original field description. In this sense, the quantization is again unique. These results confirm and extend those obtained for the first time in the context of Gowdy cosmologies with $T^3$ topology [@CMVPRD75], where the effective theory consists of a scalar field propagating on the circle but with a specific time dependent mass. In total, we have proven that, in three or less spatial dimensions, there exists a unique Fock quantization for this kind of systems, up to unitary transformations, if one demands a natural unitary implementation of the spatial symmetries of the field equations and a unitary implementation of the dynamics. This uniqueness result provides the desired robustness to the quantization process, and leads to a quantum theory whose physical predictions are, to the extent discussed in this work, uniquely determined. Finally, let us remark that the Fock quantization selected by our criteria defines a representation which is unitarily equivalent to that corresponding to the Hadamard quantization of a KG field in a closed FRW spacetime provided, of course, that the latter is reformulated in terms of the scaled field $\varphi$. Although we have proven this result only for the case in which the spatial sections are isomorphic to 3-spheres, there seems to be no serious obstruction to extend it to universes with any other compact spatial topology. Acknowledgements {#acknowledgements .unnumbered} ================ This work was supported by the research grants MICINN/MINECO FIS2011-30145-C03-02, MICINN FIS2008-06078-C03-03 and CPAN CSD2007-00042 from Spain, DGAPA-UNAM IN117012-3 from Mexico and CERN/FP/116373/2010 from Portugal. J.O. acknowledges CSIC by financial support under the grant JAE-Pre\_08\_00791. Nonzero limit of oscillatory functions {#Ap1} ====================================== In Sec. \[result1\] we made use of the fact that $\sin(\omega_n \tau)$, and more generally $\sin(\omega_n\tau+\Theta_{n,M})$ (with $M$ fixed for each $n$), cannot tend to zero in the limit $n\to\infty$ on any subsequence of the positive integers for all $t$ (or equivalently for all $\tau=t-t_0$) in a given interval. We will prove this statement in this appendix. Let $[a,b]$ be an interval of the real line with Lebesgue measure $L=b-a$ and $$\mathbb{W}=\{ w_n;\ n\in\mathbb{N}^+\}$$ be a monotonous and diverging sequence of positive real numbers; namely $w_{n+1}>w_n$ for all $n\in\mathbb{N}^+$, with $w_n$ being unbounded for large $n$. In particular, $\mathbb{W}$ may be a subsequence of the sequence of eigenvalues $\{\omega_n;\ n\in\mathbb{N}^+\}$. Besides, let $$\{\theta_n(t); n\in\mathbb{N}^+\}$$ be a sequence of twice differentiable phases, i.e., functions with values on $\mathbb{R}$ modulo $2\pi$. We also require that there exist positive numbers $X$ and $Y$ such that $$\label{bdchi} |{\dot{\theta}}_{n}|< X,\qquad |\ddot{\theta}_{n}|< Y,$$ for all $n$ (greater than a certain nonnegative integer, $n_0$) and all times $t\in [a+t_0,b+t_0]$. Under these conditions, we will now show that $$u_n(\tau)=\sin^2\left[w_n\tau+\theta_{n}(t)\right]$$ cannot tend to zero $\forall \tau\in {[a,b]}$, which obviously implies that $\sin(w_n\tau+\theta_{n})$ cannot tend to the zero function. The functions $u_n(\tau)$ are clearly integrable, and a straightforward computation shows that $$\int_a^b u_n(\tau)d\tau = \frac{L}{2}-\frac{1}{2} \int_a^b \cos\left[2 w_n\tau - 2\theta_{n}(\tau+t_0 )\right]d\tau.$$ In addition, $$\begin{aligned} \int_a^b \cos\left[2 w_n\tau - 2\theta_{n}(\tau+t_0)\right]d\tau &=& \frac{\sin\left[2 w_n b - 2\theta_{n}(b+t_0)\right]}{2 w_n - 2{\dot\theta}_{n}(b+t_0 )}- \frac{\sin\left[2 w_n a - 2\theta_{n}(a+t_0)\right]}{2 w_n - 2{\dot\theta}_{n}(a+t_0 )}\\ &+& 2 \int_a^b \frac{\ddot\theta_{n}}{\left(2 w_n+2\dot\theta_{n}\right)^2} \sin\left[2 w_n \tau - 2\theta_{n}(\tau+t_0)\right]d\tau,\end{aligned}$$ and $$\bigg|\int_a^b \frac{\ddot\theta_{n}}{\left(2w_n+2\dot\theta_{n}\right)^2}\sin\left[2 w_n \tau - 2\theta_{n}(\tau+t_0)\right]d\tau \bigg|\leq L\,\max_I\bigg|\frac{\ddot\theta_{n}}{\left(2w_n+2\dot\theta_{n}\right)^2}\bigg|.$$ Since $w_n$ is a monotonous diverging sequence, it is now straightforward to check that conditions (\[bdchi\]) are sufficient to ensure that the integral over [$[a,b]$]{} of $\cos\left[2 w_n\tau - 2\theta_{n}(\tau+t_0 )\right]$ tends to zero when $n$ goes to infinity. Therefore, the sequence of integrals $\int_a^b u_n(\tau)d\tau$ converges to $L/2$. Finally, let us suppose that the sequence of functions $u_n(\tau)$ converges to the zero function on $[a,b]$. Since the functions $|u_n(\tau)|$ are bounded from above by the constant unit function, we can apply the Lebesgue dominated convergence theorem [@simon]. This theorem ensures that the sequence of integrals $\int_a^b u_n(\tau)d\tau$ would converge indeed to the integral of the zero function, i.e. to zero. But this is incompatible with the fact, demonstrated above, that $\int_a^b u_n(\tau)d\tau$ converges to $L/2$. This contradiction shows that the values of $u_n(\tau)$ cannot converge to zero for all values of $\tau\in [a,b]$, as we wanted to prove. The phases $\Theta_{n,M}$ {#Ap2} ========================= In this appendix, we show that expression (\[sproof1\]) can be replaced by expression (\[bigsinus\]) under the assumption that $y_{n,M}$ does not tend to zero. For convenience, we repeat here the starting expression, $$\label{Appsproof1} (f_{+}-x_{n,M}f_{-})\sin(\omega_n\tau)+y_{n,M}f_{-}\cos(\omega_n\tau),$$ obtained with the hypothesis that $x^2_{n,M}+y^2_{n,M}\to 1$ for large $n$. Recall also that $M$ is fixed for each value of the positive integer $n$, and that the functions $f(t)$ and $f_+(t)$ are strictly positive. Let us introduce the definitions $$\begin{aligned} \rho_{n,M}\cos[\Theta_{n,M}]&=&f_{+}-x_{n,M}f_{-}\, ,\nonumber\\ \label{defsin}\rho_{n,M}\sin[\Theta_{n,M}]&=&y_{n,M}f_{-}\, ,\end{aligned}$$ such that $$\label{Apptheta} \cot[\Theta_{n,M}]=\frac{1}{y_{n,M}}\frac{f_{+}}{f_{-}}-\frac{x_{n,M}}{y_{n,M}}$$ and $$\label{rho}\rho^2_{n,M}=(f_{+}-x_{n,M}f_{-})^2+y^2_{n,M}f^2_{-}.$$ In particular, $\rho^2_{n,M}$ is bounded from below by $(f_{+}-|f_{-}|)^2$. Besides, since $y_{n,M}$ does not tend to zero, there exists a subsequence $S$ of values of $n$ and a number $\epsilon>0$ such that $|y_{n,M}|>\epsilon$ on $S$. For $n$ taking values in the subsequence $S$, we then conclude that $$\label{boundrho}\rho^2_{n,M}(t)\geq(f_{+}-|f_{-}|)^2+\epsilon^2 f^2_{-}= \varrho^2(t).$$ We note that the lower bound defined above is strictly positive for all values of $t$: if $f_{-}(t)\not =0$ then $\varrho^2\geq\epsilon^2 f^2_{-}(t)>0$; whereas, if $f_{-}(t) =0$, we have that $f(t) = 1$, and hence $f_{+}(t) = 1$, which implies in turn that $\varrho^2 = 1$. Employing definitions (\[defsin\]), expression (\[Appsproof1\]) reads: $$\label{bigsin}\rho_{n,M}\sin[\omega_n\tau+\Theta_{n,M}].$$ A necessary condition for the unitary implementation of the dynamics is that Eq. (\[Appsproof1\]), and therefore expression (\[bigsin\]), [tend]{} to zero for all the possible values of $t$. In particular, the above expression must tend to zero on the subsequence $S$. But, on that subsequence, which is independent of $t$, the lower bound (\[boundrho\]) is valid, leading to the conclusion that a unitary dynamics requires that the sequence formed by $\sin[\omega_n\tau+\Theta_{n,M}]$ tend to zero on $S$ at all times $t$, as claimed in Sec. \[result1\]. Let us finally show that the first and second derivatives of the functions $\Theta_{n,M}(t)$ constitute uniformly bounded (sub)sequences on $S$ (with respect to the variation of $n$; recall in this sense that the label $M$ is not free, but fixed for each value of $n$). This result shows that the conditions assumed in Appendix \[Ap1\] are actually satisfied. It is straightforward to calculate the first and second time derivatives of $\Theta_{n,M}$: $$\begin{aligned} \dot{\Theta}_{n,M}&=&\frac{y_{n,M}}{\rho^2_{n,M}}\frac{\dot f}{f},\nonumber \\ \label{ddteta} \ddot{\Theta}_{n,M}&=&\frac{y_{n,M}}{\rho^2_{n,M}f}\left(\ddot f-\frac{{\dot f}^2}{f} -\frac{2{\dot f}{\dot f_{-}}}{\rho^2_{n,M}}[(x^2_{n,M}+y^2_{n,M})f_{-}-x_{n,M}f_{+}]\right).\end{aligned}$$ Taking into account that $x^2_{n,M}+y^2_{n,M}\leq 1$ and that $\rho^2_{n,M}(t)$ is bounded from below by $\varrho^2(t)$ on $S$ \[see the bound (\[boundrho\])\], we get that, for each value of $t$, $$\begin{aligned} |\dot{\Theta}_{n,m}|&\leq &\frac{1}{\varrho^2}\frac{|\dot f|}{f},\label{d1teta} \\ \label{dd1teta} |\ddot{\Theta}_{n,m}|&\leq &\frac{1}{\varrho^2f}\left(|\ddot f|+\frac{{\dot f}^2}{f} +\frac{2{|\dot f}{\dot f_{-}}|}{\varrho^2}[|f_{-}|+|f_{+}|]\right).\end{aligned}$$ Since both $f(t)$ and $\varrho^2(t)$ are strictly positive continuous functions, the right hand side of the two inequalities (\[d1teta\]) and (\[dd1teta\]) are indeed bounded functions of $t$ on any closed interval. Hence, for any time interval [$[a,b]$]{}, there exist positive numbers $X$ and $Y$ such that $$\label{bdteta} |\dot{\Theta}_{n,M}|< X,\qquad |\ddot{\Theta}_{n,M}|< Y,$$ for all integers $n$ belonging to the subsequence $S$ and all [times]{}. This concludes our proof. [99]{} A. Ashtekar and R.S. Tate, J. Math. Phys. [**35**]{}, 6434 (1994). B. Simon, [*Topics in Functional Analysis*]{}, edited by R.F. Streater (Academic Press, London, 1972). A. Ashtekar, Gen. Rel. Grav. [**41**]{}, 707 (2009). J.M. Velhinho, Classical Quantum Gravity [**24**]{}, 3745 (2007). M. Bojowald, Living Rev. Relativity [**11**]{}, 4 (2008); G.A. Mena Marugán, J. Phys. Conf. Series [**314**]{}, 012012 (2011); K. Banerjee, G. Calcagni, and M. Martín-Benito, SIGMA [**8**]{}, 016 (2012). A. Ashtekar and P. Singh, Classical Quantum Gravity [**28**]{}, 213001 (2011). T. Thiemann, [*Modern Canonical Quantum General Relativity*]{} (Cambridge University Press, Cambridge, England, 2007). R.M. Wald, [*Quantum Field Theory in Curved Spacetime and Black Hole Thermodynamics*]{} (Chicago University Press, Chicago, 1994). A. Ashtekar and A. Magnon, Proc. R. Soc. Lond. A [**346**]{}, 375 (1975); A. Ashtekar and A. Magnon-Ashtekar, Pramana [**15**]{}, [107]{} (1980). B. Kay, Commun. Math. Phys. [**62**]{}, 55 (1978). R. Floreanini, C.T. Hill, and R. Jackiw, Ann. Phys. [**175**]{}, 345 (1987). V. Mukhanov, [*Physical Foundations of Cosmology*]{} (Cambridge University Press, Cambridge, England, 2005). M. Fernández-Méndez, G.A. Mena Marugán, J. Olmedo, and J.M. Velhinho, Phys. Rev. D [**85**]{}, 103525 (2012). M. Fernández-Méndez, G.A. Mena Marugán, and J. Olmedo, Phys. Rev. D [**86**]{}, 024003 (2012). R.H. Gowdy, Ann. Phys. (N.Y.) [**83**]{}, 203 (1974). J. Cortez and G.A. Mena Marugán, Phys. Rev. D [**72**]{}, 064020 (2005). A. Corichi, J. Cortez, and G.A. Mena Marugán, Phys. Rev. D [**73**]{}, 041502 (2006); [**73**]{}, 084020 (2006). V.F. Mukhanov, JETP Lett. [**41**]{}, 493 (1985); M. Sasaki, Prog. Theor. Phys. [**76**]{}, 1036 (1986). J.M. Bardeen, Phys. Rev. D [**22**]{}, 1882 (1980). A. Corichi, J. Cortez, and H. Quevedo, Ann. Phys. [**313**]{}, 446 (2004). D. Shale, Trans. Am. Math. Soc. [**103**]{}, 149 (1962). R. Honegger and A. Rieckers, J. Math. Phys. [**37**]{}, 4292 (1996). A. Corichi, J. Cortez, G.A. Mena Marugán, and J.M. Velhinho, Phys. Rev. D [**76**]{}, 124031 (2007). A. Corichi, J. Cortez, G.A. Mena Marugán, and J.M. Velhinho, Classical Quantum Gravity [**23**]{}, 6301 (2006). J.F. Barbero G., D.G. Vergel, and E.J.S. Villase$\tilde{\rm n}$or, Classical Quantum Gravity [**25**]{}, 085002 (2008). J. Cortez, G.A. Mena Marugán, and J.M. Velhinho, Classical Quantum Gravity [**25**]{}, 105005 (2008). J. Cortez, G.A. Mena Marugán, R. Serôdio, and J.M. Velhinho, Phys. Rev. D [**79**]{}, 084040 (2009). J. Cortez, G.A. Mena Marugán, and J.M. Velhinho, Phys. Rev. D [**81**]{}, 044037 (2010). J. Cortez, G.A. Mena Marugán, J. Olmedo, and J.M. Velhinho, Classical Quantum Gravity [**28**]{}, 172001 (2011). R.T. Jantzen, J. Math. Phys. [**19**]{}, 1163 (1978). V.F. Mukhanov, H.A. Feldman, and R.H. Bradenberger, Phys. Rep. [**215**]{}, 203 (1992). J.J. Halliwell and S.W. Hawking, Phys. Rev. D [**31**]{}, 1777 (1985). L. Castelló Gomar, J. Cortez, D. Martín-de Blas, G.A. Mena Marugán, and J.M. Velhinho, [*Uniqueness of the Fock Quantization of Scalar Fields in Spatially Flat Cosmological Spacetimes*]{}, JCAP (in press) (2012). J. Cortez, G.A. Mena Marugán, and J.M. Velhinho, Phys. Rev. D [**75**]{}, 084027 (2007). J. Cortez, G.A. Mena Marugán, J. Olmedo, and J.M. Velhinho, JCAP [**10**]{}, 030 (2010); [Phys. Rev. D [**83**]{}, 025002 (2011).]{} M. Blau, M. O’Loughlin, and L. Seri, JHEP [**1204**]{}, 098 (2012). M. Taylor, [*Partial Differential Equations II*]{} (Springer-Verlag, New York, 1996); I. Chavel, [*Eigenvalues in Riemannian Geometry*]{} (Academic Press, Orlando, 1984). A.A. Kirillov, [*Elements of the Theory of Representations*]{} (Springer-Verlag, New York, 1976). S.M. Christensen, Phys. Rev. D [**14**]{}, 2490 (1976); [**17**]{}, 976 (1978); S.L. Adler, J. Liberman, and Y.J. Ng, Ann. Phys. (N.Y.) [**[106]{}**]{}, 279 (1977); P.C.W. Davies, S.A. Fulling, S.M. Christensen, and T.S. Bunch, Ann. Phys. (N.Y.) [**[109]{}**]{}, 108 (1977); R.M. Wald, Commun. Math. Phys. [**[54]{}**]{}, 1 (1977); Phys. Rev. D [**17**]{}, 1477 (1978); T.S. Bunch, S.M. Christensen, and S.A. Fulling, Phys.Rev. D [**[18]{}**]{}, 4435 (1978). J. Hadamard, [*Lectures on Cauchy’s Problem in Linear Partial Differential Equations*]{} (Yale University Press, New Haven, 1923). W. Junker, Rev. Math. Phys.  [**14**]{}, 511 (2002). C. Lüders and J.E. Roberts, Commun. Math. Phys. [**134**]{}, 29 (1990). [^1]: An operator $T$ on a Hilbert space is called Hilbert-Schmidt if the trace of $T^{*}T$ is finite, where $T^*$ is the adjoint operator. [^2]: This is an example where Eq. (\[1new\]) is modified with terms which do not affect the asymptotics. [^3]: Remarkably, our results were recently found useful also in the context of string dynamics in arbitrary plane wave backgrounds [@string]. [^4]: Obviously, these variables are ill-defined for zero modes, i.e., when $\omega_n=0$. However, our discussion on the unitary implementation of the dynamics does not depend on a finite number of modes. So, we will analyze exclusively nonzero modes in the rest of the text. Unitarity and uniqueness for zero modes can be attained following methods and criteria of Quantum Mechanics. [^5]: In principle, Schur’s lemma can be applied only to complex representations, while we are dealing with a basis of real eigenmodes of the LB operator. Nonetheless, since the relation between real and complex eigenmodes is linear, and the dynamics is both linear and common to all the eigenmodes in the same eigenspace, the implications of the lemma can be translated to our description in terms of real modes without serious obstructions for the analysis of the evolution (see, e.g., the discussion in Ref. [@threetorus]). [^6]: See, nonetheless, the comments in the previous footnote. [^7]: A precise characterization of adiabatic states can be found, for instance, in Ref. [@luders]. [^8]: The two times are related by $\tau(t)=\int a dt$. Besides, we choose $t_0$ such that $\tau_0=\tau(t_0)$.
{ "pile_set_name": "ArXiv" }
--- abstract: | An individual has been subjected to some exposure and has developed some outcome. Using data on similar individuals, we wish to evaluate, for this case, the probability that the outcome was in fact caused by the exposure. Even with the best possible experimental data on exposure and outcome, we typically can not identify this “probability of causation” exactly, but we can provide information in the form of bounds for it. Under appropriate assumptions, these bounds can be tightened if we can make other observations (e.g., on non-experimental cases), measure additional variables (e.g., covariates) or measure complete mediators. In this work we propose new bounds for the case that a third variable mediates partially the effect of the exposure on the outcome.\ Keywords: [Probability of Causation, Mediation, Causes of effects, Bounds]{} author: - 'Rossella Murtas[^1]' - 'Alexander Philip Dawid[^2]' - 'Monica Musio[^3]' date: 'November 1, 2016' title: New bounds for the Probability of Causation in Mediation Analysis --- Introduction {#intro} ============ \[sec:introduction\] Causality is a concept very common in real life situations. Is lung cancer caused by smoking? Was contaminated water causing cholera in London in 1854? Can the court infer sex discrimination in a hiring process? However, statisticians have been very cautious in formalizing this concept. One reason may be the complex definitions and methods implemented to study causality. Another explanation may be the difficulty of translating real life problems into mathematical notations and formulas. The first step should be to identify the causal question of interest. This can be assigned to one of two main classes: questions concerning the causes of observed effects, and questions concerning the effects of applied causes. This basic distinction, all too often neglected in the causal inference literature, is fundamental to identifying the correct definition of causation. To clarify this distinction, consider the following example. An individual, Ann, might be subjected to some exposure, $X$, and might develop some outcome, $Y$. For simplicity we take $X$ to be a binary decision variable, denoting whether or not an individual is given the drug, and take the outcome variable $Y$ also to be binary, coded $1$ if the individual dies, and $0$ if not. We denote by $X_A\in\{0,1\}$ the value of Ann’s exposure, and by $Y_A\in\{0,1\}$ the value of Ann’s outcome. Questions about the effects of applied causes, “EoC”, are widely studied. For example, in medicine, randomized clinical trials are one of the most rigorous ways to assess the effect of a treatment in a population. In the EoC framework, at an individual level we would be interested in asking: “What would happen to Ann were she to be given the drug?” or “What would happen to Ann were she not to be given the drug?”. At the population level, a typical EoC query would be: “Is death caused by the drug?” In this framework, a straightforward way to assess the strenght of causality is by comparing $P_1={\mbox{\textrm P}}(Y=1 \mid X \leftarrow 1)$ and $P_0={\mbox{\textrm P}}(Y=1 \mid X\leftarrow 0)$, the two outcome probabilities under the two different interventions [@dawid2016statistical]. This can be seen as a decision problem: we can compare these two different distributions for $Y$, decide which one we prefer, and take the associated decision (give or withhold the drug). The difference $P_1 - P_0$ is known as the “Average Causal Effect”. In contrast to EoC queries, that are mostly adopted to infer knowledge in the population, CoE questions invariably require an individual investigation. For example, suppose that Ann died after being given the drug. A typical CoE question might be phrased as: “Knowing that Ann did take the drug, and died, how likely is it that she would not have died if she had not had the drug?”. In this paper we will embed such causal queries in the counterfactual framework [@rubin1974estimating]. This is based on the idea that there exist *potential variables*. If $X$ is the exposure and $Y$ the outcome, the potential variable $Y(x)$ is conceived as the value of $Y$ that would arise if, actually or hypothetically, $X$ were to be set to $x$ ($X \leftarrow x$). We denote the pair $(Y(0), Y(1))$ by ${{\textbf{Y}}}$. For an actual assignment $X \leftarrow x$, we observe $Y = Y(x)$. The potential variable $Y(x')$, with $x' \neq x$, is then not observable, but is supposed to describe what would have happened to the outcome $Y$, if, counterfactually, we had assigned the different value $x'$ to the exposure $X$. Note particularly that it is never possible to observe fully the pair ${{\textbf{Y}}}$. The definition of a CoE causal effect is completely different from the EoC definition. It is typically framed in terms of the *probability of causation* (PC), also called *probability of necessity* [@pearl1999probabilities]. Given that Ann took the drug and died, the probability of causation in Ann’s case is defined as: $$\label{def:pcA} {\mbox{\rm PC}}_A = \mbox{P}_A(Y_A(0) = 0 \mid X_A= 1, Y_A(1)=1)$$ where $\mbox{P}_A$ denotes the probability distribution over attributes of Ann. For example, suppose that Ann’s children filed a criminal lawsuit against a pharmaceutical manufacturer claiming that their drug was the [*cause*]{} of her death. Using data on similar individuals, we would wish to evaluate, for this case, the probability that the outcome was in fact caused by the exposure. In such a civil case, the required standard of proof is typically “preponderance of the evidence,” or “ balance of probabilities,” meaning that the case would succeed if it can be shown that causation is “more probable than not,” [[*i.e.*]{}]{}, ${\mbox{\rm PC}}_A > 50\%$. However, simplistic or [*ad hoc*]{} definitions and rules are widely and often wrongly applied in many courthouse. Given the possibly serious implications of the probability of causation, it is important to studying methods capable of producing accurate information. From a statistical point of view, definition involves the bivariate distribution of the two potential variables associated with the same subject. However, only one of these can ever be observed, the other then becoming counterfactual. For this reason, $\mbox{PC}_A$ is generally not fully identifiable. We can however provide useful information as bounds between which ${\mbox{\rm PC}}_A$ must lie. Under appropriate assumptions, these bounds can be tightened if we can measure additional variables, [[ *e.g.*]{}]{}, covariates [@dawid2011role], or—in the case that unobserved variables confound the exposure-outcome relationship—gather data on other, nonexperimental, cases (Tian and Pearl [@tian2000probabilities].) In this paper we propose a novel approach to bound the probability of causation in mediation analysis. Mediation aims to disentangle the extent to which the effect of $X$ on $Y$ is mediated through other pathways from the extent to which that effect is due to $X$ acting directly on $Y$. In [§$\,$\[sec:simple\]]{} we revisit the basic framework where we have information only on exposure and outcome. In [§$\,$\[sec:mediation\]]{} we focus on two different mechanisms: complete and partial mediation. In the former, the exposure is supposed to act on the outcome only through the mediator, [[*i.e.*]{}]{}, no direct effect is present. In the latter, both direct and indirect effects are considered. In [§$\,$\[sec:comparison\]]{} we compare the bounds obtained in [§$\,$\[sec:mediation\]]{} with those reviewed in [§$\,$\[sec:simple\]]{}, and in [§$\,$\[sec:conclusion\]]{} we present our conclusions. Starting Point: Simple Analysis {#sec:simple} =============================== In this Section we discuss the simple situation in which we have information, as in [Table \[tab:aspirin1\]]{}, from a randomized experimental study that tested the same drug taken by Ann. [lccc]{} & Die & Live & Total\ Exposed & 30 & 70 & 100\ Unexposed & 12 & 88 & 100\ $$\begin{aligned} \mbox{P}(Y = 1 \mid X\leftarrow 1) &=& 0.30 \label{eq:pop-rate1}\\ \mbox{P}(Y = 1 \mid X\leftarrow 0) &=& 0.12. \label{eq:pop-rate0}\end{aligned}$$ We see that, in the experimental population, individuals exposed to the drug ($X\leftarrow1$) were more likely to die than those unexposed ($X\leftarrow 0$), by $18$ percentage points. So can the court infer that it was Ann’s taking the drug that caused her death? More generally: Is it correct to use such experimental results, concerning a population, to say something about a single individual? This “Group-to-individual” (G2i) issue is discussed by Dawid [@dawid2014fitting]. The simple difference between and is not sufficient to infer causation for a single external individual. To make progress we add a further assumption that the event of Ann’s exposure, $X_A$, is independent of her potential response pair $\textbf{Y}_A$: $$X_A {\mbox{$\,\perp\!\!\!\perp\,$}}{{\textbf{Y}}}_A. \label{eq:suff}$$ Property parallels the “no-confounding” property $X_i {\mbox{$\,\perp\!\!\!\perp\,$}}\textbf{Y}_i$ that holds for individuals $i$ in the experimental study on account of randomization. We further suppose that Ann is exchangeable with the individuals in the experiment, [[*i.e.*]{}]{}, she could be considered as a subject in the experimental population. On account of and exchangeability, $\mbox{PC}_A$ in reduces to $\mbox{PC}_A = \mbox{P}(Y(0) = 0 \mid Y(1)=1)$—but we can not fully identify this from the data. In fact we can never observe the joint event $(Y(0)=0;Y(1)=1)$, since at least one of $Y(0)$ and $Y(1)$ must be counterfactual. In particular, we can never learn anything about the dependence between $Y(0)$ and $Y(1)$. However, even without making any assumptions about this dependence, we can derive the following inequalities (Dawid *et al.* [@dawid2016statistical]): $$\label{eq:generic} \max\left\{0,1 - \frac{1}{\mbox{RR}}\right\} \leq \mbox{PC}_A \leq \frac{\min\{\mbox{P}(Y = 0 \mid X\leftarrow 0),\mbox{P}(Y = 1 \mid X\leftarrow 1)\}} {\mbox{P}(Y=1 \mid X\leftarrow 1)},$$ where $$\mbox{RR} = \frac{\mbox{P}(Y = 1 \mid X\leftarrow 1)}{\mbox{P}(Y=1 \mid X\leftarrow 0)}$$ is the [*experimental risk ratio*]{} between exposed and unexposed. These bounds can be estimated from the experimental data using the population death rates in equations  and . In many cases of interest (such as [Table \[tab:aspirin1\]]{}), we have $$\mbox{P}(Y= 1 \mid X \leftarrow 0) < \mbox{P}(Y= 1 \mid X \leftarrow 1) < \mbox{P}(Y= 0 \mid X \leftarrow 0).$$ Then the lower bound in will be non-trivial, while the upper bound will be 1, so vacuous. Since, in [Table \[tab:aspirin1\]]{}, the exposed are $2.5$ times as likely to die as the unexposed ($\mbox{RR}= 30/12 = 2.5$), we have enough confidence to infer causality in Ann’s case, since $0.60 \leq \mbox{PC}_A \leq 1$. Bounds in Mediation {#sec:mediation} =================== In this Section we bound the probability of causation for a case where a third variable, $M$, is involved in the causal pathway between the exposure $X$ and the outcome $Y$. We first review the results of Dawid *et al.* [@dawid-murtas] for the case of complete mediation, where no direct effect is present between exposure and outcome but all the effect is mediated by $M$. In addition we derive new bounds for $\mbox{PC}_A$ in mediation analysis when a partial mediation mechanism is in operation. Complete mediation (Dawid *et al.* [@dawid-murtas]) {#sec:compl-mediation} --------------------------------------------------- The case of no direct effect is intuitively described by [Figure \[fig:compl-med\]]{}. Applications where this assumption might be plausible is in the treatment of ovarian cancer (Silber *et al.* [@silber2007does]), where $X$ represents management either by a medical oncologist or by a gynaecological oncologist, $M$ is the intensity of chemotherapy prescribed, and $Y$ is death within 5 years. We shall be interested in the case that $M$ is observed in the experimental data but is not observed for Ann, and see how this additional experimental evidence can be used to refine the bounds on $\mbox{PC}_A$. (0,-0.5)(3,0.5) (0,0)[ ]{} (1.5,0)[ ]{} (3,0)[ ]{} We now introduce $M(x)$ to denote the potential value of $M$ when $X\leftarrow x$, and $Y^*(m)$ to denote the potential value of $Y$ when $M\leftarrow m$. Then $Y(x):=Y^*\{M(x)\}$. We define ${{\textbf{M}}}: = (M(0),M(1))$ and ${{{\textbf{Y}}}\mbox{$^*$}}:= (Y^*(0),Y^*(1))$. We suppose that none of the causal mechanisms depicted in [Figure \[fig:compl-med\]]{} are confounded–expressed mathematically by assuming mutual independence between $X$, ${{\textbf{M}}}$ and ${{{\textbf{Y}}}\mbox{$^*$}}$ (both for experimental individuals, and for Ann). These assumptions imply no overall confounding (as in ), the Markov property $Y {\mbox{$\,\perp\!\!\!\perp\,$}}X \mid M$, and the following bounds in this case of complete mediation: $$\label{eq:pc-compl-med} \max\left\{0,{1-\frac{1}{\mbox{RR}}}\right\} \leq \mbox{PC}_A \leq \frac{\rm Num}{\mbox{P}(Y=1 \mid X\leftarrow 1)},$$ where the numerator, Num, is given in [Table \[tab:compl-med-upper\]]{}. We see from that knowing a mediator does not improve the lower bound. For the upper bound, one has to consider various scenarios according to different choices of the estimable marginal probabilities in [Table \[tab:compl-med-upper\]]{}. [l|cc]{} & $a\leq b$ & $a>b$\ $c\leq d$ & $a\cdot c+(1-d)(1-b)$ & $b\cdot c+(1-d)(1-a)$\ $c>d$ & $a\cdot d+(1-c)(1-b)$ & $b\cdot d+(1-a)(1-c)$\ Note that, given the no-confounding assumptions, the entries in [Table \[tab:compl-med-upper\]]{} are all estimable from the experimental data: $$\begin{aligned} a &=& {\mbox{\textrm P}}(M = 0 \mid X \leftarrow 0)\\ b &=& {\mbox{\textrm P}}(M = 1 \mid X \leftarrow 1)\\ c &=& {\mbox{\textrm P}}(Y = 0 \mid M \leftarrow 0)\\ d &=& {\mbox{\textrm P}}(Y = 1 \mid M \leftarrow 1).\end{aligned}$$ Partial mediation {#sec:part-mediation} ----------------- (-0.5,-0.5)(3.5,2) (0,0)[ ]{} (3,0)[ ]{} (1.5,1.5)[ ]{} The situation described by [Figure \[fig:compl-med\]]{} is unlikely to hold in many real life situations. Situations such as that represented informally by [Figure \[fig:mediation-simple\]]{}, that allow both direct and indirect effects, are more plausible. In this Section we derive new bounds for the probability of causation when a partial mediator is involved in the causal pathway. We now define $Y^*(x,m)$ as the potential value of the outcome $Y$ after setting both exposure, $X\leftarrow x$, and mediator, $M \leftarrow m$. Then $Y(x)=Y^*(x,M(x))$. We make the following assumptions, both for Ann and for the individuals in the experiment: A1: : $Y^*(x,m) {\mbox{$\,\perp\!\!\!\perp\,$}}{{\textbf{M}}}\mid X$ (no $M$–$Y$ confounding) A2: : $Y^*(x,m) {\mbox{$\,\perp\!\!\!\perp\,$}}X$ (no $X$–$Y$ confounding) A3: : $M(x) {\mbox{$\,\perp\!\!\!\perp\,$}}X$ (no $X$–$M$ confounding). Assumption **A1** expresses independence, given $X$, between a potential value $Y^*(x,m)$, that would arise on setting exposure and mediator to particular values, and the pair of potential outcomes $(M(0), M(1))$. It can be seen as a strengthening of the univariate hypothesis $Y^*(x,m) {\mbox{$\,\perp\!\!\!\perp\,$}}M(x) \mid X$. Note that **A1** and **A2** are together equivalent to the single requirement: A12: : $Y^*(x,m) {\mbox{$\,\perp\!\!\!\perp\,$}}({{\textbf{M}}}, X)$. Because we have supposed that Ann is exchangeable with the individuals in the experiment, we have $$\label{eq:pceq} \mbox{PC}_A = \mbox{P}(Y(0) = 0 \mid X\leftarrow 1, Y(1)=1)=\frac{\mbox{P}(Y(0)=0,\: Y(1)=1 \mid X\leftarrow 1)}{\mbox{P}(Y(1)=1 \mid X\leftarrow 1)}.$$ Given the no-confounding assumptions, the denominator of is ${\mbox{\textrm P}}(Y = 1 \mid X \leftarrow 1)$, which is estimable. However, the numerator of involves the joint distribution of the pair ${{\textbf{Y}}}$ of potential outcomes, and this is not estimable from the data, in view of the fact that it is never possible to observe both $Y(0)$ and $Y(1)$ simultaneously. We can however bound this numerator in terms of estimable quantities, using the fact that, for any events $A$ and $B$, and any probability distribution ${\mbox{\textrm P}}$, $$\label{eq:bound-copula} \max\{{\mbox{\textrm P}}(A)+{\mbox{\textrm P}}(B)-1, 0 \} \leq {\mbox{\textrm P}}(A \cap B) \leq \min \{{\mbox{\textrm P}}(A),{\mbox{\textrm P}}(B)\}.$$ Using , we can obtain an upper bound for the numerator as: $$\begin{aligned} &\mbox{P}(Y(0)=0,Y(1)=1 \mid X\leftarrow 1)=\mbox{P}(Y^*(0,M(0))=0,\:Y^*(1,M(1))=1 \mid X\leftarrow 1) \nonumber \\ &=\sum_{m_0} \sum_{m_1} \mbox{P}(Y^*(0,m_0)=0, Y^*(1,m_1)=1 \mid M(0)=m_0,M(1)=m_1,X\leftarrow 1) \label{eq:8a}\\ &\hspace{3cm} \times \mbox{P}(M(0)=m_0,M(1)=m_1 \mid X\leftarrow 1) \nonumber \\ &\leq \sum_{m_0} \sum_{m_1} \min\{\mbox{P}(Y^*(0,m_0)=0 \mid M(0)=m_0,M(1)=m_1,X\leftarrow 1),\nonumber\\ &\hspace{3cm} \mbox{P}(Y^*(1,m_1)=1 \mid M(0)=m_0,M(1)=m_1,X\leftarrow 1)\} \nonumber \\ & \hspace{3cm}\times \min\{\mbox{P}(M(0)=m_0 \mid X\leftarrow 1),\mbox{P}(M(1)=m_1 \mid X\leftarrow 1)\} \nonumber\\ &=\sum_{m_0} \sum_{m_1} \min\{\mbox{P}(Y^*(0,m_0)=0),\mbox{P}(Y^*(1,m_1)=1)\} \nonumber \\ & \hspace{3cm}\times \min\{\mbox{P}(M(0)=m_0),\mbox{P}(M(1)=m_1\}, \label{eq:upper-part}\end{aligned}$$ on using assumptions **A12** and **A3**. That is, $$\begin{aligned} &\mbox{P}(Y(0)=0,\:Y(1)=1 \mid X\leftarrow 1)\leq \nonumber \\ &\min\{\mbox{P}(Y^*(0,0)=0),\mbox{P}(Y^*(1,0)=1)\} \times \min\{\mbox{P}(M(0)=0),\mbox{P}(M(1)=0)\} \label{eq:upper-part1} \\ &+\min\{\mbox{P}(Y^*(0,0)=0),\mbox{P}(Y^*(1,1)=1)\} \times \min\{\mbox{P}(M(0)=0),\mbox{P}(M(1)=1)\} \label{eq:upper-part2}\\ &+\min\{\mbox{P}(Y^*(0,1)=0),\mbox{P}(Y^*(1,0)=1)\} \times \min\{\mbox{P}(M(0)=1),\mbox{P}(M(1)=0)\} \label{eq:upper-part3}\\ &+\min\{\mbox{P}(Y^*(0,1)=0),\mbox{P}(Y^*(1,1)=1)\} \times \min\{\mbox{P}(M(0)=1),\mbox{P}(M(1)=1)\} \label{eq:upper-part4}.\end{aligned}$$ It can be shown that similarly applying the lower bound of to yields the same lower bound as obtained in [§$\,$\[sec:simple\]]{} and [§$\,$\[sec:compl-mediation\]]{}, so that the lower bound is not improved by knowledge of a mediation mechanism. Assumptions **A12** and **A3** allow us to estimate the terms – in the above upper bound from the data: $$\begin{aligned} {\mbox{\textrm P}}(Y^*(x,m)=y) &=& {\mbox{\textrm P}}(Y=y \mid X\leftarrow x, M\leftarrow m)\\ {\mbox{\textrm P}}(M(x) = m) &=& {\mbox{\textrm P}}(M = m \mid X\leftarrow x).\end{aligned}$$ Comparisons {#sec:comparison} =========== In this Section we compare the bounds found in the simple analysis framework of [§$\,$\[sec:simple\]]{} with those obtained by considering a complete mediation mechanism, as in [§$\,$\[sec:compl-mediation\]]{}, and those obtained by considering a partial mediation mechanism, as in [§$\,$\[sec:mediation\]]{}. We focus on comparing these bounds to obtain the best information from the data. The numerator of the upper bound for $\mbox{PC}_A$ in , which ignores the mediator, may be written as $$\label{eq:alpha} \min\{\alpha+\beta,\gamma+\delta\},$$ where $$\begin{aligned} \alpha &=& {\mbox{\textrm P}}(Y^*(0,0)=0)\,{\mbox{\textrm P}}(M(0)=0)\\ \beta &=& {\mbox{\textrm P}}(Y^*(0,1)=0)\,{\mbox{\textrm P}}(M(0)=1)\\ \gamma &=& {\mbox{\textrm P}}(Y^*(1,0)=1)\,{\mbox{\textrm P}}(M(1)=0)\\ \delta &=& {\mbox{\textrm P}}(Y^*(1,1)=1)\,{\mbox{\textrm P}}(M(1)=1).\end{aligned}$$ We see that both and are smaller than or equal to $\alpha$, while both and are smaller than or equal to $\beta$. So the upper bound allowing for partial mediation, which is the sum of , , and , cannot exceed $2\,(\alpha+\beta) = 2\,{\mbox{\textrm P}}(Y(0)=0) = 2\, {\mbox{\textrm P}}(Y=0 \mid X \leftarrow 0)$; and similarly cannot exceed $2\,{\mbox{\textrm P}}(Y=1 \mid X \leftarrow 1)$. Thus, the upper bound for the numerator, when accounting for the mediator, can not exceed twice that obtained by ignoring it, as given by . However, as we will see in [§$\,$\[sec:example\]]{}, it could be larger or smaller than that simpler bound. On the other hand, we do not obtain a different lower bound. In the special case of complete mediation, $Y^*(0,m) = Y^*(1,m)$, $= Y^*(m)$, say. Thus the terms with $m_0\neq m_1$ in must be $0$. This leads to the following upper bound: $$\begin{aligned} &\mbox{P}(Y(0)=0,Y(1)=1 \mid X\leftarrow 1)\leq \nonumber \\ &\min\{\mbox{P}(Y^*(0)=0),\mbox{P}(Y^*(1)=1)\} \times \min\{\mbox{P}(M(0)=0),\mbox{P}(M(1)=1)\}+ \nonumber \\ &+\min\{\mbox{P}(Y^*(1)=0),\mbox{P}(Y^*(0)=1)\} \times \min\{\mbox{P}(M(0)=1),\mbox{P}(M(1)=0)\}, \nonumber $$ in agreement with [Table \[tab:compl-med-upper\]]{}. Since we have eliminated the terms and appearing in the general case, the upper bound obtained in this case of complete mediation is never bigger than that obtained from the general expression +++; nor, since $\leq \alpha$ while $\leq \beta$, can it be bigger than the bound obtained on ignoring the knowledge of the complete mediation mechanism. Examples {#sec:example} -------- To show that, in the case of partial mediation, taking account of information about the mediator may, but need not, yield a tighter upper bound, we consider two cases with different experimental data as given respectively by [Table \[tab:aspirin2\]]{} and [Table \[tab:aspirin3\]]{}. [lccc]{} & Die & Live & Total\ Exposed & 69 & 31 & 100\ Unexposed & 24 & 76 & 100\ [lccc]{} & Die & Live & Total\ Exposed & 78 & 22 & 100\ Unexposed & 32 & 68 & 100\ Suppose now we can also observe a partial mediator $M$. We might then observe the following probabilities, consistent with [Table \[tab:aspirin2\]]{}: $$\begin{aligned} & \mbox{P}(Y^*(0,0)=0)= 0.98 & \mbox{P}(Y^*(0,1)=0) = 0.165 \nonumber \\ & \mbox{P}(Y^*(1,0)=0)= 0.315 & \mbox{P}(Y^*(1,1)=0) = 0.143 \nonumber\\ & \mbox{P}(M(0)=0) = 0.73 & \mbox{P}(M(1)=0) = 0.981 \nonumber\end{aligned}$$ We then obtain: $0.65 \leq \mbox{PC}_A \leq 0.81$ when accounting for the mediator, and $0.65 \leq \mbox{PC}_A \leq 1$ when ignoring it. In this case, knowledge of the partial mediation mechanism is helpful in improving the upper bound. On the other hand, suppose we observe the following probabilities, consistent with [Table \[tab:aspirin3\]]{}: $$\begin{aligned} & \mbox{P}(Y^*(0,0)=0)= 0.98 & \mbox{P}(Y^*(0,1)=0) = 0.67 \nonumber \\ & \mbox{P}(Y^*(1,0)=0)= 0.09 & \mbox{P}(Y^*(1,1)=0) = 0.27 \nonumber\\ & \mbox{P}(M(0)=0) = 0.04 & \mbox{P}(M(1)=0) = 0.26 \nonumber\end{aligned}$$ We now obtain: $0.59 \leq \mbox{PC}_A \leq 0.95 $ when accounting for the mediator, but $0.59 \leq \mbox{PC}_A \leq 0.88$ when ignoring it. So in this case knowledge of mediation has not been helpful. Conclusions {#sec:conclusion} =========== Bounding the probability of causation in mediation analysis is an important problem for applications. By taking account of a complete mediation mechanism we can never do worse than by ignoring it. However, complete mediation is not always reasonable. In the case of partial mediation, the upper bound obtained by taking account of it may be greater or smaller than that obtained by ignoring it. We can thus compute both upper bounds and take the smaller. This work has several possible extensions. It would be interesting to extend our theoretical formulas to cases combining information on both covariates and mediators. Another promising extension arises on making connections with copula theory, where $\mbox{PC}$ can be obtained as a function of the estimable quantities ${\mbox{\textrm P}}(Y(0)=0)$ and ${\mbox{\textrm P}}(Y(1)=1)$ together with an appropriate copula. Dawid AP (2011) The role of scientific and statistical evidence in assessing causality. In: Goldberg R (ed) Perspectives on Causation. Oxford, Hart Publishing, pp 133–147 Dawid AP, Faigman DL, Fienberg SE (2014) Fitting science into legal contexts: Assessing effects of causes or causes of effects? Sociological Methods & Research 43:359–390 Dawid AP, Murtas R, Musio M (2016) Bounding the probability of causation in mediation analysis. In: Di Battista T, Moreno E, Racugno W (ed) Topics on Methodological and Applied Statistical Inference. Cham Switzerland, Springer International Publishing, pp 75–84 Dawid AP, Musio M, Fienberg SE (2016) From statistical evidence to evidence of causality. Bayesian Analysis, 11(3):725–752 Markman M (2007) Does ovarian cancer treatment and survival differ by the specialty providing chemotherapy? Journal of Clinical Oncology, 25(23):3554–3554 Pearl J (1999) Probabilities of causation: Three counterfactual interpretations and their identification. Synthèse, 121(1–2):93–149 Rubin DB (1974) Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology, 66(5):688 Tian J, Pearl J (2000) Probabilities of causation: Bounds and identification. Annals of Mathematics and Artificial Intelligence, 28(1–4):287–313 [^1]: University of Cagliari [^2]: Leverhulme Emeritus Fellow, University of Cambridge [^3]: University of Cagliari
{ "pile_set_name": "ArXiv" }
--- abstract: 'We construct universal associative envelopes for the nonassociative triple systems arising from the trilinear operations of Bremner and Peresi applied to the 2-dimensional simple associative triple system. We use noncommutative Gröbner bases to determine monomial bases, structure constants, and centers of the universal envelopes. We show that the infinite dimensional envelopes are closely related to the down-up algebras of Benkart and Roby. For the finite dimensional envelopes, we determine the Wedderburn decompositions and classify the irreducible representations.' address: 'Department of Mathematics and Statistics, University of Saskatchewan, Canada' author: - 'Hader A. Elgendy' title: | Universal associative envelopes of\ nonassociative triple systems --- Introduction ============ In this paper we use noncommutative Gröbner bases to construct the universal associative enveloping algebras of the nonassociative triple systems which arise from applying the nonassociative trilinear operations classified by Bremner and Peresi [@BremnerPeresi] to the 2-dimensional simple associative triple system of the first kind in the sense of Lister [@Lister], namely the space of $2 \times 2$ matrices $A = (a_{ij})$ with $a_{11} = a_{22} = 0$. The basic goal of this work is reduce the problem of representation of these triple systems, which are the simplest examples of nonassociative triple systems satisfying polynomial identities in degree 3 and 5 to that of associative algebras. They provide natural generalizations of Lie and Jordan triple systems, and therefore have the potential for a wide range of applications to various branches of mathematics, especially to the theory of Lie groups and non-Euclidean geometry [@faraut; @Freud; @Ros]. Indeed, they have arisen in the study of symmetric spaces [@Loos2], and have been connected with the study of the Yang-Baxter equations [@Kamiya]. By a multilinear $n$-ary operation we mean an element $\omega$ of the group algebra $\mathbb{Q} S_n$ of the symmetric group $S_n$ over the rational field $\mathbb{Q}$. Following [@BremnerPeresi] we say that two operations $\omega_1$, $\omega_2$ are *equivalent* if they generate the same left ideal. If $A$ is an associative algebra over $\mathbb{Q}$, then $\omega$ defines a multilinear $n$-ary operation $\omega(a_1,\dots,a_n)$ on the underlying vector space of $A$: $$\omega = \sum_{\sigma \in S_n} x_\sigma \sigma \quad \implies \quad \omega(a_1,\dots,a_n) = \sum_{\sigma \in S_n} x_\sigma a_{\sigma(1)} \cdots a_{\sigma(n)}.$$ In this way we obtain a nonassociative $n$-ary algebra which we denote by $A^{\omega}$. For $n = 2$, every bilinear operation is equivalent to either the zero operation, the associative operation $ab$, the Lie bracket $[a,b] = ab - ba$, or the Jordan product $a \circ b = \tfrac12 (ab + ba)$. The polynomial identities of degree $\le 3$ ($\le 4$) satisfied by the Lie bracket (Jordan product) define Lie algebras (Jordan algebras), the most important varieties of nonassociative algebras. For $n = 3$, Bremner and Peresi [@BremnerPeresi] found canonical representatives of the equivalence classes of trilinear operations, and identified 19 operations satisfying polynomial identities of degree 5 which do not follow from the identities of degree 3. These operations include the Lie, anti-Lie, Jordan, and anti-Jordan triple products. In Section \[twentytwo\], we find simpler operations equivalent to those of [@BremnerPeresi]; our operations have coefficients $\pm 1, \pm 2$ and most have coefficients $\pm 1$. We augment this list with the symmetric, alternating and cyclic sums; see Table \[matrixforms\]. ---------------------------------- -- ---------------------------------------------------------------------------------- -- -- symmetric sum $a ( b \circ c ) + b ( c \circ a ) + c ( a \circ b )$ alternating sum $a [ b, c ] + b [ c, a ] + c [ a, b ]$ cyclic sum $abc + bca + cab$ Lie family $q = \infty$ $[ a, [ b, c ] ]$ Lie family $q = \frac12$ $[ a, b \circ c ]$ Jordan family $q = \infty$ $abc + cba$ Jordan family $q = 0$ $( a \circ b ) c$ Jordan family $q = \frac12$ $a ( b \circ c ) + c (a \circ b ) + (c \circ a) b$ Jordan family $q = 1$ $a ( b \circ c )$ anti-Jordan family $q = \infty$ $a [ b, c ] + c [ a, b ] + [ c, a ] b$ anti-Jordan family $q = -1$ $a [ b, c ]$ anti-Jordan family $q = \frac12$ $abc - cba$ anti-Jordan family $q = 2$ $[ a, b ] c $ fourth family $q = \infty$ $abc - acb - bac$ fourth family $q = 0$ $abc - acb + bca$ fourth family $q = 1$ $abc - bac + cab$ fourth family $q = -1$ $abc + bac + cab$ fourth family $q = 2$ $abc + acb + bca$ fourth family $q = \frac12$ $abc + acb + bac$ cyclic commutator $[ a, bc ]$ weakly commutative $\Big\{\begin{array}{l} abc + acb + bac - cba, \\ abc - acb + 2bac \end{array}$ weakly anticommutative $\Big\{ \begin{array}{l} abc + acb - bac - cba, \\ abc + acb - 2bac \end{array}$ ---------------------------------- -- ---------------------------------------------------------------------------------- -- -- : The twenty-two trilinear operations[]{data-label="matrixforms"} In Section \[universal\] we recall the theory of noncommutative Gröbner bases for ideals in free associative algebras, and use it to define universal associative envelopes $U(A^\omega)$ of nonassociative $n$-ary algebras $A^\omega$ defined by multilinear operations $\omega$. In Section \[infinite\] we recall the down-up algebras of Benkart and Roby [@Roby]. We then consider the cases in which $U(A^{\omega})$ is infinite dimensional: we determine monomial bases and structure constants, identify the center, and determine the Gelfand-Kirillov dimension. In every case, $U(A^{\omega})$ is either a free associative algebra, a down-up algebra, or a quotient of a down-up algebra. In Section \[finite\] we consider the cases in which $U(A^{\omega})$ is finite dimensional. We use an algorithmic version of the structure theory for finite dimensional associative algebras to determine the Wedderburn decompositions and classify the irreducible representations. In most cases we obtain only the trivial 1-dimensional representation and the natural 2-dimensional representation. The results of Sections \[infinite\] and \[finite\] are summarized in Table \[summary\]. We distinguish trilinear operations of “Lie type” for which $U(A^{\omega})$ is infinite dimensional, and those of “Jordan type” for which $U(A^{\omega})$ is finite dimensional. Recall that for a finite dimensional Lie algebra $L$, the universal associative envelope $U(L)$ is infinite dimensional, and the map $L \to U(L)$ is injective; whereas for a finite dimensional Jordan algebra $J$, the universal associative envelope $U(J)$ is finite dimensional, and the map $J \to U(J)$ is injective if and only if $J$ is special. For the cases where the universal associative envelope is finite dimensional, as in the representation theory of Jordan algebras and triple systems , we obtain a complete description of the Wedderburn decomposition of the universal associative envelope, and this provides a complete classification of the finite dimensional irreducible representation of the triple system. These cases also provide natural examples where the computational approach to the Wedderburn decomposition can be applied. For the cases where the universal associative envelope is infinite dimensional, as in the representation theory of Lie algebras and triple systems [@Hodge], we provide a monomial basis for the envelope and the structures constants with respect to this basis. These cases are closely related to the theory of down-up algebras, and provide a natural way to generalize down-up algebras to structures with more than two generators and/or defined by relations of degree greater than 3. These cases also provide natural generalizations of the enveloping algebras of Lie algebras, and may therefore be connected to the theory of quantum groups [@chari]. We assume throughout that the base field ${\mathbb{F}}$ has characteristic 0; if necessary, we assume that ${\mathbb{F}}$ is algebraically closed. [lccc]{}\ & $\dim{U(A^\omega)}$ & GK-dim & $U(A^\omega)$\ $\Big\{\begin{array}{ll}\text{symmetric sum} & \\\text{cyclic sum} &\end{array}$&$\infty$&1&$A(-1,-1,1)/ \langle a^3, b^3\rangle$\ alternating sum&$\infty$&$\infty$&$F\langle a, b\rangle$\ $\Big\{\begin{array}{ll}\text{Lie $q = \infty$} & \\\text{anti-Jordan $q = \infty$} &\end{array}$&$\infty$&3&$A(2,-1,-2)$\ $\Big\{\begin{array}{ll}\text{Lie $q = \tfrac{1}{2}$} & \\\text{anti-Jordan $q = \tfrac{1}{2}$} &\end{array}$&$\infty$&3&$A(0,1,0)$\ \ \ & $\dim{U(A^\omega)}$ & GK-dim & $U(A^\omega)$\ $\left\{\begin{array}{lllllll}\text{Jordan $q = \infty$ } & \\\text{Jordan $q = \tfrac{1}{2} $} &\\\text{anti-Jordan $q = -1$}&\\\text{anti-Jordan $q = 2$}&\\\text{fourth family $q = \infty$}&\\\text{fourth family $q = -1$}&\\\text{fourth family $q = 2$}&\\\text{fourth family $q = \tfrac{1}{2}$}&\\\text{cyclic commutator }&\\\text{weakly commutative}&\\\text{weakly anticommutative}&\\\end{array}\right.$&5&0&${\mathbb{Q}}\oplus M_{2\times 2}$\ $\Big\{\begin{array}{ll}\text{Jordan $q = 0$} & \\\text{fourth family $q = 0$} &\end{array}$&9&0&$\mathfrak{R}\oplus {\mathbb{Q}}\oplus M_{2\times 2}$\ $\Big\{\begin{array}{ll}\text{Jordan $ q = 1$ } & \\\text{fourth family $q = 1$} &\end{array}$&9&0&$\mathfrak{R} \oplus {\mathbb{Q}}\oplus M_{2\times 2}$ The twenty-two trilinear operations {#twentytwo} =================================== A natural basis for $\mathbb{Q} S_3$ consists of the six permutations in lexicographical order: $\{ abc, acb, bac, bca, cab, cba \}$. Another natural basis consists of the matrix units $\{ S, E_{11}, E_{12}, E_{21}, E_{22}, A \}$ for the decomposition as a direct sum $\mathbb{Q} \oplus M_2(\mathbb{Q}) \oplus \mathbb{Q}$ of simple ideals corresponding to the partitions $3 = 2 + 1 = 1 + 1 + 1$ which label the irreducible representations of $S_3$. Bremner and Peresi [@BremnerPeresi] give the matrix $M$ whose columns express the matrix units as linear combinations of the permutations: $$M = \frac16 \left[ \begin{array}{rrrrrr} 1 &\!\! 2 &\!\! 0 &\!\! 0 &\!\! 2 &\!\! 1 \\ 1 &\!\! 0 &\!\! 2 &\!\! 2 &\!\! 0 &\!\! -1 \\ 1 &\!\! 2 &\!\! -2 &\!\! 0 &\!\! -2 &\!\! -1 \\ 1 &\!\! -2 &\!\! 2 &\!\! -2 &\!\! 0 &\!\! 1 \\ 1 &\!\! 0 &\!\! -2 &\!\! 2 &\!\! -2 &\!\! 1 \\ 1 &\!\! -2 &\!\! 0 &\!\! -2 &\!\! 2 &\!\! -1 \end{array} \right], \quad M^{-1} = \left[ \begin{array}{rrrrrr} 1 &\!\! 1 &\!\! 1 &\!\! 1 &\!\! 1 &\!\! 1 \\ 1 &\!\! 0 &\!\! 1 &\!\! 0 &\!\! -1 &\!\! -1 \\ 0 &\!\! 1 &\!\! 0 &\!\! 1 &\!\! -1 &\!\! -1 \\ 0 &\!\! 1 &\!\! -1 &\!\! -1 &\!\! 1 &\!\! 0 \\ 1 &\!\! 0 &\!\! -1 &\!\! -1 &\!\! 0 &\!\! 1 \\ 1 &\!\! -1 &\!\! -1 &\!\! 1 &\!\! 1 &\!\! -1 \end{array} \right].$$ Given a trilinear operation $\omega = x_1 abc + x_2 acb + x_3 bac + x_4 bca + x_5 cab + x_6 cba$ with coefficient vector $X = [ x_1, x_2, x_3, x_4, x_5, x_6 ]^t$, one obtains its matrix form, $$Y = \left[ \, y_1, \, \begin{bmatrix} y_2 & y_3 \\ y_4 & y_5 \end{bmatrix}\!, \, y_6 \, \right],$$ by $Y = M^{-1} X$. Two operations are equivalent if and only if their matrix forms are row-equivalent; hence canonical representatives of the equivalence classes are the operations for which each component matrix is in row canonical form. To find the simplest representative of each equivalence class, we first consider the $3^5 = 243$ operations whose coefficients in the permutation basis are $[ 1, x_2, \dots, x_6 ]$ where $x_2, \dots, x_6 \in \{ 1, 0, -1 \}$. We record the operations whose matrix forms are among the canonical representatives [@BremnerPeresi]; this gives 20 of the 22 operations. We next consider the $5^5 = 3125$ operations whose coefficients in the permutation basis satisfy $x_2, \dots, x_6 \in \{ 2, 1, 0, -1, -2 \}$; this gives the remaining two operations, and also produces alternative forms of the last two operations. In more than half of the cases, the simplified operations of Table \[matrixforms\] are more natural than the original operations, since they can be easily expressed in terms of the Lie bracket and the Jordan product. Gröbner bases and universal envelopes {#universal} ===================================== We first recall basic results about noncommutative Gröbner bases for ideals in free associative algebras. Standard references are Bergman [@Bergman], de Graaf [@deGraaf]; for another application to $n$-ary algebras, see Elgendy and Bremner [@ElgendyBremner]. \[dd\] Let $X = \{ x_1, \dots, x_n \}$ be a set of symbols with total order $x_i < x_j$ $\iff$ $i < j$. The *free monoid* $X^\ast$ on $X$ consists of all (possibly empty) words $w = x_{i_1} \cdots x_{i_k}$ ($k \ge 0$) with the associative operation of concatenation. For $w = x_{i_1} \cdots x_{i_k} \in X^*$ the *degree* is $\deg(w) = k$. The *free unital associative algebra* $F \langle X \rangle$ is the vector space with basis $X^*$ and multiplication extended bilinearly. The *deglex* (degree-lexicographical) order on $X^\ast$ is defined by: $u < v$ if and only if either ($i$) $\deg(u) < \deg(v)$ or ($ii$) $\deg(u) = \deg(v)$ and $u = w x_i u'$, $v = w x_j v'$ where $x_i < x_j$ ($w, u', v' \in X^\ast$). We say that $u \in X^*$ is a *factor* of $v \in X^*$ if there exist $w_1, w_2 \in X^*$ such that $w_1 u w_2 = v$. If $w_1$ ($w_2$) is empty then $u$ is a *left* (*right*) factor of $v$. The *support* of $f \in F\langle X \rangle$ is the set of monomials $w \in X^\ast$ that occur in $f$ with nonzero coefficient. The *leading monomial* $\mathrm{LM}(f)$ is the greatest element of the support of $f$. For an ideal $I \subseteq F\langle X \rangle$ the set of *normal words* modulo $I$ is $N(I) = \{ \, u \in X^* \mid \text{$u \ne \mathrm{LM}(f)$ for any $f \in I$} \, \}$, and $C(I)$ is the subspace of $F\langle X \rangle$ with basis $N(I)$. \[C(I)proposition\] We have $F\langle X \rangle = C(I) \oplus I$. \[No\] Let $G \subset F\langle X \rangle$ be a subset generating an ideal $I$. An element $f \in F\langle X \rangle$ is in *normal form modulo* $G$ if no monomial in the support of $f$ has $\mathrm{LM}(g)$ as a factor for any $g \in G$. A subset $G \subset I$ is a *Gröbner basis* of $I$ if for every $f \in I$ there exists $g \in G$ such that $\mathrm{LM}(g)$ is a factor of $\mathrm{LM}(f)$. A subset $G \subset F\langle X \rangle$ is *self-reduced* if every $g \in G$ is in normal form modulo $G \setminus \{g\}$ and every $g \in G$ is *monic*: the coefficient of $\mathrm{LM}(g)$ is 1. Consider elements $g, h \in F \langle X \rangle$ such that $\mathrm{LM}(g)$ is not a factor of $\mathrm{LM}(h)$ and $\mathrm{LM}(h)$ is not a factor of $\mathrm{LM}(g)$. Assume that $u, v \in X^\ast$ satisfy $\mathrm{LM}(g)\,u = v\,\mathrm{LM}(h)$, $u$ is a proper right factor of $\mathrm{LM}(h)$, and $v$ is a proper left factor of $\mathrm{LM}(g)$. Then $g u - v h$ is called a *composition* of $g$ and $h$. \[di\] If $I \subset F\langle X \rangle$ is an ideal generated by a self-reduced set $G$, then $G$ is a Gröbner basis of $I$ if and only if for all compositions $f$ of the elements of $G$ the normal form of $f$ modulo $G$ is zero. In order to compute a Gröbner basis for an ideal $I \subset F\langle X \rangle$, we start with any set $G$ of generators for $I$, and perform the following algorithm: 1. Compute all compositions of elements of $G$. Let $H$ be the set of their normal forms modulo $G$. 2. Replace $G$ by $G \cup H$, and self-reduce the new set $G$ by replacing each element by its normal form modulo the other elements. The algorithm terminates when $H = \emptyset$ or $H = \{0\}$. We now apply this theory to the construction of universal associative envelopes. Given a finite dimensional associative algebra $A$, and a multilinear $n$-ary operation $\omega$, we obtain a nonassociative $n$-ary algebra $A^{\omega}$. \[uaedefinition\] A *universal associative envelope* of $A^{\omega}$ consists of a unital associative algebra $U(A^\omega)$ and a linear map $i\colon A^{\omega} \to U(A^{\omega})$ satisfying $$i(\omega(x_1, x_2,\dots, x_n)) = {\omega} \left( i(x_1), i(x_2), \dots, i(x_n) \right),$$ for all $x_1, \dots, x_n \in A^{\omega}$, such that for any unital associative algebra ${\mathbb{A}}$ and linear map $j\colon A^{\omega} \to {\mathbb{A}}$ satisfying the same equation with $j$ in place of $i$, there is a unique homomorphism of unital associative algebras $\psi\colon U(A^\omega) \to {\mathbb{A}}$ such that $\psi \circ i = j$. \[ydefinition\] Let $B = \{ e_1, e_2, \dots,e_{m}\}$ be an ordered basis of $A^\omega$, and let $\phi\colon B \to X = \{ x_1, x_2, \dots x_m\}$ be the bijection $\phi(e_i) = x_i$. We extend $\phi$ to a linear map, denoted by the same symbol, $\phi\colon A^{\omega}\to F \langle X \rangle$. \[idealgenerators\] Consider the following elements of $F\langle X \rangle$: $$G_{i_1, \dots, i_n} = \omega ( x_{i_1}, \dots, x_{i_n} ) - \phi( \omega( e_{i_1}, \dots, e_{i_n} )), \qquad 1\leq i_1,\dots,i_n\leq m.$$ Let $I$ be the ideal generated by the set of all $G_{i_1, \dots, i_n}$, and define $U(A^\omega)= F\langle X \rangle / I$. We have the natural surjection $\pi\colon F\langle X \rangle \to U(A^\omega)$ sending $f$ to $f + I$, and the composition $i = \pi \circ \phi\colon A^{\omega} \to U(A^\omega)$. The algebra $U(A^{\omega})$ and the map $i\colon A^{\omega} \to U(A^\omega)$ form the universal associative envelope of the nonassociative $n$-ary algebra $A^{\omega}$. To obtain the elements $G_{i_1, \dots, i_n}$, we use the structure constants of $A^{\omega}$. We then use Theorem \[di\] to compute a Gröbner basis of the ideal $I$, and Proposition \[C(I)proposition\] to determine a basis of $U(A^{\omega})$. We write $\delta_{i,j}$ for the Kronecker delta, and $\widehat{\delta}_{i,j}= 1 - \delta_{i,j}$. Infinite dimensional envelopes {#infinite} ============================== In this section, we consider the trilinear operations of “Lie type”. (Benkart and Roby [@Roby]) \[doup\] Let ${\mathbb{F}}$ be a field and let $\alpha, \beta, \gamma \in {\mathbb{F}}$ be parameters. The *down-up algebra* $A(\alpha, \beta, \gamma)$ is the unital associative algebra with generators $a, b$ and relations $$b^2 a = \alpha bab + \beta a b^2 + \gamma b, \qquad ba^2 = \alpha aba + \beta a^2 b + \gamma a.$$ [@Roby Theorem 3.1, Corollary 3.2] \[bas1\] A down-up algebra has basis $$\mathfrak{B}_1 = \{ a^i (ba)^j b^k \mid i, j, k \geq 0 \},$$ and its Gelfand-Kirillov dimension is $3$. \[zhaolemma\] [@Zhao Lemma 2.2] For any $c_1, c_2 \in {\mathbb{F}}$, a down-up algebra has basis $$\mathfrak{B_2} = \{ a^i(ba + c_1 ab + c_2)^j b^k \mid i, j, k \geq 0\}.$$ For the rest of this paper, $A$ is the associative triple system with basis $$e_1 = \begin{bmatrix} 0 & 1 \\ 0 & 0 \end{bmatrix}, \qquad e_2 = \begin{bmatrix} 0 & 0 \\ 1 & 0 \end{bmatrix}.$$ We make the underlying vector space of $A$ into a nonassociative triple system $A^\omega$ in different ways corresponding to the trilinear operations $\omega$ of Table \[matrixforms\]. Let $X = \{a,b\}$ with $a < b$, and define $\phi\colon A^{\omega}\to F \langle X \rangle$ by $\phi(e_1) = a$, $\phi(e_2) = b$. The symmetric sum {#firstthree} ----------------- The structure constants for $A^\omega$ are determined by $$[e_1,e_1,e_1] = [e_2, e_2, e_2] = 0, \qquad [ e_2, e_1, e_1] = 2e_1, \qquad [e_1, e_2, e_2] = 2e_2.$$ \[smm\] A basis for $U(A^\omega)$ is the set $\{ \, a^i (ba)^j b^k \mid 0 \le i, k \le 2, j \ge 0 \, \}$. We have $U(A^\omega) = F\langle a, b \rangle / I$ where $I$ is generated by $G = \{ G_1, G_2, G_3, G_4 \}$: $$G_1= b^3, \quad G_2= b^2a + bab + ab^2 - b, \quad G_3= ba^2 + aba + a^2b - a, \quad G_4 = a^3.$$ We show that the set $G$ is a Gröbner basis of $I$. There are seven compositions: $$\begin{aligned} &S_1 = G_1 a - b G_2,\quad S_2= G_1 ba - b^2 G_2,\quad S_3 = G_1 a^2 - b^2 G_3,\quad S_4= G_2 a^2 - b^2 G_4, \\ &S_5= G_2 a - b G_3,\quad S_6= G_3 a^2 - ba G_4 ,\quad S_7= G_3 a - b G_4. \end{aligned}$$ We eliminate from $S_1, \dots, S_7$ all occurrences of the leading monomials of $G_1$, $G_2$, $G_3$, $G_4$; we write $\equiv$ to indicate congruence modulo $G$: $$\begin{aligned} S_1 &= - b^2 ab - ba b^2 + b^2 \equiv - \left(-bab - ab^2 +b\right) b - bab^2+ b^2 \equiv 0 , \\ S_2 &= - b^2 \left( bab + ab^2 - b\right) \equiv -\left(- bab - ab^2 + b \right)b^2 \equiv 0 , \\ S_3 &= -b^2aba - b^2a^2b + b^2a = -b^2 a \left( ba +ab -1\right) \\ &\equiv -\left(-bab - ab^2 + b\right) \left( ba +ab -1\right) \\ & \equiv ba\left(-bab - ab^2 + b\right) + babab - bab + a \left(-bab - ab^2 + b\right) b - ab^2\\ &\quad +bab + ab^2- b - bab +b = -ba^2 b^2 + a \left(-bab - ab^2 + b\right) b \\ & \equiv \left(aba + ab^2 - a\right)b^2 - a bab^2 + ab^2 \equiv 0, \\ S_4 &= baba^2 + ab^2a^2 - ba^2 \equiv ba\left(-aba - a^2b \right) + a\left(-bab -ab^2 +b\right) a \\& \equiv -\left(- aba - a^2b + a\right) ba + a\left(-bab - ab^2 + b\right)a \equiv 0.\end{aligned}$$ Similar calculations show that $S_5, S_6, S_7 \equiv 0$. Hence $G$ is a Gröbner basis of $I$, and Proposition \[C(I)proposition\] completes the proof. \[relations\] In $U(A^\omega)$, we have the relations $$a^3 = b^3 = 0, \qquad b^2a = - bab - ab^2 + b, \qquad ba^2 = -aba - a^2b + a.$$ Hence $U(A^\omega)$ is the quotient of $A(-1, -1,1)$ by the ideal generated by $a^3$ and $b^3$. Consider the anti-automorphism $\zeta\colon F\langle a, b \rangle \to F\langle a, b \rangle$ defined by $\zeta(a) = b$ and $ \zeta(b) = a$. Since $\zeta( G_4 ) = G_1$, $\zeta( G_1 ) = G_4$, $\zeta( G_2 ) = G_3$, $\zeta( G_3 ) = G_2$, we see that $\zeta$ induces an anti-automorphism on $U(A^\omega)$, also denoted $\zeta$. A filtration $\{ 0 \} \subseteq V^{(0)} \subseteq V^{(1)} \subseteq \cdots \subseteq \bigcup_n V^{(n)} = U(A^\omega)$ is defined by letting $V^{(n)}$ be the subspace with basis consisting of all $a^i (ba)^j b^k$ where $0 \le i, k \le 2$, $j \ge 0$, and $i+ 2j +k \le n$. The associated graded algebra is $$\mathrm{gr}(U(A^\omega)) = \bigoplus_{i \geq 0}\, \mathcal{G}^i(U(A^\omega) ), \qquad \mathcal{G}^i (U(A^\omega)) = V^{(i)}/ V^{(i -1)}, \qquad V^{-1} = \{0\}.$$ \[dimensionformula\] The dimension of $\mathcal{G}^n ( U(A^\omega) )$ is $1$ if $n = 0$, $2$ if $n = 1$, $4$ if $n = 2$ or $n \ge 3$ (odd), and $5$ if $n \ge 4$ (even). For $n = 0$, there is one monomial: $ (i, j, k) = (0, 0, 0)$. For $n = 1$, there are two: $(1, 0, 0)$, $ (0, 0,1)$. For $ n = 2$, there are four: $(0, 1, 0)$, $(0, 0, 2)$, $(1, 0,1)$, $(2, 0, 0)$. For $ n \geq 3$ and odd, four: $(0, \frac{n -1}{2}, 1)$, $(1 , \frac{n -1}{2}, 0)$, $ (2, \frac{n -3}{2 }, 1)$, $(1, \frac{n - 3}{2}, 2)$. For $n \geq 4$ and even, five: $(0, \frac{n}{2}, 0)$, $(1, \frac{n - 2}{2}, 1)$, $(2, \frac{n - 2}{2}, 0)$, $(0, \frac{n - 2}{2}, 2)$, $ ( 2, \frac{n -4}{2}, 2)$. The Gelfand-Kirillov dimension of $U(A^\omega)$ is $1$. We have $$GK\dim U(A^\omega) = \limsup_{n\to \infty} \log_n \dim V^{(n)} = \lim_{n \to \infty} \frac{\ln \dim V^{(n)}}{\ln n} = 1,$$ since Corollary \[dimensionformula\] implies that $\dim V^{(n)}$ is a polynomial of degree $1$. \[gr\] A $\mathbb{Z}$-grading of $U(A^\omega)$ is given by $$U(A^\omega) = U(A^\omega)_{-2} \oplus U(A^\omega)_{-1}\oplus U(A^\omega)_0 \oplus U(A^\omega)_1 \oplus U(A^\omega)_2,$$ where $U(A^\omega)_n = \mathrm{span}\{ \, a^i (ba)^j b^k \mid j \geq 0,\, 0\leq i, k\leq 2,\, i - k = n \, \}$. Similar to [@Roby Proposition 3.5]. Our next goal is to compute the structure constants of $U(A^\omega)$. \[sm\] For $j,\ell, r, m\geq 0 $, we define the following polynomials: $$L^m_{j, \ell ,r} = \sum^j_{t = 0}(-1)^{j+t}\binom{j}{t} a^{\ell}(ba)^{j+m-t}b^r.$$ \[SM\] We have $$-L^{j-1}_{m+1,\ell,r}+ L^{j-1}_{m,\ell,r} = L^j_{m, \ell, r}\,\, ( j> 0), \qquad -L^{m + 1}_{j, \ell ,r} + L^{m }_{j, \ell ,r} = L^{m}_{j+1,\ell,r}.$$ Use Pascal’s formula for binomial coefficients. \[a\] If $j, m \ge 0$, then in $U(A^\omega)$ we have $$\begin{aligned} \label{oone} (ba)^j \cdot a (ba)^m &= -\widehat{\delta}_{j,0}\, L^{j-1}_{m,2,1} + L^m_{j,1,0}, \\ \label{ttwo} (ba)^j b \cdot (ba)^m &= -\widehat{\delta}_{m, 0}\, L^{m-1}_{j,1,2} + L^j_{m, 0, 1}, \\ \label{tthree} (ba)^j b^2 \cdot a^2 &= -L^0_{j+1,1,1} + (ba)^{j+1} - (ba)^{j +2} \,{+}\, a^2(ba)^jb^2. \end{aligned}$$ For , we use induction on $j$. Clearly the claim is true for $j = 0$. To prove it for $j = 1$, we use induction on $m$. For $m = 0$, Lemma [\[relations\]]{} implies $$(ba) a = ba^2 = - aba - a^2b + a = -L^{0}_{0,2,1} + L^0_{1, 1, 0}.$$ By the inductive hypothesis, we have $$\begin{aligned} (ba) a (ba)^m &= (ba) a (ba)^{m-1} ba = \left(-L^{0}_{m-1,2,1} + L^{m-1}_{1, 1, 0}\right)ba \\ &= (-1)^{m}\sum^{m-1}_{t = 0}(-1)^t \binom{m{-}1}{t}a^2(ba)^{m-t-1} b^2a - a(ba)^{m +1} + a(ba)^m.\end{aligned}$$ Use the second relation of Lemma \[relations\]: (ba) a (ba)\^m = (-1)\^[m]{}\^[m-1]{}\_[t = 0]{}(-1)\^[t+1]{} a\^2(ba)\^[m-t]{}b +(-1)\^[m]{}\^[m-1]{}\_[t = 0]{}(-1)\^[t+1]{} a\^2(ba)\^[m-t-1]{}ab\^2 + (-1)\^[m]{}\^[m-1]{}\_[t = 0]{}(-1)\^t a\^2(ba)\^[m-t-1]{}b - a(ba)\^[m +1]{} + a (ba)\^m. Use Pascal’s formula in the first sum and change index in the third sum: (ba)a (ba)\^m = (-1)\^[m+1]{}( a\^2(ba)\^mb + \^[m-1]{}\_[t = 1]{}(-1)\^[t]{} a\^2(ba)\^[m-t]{}b) +(-1)\^[m]{} \^[m-1]{}\_[t = 1]{}(-1)\^[t]{} a\^2(ba)\^[m-t]{}b +(-1)\^[m]{}\^[m-1]{}\_[t = 0]{}(-1)\^[t+1]{} a\^2(ba)\^[m-t-1]{}ab\^2 + (-1)\^[m]{}\^[m]{}\_[t = 1]{}(-1)\^[t-1]{} a\^2(ba)\^[m-t]{}b - a(ba)\^[m +1]{} + a(ba)\^m. The first $m{-}1$ terms of the second and fourth sums cancel: (ba) a (ba)\^m = -L\^[0]{}\_[m,2,1]{}+ (-1)\^[m]{}\^[m-1]{}\_[t = 0]{}(-1)\^[t+1]{} a\^2(ba)\^[m-t-1]{}ab\^2 + L\^m\_[1, 1, 0]{}. To complete the proof for $j = 1$, it suffices to show that the second sum is 0. But this holds since $a^2 (ba)^\ell a = 0$ for $\ell \geq 0$. To show this, we use induction on $\ell$; the claim is true for $\ell = 0$ by the first equation of Lemma \[relations\]. For $\ell = 1$, Lemma \[relations\] implies $a^2(ba) a = a^2(-aba-a^2b+a) = 0$. For $\ell \geq 1$, Lemma \[relations\] and the inductive hypothesis give $ a^2(ba)^\ell a= a^2(ba)^{\ell -1} (-aba - a^2b+a ) =0. $ We now consider the case $j \ge 1$. Using the inductive hypothesis and Lemma \[relations\], we obtain $$\begin{aligned} & (ba)^{j+1} a (ba)^m = {ba (ba)^j a (ba)^m = ba\big(- L^{j-1}_{m,2,1} + L^m_{j, 1, 0}\big) } \\ &= (-1)^j \sum^{j}_{t = 0}(-1)^t \binom{j}{t} (ba)a(ba)^{j+m-t} \\ &= (-1)^{m+1} \sum^j_{t = 0} \sum^{j{+}m{-}t}_{s = 0}(-1)^s\binom{j}{t}\binom{j{+}m{-}t}{s}a^2(ba)^{j+m-t-s}b \\ &\quad +(-1)^{j+1}\sum^j_{t = 0}(-1)^t\binom{j}{t}a(ba)^{j+m-t+1} +(-1)^j\sum^j_{t = 0} (-1)^t\binom{j}{t} a(ba)^{j+m-t}.\end{aligned}$$ Use Pascal’s formula in the second sum and change index in the third sum: (ba)\^[j+1]{} a (ba)\^m = (-1)\^[m+1]{}\^j\_[t = 0]{}\^[j+m-t]{}\_[s = 0]{}(-1)\^sa\^2(ba)\^[j+m-t-s]{}b + (-1)\^[j+1]{}\^j\_[t = 0]{}(-1)\^ta(ba)\^[j+m-t+1]{} -(-1)\^[j+1]{}\^j\_[t=1]{}(-1)\^ta(ba)\^[j+m-t+1]{} -(-1)\^[j]{}\^j\_[t=1]{}(-1)\^ta(ba)\^[j+m-t+1]{}+a(ba)\^m. The last two sums cancel and the previous expression simplifies to $$(-1)^{m+1}\sum^j_{t = 0}\sum^{j+m-t}_{s = 0}(-1)^s\binom{j}{t}\binom{j{+}m{-}t}{s}a^2(ba)^{j+m-t-s}b + L^m_{j +1, 1, 0}.$$ To complete the proof, it suffices to show that \^j\_[t=0]{} \^[j+m-t]{}\_[s=0]{} (-1)\^s a\^2(ba)\^[j+m-t-s]{}b = \^m\_[t=0]{} (-1)\^t a\^2(ba)\^[j+m-t]{}b. Let $C_r$ denote the coefficient of $a^2(ba)^{j+m-r}b$ in the left side: $$\begin{aligned} C_r &= \sum^j_{t=0} \sum^{j+m-t}_{s=0} \delta_{s+t,r} (-1)^s \binom{j}{t} \binom{j{+}m{-}t}{s} = \sum^j_{t=0} \binom{j}{t} (-1)^{r-t} \binom{j{+}m{-}t}{r{-}t} \\ & = (-1)^r \sum^j_{t=0} (-1)^{t} \binom{j}{t} \binom{j{+}m{-}t}{r{-}t} = (-1)^r \binom{m}{r}. \end{aligned}$$ For the last equality, see [@vanLintWilson Example 10.3]. This completes the proof of . The proof of is obvious by using the anti-automorphism $\zeta$: $$(ba)^j b (ba)^m = \zeta ( (ba)^m a (ba)^j ) = -\widehat{\delta}_{m,0}\, L^{m-1}_{j, 1, 2} + L^j_{m, 0, 1}.$$ For , we use Lemma \[relations\] and get \[al\] [(ba)\^j b\^2 a\^2 = (ba)\^j (-bab - ab\^2 +b)a = -(ba)\^[j+2]{} - (ba)\^j ab\^2 a + (ba)\^[j +1]{}]{}. Write $ T = - (ba)^j ab^2 a$. Lemma \[relations\] implies [T = - (ba)\^j a(-bab - ab\^2 +b) = (ba)\^j a (ba) b + \_[j, 0]{} a\^2 b\^2 - (ba)\^j ab.]{} Using and Lemma \[SM\] we get $$\begin{aligned} T &= \big( -\widehat{\delta}_{j,0}\, L^{j-1}_{1,2,1} + L^1_{j, 1, 0} + \widehat{ \delta}_{j,0}\, L^{j-1}_{0,2,1}-L^0_{j,1,0} \big) b + \delta_{j,0}\, a^2 b^2 \\ &= \widehat{ \delta}_{j,0}\, L^j_{0, 2, 1}\, b - L^{0}_{j+1,1,0}\, b+ \delta_{j, 0}\, a^2 b^2 = a^2 (ba)^j b^2 - L^0_{j+1,1,1}.\end{aligned}$$ Using $T$ in completes the proof of . \[strc\] The structure constants of $U(A^\omega)$ are $$\begin{aligned} a^i (ba)^j b^k \cdot a^\ell(ba)^m b^n &= a^i(ba)^{j+k+\ell +m} b^n, \; \text{if $(k, \ell) = (0,0)$ \text{or} $(k,\ell)=(1,1)$},\label{AA} \\ a^i(ba)^j \cdot a (ba)^mb^n &= -\delta_{i,0} \widehat{ \delta}_{n,2}\widehat{ \delta}_{j,0}\,L^{j-1}_{m,2,n+1} + \widehat{ \delta}_{i,2}L^m_{j, i+1,n},\label{BB} \\ a^i(ba)^jb \cdot (ba)^mb^n &= -\delta_{n,0} \widehat{\delta}_{i,2}\widehat{ \delta}_{m,0}\, L^{m-1}_{j,i+1,2} +\widehat{ \delta}_{n,2} \, L^j_{m, i, n+1},\label{CC} \\ a^i(ba)^jb^2 \cdot a(ba)^mb^n &= - \delta_{n,0}\widehat{ \delta}_{i,2}\, L^m_{j, i+1, 2} + \widehat{ \delta}_{n,2}\, L^j_{m +1,i, n+1}, \label{DD} \\ a^i(ba)^jb \cdot a^2 (ba)^mb^n &= - \delta_{i,0}\widehat{ \delta}_{n,2} L^j_{m, 2,n+1} + \widehat{ \delta}_{i,2} \, L^{m}_{j +1,i + 1, n}, \label{EE} \\ a^i(ba)^j b^2 \cdot (ba)^m b^n &=\delta_{m,0}\delta_{n,0}\, a^i(ba)^j b^2, \label{FF} \\ a^i(ba)^j \cdot a^2 (ba)^mb^n &= \delta_{i,0}\delta_{j,0}\, a^2 (ba)^mb^n, \label{GG}\end{aligned}$$ together with $$\begin{aligned} & a^i (ba)^j b^2 \cdot a^2 (ba)^m b^n \notag \\ & = \sum^{j+1}_{k = 0} (-1)^{k+j} \binom{j+1}{k} \left[ -\delta_{n,0} \delta_{i,0}\widehat{\delta}_{m,0}\, L^{m-1}_{j-k+1,2,2} + \widehat{ \delta}_{n,2}\widehat{ \delta}_{i,2} \, L^{j - k +1}_{m, i+1, n+1} \right] \label{HH} \\ &\quad + a^i(ba)^{j+m+1}b^n - a^i(ba)^{j+m + 2}b^n +\delta_{i,0} \delta_{m, 0}\delta_{n,0} \, a^{2}(ba)^jb^2. \notag\end{aligned}$$ For , use the associativity of $U(A^\omega)$. For and use Lemma [\[relations\]]{} and equations and of Lemma \[a\]. For , Lemma [\[relations\]]{} implies a\^i (ba)\^jb\^2 a(ba)\^mb\^n = a\^i(ba)\^j (-bab - ab\^2 + b)(ba)\^m b\^n = - a\^i(ba)\^[j+1]{}b (ba)\^m b\^n - \_[m,0]{}\_[n,0]{} a\^i (ba)\^j a b\^2+ a\^i(ba)\^jb(ba)\^m b\^n. Using and and Lemma \[SM\] we obtain : a\^i (ba)\^j b\^2 a(ba)\^mb\^n = \_[n,0]{} \_[i,2]{}\_[m, 0]{} L\^[m-1]{}\_[j+1, i+1, 2]{} -\_[n, 2]{} L\_[m,i, n+1]{}\^[j+1]{} - \_[m,0]{} \_[n,0]{} \_[i, 2]{} L\^0\_[j, i+1,2]{} - \_[n, 0]{}\_[i, 2]{}\_[m, 0]{}L\^[m-1]{}\_[j, i+1, 2]{} +\_[n, 2]{}L\^j\_[m,i,n+1]{} = -\_[n, 0]{} \_[i, 2]{}\_[m, 0]{} L\^m\_[j, i+1, 2]{} + \_[n, 2]{} L\^[j]{}\_[m+1,i,n+1]{} - \_[m,0]{} \_[n, 0]{} \_[i,2]{} L\^0\_[j, i+1,2]{}. For use and the anti-automorphism $\zeta$. The proofs of and are obvious by Lemma \[relations\]. For , we use of Lemma \[a\] and obtain a\^i (ba)\^j b\^2a\^2 (ba)\^m b\^n= - a\^i L\^0\_[j+1,1,1]{}(ba)\^m b\^n + a\^i (ba)\^[j+1]{}(ba)\^m b\^n -a\^i (ba)\^[j +2]{} (ba)\^m b\^n [+]{}a\^[i+2]{}(ba)\^jb\^2 (ba)\^m b\^n. Using Lemma \[relations\], we get \[co\] a\^i (ba)\^j b\^2a\^2 (ba)\^m b\^n = -\_[i, 2]{} L\^0\_[j+1,i+1,1]{}(ba)\^m b\^n + a\^i(ba)\^[j+1+m]{} b\^n - a\^i(ba)\^[j+m + 2]{} b\^n +\_[i, 0]{} \_[m, 0]{}\_[n, 0]{} a\^[2]{}(ba)\^jb\^2. Write $ A = L^0_{j+1,i+1,1}(ba)^m b^n$ and use to obtain A= \^[j+1]{}\_[k = 0]{}(-1)\^[k+j + 1]{}a\^[i+1]{}(ba)\^[j-k+1]{}b(ba)\^m b\^n =\^[j+1]{}\_[k = 0]{}(-1)\^[k+j+ 1]{}. Using $A$ in completes the proof of . Our next goal is to describe the center $Z(U(A^\omega))$ of $U(A^\omega)$. We consider the following functions: $$\begin{array}{ccccc} & \gamma_1(m) & \gamma_2(m) & \gamma_3(m) & \gamma_4(m) \\ \text{$m$ even} & m+1 & -3 & 0 & -m+2 \\ \text{$m$ odd} & -(m-3) & -1 & -2 & m \end{array}$$ We consider the following elements: \(m) = \_[m2]{} \^[m -2]{}\_[j = 1]{} (-1)\^[j+1]{} ( (ba)\^j- a(ba)\^[j-1]{} b)+ \_1(m) (ba)\^[m-1]{} + \_2(m) (ba)\^m + \_3(m) a(ba)\^[m-1]{}b + \_4(m) a(ba)\^[m-2]{}b + 3 a\^2 (ba)\^[m-2]{} b\^2. The center of $U(A^\omega)$ is the polynomial algebra in $\mathcal{Z}(m)$, $m \ge 2$: $$Z(U(A^\omega)) = {\mathbb{F}}[ \, \mathcal{Z}(m) \mid m \geq 2 \, ].$$ By Corollary \[gr\] we know $Z(U(A^\omega))$ is graded. Thus if $z$ is central and $z = z_{-2} + z_{-1}+z_{0}+z_{1}+ z_{2}$ is its decomposition into homogenous components, then each $z_i$ is itself central. We now show that $z\in U(A^\omega)_0$. First assume $$0 \neq z_{-2}= \sum_{j \geq 0} s_j (ba)^j b^2 \in Z(U(A^\omega)), \quad s_j \in {\mathbb{F}}.$$ It follows that $$0 = z_{-2}\, a - a\, z_{-2} = \sum_{j \geq 0} s_j (ba)^j b^2a - \sum_{j \geq 0} s_j a (ba)^j b^2.$$ Using of Theorem \[strc\], we see that this element is 0: $$\sum_{j \geq 0}\sum^j_{t = 0}(-1)^{j + t + 1} \binom{j}{t} s_j a (ba)^{j - t} b^2 -\sum_{j \geq 0}s_j (ba)^{j +1}b +\sum_{j \geq 0} s_j (ba)^jb - \sum_{j \geq 0} s_j a (ba)^j b^2.$$ Comparing the coefficients on both sides gives $ s_j = 0$ for all $j$. Now assume $$0 \neq z_2 = \sum_{j\geq 0} s_j a^2(ba)^j \in Z(U(A^\omega)), \quad s_j \in {\mathbb{F}}.$$ It follows that $$0 = b\, z_2 - z_2\, b = \sum_{j\geq 0} s_j b a^2(ba)^j - \sum_{j\geq 0} s_j a^2(ba)^jb.$$ Applying the anti-automorphism $\zeta$ to both sides gives $$\begin{aligned} &\sum_{j\geq 0} s_j (ba)^j b^2 a = \sum_{j\geq 0} s_ja(ba)^j b^2. \end{aligned}$$ Hence $[ \, \sum_{j \geq 0} s_j (ba)^j b^2, \, a \, ] = 0$, contradicting the previous case. Next assume $$0\neq z_1 = \sum_{j\geq 0} s_j\, a (ba)^j + \sum_{\ell\geq 0} t_{\ell}\, a^2(ba)^{\ell} b \in Z(U(A^\omega)), \quad s_j, t_{\ell} \in {\mathbb{F}}.$$ It follows that $$0 = b z_1 - z_1 b = \sum_{j \geq 0} s_j \, (ba)^{j +1} + \sum_{\ell\geq 0} t_{\ell}\, b a^2(ba)^{\ell} b - \sum_{j\geq 0} s_j\, a (ba)^j b- \sum_{\ell\geq 0} t_{\ell}\, a^2(ba)^{\ell} b^2.$$ Using of Theorem \[strc\] gives 0 = \_[j0]{} s\_j (ba)\^[j +1]{} - \_[0]{}\^\_[t = 0]{}(-1)\^[+t]{} t\_ a\^2(ba)\^[- t]{}b\^2 -\_[0]{} t\_ a(ba)\^[+1]{} b + \_[0]{} t\_ a(ba)\^b - \_[j 0]{} s\_j a (ba)\^j b- \_[0]{} t\_ a\^2(ba)\^ b\^2. Comparing the coefficients on both sides gives $ s_j= 0 = t_{\ell}$ for all $j, \ell$. Similarly we can show that $Z(U(A^\omega))\cap U(A^\omega)_{-1} = 0$. Therefore $z\in U(A^\omega)_{0}$. Now $U(A^\omega)_{0}$ is a commutative subalgebra: any element in $U(A^\omega)_0$ is a linear combination of $(ba)^j$, $a (ba)^k b$, $a^2(ba)^\ell b^2$ for $j, k, \ell \geq 0$, and these elements commute by Theorem \[strc\]. Using the anti-automorphism $\zeta$, we see that an element in $U(A^\omega)_0$ commutes with $a$ if and only if it commutes with $b$. So it suffices to determine the elements that commute with $a$. Without loss of generality, we choose $$z = \sum^m_{j = 0}\sum^2_{i=0} s_{i, j}\, a^i (ba)^j b^i \in U(A^\omega)_0, \quad m\geq 0,\,\, s_{i,j} \in {\mathbb{F}}.$$ By the relations of Lemma \[relations\] we have $$\begin{aligned} \label{ax} az = \sum^m_{j = 0} s_{0, j}\, a(ba)^j + \sum^m_{j = 0} s_{1,j}\, a^2(ba)^jb.\end{aligned}$$ On the other hand, za = \^m\_[j = 0]{} s\_[0,j]{}(ba)\^ja +\^m\_[j = 0]{} s\_[1,j]{} a(ba)\^[j+1]{}+\^m\_[j = 0]{}s\_[2,j]{} a\^2(ba)\^j b\^2a. Using and of Theorem \[strc\] we obtain \[xaa\] za = s\_[0, 0]{} a + \^m\_[j = 1]{}s\_[0, j]{}(-a\^2(ba)\^[j-1]{}b +\^j\_[t = 0]{}(-1)\^[j+t]{} a(ba)\^[j-t]{})+\^m\_[j = 0]{} s\_[1, j]{} a(ba)\^[j+1]{}- \^m\_[j = 0]{}s\_[2, j]{} a\^2(ba)\^[j+1]{}b + \^m\_[j=0]{}s\_[2, j]{}a\^2(ba)\^j b. We write $$A = \sum^m_{j = 1} \sum^j_{t = 0}(-1)^{j+t} \binom{j}{t} s_{0, j}\, a(ba)^{j-t}, \qquad E=\sum^m_{j = 0} s_{1, j}\, a(ba)^{j+1}.$$ We obtain $$\begin{aligned} \label{AE} & A + E = \sum^m_{r = 0} \bigg( \sum^m_{j = 1} (-1)^r \binom{j}{r} s_{0,j} \bigg) a(ba)^r + \sum^{m+1}_{j = 1} s_{1,j-1}\,a(ba)^{j} \\ &= \sum^m_{j = 1} s_{0, j}\, a + \sum^m_{r = 1}\bigg( \sum^m_{j = 1} (-1)^r \binom{j}{r}s_{0, j} + s_{1,r-1}\,\bigg)a(ba)^r + s_{1, m}\, a(ba)^{m+1}. \notag\end{aligned}$$ Using in gives $$\begin{aligned} za &= \bigg(s_{0, 0} \,+ \sum^m_{j = 1} s_{0, j}\bigg) a - \sum^m_{j = 1}s_{0, j}\, a^2(ba)^{j-1}b \\ &\quad + \sum^m_{r = 1} \bigg( \sum^m_{j = 1} s_{0, j}\,(-1)^r \binom{j}{r} + s_{1,r-1}\,\bigg)a(ba)^r + s_{1, m}\, a(ba)^{m+1} \\ &\quad - \sum^m_{j = 0}s_{2, j}\, a^2(ba)^{j+1}b + \sum^m_{j=0}s_{2, j}\, a^2(ba)^j b.\end{aligned}$$ Changing index in the second and fourth sums and combining coefficients gives $$\begin{aligned} za &= \Big(s_{0, 0}+ \sum^m_{j = 1} s_{0, j}\Big) a + \Big( {-}s_{0, 1} {+} s_{2, 0}\Big)a^2b + \sum^{m-1}_{j = 1} \Big( {-}s_{0, j+1} {-} s_{2, j-1} {+} s_{2, j} \Big) a^2(ba)^{j}b \notag \\ &\quad + \sum^m_{r = 1}\Big( \sum^m_{j = 1} s_{0, j}(-1)^r \binom{j}{r} + s_{1, r-1}\Big)a(ba)^r + s_{1, m}\, a(ba)^{m+1} \\ &\quad +\left(- s_{2, m-1} + s_{2, m}\right) a^2(ba)^m b - s_{2, m} \, a^2(ba)^{m +1}b. \notag\end{aligned}$$ Comparing the coefficients in this expression with , we get this linear system: $$\begin{aligned} \label{*} \begin{cases} &\sum^m_{j = 1} s_{0, j} = 0,\\& -s_{0, 1} + s_{2, 0} - s_{1, 0} = 0,\\ &\notag -s_{0, j+1} - s_{2, j-1}+s_{2, j}- s_{1, j} = 0\quad (1\leq j \leq m-1), \\&\tag{$\mathcal{T}$} (-1)^r \sum^m_{j = 1} \binom{j}{r}s_{0, j} + s_{1, r-1} - s_{0, r} = 0\quad (1\leq r \leq m),\notag \\ & s_{2, m} = s_{1, m}= s_{2, m-1} = 0.\notag\end{cases}\end{aligned}$$ For $m < 2$, the only solution is trivial. For $m \geq 2$, a calculation (details omitted) shows that has $m{-}1$ linearly independent solutions. For each $m$, we have the following solution: $$\begin{aligned} {2} \label{**} s_{2,m-2} &= 1, \quad s_{2,j} = 0 \; (j\neq m-2), &\quad s_{0, j} &= \frac{1}{3}(-1)^{j+1} \binom{m{-}1}{j{-}1} \; (1 \leq j \leq m-2), \\ s_{0, m-1} &= \begin{cases} (m{+}1)/3 &\text{if $m$ is even} \\ (3{-}m)/3 & \text{if $m$ is odd} \end{cases}, &\quad s_{0,m} &= \begin{cases} -1 &\text{if $m$ is even} \\ -1/3 & \text{if $m$ is odd} \end{cases}, \\ s_{1,i-1} &= -s_{0,i} \; (1\leq i\leq m-2), &\quad s_{1, m-2} &= -s_{0, m-1}+1, \\ s_{1, m-1} &= \begin{cases} 0 & \text{if $m$ is even} \\ -2/3 & \text{if $m$ is odd} \end{cases}, &\quad s_{1,m} &=0.\end{aligned}$$ Using this solution for $z$, and observing that any solution for $m{-}1$ is also a solution for $m$, we obtain a complete list of linearly independent solutions for . The alternating sum ------------------- The structure constants for $A^\omega$ are $0$, the set of ideal generators is empty, and hence $U(A^\omega)$ is the free associative algebra on $a$ and $b$. The Gelfand-Kirillov dimension of $U(A^\omega)$ is $\infty$. The cyclic sum --------------- The results are identical to those for the symmetric sum, since the structure constants are $$[e_2, e_1, e_1] = e_1, \qquad [e_1, e_2, e_2] = e_2, \qquad [e_1, e_1,e_1] =[e_2, e_2, e_2]= 0.$$ The Lie family, $q = \infty$ {#Lie} ---------------------------- The structure constants for $A^{\omega^{\tiny{\infty}}_{\tiny{L}}}$ are determined by $$[e_1, e_2, e_1 ] = 2e_1, \qquad [e_2, e_2, e_1] = -2 e_2.$$ \[do\] The universal associative envelope $U(A^{\omega^{\tiny{\infty}}_{\tiny{L}}})$ is isomorphic to the down-up algebra $A(2, -1, -2)$. We have $U\big(A^{\omega^{\tiny{\infty}}_{\tiny{L}}}\big) = F\langle a, b \rangle / I$, where $I$ is the ideal generated by these two elements, which form a Gröbner basis: $b^2a -2 bab + ab^2 +2 b$, $ba^2 -2 aba + a^2b +2 a$. \[sl\] If we replace ${\omega^{\tiny{\infty}}_{\tiny{L}}}$ by $\omega^{\prime} = [[-, - ], -]$ then we get the 2-dimensional simple Lie triple system $A^{\omega^{\prime}}$ with relations $[e_1, e_2, e_1] = 2e_1$, $[e_1, e_2, e_2] = -2 e_2$; the results for $U(A^{\omega^{\prime}})$ are identical to those for ${\omega^{\tiny{\infty}}_{\tiny{L}}}$. Benkart and Roby [@Roby] showed that the down-up algebra $A(2, -1, -2)$ is isomorphic to the universal associative envelope $U({\mathfrak{sl}_2})$ of the simple Lie algebra of $2 \times 2$ matrices of trace $0$ with basis $\{h, e, f \}$ and relations $[e, f] = h$, $[h, e] = 2e$, and $[h, f]= -2f$. In $U({\mathfrak{sl}_2})$ we have $ef - fe = h$, $he - eh= 2e$, $hf-fh= -2f$. \[helemma1\] If $\ell, k, m, j \ge 0$ then in $U({\mathfrak{sl}_2})$ we have $$\begin{aligned} e^\ell \cdot h^k &= \sum_{q = 0}^k (-1)^q 2^q \binom{k}{q} \ell^q h^{k-q} e^\ell, \label{one} \\ h^k \cdot f^m &= \sum_{q=0}^k (-1)^q 2^q \binom{k}{q} m^q f^m h^{k-q}, \label{two} \\ e^\ell \cdot f^j &= \ell! j!\sum_{r=0}^{\min(j,\,\ell)} \frac{f^{j-r}}{(j - r)!}\binom{h{-}j{-}\ell{+}2r}{r} \frac{e^{\ell-r}}{(\ell -r)!}. \label{three} \end{aligned}$$ For , we use induction on $k$. The claim is clear for $k= 0$. To prove the claim for $k =1$, we use induction on $\ell$. For $\ell = 1$, the claim holds since $eh = he- 2e$. Assume that $\ell \ge 1$. By the inductive hypothesis we have $$\begin{aligned} e^{\ell+1} h = e e^\ell h = \big( e h e^\ell - 2\ell e^{\ell +1} \big) = h e^{\ell +1} -2e^{\ell+1} - 2\ell e^{\ell+1}, \end{aligned}$$ so the claim is true for $k = 1$. For $k \ge 1$, the inductive hypothesis implies $$\begin{aligned} &e^\ell h^{k+1} = \sum_{q = 0}^k (-1)^q 2^q \binom{k}{q} \ell^q h^{k-q} e^\ell h = \sum_{q = 0}^k (-1)^q 2^q \binom{k}{q} \ell^q h^{k - q} \big ( he^{\ell} - 2\ell e^\ell\big) \\ &= \sum_{q=0}^k (-1)^q 2^q \binom{k}{q} \ell^q h^{k-q+1}e^{\ell} + \sum_{q=1}^{k +1} (-1)^{q} 2^{q} \binom{k}{q-1} \ell^{q}h^{k-q+1}e^{\ell} \\ &= h^{k+1}e^{\ell} + \sum_{q=1}^k (-1)^q 2^q \left[ \binom{k}{q} {+} \binom{k}{q{-}1} \right] \ell^q h^{k+1-q} e^{\ell} + (-1)^{k +1} 2^{k+1} \ell^{k+1} e^\ell. \end{aligned}$$ Using Pascal’s formula for binomial coefficients we obtain $$\begin{aligned} e^\ell h^{k+1}&= h^{k+1} e^\ell + \sum_{q=1}^k (-1)^q 2^q \binom{k{+}1}{q} \ell^q h^{k+1-q} e^\ell +(-1)^{k + 1} 2^{k+1} \ell^{k+1} e^\ell. \end{aligned}$$ This proves , and is similar; for , see Humphreys [@Humphreys Lemma 26.2]. \[th\] The structure constants of $U({\mathfrak{sl}_2})$ are ( f\^i h\^j e\^k )( f\^h\^m e\^n ) = k! ! \_[r = 0]{}\^[(,k)]{} \_[q = 0]{}\^j\^m\_[i = 0]{} (-1)\^[q + i]{} 2\^[q+i]{} h\^[j-q]{} h\^[m-i]{} e\^[k - r + n]{}. Using Theorem \[th\] and the homomorphism $\psi\colon {\mathfrak{sl}_2}\to A(2, -1, 1)$ from [@Roby], we obtain the structure constants of $U\big(A^{\omega^{\tiny{\infty}}_{\tiny{L}}}\big)$ with respect to the basis $\mathfrak{B}_2$ with $(c_1, c_2)= (-1,0)$ (see Lemma \[zhaolemma\]). The Lie family, $q = \frac12$ ------------------------------ The structure constants for $A^{\omega^{\tiny{1/2}}_{\tiny{L}}}$ are zero. \[lj\] The universal associative envelope $U(A^{\omega^{\tiny{1/2}}_{\tiny{L}}})$ is isomorphic to the down-up algebra $A(0, 1, 0)$. We have $U(A^{\omega^{\tiny{1/2}}_{\tiny{L}}}) = F\langle a, b\rangle / I$, where $I$ is the ideal generated by these two elements, which form a Gröbner basis: $b^2a - ab^2$, $ba^2 - a^2b$. If we replace ${\omega^{\tiny{1/2}}_{\tiny{L}}}$ by ${\omega^{\prime\prime}} = [- \circ-, -]$ then we get an anti-Lie triple system $A^{\omega^{\prime\prime}}$, and the results for $U(A^{\omega^{\prime\prime}})$ are the same as those for $A^{\omega^{\tiny{1/2}}_{\tiny{L}}}$. \[rr\] If $i, j \ge 0$, then in $U(A^{\omega^{\tiny{1/2}}_{\tiny{L}}})$ we have $$\begin{aligned} b^i\cdot a^j &=\begin{cases} a^{j-1} (ba) b^{i-1} &\text{if $i, j$ are both odd,} \\ a^jb^i &\text{otherwise.} \label{ba} \end{cases} \\ (ba)^j \cdot a^i &= \begin{cases} a^{i+1} (ba)^{j-1} b &\text{if $i$ is odd, $j \neq 0$,} \\ a^i(ba)^j &\text{otherwise.} \label{baa} \end{cases} \\ b^i \cdot (ba)^j &= \begin{cases} a (ba)^{j-1} b^{i+1} &\text{if $i$ is odd, $j \neq 0$, } \\ (ba)^j b^i &\text{otherwise.} \end{cases}\label{bba} \\ (ba)^i \cdot a(ba)^j &= \begin{cases} a^{2j+2}(ba)^{i-j-1}b^{2j+1} &\text{if $i > j$,} \\ a^{2i+1}(ba)^{j-i}b^{2i} &\text{otherwise.}\label{haa} \end{cases} \end{aligned}$$ For , we use induction on $i$. The claim is clear for $i = 0$. To prove the claim for $i = 1$ we use induction on $j$. For $j = 0$ or $1$, the claim is obvious. For $j = 2$ the claim holds since $ba^2 = a^2 b $ . We now prove the claim for $j\geq 2$. By the inductive hypothesis, we have $$\begin{aligned} ba^j &= \begin{cases} a^{j-2} baa &\text{if $j{-}1$ is odd} \\ a^{j-1} ba &\text{otherwise} \end{cases} =\begin{cases} a^{j-1} ba &\text{if $j$ is odd} \\ a^{j-2}baa &\text{otherwise} \end{cases} = \begin{cases} a^{j-1} ba &\text{if $j$ is odd} \\ a^jb &\text{otherwise} \end{cases}.\end{aligned}$$ So the claim holds for $i = 1$. We now consider the case $i\geq 1$. If $i+1$ is odd, then the inductive hypothesis implies $$b^{i+1}a^j = bb^ia^j = b a^j b^i = \begin{cases} a^{j-1}(ba)b^i &\text{if $j$ is odd}\\ a^{j}b^{i+1} &\text{otherwise} \end{cases}.$$ If $i+1$ is even, then the inductive hypothesis gives $$\begin{aligned} b^{i+1}a^j &= bb^ia^j =\begin{cases} b a^{j-1}(ba)b^{i-1} &\text{if $j$ is odd} \\ ba^jb^i &\text{otherwise}\end{cases} = \begin{cases} a^{j-1}b(ba)b^{i-1} &\text{if $j$ is odd} \\ a^j b^{i+1} &\text{otherwise} \end{cases}. \end{aligned}$$ Using $b^2a = ab^2$ we get $b^{i+1}a^j = a^{j-1}ab^2 b^{i-1}$ if $j$ is odd, $a^j b^{i+1}$ otherwise; in both cases the result is $a^jb^{i+1}$. This completes the proof of . For we use induction on $i$. The claim is obvious for $i = 0$. To prove the claim for $i = 1$, we use induction on $j$. For $j = 0$, the claim is obvious. For $j = 1$, the claim holds by using $ba^2 = a^2b$. We now consider the case of general $j$. By the inductive hypothesis, we have $(ba)^ja = ba (ba)^{j-1}a = ba a^{2} (ba)^{j-2} b = a^2 ba(ba)^{j-2} = a^2(ba)^{j-1}b.$ So the claim is true for $i =1$. We now consider the case $i\geq 1$. The claim is obvious for $j= 0$, so we assume that $j\neq 0$. If $i+1$ is odd, then the inductive hypothesis implies $ (ba)^ja^{i+1} = a^i(ba)^ja = a^{i+2}(ba)^{j-1}b$. If $i+1$ is even, then $ (ba)^ja^{i+1} = a^{i+1}(ba)^{j-1}ba = a^{i+1}(ba)^j. $ This completes the proof of . The proof of follows by using the anti-automorphism $\eta$ from [@Roby] of the down-up algebra $A(0, 1, 0)$. For , we use induction on $i$. The claim is obvious for $i = 0$. To prove the claim for $i = 1$, we use induction on $j$. The claim holds for $j = 0$ by using $(ba) a = a^2b$. For $j \geq 1$, $ba^2 = a^2b$ and imply $ ba a (ba)^j = a^2b(ba)^j = a^3(ba)^{j-1}b^2. $ So the claim is true for $i = 1$. We now consider the case of $i> 1$. By the inductive hypothesis, we have $$\begin{aligned} (ba)^{i+1} a (ba)^j &= ba (ba)^ia(ba)^j = \begin{cases} ba a^{2j+2}(ba)^{i-j-1}b^{2j+1} &\text{if $i > j$ } \\ ba a^{2i+1}(ba)^{j-i}b^{2i} &\text{if $i \leq j$} \end{cases}.\notag \end{aligned}$$ Therefore $$(ba)^{i+1} a (ba)^j = \begin{cases} b a^{2j+3}(ba)^{i-j-1}b^{2j+1} &\text{if $i > j$ } \\ b a^{2i+2}(ba)^{j-i}b^{2i} &\text{if $i \leq j$} \end{cases}.\label{i}$$ Two cases need to be considered. (I) If $i+1 > j$, then $ i = j$ or $i >j $. Hence, $$\begin{aligned} (ba)^{i+1} a (ba)^j = \begin{cases} b a^{2j+3}(ba)^{i-j-1}b^{2j+1} &\text{if $i > j$ } \\ b a^{2i+2}b^{2i} &\text{if $i = j$} \end{cases}.\end{aligned}$$ Using we obtain $$\begin{aligned} (ba)^{i+1} a (ba)^j = \begin{cases} a^{2j} ba^3(ba)^{i-j-1}b^{2j+1} = a^{2j+2} (ba)^{i - j}b^{2j+1} &\text{if $i > j$} \\ a^{2i + 2}b^{2i +1} &\text{if $i = j$} \end{cases}.\end{aligned}$$ Therefore, $ (ba)^{i+1} a (ba)^j = a^{2j+2}(ba)^{i-j}b^{2j+1}. $ (II) If $ i +1 \leq j$, then $i < j$. Hence $ (ba)^{i+1} a (ba)^j = b a^{2i+2}(ba)^{j-i}b^{2i}.$ Using and we obtain $ (ba)^{i+1} a (ba)^j = a^{2i+2}b(ba)^{j-i}b^{2i} = a^{2i+3}(ba)^{j-i-1}b^{2(i+1)}. $ Combining the results of (I) and (II) completes the proof of . \[stanti\] The structure constants of $U(A^{\omega^{\tiny{1/2}}_{\tiny{L}}})$ are a\^i (ba)\^j b\^k a\^(ba)\^m b\^n = a\^[i+-1]{}(ba)\^[j+m+1]{}b\^[k-1+n]{} &\ a\^[i+]{}(ba)\^[j+m]{} b\^[k+n]{} &\ \_[j, m]{} a\^[2m + i++1]{} (ba)\^[j-m-1]{} b\^[2m+ k+n+1 ]{}\ + (1- \_[j, m]{}) a\^[2j+i+]{}(ba)\^[m-j]{}b\^[2j+k+n]{} &\ \_[j, m-1]{} a\^[2m+i+]{}(ba)\^[j-m]{} b\^[2m+k+n]{}\ + (1 - \_[j, m-1]{}) a\^[2j+i++1]{}(ba)\^[m-j-1]{}b\^[2j+k+n+1]{} & where $\chi_{\ell, t} = 1$ if $\ell > t$ and 0 otherwise. We use equations , and . If $k$ and $\ell $ are odd, then [a\^i (ba)\^j b\^k a\^ (ba)\^m b\^n = a\^i (ba)\^ja\^[-1]{} (ba)b\^[k-1]{}(ba)\^mb\^n = a\^[i+-1]{}(ba)\^[j+1+m]{}b\^[k-1+n]{}]{}. If $k$ and $\ell$ are even, then [a\^i (ba)\^j b\^k a\^ (ba)\^m b\^n = a\^i [(ba)\^ja\^]{} b\^k(ba)\^mb\^n = a\^[i+]{}(ba)\^[j+m]{} b\^[k+n]{}]{}. If $k$ is even and $\ell$ is odd, then $$\begin{aligned} & a^i(ba)^j b^k \cdot a^{\ell} (ba)^mb^n = a^i (ba)^j a^{\ell} b^k (ba)^m b^n \\ &= \widehat{\delta}_{j,0} \, a^{i+ \ell +1} (ba)^{j-1}(b^{k+1}(ba)^m)b^n + \delta_{j,0}\, a^{i+\ell} b^k (ba)^mb^n \\ &= \widehat{\delta}_{j,0}\left[ \widehat{\delta}_{m,0}\, a^{i+\ell+1} {(ba)^{j-1}a (ba)^{m-1}} b^{k+2+n} +\delta_{m,0}\, a^{i+\ell+1}(ba)^{j-1}b^{k+n+1}\right] \\ &\qquad +\delta_{j,0}\, a^{i+\ell} (ba)^mb^{k+n}.\end{aligned}$$ Using completes the proof. If $k$ is odd and $\ell$ is even, then [a\^i (ba)\^j [b\^k a\^]{}(ba)\^m b\^n = a\^i (ba)\^j a\^ b\^k(ba)\^m b\^n = a\^[i+]{}(ba)\^j b\^k (ba)\^m b\^n]{} = \_[m,0]{} a\^[i+]{} (ba)\^[j]{} a (ba)\^[m-1]{} b\^[k+n+1]{} + \_[m, 0]{} a\^[i+]{}(ba)\^j b\^[k + n]{}. Using again completes the proof. The anti-Jordan family, $q = \infty$ ------------------------------------ The structure constants for $A^{\omega^{\tiny{\infty}}_{\tiny{AJ}}}$ are $$[e_1, e_1, e_2 ] = -2 e_1, \quad [e_2, e_1, e_1]= 2e_1, \quad [ e_1, e_2, e_2] = 2e_2, \quad [e_2, e_2, e_1] = -2e_2.$$ The universal associative envelope $U(A^{\omega^{\tiny{\infty}}_{\tiny{AJ}}})$ is isomorphic to the down-up algebra $A(2, -1, -2)$, so we have $U(A^{\omega^{\tiny{\infty}}_{\tiny{AJ}}}) \cong U(A^{\omega^{\tiny{\infty}}_{\tiny{L}}})$. Similar to the proof of Lemma \[do\]. The anti-Jordan family, $q = \frac12$ ------------------------------------- The structure constants for $A^{\omega^{\tiny{1/2}}_{\tiny{AJ} }} $ are zero. The universal associative envelope $ U\big(A^{\omega^{\tiny{1/2}}_{\tiny{AJ} }}\big)$ is isomorphic to the down-up algebra $A(0, 1, 0)$, so we have $U\big(A^{\omega^{\tiny{1/2}}_{\tiny{AJ} }}\big) \cong U\big(A^{\omega^{\tiny{1/2}}_{\tiny{L} }}\big)$. Similar to the proof of Lemma \[lj\]. The structure $A^{\omega^{\tiny{1/2}}_{\tiny{AJ} }} $ is an example of non-simple anti-Jordan triple systems. For the classification of simple finite-dimensional anti-Jordan triple systems see [@simen Theorem 6]. For the universal associative envelope of the anti-Jordan triple system of all $n \times n$ matrices see [@Elgendy Theorem 6.2.12]. Finite dimensional envelopes {#finite} ============================ In this section, we consider the trilinear operations of “Jordan type”. The Jordan family, $q = \infty$ {#Jordan} ------------------------------- The structure constants for $A^{\omega^{\tiny{\infty}}_{\tiny{J}}}$ are $$[e_1, e_2, e_1] = 2 e_1, \qquad [ e_2, e_1, e_2 ] = 2 e_2.$$ \[Jinf\] A basis for $U(A^{\omega^{\tiny{\infty}}_{\tiny{J}}})$ consists of the elements 1, $a$, $b$, $ab$, $ba$. The structure constants are $a \cdot b = ab$, $a \cdot ba = a$, $b \cdot a = ba$, $b \cdot ab = b$, $ab \cdot a = a$, $ab \cdot ab = ab$, $ba \cdot b = b$, $ba \cdot ba = ba$. The Wedderburn decomposition is $U(A^{\omega^{\tiny{\infty}}_{\tiny{J}}}) = {\mathbb{Q}}\oplus M_{2\times 2}$. The only finite dimensional irreducible representations are the trivial 1-dimensional representation and the natural 2-dimensional representation. We have $U = U(A^{\omega^{\tiny{\infty}}_{\tiny{J}}}) = F\langle a, b \rangle / I$ where $I$ is generated by $b^3$, $b^2a + ab^2$, $bab - b$, $ba^2 + a^2b$, $aba - a$, $a^3$. We compute a Gröbner basis of $I$. There are four compositions with normal forms $ab^2$, $a^2b$, $b^2$, $a^2$. Including these with the original generators and self-reducing the resulting set produces the four generators $bab - b$, $aba - a$, $b^2$, $a^2$. All compositions of these elements reduce to $0$, and so we have a Gröbner basis. A basis for the quotient algebra consists of the cosets of the monomials which are not divisible by the leading monomial of any element of the Gröbner basis. This gives the stated basis for $U$. It follows that $U$ satisfies $a^2 = 0$, $b^2 = 0$, $aba = a$, $bab = b$ and these give the stated structure constants. To decompose $U$ we follow [@Bremner]. Using [@Bremner Corollary 12] we verify that the radical is zero, and hence $U$ is semisimple. By [@Bremner Corollary 15] the center $Z(U)$ has dimension 2, basis $z_1 = 1$, $z_2 = ab + ba$, and structure constants $z_1 \cdot z_1 = z_1$, $z_1 \cdot z_2 = z_2$, $z_2 \cdot z_2 = z_2$. Since $z^2_2 = z_2$, the minimal polynomial of $z_2$ is $t^2 - t$. Thus $Z(U)$ splits in two 1-dimensional ideals with bases $z_2 - z_1$ and $z_2$. Scaling these basis elements to obtain idempotents gives $e_1 = -z_2 +z_1$, $e_2 = z_2$. The corresponding elements in $U$ are $e_1 = -ab-ba+1$, $e_2 = ab+ba$. The ideals in $U$ generated by $e_1$ and $e_2$ have dimensions 1 and 4 respectively, and this gives the Wedderburn decomposition. The Jordan family, $q = 0$ -------------------------- The structure constants for $A^{\omega^{\tiny{0}}_{\tiny{J}}}$ are $$[e_1, e_2, e_1] = [e_2, e_1, e_1] = e_1, \qquad [ e_2, e_1, e_2] = [e_1, e_2, e_2]= e_2.$$ \[jq0\] A basis for $U(A^{\omega^{\tiny{0}}_{\tiny{J}}})$ consists of the elements $1$, $a$, $b$, $a^2$, $ab$, $ba$, $b^2$, $aba$, $ab^2$. The structure constants are $a\cdot a = a^2$, $a \cdot b = ab$, $a \cdot ba = aba$, $a \cdot b^2 = ab^2$, $b \cdot a = ba$, $b \cdot b = b^2$, $b \cdot a^2 = a-aba$, $b \cdot ab = b-ab^2$, $b \cdot aba = ba$, $b \cdot ab^2 = b^2$, $ab \cdot a = aba$, $ab \cdot b = ab^2$, $ab \cdot a^2 = a^2$, $ab \cdot ab = ab$, $ab \cdot aba = aba$, $ab \cdot ab^2 = ab^2$, $ba \cdot a = a-aba$, $ba \cdot b = b-ab^2$, $ba \cdot ba = ba$, $ba \cdot b^2 = b^2$, $aba \cdot a = a^2$, $aba \cdot b = ab$, $aba \cdot ba = aba$, $aba \cdot b^2 = ab^2$. The Wedderburn decomposition is $U(A^{\omega^{\tiny{0}}_{\tiny{J}}}) = \mathfrak{R} \oplus {\mathbb{Q}}\oplus M_{2\times 2}$ where $\mathfrak{R}$ is the radical of dimension 4. There are only two finite dimensional irreducible representations. We have $U = U(A^{\omega^{\tiny{0}}_{\tiny{J}}}) = F\langle a, b \rangle / I$ where $I$ is generated by $b^3$, $b^2a$, $bab + ab^2 - b$, $ba^2 + aba - a$, $a^2b$, $a^3$. This set is a Gröbner basis for $I$. Hence $U$ is finite dimensional and has the stated basis. The following relations hold in $U$: $b^3 = 0$, $b^2a = 0$, $bab = -ab^2 + b$, $ba^2=- aba + a$, $a^2b = 0$, $a^3=0.$ These imply the stated structure constants. Using [@Bremner Corollary 12], a basis of the radical $\mathfrak{R} = \mathfrak{R}(U)$ consists of the elements $\xi_1 = a - aba,\,\, \xi_2 = a^2,\,\, \xi_3 = b^2,\,\, \xi_4= ab^2.$ Hence we have these relations in $Q = U/\mathfrak{R}$: $a = aba,\,\, a^2 = b^2 = ab^2 = 0.$ The semisimple quotient $Q$ has dimension $5$, and a basis consists of the cosets of $\eta_1 = 1$, $\eta_2 = b$, $\eta_3 = ab$, $\eta_4 = ba$, $\eta_5= aba$. The center $Z(Q)$ has dimension $2$, basis $z_1=\eta_1,\,\, z_2=\eta_3 +\eta_4,$ and structure constants $z_1\cdot z_1= z_1$, $z_1\cdot z_2 = z_2 \cdot z_1 = z_2$, $z_2 \cdot z_2 = z_2$. Since $ z^2_2 = z_2$, the minimal polynomial of $z_2$ is $t^2 - t$. Thus $ Z(Q) = J \oplus K$ where $J= \langle z_2-z_1 \rangle$ and $K= \langle z_2\rangle$ and both ideals are 1-dimensional. Scaling the basis elements to obtain idempotents gives $e_1 = z_1 - z_2$, $e_2 = z_2$. The corresponding elements in $Q$ are $e_1 = \eta_1 - \eta_3 - \eta_4, \,\, e_2 = \eta_3 +\eta_4.$ The ideals in $Q$ generated by $e_1$ and $e_2$ have dimensions 1 and 4 respectively, and this gives the Wedderburn decomposition. The Jordan family, $q = \frac12$ -------------------------------- The structure constants for $A^{\omega^{\tiny{1/2}}_{\tiny{J}}}$ are $$\begin{aligned} &[e_1, e_1, e_2] = [e_1, e_2, e_1]= [e_2, e_1, e_1]= 2 e_1, \\& [e_2, e_2, e_1] = [ e_2, e_1, e_2] = [e_1, e_2, e_2]= 2e_2.\end{aligned}$$ We have the isomorphism $U(A^{\omega^{\tiny{1/2}}_{\tiny{J}}}) \cong U(A^{\omega^{\tiny{\infty}}_{\tiny{J}}})$. We have $U = U(A^{\omega^{\tiny{1/2}}_{\tiny{J}}}) = F\langle a, b\rangle / I$ where $I$ is generated by $b^3$, $b^2a + \tfrac12 bab - \tfrac12 b$, $b^2a +2 bab +3 ab^2 -2 b$, $ba^2 + \tfrac23 aba + \tfrac13 a^2b - \tfrac23 a$, $aba +2 a^2b - a$, $a^3$. The first iteration of the Gröbner basis algorithm produces the seven compositions $bab^2 - 2 b^2$, $bab^2 - \tfrac12 b^2$, $bab - b$, $bab +2 ab^2 - b$, $bab + \tfrac{3}{2} ab^2 -b$, $a^2b$, $a^2$. Including these with the original generators, and self-reducing the resulting set, produces the same ideal generators as for $q = \infty$; hence the two quotient algebras are isomorphic. The Jordan family, $q = 1$ -------------------------- The structure constants for $A^{\omega^{\tiny{1}}_{\tiny{J}}}$ are $$[e_1, e_1, e_2] = [e_1, e_2, e_1]= e_1, \qquad [e_2, e_2, e_1] = [e_2, e_1, e_2]= e_2.$$ \[jor1\] A basis for $U(A^{\omega^{\tiny{1}}_{\tiny{J}}})$ consists of the elements $1$, $a$, $b$, $a^2$, $ab$, $ba$, $b^2$, $a^2b$, $bab$. The structure constants of $U(A^{\omega^{\tiny{1}}_{\tiny{J}}})$ are $a \cdot a = a^2$, $a\cdot b = ab$, $a\cdot ab = a^2b$, $a\cdot ba = a-a^2b$, $a\cdot bab = ab$, $b \cdot a = ba$, $b\cdot b =b^2$, $b\cdot ab = bab$, $b\cdot ba = b-bab$, $b\cdot bab = b^2$, $a^2\cdot b = a^2b$, $a^2\cdot ba = a^2$, $a^2\cdot bab = a^2b$, $ab\cdot a = a - a^2b$, $ab\cdot ab = ab$, $ba\cdot b=bab$, $ba\cdot ba = ba$, $ba\cdot bab=bab$, $b^2\cdot a = b - bab$, $b^2\cdot ab = b^2$, $a^2b\cdot a = a^2$, $a^2b\cdot ab = a^2b$, $bab\cdot a = ba$, $bab\cdot ab = bab$. The Wedderburn decomposition is $U(A^{\omega^{\tiny{1}}_{\tiny{J}}}) = \mathfrak{R} \oplus {\mathbb{Q}}\oplus M_{2\times 2}$ where $\mathfrak{R}$ is the radical of dimension 4. There are two finite dimensional irreducible representations. The original set of generators of the ideal $I$ is a Gröbner basis and consists of the six elements $b^3$, $b^2a + bab - b$, $ba^2$, $ab^2$, $aba + a^2b - a$, $a^3$. Hence $U = U(A^{\omega^{\tiny{1}}_{\tiny{J}}})$ is finite dimensional with the stated basis. The following relations hold in $U$: $b^3 = 0$, $b^2a =- bab + b$, $ba^2 = 0$, $ab^2 = 0$, $aba =- a^2b + a$, $a^3 = 0$. These give the stated structure constants. A basis of the radical $\mathfrak{R} = \mathfrak{R}(U)$ consists of the elements $\xi_1 = b - bab$, $\xi_2 = a^2$, $\xi_3 = b^2$, $\xi_4= a^2b$ which give these relations in $Q = U/\mathfrak{R}$: $ b = bab,\, a^2 = b^2 = a^2b = 0.$ The semisimple quotient $Q$ has dimension $5$ and a basis consists of the cosets of $\eta_1= 1$, $\eta_2= a$, $\eta_3= ab$, $\eta_4= ba$, $\eta_5 = bab$. The center $Z(Q)$ has dimension 2 with basis $z_1=\eta_1$, $z_2=\eta_3 +\eta_4$ and structure constants $z_1\cdot z_1= z_1,\, z_1\cdot z_2 = z_2 \cdot z_1 = z_2,\,z_2 \cdot z_2 = z_2.$ Since $ z^2_2 = z_2$, the minimal polynomial of $z_2$ is $t^2 - t$. Thus $ Z(Q) = J \oplus K$, where $J= \langle z_2-z_1 \rangle$ and $K = \langle z_2\rangle$; both ideals are 1-dimensional. Scaling these basis elements to obtain idempotents gives $e_1 = z_1 - z_2$, $e_2 = z_2$. The corresponding elements of $Q$ are $e_1 = \eta_1 - \eta_3 - \eta_4$, $e_2 = \eta_3 +\eta_4$. The ideals in $Q$ generated by $e_1$ and $e_2$ have dimensions 1 and 4 respectively, and this gives the Wedderburn decomposition. The anti-Jordan family, $q = -1$ {#anti-Jordan} -------------------------------- The structure constants for $A^{ \omega^{\tiny{-1}}_{\tiny{AJ}} }$ are $$[e_1, e_1, e_2] = -e_1, \qquad [e_1, e_2, e_1 ] = e_1, \qquad [e_2, e_1, e_2]= e_2, \qquad [e_2, e_2, e_1] = -e_2.$$ \[AJ-1\] We have the isomorphisms $U\big(A^{ \omega^{\tiny{-1}}_{\tiny{AJ}} }\big) \cong U\big(A^{ \omega^{\tiny{1/2}}_{\tiny{J}} }\big) \cong U\big(A^{ \omega^{\tiny{\infty}}_{\tiny{J}} }\big)$. We have $U = U\big(A^{ \omega^{\tiny{-1}}_{\tiny{AJ}} }\big) = F\langle a, b\rangle / I$ where $I$ is generated by $b^2a - bab + b$ and $aba - a^2b - a$. The first iteration of the Gröbner basis algorithm produces one composition, $bab - b$. Including this element with the original generators, and self-reducing the resulting set, produces a new set of three generators: $b^2a$, $bab - b$, $aba - a^2b - a$. The second iteration produces three compositions: $ba^2b$, $a^2b^2$, $b^2$. Including these elements with the previous generators, and self-reducing the resulting set, produces a new set of four generators: $ba^2b$, $bab - b$, $aba - a^2b - a$, $b^2$. The third iteration produces two compositions: $ba^3b + ba^2$, $a^2b$. Including these elements with the previous generators, and self-reducing the resulting set, produces a new set of five generators: $bab - b$, $ba^2$, $aba - a$, $a^2b$, $b^2$. The fourth iteration produces one composition, $a^2$. Including this element with the previous generators, and self-reducing the resulting set, produces a new set of four generators: $bab - b$, $aba - a$, $b^2$, $a^2$. This is a Gröbner basis for the ideal, and is the same Gröbner basis as for the Jordan cases $q = \infty$, $q = \frac12$; hence the quotient algebras are isomorphic. The anti-Jordan family, $q = 2$ ------------------------------- The structure constants for $A^{\omega^{\tiny{2}}_{\tiny{AJ} }}$ are $$[e_1, e_2, e_1]= e_1, \qquad [e_2, e_1, e_1]= -e_1, \qquad [e_2, e_1, e_2]= e_2, \qquad [e_1, e_2, e_2] = -e_2.$$ We have $U\big(A^{ \omega^{\tiny{2}}_{\tiny{AJ}} }\big)\cong U(A^{ \omega^{\tiny{-1}}_{\tiny{AJ}} })$. Similar to the proof of Theorem \[AJ-1\]. The last nine operations {#fourth} ------------------------ We first consider the fourth family with $q = \infty$. The structure constants for $A^{\omega^{\tiny{\infty}}_{\tiny{F}}}$ are $$\begin{aligned} & [e_1, e_1, e_2] = [e_2, e_1, e_1]= -e_1, \qquad [e_1, e_2, e_1] = e_1, \\ & [e_2, e_2, e_1] = [e_1, e_2, e_2]= -e_2, \qquad [e_2, e_1, e_2] = e_2. \end{aligned}$$ \[fourthinfinity\] We have $U(A^{\omega^{\tiny{\infty}}_{\tiny{F} }}) \cong U(A^{\omega^{\tiny{\infty}}_{\tiny{J}}})$. We have $U\big(A^{ \omega^{\infty}_{\tiny{F}} }\big) = F\langle a, b \rangle / J$ where $J$ is generated by $b^3$, $b^2a - bab + ab^2 + b$, $bab-b$, $ba^2-aba+a^2b+a$, $aba-a$, $a^3$. Self-reducing this set of generators gives the set of generators for the Jordan case, $q =\infty$ (see the proof of Theorem \[Jinf\]). For the fourth family with $q = 0$, the structure constants for $A^{\omega^{\tiny{0}}_{\tiny{F} }}$ are $$[e_1, e_2, e_1] = e_1, \qquad [ e_2, e_1, e_2]= e_2.$$ We have $U( A^{\omega^{\tiny{0}}_{\tiny{F} }} )\cong U( A^{\omega^{\tiny{0}}_{\tiny{J} }})$. Similar to the proof of Proposition \[fourthinfinity\]. For the fourth family with $q = 1$, the structure constants for $A^{\omega^{\tiny{1}}_{\tiny{F} }}$ are $$[e_1, e_2, e_1] = e_1, \qquad [e_2, e_1, e_2] = e_2.$$ We have $U( A^{\omega^{\tiny{1}}_{\tiny{F} }} ) \cong U( A^{\omega^{\tiny{1}}_{\tiny{J} }})$. Similar to the proof of Proposition \[fourthinfinity\]. We consider the last six operations together. Fourth family, $q = -1$: The structure constants for $A^{\omega^{\tiny{-1}}_{\tiny{F} }}$ are $$[e_1, e_2, e_1]= e_1, \qquad [e_2, e_1, e_1] = 2e_1, \qquad [e_2, e_1, e_2]= e_2, \qquad [e_1, e_2, e_2]= 2e_2.$$ Fourth family, $q = 2$: The structure constants for $A^{\omega^{\tiny{2}}_{\tiny{F} }}$ are $$[e_1, e_1, e_2] = 2e_1, \qquad [e_1, e_2, e_1] = e_1, \qquad [e_2, e_2, e_1]= 2e_2, \qquad [e_2, e_1, e_2]=e_2.$$ Fourth family, $q = \tfrac{1}{2}$: The structure constants for $A^{\omega^{\tiny{1/2}}_{\tiny{F} }}$ are $$\begin{aligned} & [e_1, e_1, e_2] = [e_1, e_2, e_1] = [e_2, e_1, e_1] = e_1, \\ & [e_2, e_2, e_1]=[e_2, e_1, e_2] = [e_1, e_2, e_2] = e_2. \end{aligned}$$ Cyclic commutator: The structure constants for $A^{\omega_{\tiny{cc} }}$ are $$[e_1, e_1, e_2] = -e_1, \qquad [e_1, e_2, e_1]= e_1, \qquad [e_2,e_2, e_1] = -e_2, \qquad [e_2, e_1, e_2]= e_2.$$ Weakly commutative operation: The structure constants for $A^{\omega_{wc}}$ are $$\begin{aligned} & [e_1, e_1, e_2] = -e_1, \qquad [e_1, e_2, e_1] = e_1, \qquad [e_2, e_1, e_1] = 2e_1, \\ & [e_2, e_2, e_1] = -e_2, \qquad [e_2, e_1, e_2] = e_2, \qquad [e_1, e_2, e_2] = 2e_2. \end{aligned}$$ Weakly anti-commutative operation: The structure constants for $A^{\omega_{wa}}$ are $$\begin{aligned} & [e_1, e_2, e_1] = [e_1, e_1, e_2]= e_1, \qquad [e_2, e_1, e_1] = -2e_1, \\ & [e_2, e_1, e_2] = [e_2, e_2, e_1] = e_2, \qquad [e_1, e_2, e_2] = -2e_2. \end{aligned}$$ \[F-1\] We have the following isomorphisms: $$U(A^{\omega^{\tiny{-1}}_{\tiny{F} }}) \cong U(A^{\omega^{\tiny{2}}_{\tiny{F} }}) \cong U(A^{\omega^{\tiny{1/2}}_{\tiny{F} }}) \cong U(A^{\omega_{\tiny{cc} }}) \cong U(A^{\omega_{\tiny{wc} }}) \cong U(A^{\omega_{wa}}) \cong U(A^{\omega^{\tiny{\infty}}_{\tiny{J} }}).$$ We have $U(A^{\omega^{\tiny{-1}}_{\tiny{F} }}) = F\langle a, b \rangle / I$ where $I$ is generated by $b^3$, $b^2a+bab+ab^2-b$, $b^2a+\tfrac{1}{2}ab^2$, $bab+\tfrac{1}{2}ab^2-b$, $ba^2+2aba-2a$, $ba^2+aba+a^2b-a$, $ba^2+2a^2b$, $a^3$. We compute a Gröbner basis for $I$. The first iteration produces eight compositions with the normal forms $a^2b^2$, $a^2ba-a^2$, $a^2ba$, $aba-a^2b-a$, $ab^2$, $a^2b$, $b^2$, $a^2$. Including these elements with the original generators, and then self-reducing the resulting set, produces a new set of four ideal generators: $bab-b$, $aba-a$, $a^2$, $b^2$. This is a Gröbner basis $I$. In fact, this is the same Gröbner basis as in the Jordan case, $q = \infty$ (see the proof of Theorem \[Jinf\]). The other cases are similar. Acknowledgements {#acknowledgements .unnumbered} ================ The author thanks the anonymous referee for helpful comments. This work forms part of the author’s doctoral thesis. The author was supported by a Teacher Scholar Doctoral Fellowship from the University of Saskatchewan. She thanks her supervisor, Prof. Murray Bremner, for many suggestions and insights which greatly improved this paper, and Prof. Chris Soteros for the reference to [@vanLintWilson Example 10.3]. [99]{} <span style="font-variant:small-caps;">G. M. Bergman</span>: The diamond lemma for ring theory. *Adv. Math.* 29 (1978) 178–218. <span style="font-variant:small-caps;">S. Bashir</span>: *Automorphisms of Simple Anti-Jordan Pairs.* Ph.D. thesis, University of Ottawa, 2008. <span style="font-variant:small-caps;">G. Benkart, T. Roby</span>: Down-up algebras. *J. Algebra* 209 (1998) 305–344. <span style="font-variant:small-caps;">M. R. Bremner</span>: How to compute the Wedderburn decomposition of a finite dimensional associative algebra. *Groups Complex. Cryptol* 3 (2011) 47–66. <span style="font-variant:small-caps;">M. R. Bremner, L. A. Peresi</span>: Classification of trilinear operations. *Comm. Algebra* 35 (2007) 2932–2959. <span style="font-variant:small-caps;">T. Cassidy, B. Shelton</span>: Basic properties of generalized down-up algebras. *J. Algebra* 279 (2004), no. 1, 402-–421. <span style="font-variant:small-caps;">V. Chari, A. Pressley</span>: *A Guide to Quantum Groups*. Corrected reprint of the 1994 original. Cambridge University Press, Cambridge, 1995. xvi+651 pp. <span style="font-variant:small-caps;">W. A. de Graaf</span>: *Lie Algebras: Theory and Algorithms*. North-Holland, Amsterdam, 2000. <span style="font-variant:small-caps;">H. A. Elgendy, M. R. Bremner</span>: Universal associative envelopes of $(n{+}1)$-dimensional $n$-Lie algebras. *Comm. Algebra* 40 (2012), no. 5, 1827–1842. <span style="font-variant:small-caps;">H. A. Elgendy</span>: *Polynomial Identities and Enveloping Algebras for n-ary Structures.* Ph.D. thesis, University of Saskatchewan, 2012. <span style="font-variant:small-caps;">J. Faraut, S. Kaneyuki, A. Koranyi, Q.-K. Lu, G. Roos</span>: Analysis and Geometry on Complex Homogeneous Domains. Progress in Mathematics, 185. Birkhauser Boston, Inc., Boston, MA, 2000. <span style="font-variant:small-caps;">H. Freudenthal</span>: Lie groups in the foundations of geometry. *Advances in Math*. 1 1964 fasc. 2, 145-190(1964). <span style="font-variant:small-caps;">T. L. Hodge, B. J. Parshall</span>: On the representations theory of Lie triple systems. *Trans. Amer. Math. Soc*. 354(2002), no. 11, 4359-4391. <span style="font-variant:small-caps;">J. E. Humphreys</span>: *Introduction to Lie Algebras and Representation Theory*. Springer, New York, 1972. <span style="font-variant:small-caps;">N. Jacobson</span>: *Structure and Representations of Jordan Algebras*. American Mathematical Society, Providence, 1968. <span style="font-variant:small-caps;">N. Kamiya, S. Okubo</span>: On triple systems and Yang-Baxter equations. *Proceedings of the Seventh International Colloquium on Differential Equations (Plovdiv, 1996)*, 189-–196, VSP, Utrecht, 1997. <span style="font-variant:small-caps;">W. G. Lister</span>: Ternary rings. *Trans. Amer. Math. Soc* 154 (1971) 37–55. <span style="font-variant:small-caps;">O. Loos</span>: Symmetric spaces. I: General theory. *W. A. Benjamin, Inc., New York-Amsterdam* 1969 viii+198 pp. 22.70 (53.00) <span style="font-variant:small-caps;">O. Loos</span>: *Lectures on Jordan Triples*. The University of British Columbia, Vancouver, B.C., 1971. iii+67 pp. <span style="font-variant:small-caps;">K. Meyberg</span>: *Lectures on Algebras and Triple Systems*. Notes on a course of Lectures given during the academic year 1971-1972. The University of Virgina, Charlottesville, Va., 1972. v+226 pp. <span style="font-variant:small-caps;">B. Rosenfeld</span>: *Geometry of Lie Groups*. Mathematics and its Applications, 393. Kluwer Academic Publishers Group, Dordrecht, 1997. xviii+393 pp. <span style="font-variant:small-caps;">J. H. van Lint, R. M. Wilson</span>: *A Course in Combinatorics*. Cambridge University Press, 1992. <span style="font-variant:small-caps;">K. Zhao</span>: Centers of down-up algebras. *J. Algebra* 214 (1999) 103–121.
{ "pile_set_name": "ArXiv" }
--- abstract: 'As a preparation for a mathematically consistent study of the physics of symmetric spacetimes in a noncommutative setting, we study symmetry reductions in deformed gravity. We focus on deformations that are given by a twist of a Lie algebra acting on the spacetime manifold. We derive conditions on those twists that allow a given symmetry reduction. A complete classification of admissible deformations is possible in a class of twists generated by commuting vector fields. As examples, we explicitly construct the families of vector fields that generate twists which are compatible with Friedmann-Robertson-Walker cosmologies and Schwarzschild black holes, respectively. We find nontrivial isotropic twists of FRW cosmologies and nontrivial twists that are compatible with all classical symmetries of black hole solutions.' author: - | Thorsten Ohl [^1]\ Alexander Schenkel [^2]\ \ Institut für Theoretische Physik und Astrophysik\ Universität Würzburg, Am Hubland, 97074 Würzburg, Germany date: October 2008 title: Symmetry Reduction in Twisted Noncommutative Gravity with Applications to Cosmology and Black Holes --- Introduction ============ The study of noncommutative geometry is an active topic in both theoretical physics and mathematics. From the mathematical perspective it is a generalization of classical (commutative) geometry. From the physics perspective it is suggested by the *Gedankenexperiment* of localizing events in spacetime with a Planck scale resolution [@Doplicher:1994zv]. In this *Gedankenexperiment*, a sharp localization induces an uncertainty in the spacetime coordinates, which can naturally be described by a noncommutative spacetime. Furthermore, noncommutative geometry and quantum gravity appear to be connected strongly and one can probably model “low energy” effects of quantum gravity theories using noncommutative geometry. There have been many attempts to formulate scalar, gauge and gravity theories on noncommutative spacetime, in particular using the simplest example of a Moyal-Weyl spacetime having constant noncommutativity between space and time coordinates, see [@Szabo:2001kg; @MullerHoissen:2007xy] for reviews. Furthermore, this framework had been applied to phenomenological particle physics with [@NCSM; @NCSM-Pheno] and without Seiberg-Witten maps (see the review [@NC-Pheno] and references therein), cosmology [@NC-Cosmo] and black hole physics (see the review [@Nicolini:2008aj] and references therein). Our work is based on the approach outlined in [@Aschieri:2005yw; @Aschieri:2005zs; @Aschieri:2006kc], where a noncommutative gravity theory based on an arbitrary twist deformation is established. This approach has the advantages of being formulated using the symmetry principle of deformed diffeomorphisms, being coordinate independent and applicable to nontrivial topologies. However, there is also the disadvantage that it does not match the Seiberg-Witten limit of string theory [@AlvarezGaume:2006bn]. Nevertheless, string theory is not the only candidate for a fundamental theory of quantum gravity. Therefore, the investigation of deformed gravity remains interesting on its own terms and it could very well emerge from a fundamental theory of quantum gravity different from string theory. The outline of this paper is as follows. In section \[sec:basics\] we review the basics of the formalism of twisted noncommutative differential geometry. For more details and the proofs we refer to the original paper [@Aschieri:2005zs] and the review [@Aschieri:2006kc]. We will work with a general twist and do not restrict ourselves to the Moyal-Weyl deformation. In section \[sec:symred\] we will study symmetry reduction in theories based on twisted symmetries, such as the twisted diffeomorphisms in our theory of interest. The reason is that we aim to investigate which deformations of cosmological and black hole symmetries are possible. We will derive the conditions that the twist has to satisfy in order to be compatible with the reduced symmetry. In section \[sec:jambor\] we restrict the twists to the class of Reshetikhin-Jambor-Sykora twists [@Reshetikhin:1990ep; @Jambor:2004kc], that are twists generated by commuting vector fields and are convenient for practical applications. Within this restricted class of twists we can classify more explicitly the possible deformations of Lie algebra symmetries acting on a manifold $\mathcal{M}$. In section \[sec:cosmo\] and \[sec:blackhole\] we apply the formalism to cosmological symmetries as well as the black hole. We classify the possible Reshetikhin-Jambor-Sykora deformations of these models and obtain physically interesting ones. In section \[sec:conc\] we conclude and give an outlook to possible further investigations. In particular possible applications to phenomenological cosmology and black hole physics will be discussed. \[sec:basics\]Basics of Twisted Differential Geometry and Gravity ================================================================= In order to establish notation, we will give a short summary of the framework of twisted differential geometry and gravity. More details can be found in [@Aschieri:2005yw; @Aschieri:2005zs; @Aschieri:2006kc]. There is a quite general procedure for constructing noncommutative spaces and their corresponding symmetries by using a twist. For this we require the following ingredients [@Aschieri:2006kc]: 1. a Lie algebra $\mathfrak{g}$ 2. an action of the Lie algebra on the space we want to deform 3. a twist element $\mathcal{F}$, constructed from the generators of the Lie algebra $\mathfrak{g}$ By a twist element we denote an invertible element of $U\mathfrak{g}\otimes U\mathfrak{g}$, where $U\mathfrak{g}$ is the universal enveloping algebra of $\mathfrak{g}$. $\mathcal{F}$ has to fulfill some conditions, which will be specified later. The basic idea in the following is to combine any bilinear map with the inverse twist and therefore deform these maps. This leads to a mathematically consistent deformed theory covariant under the deformed transformations. We will show this now for the deformation of diffeomorphisms. For our purpose we are interested in the Lie algebra of vector fields $\Xi$ on a manifold $\mathcal{M}$. The transformations induced by $\Xi$ can be seen as infinitesimal diffeomorphisms. A natural action of these transformations on the algebra of tensor fields $\mathcal{T}:=\bigoplus\limits_{n,m} \bigotimes^{n} \Omega \otimes \bigotimes^m \Xi $ is given by the Lie derivative $\mathcal{L}$. $\Omega$ denotes the space of one-forms. In order to deform this Lie algebra, as well as its action on tensor fields and the tensor fields themselves, we first have to construct the enveloping algebra $U\Xi$. This is the associative tensor algebra generated by the elements of $\Xi$ and the unit $1$, modulo the left and right ideals generated by the elements ${[v,w]}-v w + w v$. This algebra can be seen as a Hopf algebra by using the following coproduct $\Delta$, antipode $S$ and counit $\epsilon$ defined on the generators $u\in \Xi$ and $1$ by: [ll]{} (u) = u1 + 1u, & (1) = 11 ,\ (u) = 0, & (1) = 1 ,\ S(u) = -u, & S(1) = 1 . These definitions can be consistently carried over to the whole enveloping algebra demanding $\Delta$ and $\epsilon$ to be algebra homomorphisms and $S$ to be an anti-homomorphism, i.e. for any two elements $\eta,\xi\in U\Xi$, we require &() = ()() ,\ &() = () () ,\ &S()= S() S() . The action of the enveloping algebra on the tensor fields can be defined by extending the Lie derivative \_ () := \_(\_ ()) , ,U ,  . This action is consistent with the Lie algebra properties, since $\mathcal{L}_{{[u,v]}}(\tau) = \mathcal{L}_{uv}(\tau)-\mathcal{L}_{vu}(\tau)$ for all $u,v\in\Xi$ by the properties of the Lie derivative. The extension of the Lie algebra $\Xi$ to the Hopf algebra $(U\Xi,\cdot,\Delta,S,\epsilon)$, where $\cdot$ is the multiplication in $U\Xi$, can now be used in order to construct deformations of it. For the deformations we restrict ourselves to twist deformations, which is a wide class of possible deformations. The reason is that for twist deformations the construction of deformed differential geometry and gravity can be performed explicitly by only using properties of the twist, see [@Aschieri:2005zs]. Other deformations require further investigations. In order to perform the deformation we require a twist element $\mathcal{F}= f^{\alpha}\otimes f_{\alpha}\in U\Xi\otimes U\Xi$ (the sum over $\alpha$ is understood) fulfilling the following conditions &\_[12]{}()=\_[23]{}() ,\ &() = 1 = () ,\ &=11 +() , where $\mathcal{F}_{12}:= \mathcal{F}\otimes 1 $, $\mathcal{F}_{23}:=1\otimes\mathcal{F}$ and $\lambda$ is the deformation parameter. The first condition will assure the associativity of the deformed products, the second will assure that deformed multiplications with unit elements will be trivial and the third condition assures the existence of the undeformed classical limit $\lambda\to 0$. Furthermore, we can assume without loss of generality that $f_\alpha$ (and also $f^\alpha$) are linearly independent for all $\alpha$, what can be assured by combining linearly dependent $f$. Note that $\mathcal{F}$ is regarded as formal power series in $\lambda$, such as the deformation itself. Strict (convergent) deformations will not be regarded here. The simplest example is the twist on $\mathbb{R}^n$ given by $\mathcal{F}_\theta:= \exp{\bigl(-\frac{i\lambda}{2}\theta^{\mu\nu} \partial_\mu\otimes\partial_\nu \bigr)}$ with $\theta^{\mu\nu}=\mathrm{const.}$ and antisymmetric, leading to the Moyal-Weyl deformation, but there are also more complicated ones. From a twist, one can construct the twisted triangular Hopf algebra $(U\Xi_\mathcal{F},\cdot,\Delta_\mathcal{F},S_\mathcal{F},\epsilon_\mathcal{F})$ with $R$-matrix $R:=\mathcal{F}_{21}\mathcal{F}^{-1}=:R^\alpha \otimes R_\alpha$, inverse $R^{-1} =: \bar R^\alpha \otimes \bar R_\alpha = R_{21}$ and &\_():= ()\^[-1]{} ,\_():= () ,S\_():= S()\^[-1]{} , where $\chi := f^\alpha S(f_\alpha)$, $\chi^{-1}:= S(\bar f^\alpha) \bar f_\alpha$ and $\bar f^\alpha \otimes \bar f_\alpha := \mathcal{F}^{-1}$. Furthermore, $\mathcal{F}_{21}:= f_\alpha\otimes f^\alpha$ and $R_{21}:=R_\alpha \otimes R^\alpha$. Again, we can assume without loss of generality that all summands of $\mathcal{F}^{-1}$, $R$ and $R^{-1}$ are linearly independent. However, as explained in [@Aschieri:2005zs], it is simpler to use the triangular $\star$-Hopf algebra $\mathcal{H}_\Xi^\star=(U\Xi_\star,\star,\Delta_\star,S_\star,\epsilon_\star)$, isomorphic to $(U\Xi_\mathcal{F},\cdot,\Delta_\mathcal{F},S_\mathcal{F},\epsilon_\mathcal{F})$. The operations in this algebra on its generators $u,v\in\Xi$ (note that this algebra has the same generators as the classical Hopf algebra) are defined by \[eqn:defstarhopfactions\] &uv := |f\^(u)|f\_(v) ,\ &\_(u) := u1 + X\_[|R\^]{} |R\_(u) ,\ &\_(u):=(u)=0 ,\ &S\_\^[-1]{}(u) := - |R\^(u) X\_[|R\_]{} , where for all $\xi\in U\Xi$ we define $X_\xi := \bar f^\alpha \xi \chi S^{-1}(\bar f_\alpha)$. The action of the twist on the elements of $U\Xi$ is defined by extending the Lie derivative to the adjoint action [@Aschieri:2005zs]. Note that $U\Xi=U\Xi_\star$ as vector spaces. The $R$-matrix is given by $R_\star := X_{R^\alpha} \otimes X_{R_\alpha}$ and is triangular. The coproduct and antipode (\[eqn:defstarhopfactions\]) is defined consistently on $U\Xi_\star$ by using for all $\xi,\eta\in U\Xi_\star$ the definitions \_():= \_()\_() , S\_():= S\_()S\_() . The next step is to define the $\star$-Lie algebra of deformed infinitesimal diffeomorphisms. It has been shown [@Aschieri:2005zs] that for the twist deformation case the choice $(\Xi_\star,{[~,~]}_\star)$, where $\Xi_\star=\Xi$ as vector spaces and [\[u,v\]]{}\_:= [\[|f\^(u),|f\_(v)\]]{} is a natural choice for a $\star$-Lie algebra. It fulfills all conditions which are necessary for a sensible $\star$-Lie algebra given by 1. $\Xi_\star\subset U\Xi_\star$ is a linear space, which generates $U\Xi_\star$ 2. $\Delta_\star(\Xi_\star) \subseteq \Xi_\star\otimes 1 + U\Xi_\star\otimes\Xi_\star$ 3. ${[\Xi_\star,\Xi_\star]}_\star\subseteq\Xi_\star$ The advantage of using the $\star$-Hopf algebra $(U\Xi_\star,\star,\Delta_\star,S_\star,\epsilon_\star)$ instead of the $\mathcal{F}$-Hopf algebra $(U\Xi_\mathcal{F},\cdot,\Delta_\mathcal{F},S_\mathcal{F},\epsilon_\mathcal{F})$ is that the $\star$-Lie algebra of vector fields is isomorphic to $\Xi$ as a vector space. For the $\mathcal{F}$-Hopf algebra this is not the case and the $\mathcal{F}$-Lie algebra consists in general of multidifferential operators. The algebra of tensor fields $\mathcal{T}$ is deformed by using the $\star$-tensor product [@Aschieri:2005zs] \_\^:= |f\^() |f\_(\^) , where as basic ingredients the deformed algebra of functions $A_\star:=(C^\infty(M),\star)$ as well as the $A_\star$-bimodules of vector fields $\Xi_\star$ and one-forms $\Omega_\star$ enter. We call $\mathcal{T}_\star$ the deformed algebra of tensor fields. Note that $\mathcal{T}_\star = \mathcal{T}$ as vector spaces. The action of the deformed infinitesimal diffeomorphisms on $\mathcal{T}_\star$ is defined by the $\star$-Lie derivative \[eqn:starliederivative\] \^\_[u]{}():= \_[|f\^(u)]{}(|f\_()) , \_ , u\_ , which can be extended to all of $U\Xi_\star$ by $\mathcal{L}^\star_{\xi\star\eta}(\tau):= \mathcal{L}^\star_\xi(\mathcal{L}^\star_\eta (\tau)) $. Furthermore, we define the $\star$-pairing $\langle \cdot,\cdot\rangle_\star:\Xi_\star\otimes_\mathbb{C} \Omega_\star\to A_\star$ between vector fields and one-forms as \[eqn:starpairing\] v,\_:= |f\^(v),|f\_() , v\_, \_ , where $\langle \cdot,\cdot\rangle$ is the undeformed pairing. Based on the deformed symmetry principle one can define covariant derivatives, torsion and curvature. This leads to deformed Einstein equations, see [@Aschieri:2005zs], which we do not have to review here, since we do not use them in the following. \[sec:symred\]Symmetry Reduction in Twisted Differential Geometry ================================================================= Assume that we have constructed a deformed gravity theory based on a twist $\mathcal{F}\in U\Xi\otimes U\Xi$. Like in Einstein gravity, the physical applications of this theory is strongly dependent on symmetry reduction. In this section we first define what we mean by symmetry reduction of a theory covariant under a Lie algebraic symmetry (e.g. infinitesimal diffeomorphisms) and then extend the principles to deformed symmetries and $\star$-Lie algebras. In undeformed general relativity we often face the fact that the systems we want to describe have certain (approximate) symmetries. Here we restrict ourselves to Lie group symmetries. For example in cosmology one usually constrains oneself to fields invariant under certain symmetry groups $G$, like e.g. the euclidian group $E_3$ for flat universes or the $SO(4)$ group for universes with topology $\mathbb{R}\times S_3$, where the spatial hypersurfaces are 3-spheres. For a non rotating black hole one usually demands the metric to be stationary and spherically symmetric. Practically, one uses the corresponding Lie algebra $\mathfrak{g}$ of the symmetry group $G$, represents it faithfully on the Lie algebra of vector fields $\Xi$ on the manifold $\mathcal{M}$ and demands the fields $\tau\in\mathcal{T}$, which occur in the theory, to be invariant under these transformations, i.e. we demand \_v()=0 , v . Since the Lie algebra $\mathfrak{g}$ is a linear space we can choose a basis $\lbrace t_i :i=1,\cdots,\mathrm{dim}(\mathfrak{g})\rbrace$ and can equivalently demand \_[t\_i]{}()=0 , i=1,2,,() . The Lie bracket of the generators has to fulfill [\[t\_i,t\_j\]]{}=f\_[ij]{}\^[  k]{}t\_k , where $f_{ij}^{~~k}$ are the structure constants. One can easily show that if we combine two invariant tensors with the tensor product, the resulting tensor is invariant too because of the trivial coproduct \_[t\_i]{}(\^) = \_[t\_i]{}() \^+ \_[t\_i]{}(\^) . The same holds true for pairings $\langle v,\omega\rangle$ of invariant objects $v\in\Xi$ and $\omega\in\Omega$. Furthermore, if a tensor is invariant under infinitesimal transformations, it is also invariant under (at least small) finite transformations, since they are given by exponentiating the generators. The exponentiated generators are part of the enveloping algebra, i.e. $\exp(\alpha^i t_i)\in U\mathfrak{g}$, where $\alpha^i$ are parameters. For large finite transformations the topology of the Lie group can play a role, such that the group elements may not simply be given by exponentiating the generators. In the following we will focus only on small finite transformations in order to avoid topological effects. We now generalize this to the case of $\star$-Hopf algebras and their corresponding $\star$-Lie algebras. Our plan is as follows: we start with a suitable definition of a $\star$-Lie subalgebra constructed from the Lie algebra $(\mathfrak{g},{[~,~]})$. This definition is guided by conditions, which allow for deformed symmetry reduction using infinitesimal transformations. Then we complete this $\star$-Lie subalgebra in several steps to a $\star$-enveloping subalgebra, a $\star$-Hopf subalgebra and a triangular $\star$-Hopf subalgebra. We will always be careful that the dimension of the $\star$-Lie subalgebra remains the same as the dimension of the corresponding classical Lie algebra. At each step we obtain several restrictions between the twist and $(\mathfrak{g},{[~,~]})$. We start by taking the generators $\lbrace t_i\rbrace$ of $\mathfrak{g}\subseteq\Xi$ and representing their deformations in the $\star$-Lie algebra $(\Xi_\star,{[~,~]}_\star)$ as \[eqn:stargenerators\] t\_i\^= t\_i +\_[n=1]{}\^ \^n t\_i\^[(n)]{} , where $\lambda$ is the deformation parameter and $t_i^{(n)}\in\Xi_\star$. The span of these deformed generators, together with the $\star$-Lie bracket, should form a $\star$-Lie subalgebra $(\mathfrak{g}_\star,{[~,~]}_\star) := (\mathrm{span}(t_i^\star),{[~,~]}_\star )$. Therefore $(\mathfrak{g}_\star,{[~,~]}_\star)$ has to obey certain conditions. Natural conditions are \[eqn:infinitesimalconditions\] $$\begin{aligned} \label{eqn:infconda}{[\mathfrak{g_\star},\mathfrak{g}_\star]}_\star \subseteq \mathfrak{g}_\star, \quad &\text{i.e. ${[t_i^\star,t_j^\star]}_\star = f_{ij}^{\star~k}t^\star_k$ with $f_{ij}^{\star~k} = f_{ij}^{~~k} + \mathcal{O}(\lambda)$} \\ \label{eqn:infcondb} \Delta_\star(\mathfrak{g}_\star) \subseteq \mathfrak{g}_\star \otimes 1 + U\Xi_\star \otimes \mathfrak{g}_\star, \quad &\text{which is equivalent to $\bar R_\alpha(\mathfrak{g}_\star)\subseteq \mathfrak{g}_\star~\forall_\alpha$}\end{aligned}$$ The first condition is a basic feature of a $\star$-Lie algebra. The second condition implies that if we have two $\mathfrak{g}_\star$ invariant tensors $\tau,\tau^\prime\in\mathcal{T}_\star$, the $\star$-tensor product of them is invariant as well \^\_[t\_i\^]{}(\_\^) = \^\_[t\_i\^]{}()\_\^+ |R\^()\_\^\_[|R\_(t\_i\^)]{}(\^) =0 , since $\bar R_\alpha(t_i^\star) \in \mathfrak{g}_\star$. The $\star$-pairings $\langle v,\omega\rangle_\star$ of two invariant objects $v\in\Xi_\star$ and $\omega\in\Omega_\star$ are also invariant under the $\star$-action of $\mathfrak{g}_\star$. These are important features if one wants to combine invariant objects to e.g. an invariant action. Furthermore, the conditions are sufficient such that the following consistency relation is fulfilled for any invariant tensor $\tau\in\mathcal{T}_\star$ 0=f\_[ij]{}\^[ k]{} \^\_[t\^\_k]{}() = \^\_[[\[t\^\_i,t\^\_j\]]{}]{}()= \^\_[t\^\_[i]{}]{}(\^\_[t\^\_j]{}()) - \^\_[|R\^(t\^\_[j]{})]{}(\^\_[|R\_(t\^\_i)]{}()) , since $\bar R_\alpha(t_i^\star) \in \mathfrak{g}_\star$. Hence by demanding the two conditions (\[eqn:infinitesimalconditions\]) for the $\star$-Lie subalgebra $(\mathrm{span}(t_i^\star),{[~,~]}_\star )$ we can consistently perform symmetry reduction by using deformed [*infinitesimal*]{} transformations. In the classical limit $\lambda\to 0$ we obtain the classical Lie algebra $(\mathfrak{g}_\star,{[~,~]}_\star ) \stackrel{\lambda\to 0}{\longrightarrow} (\mathfrak{g},{[~,~]})$. Next, we consider the extension of the $\star$-Lie subalgebra $(\mathfrak{g}_\star,{[~,~]}_\star)\subseteq(\Xi_\star,{[~,~]}_\star)$ to the triangular $\star$-Hopf subalgebra $\mathcal{H}_\mathfrak{g}^\star=(U\mathfrak{g}_\star,\star,\Delta_\star,S_\star,\epsilon_\star)\subseteq\mathcal{H}_\Xi^\star$. This can be seen as extending the infinitesimal transformations to a quantum group. We will divide this path into several steps, where in every step we have to demand additional restrictions on the twist. Firstly, we construct the $\star$-tensor algebra generated by the elements of $\mathfrak{g}_\star$ and $1$. We take this tensor algebra modulo the left and right ideals generated by the elements ${[u,v]}_\star - u\star v + \bar R^\alpha(v)\star \bar R_\alpha(u)$. It is necessary that these elements are part of $U\mathfrak{g}_\star$, i.e. we require \[eqn:envelopcond\] |R\^(\_)|R\_(\_)U\_ . This leads to the algebra $(U\mathfrak{g}_\star,\star)$, which is a subalgebra of $(U\Xi_\star,\star)$. Secondly, we extend this subalgebra to a $\star$-Hopf subalgebra. Therefore we additionally have to require that \[eqn:hopfconditions\] $$\begin{aligned} \label{eqn:hopfconditions1}\Delta_\star(U\mathfrak{g}_\star) &\subseteq U\mathfrak{g}_\star \otimes U\mathfrak{g}_\star~,\\ \label{eqn:hopfconditions2}S_\star(U\mathfrak{g}_\star)&\subseteq U\mathfrak{g}_\star~.\end{aligned}$$ Note that we do not demand that $S^{-1}_\star$ (defined on $U\Xi_\star$) closes in $U\mathfrak{g}_\star$, since this is in general not the case for a nonquasitriangular Hopf algebra and we do not want to demand quasitriangularity at this stage. Then the $\star$-Hopf algebra $\mathcal{H}_\mathfrak{g}^\star$ is a Hopf subalgebra of $\mathcal{H}_\Xi^\star$. Thirdly, we additionally demand that there exists an $R$-matrix $R_\star\in U\mathfrak{g}_\star\otimes U\mathfrak{g}_\star$. It is natural to take the $R$-matrix of the triangular $\star$-Hopf algebra $\mathcal{H}_\Xi^\star$ defined by $R_\star := X_{R^\alpha}\otimes X_{R_\alpha}$. This leads to the restrictions \[eqn:triangularcond\] X\_[R\^]{},X\_[R\_]{}U\_ , \_. Since $R_\star$ is triangular, i.e. $R_\star^{-1 }=\bar R_\star^\alpha \otimes \bar R_{\star\alpha} = R_{\star21} = R_{\star\alpha} \otimes R_\star^\alpha$, we also have $X_{\bar R^\alpha},X_{\bar R_\alpha}\in U\mathfrak{g}_\star~,~\forall_\alpha$. If these conditions are fulfilled, $\mathcal{H}_\mathfrak{g}^\star$ is a triangular $\star$-Hopf subalgebra of $\mathcal{H}_\Xi^\star$ with the same $R$-matrix. As we have seen, extending the $\star$-Lie subalgebra to a (triangular) $\star$-Hopf subalgebra gives severe restrictions on the possible deformations, more than just working with the deformed infinitesimal transformations given by a $\star$-Lie subalgebra or the finite transformations given by the $\star$-enveloping subalgebra $(U\mathfrak{g}_\star,\star)$. Now the question arises if we actually require the deformed finite transformations to form a (triangular) $\star$-Hopf algebra in order to use them for a sensible symmetry reduction. Because $(U\mathfrak{g}_\star,\star)$ describes deformed finite transformations and we have the relation \[eqn:inf-fin\] \^\_[U\_\\1]{}()=0\^\_[\_]{}()=0 , we can consistently demand tensors to be invariant under $(U\mathfrak{g}_\star,\star)$, since we require tensors to be invariant under $(\mathfrak{g}_\star,{[~,~]}_\star)$. Therefore, a well defined $(U\mathfrak{g}_\star,\star)$ leads to a structure sufficient for symmetry reduction. The equivalence (\[eqn:inf-fin\]) can be shown by using linearity of the $\star$-Lie derivative and the property $\mathcal{L}^\star_{\xi\star\eta}(\tau)=\mathcal{L}^\star_{\xi}(\mathcal{L}^\star_\eta(\tau))$. In order to better understand the different restrictions necessary for constructing the $\star$-Lie subalgebra $(\mathfrak{g}_\star,{[~,~]}_\star)$, the $\star$-enveloping subalgebra and the (triangular) $\star$-Hopf subalgebra $(U\mathfrak{g}_\star,\star,\Delta_\star,S_\star,\epsilon_\star)$, we restrict ourselves in the following sections to the class of Reshetikhin-Jambor-Sykora twists [@Reshetikhin:1990ep; @Jambor:2004kc]. This is a suitable nontrivial generalization of the Moyal-Weyl product, also containing e.g. $\kappa$ and $q$ deformations when applied to Poincaré symmetry. \[sec:jambor\]The Case of Reshetikhin-Jambor-Sykora Twists ========================================================== Let $\lbrace V_a\in\Xi\rbrace$ be an arbitrary set of mutually commuting vector fields, i.e. ${[V_a,V_b]}=0~,~\forall_{a,b}$, on an $n$ dimensional manifold $\mathcal{M}$. Then the object \[eqn:jstwist\] \_[V]{} := (- \^[ab]{} V\_aV\_b )UU is a twist element, if $\theta$ is constant and antisymmetric [@Aschieri:2005zs; @Reshetikhin:1990ep; @Jambor:2004kc]. We call (\[eqn:jstwist\]) a Reshetikhin-Jambor-Sykora twist. Note that this twist is not restricted to the topology $\mathbb{R}^n$ for the manifold $\mathcal{M}$. Furthermore, we can restrict ourselves to $\theta$ with maximal rank and an even number of vector fields $V_a$, since we can lower the rank of the Poisson structure afterwards by choosing some of the $V_a$ to be zero. We can therefore without loss of generality use the standard form \[eqn:theta\] = 0 & 1 & 0 & 0 &\ -1 & 0 & 0 & 0 &\ 0 & 0 & 0 & 1 &\ 0 & 0 & -1 & 0 &\ &&&& by applying a suitable $GL(n)$ transformation on the $V_a$. This twist element is easy to apply and in particular we have for the inverse and the $R$-matrix \[eqn:jsrmatrix\] \_V\^[-1]{} = ( \^[ab]{} V\_aV\_b ) , R=\_[V,21]{}\_V\^[-1]{}=\_V\^[-2]{}=(i \^[ab]{} V\_aV\_b ) . Now let $(\mathfrak{g},{[~,~]})\subseteq(\Xi,{[~,~]})$ be the Lie algebra of the symmetry we want to deform. We choose a basis of this Lie algebra $\lbrace t_i : i=1,\cdots ,\mathrm{dim}(\mathfrak{g})\rbrace$ with ${[t_i,t_j]}=f_{ij}^{~~k}t_k$. Next, we discuss the symmetry reduction based on the $\star$-Lie subalgebra, as explained in section \[sec:symred\]. Therefore we make the ansatz (\[eqn:stargenerators\]) for the generators $t_i^\star$. Furthermore, we evaluate the two conditions (\[eqn:infinitesimalconditions\]) the $t_i^\star$ have to satisfy. We start with the coproduct condition (\[eqn:infcondb\]), which is equivalent to $\bar R_\alpha(t_i^\star)\in \mathrm{span}(t_i^\star),~ \forall_\alpha$, where $\alpha$ is a multi index. Using the explicit form of the inverse $R$-matrix (\[eqn:jsrmatrix\]) we arrive at the conditions \[eqn:condition1\] [\[V\_[a\_1]{},\]]{} = \_[a\_1a\_n i]{}\^[j]{} t\_j\^ , where $\mathcal{N}_{a_1\cdots a_n i}^{\star j} :=\mathcal{N}_{a_1\cdots a_n i}^j + \sum\limits_{k=1}^{\infty} \lambda^k \mathcal{N}_{a_1\cdots a_n i}^{(k)~j}$ are constants. The only independent condition in (\[eqn:condition1\]) is given by \[eqn:condition2\] [\[V\_a,t\_i\^\]]{} = \^[j]{}\_[a i]{} t\^\_j  , since it implies all the other ones by linearity. In particular, the zeroth order in $\lambda$ of (\[eqn:condition2\]) yields \[eqn:condition3\] [\[V\_a,t\_i\]]{} = \_[ai]{}\^[j]{} t\_j . This leads to the following \[propo:ideal\] Let $(\mathfrak{g},{[~,~]})\subseteq(\Xi,{[~,~]})$ be a classical Lie algebra and $(\Xi_\star,{[~,~]}_\star)$ the $\star$-Lie algebra of vector fields deformed by a Reshetikhin-Jambor-Sykora twist, constructed with vector fields $V_a$. Then for a symmetry reduction respecting the minimal axioms (\[eqn:infinitesimalconditions\]), it is necessary that the following Lie bracket relations hold true [\[V\_a,\]]{} ,\_a . In other words, $(\mathrm{span}(t_i,V_a),{[~,~]})\subseteq(\Xi,{[~,~]})$ forms a Lie algebra with ideal $\mathfrak{g}$. Here $t_i$ are the generators of $\mathfrak{g}$. Note that this gives conditions relating the [*classical*]{} Lie algebra $(\mathfrak{g},{[~,~]})$ with the twist. Next, we evaluate the $\star$-Lie bracket condition (\[eqn:infconda\]). Using the explicit form of the inverse twist (\[eqn:jsrmatrix\]) and (\[eqn:condition2\]) we obtain &|f\_[\_[(n)]{}]{}(t\_i\^) = [\[V\_[a\_1]{},\]]{} = (\^\_[a\_n]{}\^\_[a\_1]{} )\_i\^j t\_j\^=:\^[j]{}\_[\_[(n)]{}i]{} t\_j\^ ,\ &|f\^[\_[(n)]{}]{}(t\_i\^) = \^[\_[(n)]{}\_[(n)]{}]{} |f\_[\_[(n)]{}]{}(t\_i\^) ,\ &\^[\_[(n)]{}\_[(n)]{}]{} := ()\^n \^[b\_1 a\_1]{} \^[b\_n a\_n]{} , where $\alpha_{(n)},\beta_{(n)}$ are multi indices. This leads to \[eqn:prestarcommutator\] [\[t\_i\^,t\_j\^\]]{}\_= \^[\_[(n)]{}\_[(n)]{}]{} \^[k]{}\_[\_[(n)]{} i]{} \^[l]{}\_[\_[(n)]{}j]{}[\[t\_k\^,t\_l\^\]]{} . Note that in particular for the choice $t_i^\star = t_i, ~\forall_i,$ the $\star$-Lie subalgebra closes with structure constants \[eqn:starliealgebra\] [\[t\_i,t\_j\]]{}\_= \^[\_[(n)]{}\_[(n)]{}]{} \^[ k]{}\_[\_[(n)]{} i]{} \^[ l]{}\_[\_[(n)]{}j]{}[\[t\_k,t\_l\]]{} = \^[\_[(n)]{}\_[(n)]{}]{} \^[ k]{}\_[\_[(n)]{} i]{} \^[ l]{}\_[\_[(n)]{}j]{} f\_[kl]{}\^[  m]{} t\_m =: f\_[ij]{}\^[ m]{} t\_m , where we have used the $\mathcal{N}$ defined in (\[eqn:condition3\]). This leads to the following \[propo:starliealgebra\] Let ${[V_a,\mathfrak{g}_\star]}\subseteq\mathfrak{g}_\star~,\forall_a$. Then we can always construct a $\star$-Lie subalgebra $(\mathfrak{g}_\star,{[~,~]}_\star)\subseteq(\Xi_\star,{[~,~]}_\star)$ by choosing the generators as $t_i^\star = t_i$ for all $i$. With this we have $\mathfrak{g}_\star = \mathfrak{g}$ as vector spaces and the structure constants are deformed as f\_[ij]{}\^[ m]{}=\^[\_[(n)]{}\_[(n)]{}]{} \^[ k]{}\_[\_[(n)]{} i]{} \^[ l]{}\_[\_[(n)]{}j]{} f\_[kl]{}\^[  m]{} . Since the condition (\[eqn:infcondb\]) together with the requirement $t_i^\star=t_i$, for all $i$, automatically fulfills (\[eqn:infconda\]), we choose $t_i^\star=t_i$, for all $i$, as a canonical embedding. In general, other possible embeddings require further constructions to fulfill condition (\[eqn:infconda\]) and are therefore less natural. We will discuss possible differences between this and other embeddings later on, when we construct the $\star$-Hopf subalgebra and the $\star$-Lie derivative action on $\star$-tensor fields. In addition, we obtain that the necessary condition (\[eqn:envelopcond\]) for extending $\mathfrak{g}_\star$ to the $\star$-enveloping subalgebra $(U\mathfrak{g}_\star,\star)\subseteq(U\Xi_\star,\star)$ is automatically fulfilled, since we have $\bar R_{\alpha_{(n)}}(\mathfrak{g}_\star)\subseteq \mathfrak{g}_\star$ for all $\alpha_{(n)}$ and additionally |R\^[\_[(n)]{}]{}(\_)=(-2)\^n \^[\_[(n)]{}\_[(n)]{}]{}|R\_[\_[(n)]{}]{}(\_)\_ , \_[(n)]{}. Next, we evaluate the conditions (\[eqn:hopfconditions\]), which have to be fulfilled in order to construct the $\star$-Hopf subalgebra $\mathcal{H}_\mathfrak{g}^\star\subseteq\mathcal{H}_{\Xi}^\star$. For the particular choice of the twist (\[eqn:jstwist\]) we obtain the following \[propo:starhopf\] Let $(U\mathfrak{g}_\star,\star)\subseteq(U\Xi_\star,\star)$ be a $\star$-enveloping subalgebra and let the deformation parameter $\lambda\neq 0$. Then in order to extend $(U\mathfrak{g}_\star,\star)$ to the $\star$-Hopf subalgebra $\mathcal{H}_\mathfrak{g}^\star=(U\mathfrak{g}_\star,\star,\Delta_\star,S_\star,\epsilon_\star)\subseteq\mathcal{H}_\Xi^\star$ the condition V\_[a\_1]{} \_ ,  [\[V\_[a\_2]{},\_\]]{}0 has to hold true for all pairs of indices $(a_1,a_2)$ connected by the antisymmetric matrix $\theta$ (\[eqn:theta\]), i.e. $(a_1,a_2)\in \big\lbrace(1,2),(2,1),(3,4),(4,3),\dots\big\rbrace$. Note that these conditions depend on the embedding $t_i^\star=t_i^\star(t_j)$. The proof of this proposition is shown in the appendix \[app:proof\]. Finally, if we demand $\mathcal{H}_\mathfrak{g}^\star$ to be a triangular $\star$-Hopf algebra (\[eqn:triangularcond\]) we obtain the stringent condition V\_a \_ , \_a . This can be shown by using $X_{R_{\alpha}} = R_\alpha$ and $V_a \star V_b = V_a V_b$, which holds true for the class of Reshetikhin-Jambor-Sykora twists. As we have seen above, there are much stronger restrictions on the Lie algebra $(\mathfrak{g},{[~,~]})$ and the twist, if we want to extend the deformed infinitesimal transformations $(\mathfrak{g}_\star,{[~,~]}_\star)$ to the (triangular) $\star$-Hopf subalgebra $\mathcal{H}_\mathfrak{g}^\star$. In particular this extension restricts the $V_a$ themselves, while for infinitesimal transformations and the finite transformations $(U\mathfrak{g}_\star,\star)$ only the images of $V_a$ acting on $\mathfrak{g}_\star$ are important. Next, we study the $\star$-action of the $\star$-Lie and Hopf algebra on the deformed tensor fields. The $\star$-action of the generators $t_i^\star$ on $\tau\in\mathcal{T}_\star$ is defined by (\[eqn:starliederivative\]) and simplifies to \^\_[t\_i\^]{}() = \^[\_[(n)]{}\_[(n)]{}]{} \^[j]{}\_[\_[(n)]{}i]{}  \_[t\_j\^]{} (|f\_[\_[(n)]{}]{}()) . For invariant tensors, the $\star$-Lie derivative has to vanish to all orders in $\lambda$, since we work with formal power series. If we now for explicitness take the natural choice $t_i^\star=t_i$ we obtain the following \[propo:starinvariance\] Let ${[V_a,\mathfrak{g}_\star]}\subseteq\mathfrak{g}_\star~,\forall_a$ and $t_i^\star=t_i,~\forall_i$. Then a tensor $\tau\in\mathcal{T}_\star$ is $\star$-invariant under $(\mathfrak{g}_\star,{[~,~]}_\star)$, if and only if it is invariant under the undeformed action of $(\mathfrak{g},{[~,~]})$, i.e. \^\_[\_]{}()=0  \_()=0 . For the proof we make the ansatz $\tau=\sum\limits_{n=0}^\infty \lambda^n \tau_{n}$ and investigate $\mathcal{L}^\star_{t_i}(\tau)$ order by order in $\lambda$, since we work with formal power series. By using (\[eqn:condition3\]) to reorder the Lie derivatives such that $t_i$ is moved to the right, it can be shown recursively in powers of $\lambda$ that the proposition holds true. Note that for $t_i^\star\neq t_i$ this does not necessarily hold true. We can not make statements for this case, since we would require a general solution of (\[eqn:prestarcommutator\]), which we do not have yet. But we mention again that we consider choosing $t_i^\star$ different from $t_i$ quite unnatural. This proposition translates to the case of finite symmetry transformations with $t_i^\star=t_i$ because of the properties of the $\star$-Lie derivative. The framework developed in this section will now be applied to cosmology and black holes in order to give some specific examples and discuss possible physical implications. \[sec:cosmo\]Application to Cosmology ===================================== In this section we will investigate models with symmetry group $E_3$ in four spacetime dimensions with topology $\mathbb{R}^4$. These are flat Friedmann-Robertson-Walker (FRW) universes. The undeformed Lie algebra of this group is generated by the “momenta” $p_i$ and “angular momenta” $L_i$, $i\in\lbrace1,2,3\rbrace$, which we can represent in the Lie algebra of vector fields as p\_i = \_i,L\_i = \_[ijk]{}x\^j \_k , where $\epsilon_{ijk}$ is the Levi-Civita symbol. The undeformed Lie bracket relations are [\[p\_i,p\_j\]]{}=0 ,[\[p\_i,L\_j\]]{}=-\_[ijk]{}p\_k ,[\[L\_i,L\_j\]]{}=-\_[ijk]{}L\_k . We will work with the natural embedding $t_i^\star=t_i$, and therefore the $\star$-Lie subalgebra is given by $\mathfrak{g}_\star=\mathfrak{e}_{3\star}=\mathfrak{e}_3=\mathrm{span}(p_i,L_i)$. We can now explicitly evaluate the condition each twist vector field $V_a$ has to satisfy given by ${[V_a,\mathfrak{e}_{3}]}\subseteq\mathfrak{e}_{3}$ (cf. proposition \[propo:ideal\]). Since the generators are at most linear in the spatial coordinates, $V_a$ can be at most quadratic in order to fulfill this condition. If we make a quadratic ansatz with time dependent coefficients we obtain that each $V_a$ has to be of the form \[eqn:FRWV\] V\_a = V\_a\^0(t) \_t + c\_a\^i\_i + d\_a\^i L\_i + f\_a x\^i\_i , where $c_a^i,~d_a^i,~f_a\in\mathbb{R}$ and $V_a^0(t)\in C^\infty(\mathbb{R})$ in order to obtain hermitian deformations. If all $V_a$ have the form (\[eqn:FRWV\]), the $\star$-Lie algebra closes (cf. proposition \[propo:starliealgebra\]). Next, we have to find conditions such that the $V_a$ are mutually commuting. A brief calculation shows that the following conditions have to be fulfilled: \[eqn:frwcond\] $$\begin{aligned} \label{eqn:frwcond1}&d_a^i d_b^j \epsilon_{ijk} = 0 ~,\forall_k~,\\ \label{eqn:frwcond2}&c_a^i d_b^j \epsilon_{ijk} - c_b^i d_a^j \epsilon_{ijk} + f_a c_b^k - f_b c_a^k =0~,\forall_k ~,\\ \label{eqn:frwcond3}&{[V_a^0(t)\partial_t,V_b^0(t)\partial_t]}=0~.\end{aligned}$$ As a first step, we will now work out all possible deformations of $\mathfrak{e}_{3}$ when twisted with two commuting vector fields. We will classify the possible solutions. Therefore we divide the solutions into classes depending on the value of $d_a^i$ and $f_a$. We use as notation for our cosmologies $\mathfrak{C}_{AB}$, where $A\in\lbrace1,2,3\rbrace$ and $B\in\lbrace1,2\rbrace$, which will become clear later on, when we sum up the results in table \[tab:frw\]. Type $\mathfrak{C}_{11}$ is defined to be vector fields with $d_1^i=d_2^i=0$ and $f_1=f_2=0$, i.e.  V\_[1(\_[11]{})]{} = V\_[1]{}\^0(t)\_t +c\_[1]{}\^i\_i ,V\_[2(\_[11]{})]{} = V\_[2]{}\^0(t)\_t +c\_[2]{}\^i\_i . These vector fields fulfill the first two conditions (\[eqn:frwcond1\]) and (\[eqn:frwcond2\]). The solutions of the third condition (\[eqn:frwcond3\]) will be discussed later, since this classification we perform now does not depend on it. Type $\mathfrak{C}_{21}$ is defined to be vector fields with $d_1^i=d_2^i=0$, $f_1\neq0$ and $f_2=0$. The first condition (\[eqn:frwcond1\]) is trivially fulfilled and the second (\[eqn:frwcond2\]) is fulfilled, if and only if $c_{2}^i=0,~\forall_i$, i.e. type $\mathfrak{C}_{21}$ is given by the vector fields V\_[1(\_[21]{})]{} = V\_[1]{}\^0(t)\_t + c\_[1]{}\^i\_i + f\_[1]{} x\^i\_i ,V\_[2(\_[21]{})]{} = V\_[2]{}\^0(t)\_t . These vector fields can be simplified to V\_[1(\_[21]{})]{} = c\_[1]{}\^i\_i + f\_[1]{} x\^i\_i ,V\_[2(\_[21]{})]{} = V\_[2]{}\^0(t)\_t , since both lead to the same twist (\[eqn:jstwist\]). Solutions with $d_1^i=d_2^i=0$, $f_1\neq0$ and $f_2\neq0$ lie in type $\mathfrak{C}_{21}$, since we can perform the twist conserving map $V_2\to V_2 - \frac{f_2}{f_1} V_1$, which transforms $f_2$ to zero. Furthermore $\mathfrak{C}_{31}$ is defined by $d_1^i=d_2^i=0$, $f_1=0$ and $f_2\neq0$ and is equivalent to $\mathfrak{C}_{21}$ by interchanging the labels of the vector fields. Next, we go on to solutions with without loss of generality $\mathbf{d}_1\neq0$ and $\mathbf{d}_2=0$ ($\mathbf{d}$ denotes the vector). Note that this class contains also the class with $\mathbf{d}_1\neq0$ and $\mathbf{d}_2\neq0$. To see this, we use the first condition (\[eqn:frwcond1\]) and obtain that $\mathbf{d}_1$ and $\mathbf{d}_2$ have to be parallel, i.e. $d^i_2=\kappa d^i_1$. Then we can transform $\mathbf{d}_2$ to zero by using the twist conserving map $V_2\to V_2 - \kappa V_1$. Type $\mathfrak{C}_{12}$ is defined to be vector fields with $\mathbf{d}_1\neq0$, $\mathbf{d}_2=0$ and $f_1=f_2=0$. The first condition (\[eqn:frwcond1\]) is trivially fulfilled, while the second condition (\[eqn:frwcond2\]) requires that $\mathbf{c}_2$ is parallel to $\mathbf{d}_1$, i.e. we obtain V\_[1(\_[12]{})]{} = V\_[1]{}\^0(t)\_t + c\_[1]{}\^i\_i + d\_[1]{}\^i L\_i ,V\_[2(\_[12]{})]{} = V\_[2]{}\^0(t)\_t +  d\_[1]{}\^i\_i , where $\kappa\in\mathbb{R}$ is a constant. Type $\mathfrak{C}_{22}$ is defined to be vector fields with $\mathbf{d}_1\neq0$, $\mathbf{d}_2=0$, $f_1\neq0$ and $f_2=0$. Solving the second condition (\[eqn:frwcond2\]) (therefore we have to use that the vectors are real!) we obtain V\_[1(\_[22]{})]{} = c\_[1]{}\^i\_i + d\_[1]{}\^i L\_i + f\_[1]{} x\^i\_i ,V\_[2(\_[22]{})]{} = V\_[2]{}\^0(t)\_t , where we could set without loss of generality $V_{1}^0(t)$ to zero, as in type $\mathfrak{C}_{21}$. Note that $\mathfrak{C}_{21}$ is contained in $\mathfrak{C}_{22}$ by violating the condition $\mathbf{d}_1\neq0$. Finally, we come to the last class, type $\mathfrak{C}_{32}$, defined by $\mathbf{d}_1\neq0$, $\mathbf{d}_2=0$, $f_1=0$ and $f_2\neq0$. This class contains also the case $\mathbf{d}_1\neq0$, $\mathbf{d}_2=0$, $f_1\neq0$ and $f_2\neq0$ by using the twist conserving map $V_1\to V_1 - \frac{f_1}{f_2} V_2$. The vector fields are given by V\_[1(\_[32]{})]{} = V\_[1]{}\^0(t)\_t + \_i + d\_[1]{}\^i L\_i  ,V\_[2(\_[32]{})]{} = V\_[2]{}\^0(t)\_t + c\_[2]{}\^i\_i + f\_[2]{} x\^i\_i . Note that type $\mathfrak{C}_{11}$ and $\mathfrak{C}_{12}$ can be extended to a triangular $\star$-Hopf subalgebra by choosing $V_1^0(t)=V_2^0(t)=0$ in each case. For a better overview we additionally present the the results in table \[tab:frw\], containing all possible two vector field deformations $\mathfrak{C}_{AB}$ of the Lie algebra of the euclidian group. From this table the notation $\mathfrak{C}_{AB}$ becomes clear. [$~~\mathfrak{C}_{AB}$]{} $\mathbf{d}_1=\mathbf{d}_2=0$ $\mathbf{d}_1\neq0$ , $\mathbf{d}_2=0$ --------------------------- ------------------------------------------------ ---------------------------------------------------------------------------------------------------- $f_1=0$, $V_1= V_{1}^0(t)\partial_t +c_{1}^i\partial_i$ $V_1= V_{1}^0(t)\partial_t + c_{1}^i\partial_i + d_{1}^i L_i$ $f_2=0$ $V_2= V_{2}^0(t)\partial_t +c_{2}^i\partial_i$ $V_2= V_{2}^0(t)\partial_t + \kappa~d_{1}^i\partial_i$ $f_1\neq0$, $V_1= c_{1}^i\partial_i + f_{1} x^i\partial_i$ $V_1= c_{1}^i\partial_i + d_{1}^i L_i + f_{1} x^i\partial_i$ $f_2=0$ $V_2= V_{2}^0(t)\partial_t$ $V_2= V_{2}^0(t)\partial_t$ $f_1=0$, $V_1= V_{1}^0(t)\partial_t$ $V_1= V_{1}^0(t)\partial_t + \frac{1}{f_2}d_{1}^j c_{2}^k\epsilon_{jki} \partial_i + d_{1}^i L_i$ $f_2\neq0$ $V_2= c_{2}^i\partial_i + f_{2} x^i\partial_i$ $V_2= V_{2}^0(t)\partial_t + c_{2}^i\partial_i + f_{2} x^i\partial_i$ : \[tab:frw\]Two vector field deformations of the cosmological symmetry group $E_3$. Next, we discuss solutions to the third condition (\[eqn:frwcond3\]) ${[V_1^0(t)\partial_t,V_2^0(t)\partial_t]}=0$. It is obvious that choosing either $V_1^0(t)=0$ or $V_2^0(t)=0$ and the other one arbitrary is a solution. Additionally, we consider solutions with $V_1^0(t)\neq0$ and $V_2^0(t)\neq0$. Therefore there has to be some point $t_0\in\mathbb{R}$, such that without loss of generality $V_1^0(t)$ is unequal zero in some open region $U\subseteq\mathbb{R}$ around $t_0$. In this region we can perform the diffeomorphism $t\to \tilde t(t):= \int\limits_{t_0}^t dt^\prime \frac{1}{V_1^0(t^\prime)}$ leading to $\tilde V_1^0(\tilde t)=1$. With this the third condition (\[eqn:frwcond3\]) becomes 0=[\[V\_1\^0(t)\_t,V\_2\^0(t)\_t\]]{}=[\[V\_1\^0(t)\_[t]{},V\_2\^0(t)\_[t]{}\]]{} = (\_[t]{}V\_2\^0(t)) \_[t]{} . This condition is solved if and only if $\tilde V_2^0(\tilde t) =\mathrm{const.}$ for $t\in U\subseteq\mathbb{R}$. For the subset of analytical functions $C^\omega(\mathbb{R})\subset C^\infty(\mathbb{R})$ we can continue this condition to all $\mathbb{R}$ and obtain the global relation $V_2^0(t)= \kappa V_1^0(t)$, with some constant $\kappa\in\mathbb{R}$. For non analytic, but smooth functions, we can not continue these relations to all $\mathbb{R}$ and therefore only obtain local conditions restricting the functions in the overlap of their supports to be linearly dependent. In particular non analytic functions with disjoint supports fulfill the condition (\[eqn:frwcond3\]) trivially. After characterizing the possible two vector field deformations of $\mathfrak{e}_3$ we briefly give a method how to obtain twists generated by a larger number of vector fields. For this purpose we use the canonical form of $\theta$ (\[eqn:theta\]). Assume that we want to obtain deformations with e.g. four vector fields. Then of course all vector fields have to be of the form (\[eqn:FRWV\]). According to the form of $\theta$ we have two blocks of vector fields $(a,b)=(1,2)$ and $(a,b)=(3,4)$, in which the classification described above for two vector fields can be performed. This means that all four vector field twists can be obtained by using two types of two vector field twists. We label the twist by using a tuple of types, e.g. $(\mathfrak{C}_{11},\mathfrak{C}_{22})$ means that $V_1,V_2$ are of type $\mathfrak{C}_{11}$ and $V_3,V_4$ of type $\mathfrak{C}_{22}$. But this does only assure that ${[V_a,V_b]}=0$ for $(a,b)\in\lbrace(1,2),(3,4)\rbrace$ and we have to demand further restrictions in order to fulfill ${[V_a,V_b]}=0$ for all $(a,b)$ and that all vector fields give independent contributions to the twist. In particular twists constructed with linearly dependent vector fields can be reduced to a twist constructed by a lower number of vector fields. This method naturally extends to a larger number of vector fields, until we cannot find anymore independent and mutually commuting vector fields. We will now give two examples for the $\mathfrak{e}_3$ case in order to clarify the construction. As a first example we construct the four vector field twist $(\mathfrak{C}_{11},\mathfrak{C}_{11})$. In this case all four vector fields commute without imprinting further restrictions. We assume that three of the four vectors $\mathbf{c}_a$ are linearly independent, such that the fourth one, say $\mathbf{c}_4$, can be decomposed into the other ones. If we now choose four linearly independent functions $V_a^0(t)$ (this means that they are non analytic) leads to a proper four vector field twist. As a second simple example we construct the four vector field twist $(\mathfrak{C}_{21},\mathfrak{C}_{21})$. In order to have commuting vector fields we obtain the condition $c_3^i=\frac{f_3}{f_1}c_1^i$. We therefore have $V_3=\frac{f_3}{f_1}V_1$ and the four vector field twist can be reduced to the two vector field twist of type $\mathfrak{C}_{21}$ with $\tilde V_1 =V_1$ and $\tilde V_2 = V_2 +\frac{f_3}{f_1}V_4$. This is an example of an improper four vector field twist. This method can be applied in order to investigate general combinations of two vector field twists, if one requires them. Because this construction is straightforward and we do not require these twists for our discussions, we do not present them here. At the end we calculate the $\star$-commutator of the linear coordinate functions $x^\mu\in A_\star$ for the various types of models in first order in the deformation parameter $\lambda$. It is given by c\^:=[\[x\^x\^\]]{} := x\^x\^-x\^x\^= i\^[ab]{} V\_a(x\^) V\_b(x\^) +(\^2) . The results are given in appendix \[app:starcom\] and show that these commutators can be at most quadratic in the spatial coordinates $x^i$. Possible applications of these models will be discussed in the outlook, see section \[sec:conc\]. \[sec:blackhole\]Application to Black Holes =========================================== In this section we investigate possible deformations of non rotating black holes. We will do this in analogy to the cosmological models and therefore do not have to explain every single step. The undeformed Lie algebra of the symmetry group $\mathbb{R}\times SO(3)$ of a non rotating black hole is generated by the vector fields p\^0=\_t,L\_i = \_[ijk]{} x\^j\_k , given in cartesian coordinates. We choose $t_i^\star =t_i$ for all $i$ and define $\mathfrak{g}_\star=\mathfrak{g} = \mathrm{span}(p^0,L_i)$. It can be shown that each twist vector field $V_a$ has to be of the form V\_a = (c\^0\_a(r)+N\_a\^0 t)\_t + d\_a\^i L\_i + f\_a(r) x\^i \_i in order to fulfill ${[V_a,\mathfrak{g}]}\subseteq\mathfrak{g}$. Here $r=\Vert\mathbf{x}\Vert$ is the euclidian norm of the spatial position vector. The next task is to construct the two vector field deformations. Therefore we additionally have to demand ${[V_a,V_b]} =0,~\forall_{a,b}$, leading to the conditions \[eqn:blackcond1\]&d\_a\^i d\_b\^j\_[ijk]{}=0 , \_k ,\ \[eqn:blackcond2\]&(f\_a(r) x\^j \_j - N\_a\^0) c\^0\_b(r) - (f\_b(r) x\^j\_j -N\_b\^0) c\^0\_a(r) =0 ,\ \[eqn:blackcond3\]& f\_a(r)f\_b\^(r)-f\_a\^(r) f\_b(r) = 0 , where $f_a^\prime(r)$ means the derivative of $f_a(r)$. Note that (\[eqn:blackcond3\]) is a condition similar to (\[eqn:frwcond3\]), and therefore has the same type of solutions. Because of this, the functions $f_1(r)$ and $f_2(r)$ have to be parallel in the overlap of their supports. From this we can always eliminate locally one $f_a(r)$ by a twist conserving map and simplify the investigation of the condition (\[eqn:blackcond2\]). At the end, the local solutions have to be glued together. We choose without loss of generality $f_1(r)=0$ for our classification of local solutions. The solution to (\[eqn:blackcond1\]) is that the $\mathbf{d}_a$ have to be parallel. We use \_a=\_a with constants $\kappa_a\in\mathbb{R}$ and some arbitrary vector $\mathbf{d}\neq0$. We now classify the solutions to (\[eqn:blackcond2\]) according to $N_a^0$ and $f_2(r)$ and label them by $\mathfrak{B}_{AB}$. We distinguish between $f_2(r)$ being the zero function or not. The result is shown in table \[tab:blackhole\]. Other choices of parameters can be mapped by a twist conserving map into these classes. Note that in particular for analytical functions $f_a(r)$ the twist conserving map transforming $f_1(r)$ to zero can be performed globally, and with this also the classification of twists given in table \[tab:blackhole\]. [$~~\mathfrak{B}_{AB}$]{} $f_2(r)=0$ $f_2(r)\neq0$ --------------------------- ------------------------------------------- ------------------------------------------------------------------------------------------------- $N_1^0=0$, $V_1=c_1^0(r)\partial_t +\kappa_1 d^iL_i$ $V_1=c_1^0\partial_t +\kappa_1 d^i L_i $ $N_2^0=0$ $V_2=c_2^0(r)\partial_t+\kappa_2 d^i L_i$ $V_2=c_2^0(r)\partial_t +\kappa_2 d^i L_i + f_2(r) x^i\partial_i$ $N_1^0\neq0$, $V_1=(c_1^0(r)+N_1^0 t)\partial_t $ $V_1=(c_1^0(r)+N_1^0 t)\partial_t+\kappa_1 d^i L_i$ $N_2^0=0$ $V_2=\kappa_2 d^i L_i$ $V_2= -\frac{1}{N_1^0}f_2(r)r c_1^{0\prime}(r)\partial_t +\kappa_2 d^i L_i+f_2(r)x^i\partial_i$ $N_1^0=0$, $V_1=\kappa_1 d^i L_i$ $V_1=c_1^0(r) \partial_t + \kappa_1 d^i L_i$, $N_2^0\neq0$ $V_2=(c_2^0(r)+N_2^0 t)\partial_t $ $V_2= (c_2^0(r)+N_2^0 t)\partial_t +\kappa_2 d^i L_i +f_2(r) x^i\partial_i $ : \[tab:blackhole\]Two vector field deformations of the black hole symmetry group $\mathbb{R}\times SO(3)$. Note that $c_1^0(r)=c_1^0$ has to be constant in type $\mathfrak{B}_{12}$. In type $\mathfrak{B}_{32}$ we still have to solve a differential equation for $c_1^0(r)$ given by \[eqn:blackode\] c\_1\^0(r) = r c\_1\^[0]{}(r) , for an arbitrary given $f_2(r)$. We will not work out the solutions to this differential equation, since type $\mathfrak{B}_{32}$ is a quite unphysical model, in which the noncommutativity is increasing linear in time due to $N_2^0\neq 0$. Note that $\mathfrak{B}_{11}$ can be extended to a triangular $\star$-Hopf algebra by choosing $c_a^0(r)=c_a^0$, for $a\in\lbrace1,2\rbrace$. In addition, $\mathfrak{B}_{12}$ is a $\star$-Hopf algebra for $\kappa_1=\kappa_2=0$. The $\star$-commutators $c^{\mu\nu}={[x^\mu\stackrel{\star}{,}x^\nu]}$ of the coordinate functions $x^\mu\in A_\star$ in order $\lambda^1$ for these models are given in the appendix \[app:starcom\]. They can be used in order to construct sensible physical models of a noncommutative black hole. By using the method explained in the previous section, the two vector field twists can be extended to multiple vector field twists. Since we do not require these twists in our work and their construction is straightforward, we do not present them here. \[sec:conc\]Conclusion and Outlook ================================== We have discussed symmetry reduction in noncommutative gravity using the formalism of twisted noncommutative differential geometry. Our motivation for these investigations derives from the fact that, for most physical applications of gravity theories, including cosmology, symmetry reduction is required due to the complexity of such models, already in the undeformed case. In section \[sec:symred\] we have presented a general method for symmetry reduction in twisted gravity theories. As a result we have obtained restrictions on the twist, depending on the structure of the twisted symmetry group. In particular, we find that deforming the infinitesimal symmetry transformations results in weaker restrictions than deforming the finite transformations and demanding a quantum group structure. In section \[sec:jambor\] we have applied this general method to gravity theories twisted by Reshetikhin-Jambor-Sykora twists. These are twists constructed from commuting vector fields. In this case we could give explicit conditions, which have to be fulfilled in order to allow symmetry reduction of a given Lie group. In sections \[sec:cosmo\] and \[sec:blackhole\] we have investigated admissible deformations of FRW and black hole symmetries by a Reshetikhin-Jambor-Sykora twist. In this class we have classified all possible deformations. This lays the foundation for phenomenological studies of noncommutative cosmology and black hole physics based on twisted gravity. In a forthcoming work [@Ohl/Schenkel:Cosmo:2009] we will investigate cosmological implications of twisted FRW models by studying fluctuations of quantum fields living on twisted FRW backgrounds. Quantum fields were already introduced in a twisted framework in [@Aschieri:2007sq]. As we see from proposition \[propo:starinvariance\], the noncommutative backgrounds are also invariant under the undeformed action of the classical symmetry. This means that they have the same coordinate representations with respect to the undeformed basis vectors as the commutative fields in Einstein gravity. With this we have a construction principle for noncommutative backgrounds, in their natural basis, by representing the classical fields in the deformed basis. A class of models of particular interest is type $\mathfrak{C}_{22}$ in section \[sec:cosmo\] (cf. table \[tab:frw\]). These twists break classical translation invariance, but classical rotation invariance can be retained by tuning $\mathbf{d}_1$ and $\mathbf{c}_1$ to small values. Furthermore, the global factor $V_2^0(t)$ in the exponent of the twist can be used in order to tune noncommutativity effects depending on time. Obviously, enforcing a suitable $V_2^0(t)$ by hand leads to phenomenologically valid models. Since there is no natural choice of $V_2^0(t)$, it is interesting to investigate the dynamics of $V_2^0(t)$ in a given field configuration and study if it leads to a model consistent with cosmological observations. In this case, the model would be physically attractive. This will also be subject of future work [@Ohl/Schenkel:Cosmo:2009]. Dynamical noncommutativity has already been studied in the case of scalar field theories on Minkowski spacetime [@Aschieri:2008zv]. In the case of black hole physics, models of particular interest would be $\mathfrak{B}_{11}$ with functions $c^0_a(r)$ decreasing sufficiently quickly with $r$ and $\mathfrak{B}_{12}$ with $f_2(r)$ and $c_2^0(r)$ decreasing sufficiently quickly with $r$ (cf. table \[tab:blackhole\]). It will again be interesting to investigate the dynamics of these functions on a given field configuration. Note that the type $\mathfrak{B}_{12}$ with $\kappa_1=\kappa_2=0$ is invariant under the classical black hole symmetries, and therefore particularly interesting for physical applications. On the other hand, models with nonvanishing $N_a^0$ are of little physical interest, because the noncommutativity is growing linearly in time, which would be unphysical. Other avenues for future work are the classification of models on nontrivial topologies (like, e.g., $\mathbb{R}\times S_3$ in cosmology), investigating nontrivial embeddings $t_i^\star=t_i^\star(t_j)$ and using a wider class of twist elements. Acknowledgements {#acknowledgements .unnumbered} ================ AS thanks Christoph Uhlemann and Julian Adamek for discussions and comments on this work. This research is supported by Deutsche Forschungsgemeinschaft through the Research Training Group 1147 *Theoretical Astrophysics and Particle Physics*. \[app:proof\]Proof of Proposition \[propo:starhopf\] ==================================================== In this appendix we show that for Reshetikhin-Jambor-Sykora twists (\[eqn:jstwist\]) the conditions (\[eqn:hopfconditions\]) necessary for extending the $\star$-enveloping subalgebra $(U\mathfrak{g}_\star,\star)$ to a $\star$-Hopf subalgebra are equivalent to the simplified conditions of proposition \[propo:starhopf\]. The plan is as follows: we use (\[eqn:hopfconditions1\]) and show that it is equivalent to the conditions of proposition \[propo:starhopf\]. In a second step, we show that (\[eqn:hopfconditions2\]) is automatically satisfied if (\[eqn:hopfconditions1\]) is fulfilled, and thus does not lead to additional conditions. We start with (\[eqn:hopfconditions1\]) and show the identity \_(U\_)U\_U\_  \_(\_)U\_U\_ . The direction $\Rightarrow$ is trivial, since $\mathfrak{g}_\star\subset U\mathfrak{g}_\star$, and the direction $\Leftarrow$ can be shown using that $\Delta_\star$ is a $\star$-algebra homomorphism and that $U\mathfrak{g}_\star$ closes under $\star$-multiplication. Furthermore, using (\[eqn:defstarhopfactions\]) we obtain \_(\_)U\_U\_  X\_[|R\^]{}U\_ ,  |R\_(\_)0 . Therefore we have to use that all $\bar R^\alpha$ are linearly independent, that $X$ is a vector space isomorphism [@Aschieri:2005zs] and that we have $\bar R_\alpha(\mathfrak{g}_\star)\subseteq \mathfrak{g}_\star$, due to the minimal axioms (\[eqn:infinitesimalconditions\]). Additionally, we can show that $X_{\bar R^\alpha}=\bar f^\beta \bar R^\alpha \chi S^{-1}(\bar f_\beta)= \bar R^\alpha$. This is done by applying the explicit form of the twist (\[eqn:jstwist\]) and using that the $V_a$ mutually commute. Next, we show that \[eqn:prepropo\] |R\^U\_ ,  |R\_(\_)0    \^[ba]{}V\_b\_ ,   a[\[V\_a,\_\]]{}0 . The direction $\Rightarrow$ is trivial, since the RHS is a special case of the LHS. The direction $\Leftarrow$ can be shown by using that the $V_a$ mutually commute and the explicit expression of the $R$-matrix (\[eqn:jsrmatrix\]). Finally, the RHS of (\[eqn:prepropo\]) is equivalent to the condition of proposition \[propo:starhopf\] by using the canonical form of $\theta$ (\[eqn:theta\]). Next, we show that (\[eqn:hopfconditions2\]) is satisfied, if (\[eqn:hopfconditions1\]) is fulfilled. For this we use that for Reshetikhin-Jambor-Sykora twists we have $\chi=f^\alpha S(f_\alpha)=1$, which leads to the identity S\_() = S()\^[-1]{} =S() = S\^[-1]{}() = S\_\^[-1]{}() , U  for the antipode in the $\mathcal{F}$-Hopf algebra. This property translates to the $\star$-Hopf algebra, since it is isomorphic to the $\mathcal{F}$-Hopf algebra and we obtain the following equivalences of (\[eqn:hopfconditions2\]) \[eqn:antipodesimple\] S\_(U\_)U\_  S\_(\_)U\_  S\^[-1]{}\_(\_)U\_ . For the first equivalence we had to use that $S_\star$ is a $\star$-anti homomorphism. Using the RHS of (\[eqn:prepropo\]), which is equivalent to (\[eqn:hopfconditions1\]), and the definition of $S_\star^{-1}$ (\[eqn:defstarhopfactions\]), we obtain S\^[-1]{}\_(\_) = -\_[n=0]{}\^ \^[a\_1b\_1]{}\^[a\_n b\_n]{} [\[V\_[a\_1]{},,[\[V\_[a\_n]{},\_\]]{}\]]{}  V\_[b\_1]{}V\_[b\_n]{} U\_ , where we have used $\xi \star V_a = \xi V_b$ for all $\xi\in U\Xi_\star$, since the action of the twist on $V_a$ is trivial. \[app:starcom\]$\star$-Commutators of the Coordinate Functions in FRW and Black Hole Models =========================================================================================== In tables \[tab:cosmocom\] and \[tab:blackcom\], we list the $\star$-commutators among the linear coordinate functions to order $\lambda^1$ in the FRW and black hole models. In these expressions, $(i\leftrightarrow j)$ denotes the same term with $i$ and $j$ interchanged. Type $c^{\mu\nu}:={[x^\mu\stackrel{\star}{,}x^\nu]}$ in $\mathcal{O}(\lambda^1)$ --------------------- ------------------------------------------------------------------------------------------------------------------------------------------------ $\mathfrak{C}_{11}$ $c^{0i}=i\lambda \bigl( V_{1}^0(t) c_{2}^i - V_{2}^0(t) c_{1}^i \bigr)$   $c^{ij}=i\lambda \bigl( c_{1}^i c_{2}^j - (i\leftrightarrow j) \bigr)$ $\mathfrak{C}_{21}$ $c^{0i}=-i\lambda V_{2}^0(t) \bigl( c_{1}^i + f_{1} x^i \bigr)$   $c^{ij}=0$ $\mathfrak{C}_{12}$ $c^{0i}=i\lambda \bigl( V_{1}^0(t) \kappa d_{1}^i - V_{2}(t) (c_{1}^i+d_{1}^k\epsilon_{kli}x^l ) \bigr)$   $c^{ij}=i\lambda \kappa \bigl( (c_{1}^i +d_{1}^k \epsilon_{kli}x^l ) d_{1}^j - (i\leftrightarrow j) \bigr)$ $\mathfrak{C}_{22}$ $c^{0i}=-i\lambda V_{2}^0(t) \bigl( c_{1}^i + d_{1}^j\epsilon_{jki}x^k + f_{1} x^i \bigr)$   $c^{ij}=0$ $\mathfrak{C}_{32}$ $c^{0i}=i\lambda \bigl( V_1^0(t) (c_2^i + f_2 x^i) - V_2^0(t) (\frac{1}{f_2}d_1^j c_2^k\epsilon_{jki} + d_1^j\epsilon_{jki}x^k) \bigr)$   $c^{ij}=i\lambda \bigl( (\frac{1}{f_2} d_1^k c_2^l \epsilon_{kli} +d_1^k\epsilon_{kli} x^l )~(c_2^j + f_2 x^j) - (i\leftrightarrow j) \bigr)$ : \[tab:cosmocom\] $\star$-commutators in the cosmological models $\mathfrak{C}_{AB}$. Type $c^{\mu\nu}:={[x^\mu\stackrel{\star}{,}x^\nu]}$ in $\mathcal{O}(\lambda^1)$ --------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------- $\mathfrak{B}_{11}$ $c^{0i}=i\lambda \bigl(c_1^0(r) \kappa_2 -c_2^0(r) \kappa_1\bigr) d^j\epsilon_{jki}x^k$   $c^{ij}=0$ $\mathfrak{B}_{21}$ $c^{0i}=i\lambda \bigl(c_1^0(r)+N_1^0 t\bigr) d^j \epsilon_{jki}x^k$   $c^{ij}=0$ $\mathfrak{B}_{12}$ $c^{0i} =i\lambda \bigl( c_1^0 (\kappa_2 d^j\epsilon_{jki} x^k+f_2(r) x^i) -\kappa_1 c_2^0(r) d^j\epsilon_{jki}x^k\bigr)$   $c^{ij}= i\lambda\bigl(\kappa_1 d^k\epsilon_{kli}x^l (\kappa_2 d^m\epsilon_{mnj} x^n +f_2(r)x^j) -(i\leftrightarrow j)\bigr)$ $\mathfrak{B}_{22}$ $c^{0i}=i\lambda \Bigl((c_1^0(r)+N_1^0 t)(\kappa_2 d^j\epsilon_{jki} x^k +f_2(r)x^i) + \frac{1}{N_1^0} f_2(r)rc_1^{0\prime}(r)\kappa_1d^j\epsilon_{jki}x^k \Bigr)$   $c^{ij}= i\lambda\bigl(\kappa_1 d^k\epsilon_{kli}x^l (\kappa_2 d^m\epsilon_{mnj} x^n +f_2(r)x^j) -(i\leftrightarrow j)\bigr)$ $\mathfrak{B}_{32}$ $c^{0i}=i\lambda \bigl( c_1^0(r)~(\kappa_2 d^j\epsilon_{jki} x^k+f_2(r) x^i) -(c_2^0(r) +N_2^0 t) \kappa_1 d^j\epsilon_{jki}x^k \bigr)$   $c^{ij}=i\lambda\bigl(\kappa_1 d^k\epsilon_{kli}x^l (\kappa_2 d^m\epsilon_{mnj} x^n +f_2(r)x^j) -(i\leftrightarrow j)\bigr)$ : \[tab:blackcom\]$\star$-commutators in the black hole models $\mathfrak{B}_{AB}$. [10]{} S. Doplicher, K. Fredenhagen and J. E. Roberts, Phys. Lett. **B331**, 39 (1994); S. Doplicher, K. Fredenhagen and J. E. Roberts, Commun. Math. Phys. **172**, 187 (1995) \[arXiv:hep-th/0303037\]. R. J. Szabo, Phys. Rept. **378**, 207 (2003) \[arXiv:hep-th/0109162\]; R. J. Szabo, Class. Quant. Grav. **23**, R199 (2006) \[arXiv:hep-th/0606233\]. F. Muller-Hoissen, AIP Conf. Proc. **977**, 12 (2008) \[arXiv:0710.4418 \[gr-qc\]\]. X. Calmet, B. Jur[č]{}o, P. Schupp, J. Wess and M. Wohlgenannt, Eur. Phys. J. **C23**, 363 (2002) \[arXiv:hep-ph/0111115\]; B. Melic, K. Passek-Kumericki, J. Trampetic, P. Schupp and M. Wohlgenannt, Eur. Phys. J. C **42**, 483 (2005) \[arXiv:hep-ph/0502249\]; B. Melic, K. Passek-Kumericki, J. Trampetic, P. Schupp and M. Wohlgenannt, Eur. Phys. J. C **42**, 499 (2005) \[arXiv:hep-ph/0503064\]. W. Behr, N. G. Deshpande, G. Duplancic, P. Schupp, J. Trampetic and J. Wess, Eur. Phys. J. C **29**, 441 (2003) \[arXiv:hep-ph/0202121\]; G. Duplancic, P. Schupp and J. Trampetic, Eur. Phys. J. C **32**, 141 (2003) \[arXiv:hep-ph/0309138\]; P. Schupp, J. Trampetic, J. Wess and G. Raffelt, Eur. Phys. J. C **36**, 405 (2004) \[arXiv:hep-ph/0212292\]; P. Minkowski, P. Schupp and J. Trampetic, Eur. Phys. J. C **37**, 123 (2004) \[arXiv:hep-th/0302175\]; B. Melic, K. Passek-Kumericki and J. Trampetic, Phys. Rev. **D72**, 054004 (2005) \[arXiv:hep-ph/0503133\]; Phys. Rev. **D72**, 057502 (2005) \[arXiv:hep-ph/0507231\]; M. Buric, D. Latas, V. Radovanovic and J. Trampetic, Phys. Rev. **D75**, 097701 (2007); T. Ohl and J. Reuter, Phys. Rev. **D70**, 076007 (2004) \[arXiv:hep-ph/0406098\]; A. Alboteanu, T. Ohl and R. Rückl, Phys. Rev. **D74**, 096004 (2006) \[arXiv:hep-ph/0608155\]; eConf **C0705302**, TEV05 (2007) \[Acta Phys. Polon. **B38**, 3647 (2007)\] \[arXiv:0709.2359 \[hep-ph\]\]; Phys. Rev. **D76**, 105018 (2007) \[arXiv:0707.3595 \[hep-ph\]\]. I. Hinchliffe, N. Kersting and Y. L. Ma, Int. J. Mod. Phys. **A19**, 179 (2004) \[arXiv:hep-ph/0205040\]. C. S. Chu, B. R. Greene and G. Shiu, Mod. Phys. Lett. **A16**, 2231 (2001) \[arXiv:hep-th/0011241\]; F. Lizzi, G. Mangano, G. Miele and M. Peloso, JHEP **0206**, 049 (2002) \[arXiv:hep-th/0203099\]; R. Brandenberger and P. M. Ho, Phys. Rev. **D66**, 023517 (2002) \[AAPPS Bull. **12N1**, 10 (2002)\] \[arXiv:hep-th/0203119\]; Q. G. Huang and M. Li, JHEP **0306**, 014 (2003) \[arXiv:hep-th/0304203\]; Q. G. Huang and M. Li, JCAP **0311**, 001 (2003) \[arXiv:astro-ph/0308458\]; S. Tsujikawa, R. Maartens and R. Brandenberger, Phys. Lett. **B574**, 141 (2003) \[arXiv:astro-ph/0308169\]; A. H. Fatollahi and M. Hajirahimi, Europhys. Lett. **75**, 542 (2006) \[arXiv:astro-ph/0607257\]; E. Akofor, A. P. Balachandran, S. G. Jo, A. Joseph and B. A. Qureshi, JHEP **0805**, 092 (2008) \[arXiv:0710.5897 \[astro-ph\]\]; E. Akofor, A. P. Balachandran, A. Joseph, L. Pekowsky and B. A. Qureshi, arXiv:0806.2458 \[astro-ph\]; S. Fabi, B. Harms and A. Stern, Phys. Rev. **D78**, 065037 (2008) \[arXiv:0808.0943 \[hep-th\]\]. P. Nicolini, arXiv:0807.1939 \[hep-th\]. P. Aschieri, C. Blohmann, M. Dimitrijevic, F. Meyer, P. Schupp and J. Wess, Class. Quant. Grav. **22**, 3511 (2005) \[arXiv:hep-th/0504183\]. P. Aschieri, M. Dimitrijevic, F. Meyer and J. Wess, Class. Quant. Grav. **23**, 1883 (2006) \[arXiv:hep-th/0510059\]. P. Aschieri, J. Phys. Conf. Ser. **53**, 799 (2006) \[arXiv:hep-th/0608172\]. L. Alvarez-Gaume, F. Meyer and M. A. Vazquez-Mozo, Nucl. Phys. B **753**, 92 (2006) \[arXiv:hep-th/0605113\]. N. Reshetikhin, Lett. Math. Phys.  [**20**]{}, 331 (1990). C. Jambor and A. Sykora, \[arXiv:hep-th/0405268\]. T. Ohl and A. Schenkel, *Cosmological and Black Hole Solutions in Twisted Noncommutative Gravity*, to be published; *Classical and Quantum Field Fluctuations on Noncommutative Backgrounds*, in preparation. P. Aschieri, F. Lizzi and P. Vitale, Phys. Rev. **D77**, 025037 (2008) \[arXiv:0708.3002 \[hep-th\]\]. P. Aschieri, L. Castellani and M. Dimitrijevic, Lett. Math. Phys. **85**, 39 (2008) \[arXiv:0803.4325 \[hep-th\]\]. [^1]: e-mail: `[email protected]` [^2]: e-mail: `[email protected]`
{ "pile_set_name": "ArXiv" }
--- author: - Rudy Wijnands title: 'Accretion-driven millisecond X-ray pulsars' --- [**Abstract:**]{} I present an overview of our current observational knowledge of the six known accretion-driven millisecond X-ray pulsars. A prominent place in this review is given to SAX J1808.4–3658; it was the first such system discovered and currently four outbursts have been observed from this source, three of which have been studied in detail using the [*Rossi X-ray Timing Explorer*]{} satellite. This makes SAX J1808.4–3658 the best studied example of an accretion-driven millisecond pulsar. Its most recent outburst in October 2002 is of particular interest because of the discovery of two simultaneous kilohertz quasi-periodic oscillations and nearly coherent oscillations during type-I X-ray bursts. This is the first (and so far only) time that such phenomena are observed in a system for which the neutron star spin frequency is exactly known. The other five systems were discovered within the last three years (with IGR J00291+5934 only discovered in December 2004) and only limited results have been published. Introduction \[section:intro\] ============================== Ordinary pulsars are born as highly-magnetized (B $\sim 10^{12}$ G), rapidly rotating (P $\sim$ 10 ms) neutron stars which spin down on timescales of 10 to 100 million years due to magnetic dipole radiation. However, a number of millisecond (P $<$ 10 ms) radio pulsars is known with ages of billions of years and weak (B $\sim 10^{8-9}$ G) surface magnetic fields. Since many of these millisecond pulsars are in binaries, it has long been suspected (see, e.g., Bhattacharya & van den Heuvel 1991 for an extended review) that the neutron stars were spun up by mass transfer from a stellar companion in a low-mass X-ray binary (LMXB), but years of searching for coherent millisecond pulsations in LMXBs failed to yield a detection (Vaughan et al. 1994 and references therein). The launch of the NASA [*Rossi X-ray Timing Explorer*]{} ([*RXTE*]{}) brought the discovery of kilohertz quasi-periodic oscillations (kHz QPOs; Strohmayer et al. 1996; Van der Klis et al. 1996) as well as nearly coherent oscillations (’burst oscillations’) during type-I X-ray bursts in a number of LMXBs (e.g., Strohmayer et al. 1996), providing tantalizingly suggestive evidence for weakly magnetic neutron stars with millisecond spin periods (see Van der Klis 2000, 2004 and Strohmayer & Bildsten 2003 for more details about kHz QPOs and burst oscillations in LMXBs). In April 1998 the first accretion-driven millisecond X-ray pulsar (SAX J1808.4–3658) was discovered (Wijnands & van der Klis 1998a) proving that indeed neutron stars in LMXBs can spin very rapidly. This conclusion was further strengthened by the discovery of four additional systems in 2002 and 2003 (Markwardt et al. 2002a, 2003a, 2003b, Galloway et al. 2002), and recently, in December 2004, with the discovery of IGR J00291+5934 as a millisecond X-ray pulsar (Markwardt et al. 2004a). Here, I will give a brief summary of our current observational knowledge of those accretion-driven millisecond X-ray pulsars. Preliminary versions of this review were published by Wijnands (2004a, 2004b). SAX J1808.4–3658 \[section:1808\] ================================= [c]{} The September 1996 Outburst \[subsection:1996outburst\] ------------------------------------------------------- In September 1996, a new X-ray transient and LMXB was discovered with the Wide Field Cameras (WFCs) aboard the Dutch-Italian [*BeppoSAX*]{} satellite and the source was designated SAX J1808.4–3658 (In ’t Zand et al. 1998). Three type-I X-ray bursts were detected, demonstrating that the compact object in this system is a neutron star. From those bursts, a distance estimate of 2.5 kpc was determined (In ’t Zand et al. 1998, 2001). The maximum luminosity during this outburst was $\sim 10^{36}$ erg s$^{-1}$, significantly lower than the peak outburst luminosity of ’classical’ neutron star transients (which typically can reach a luminosity of $10^{37}$ to $10^{38}$ ergs s$^{-1}$). This low peak luminosity showed that the source was part of the growing group of faint neutron-star X-ray transients (Heise et al. 1999). The outburst continued for about three weeks (see Fig. \[fig:1808\_asm\]) after which the source was thought to have returned to quiescence. However, it was found (Revnivtsev 2003) that the source was detected on October 29, 1996 (using slew data obtained with the proportional counter array \[PCA\] aboard [*RXTE*]{}) with a luminosity of about a tenth of the outburst peak luminosity. This demonstrates that six weeks after the main outburst the source was still active (possibly only sporadically), which might indicate that at the end of this outburst the source behaved in a manner very similar to what was seen during its 2000 and 2002 outbursts (see § \[subsection:2000outburst\] and § \[subsection:2002outburst\]). After it was found that SAX J1808.4–3658 harbors a millisecond pulsar (§ \[subsection:1998outburst\]), the three observed X-ray bursts seen with [*BeppoSAX*]{}/WFC were scrutinized for potential burst oscillations (In ’t Zand et al. 2001). A marginal detection of a 401 Hz oscillation was made in the third burst. This result suggested that the burst oscillations observed in the other, non-pulsating, neutron-star LMXBs occur indeed at their neutron-star spin frequencies. This result has been confirmed by the recent detection of burst oscillations during the 2002 outburst of SAX J1808.4–3658 (§ \[subsubsection:bursts\]). [c]{} The April 1998 outburst \[subsection:1998outburst\] --------------------------------------------------- On April 9, 1998, [*RXTE*]{}/PCA slew observations indicated that SAX J1808.4–3658 was active again (Marshall 1998; see Fig. \[fig:1808\_asm\] for the [*RXTE*]{}/ASM light curve during this outburst). Using public TOO observations of this source from April 11, it was discovered (Wijnands & Van der Klis 1998a) that coherent 401 Hz pulsations (Fig. \[fig:pulsations\]) were present in the persistent X-ray flux of the source, making it the first accretion-driven millisecond X-ray pulsar discovered. After this discovery, several more public [*RXTE*]{} observations were made (using the PCA) which were used by several groups to study different aspects of the source. I will briefly mention those results and I point to references for the details. A detailed analysis of the coherent timing behavior (Chakrabarty & Morgan 1998) showed that the neutron star was in a tight binary with a very low-mass companion star in a $\sim$2-hr orbital period. Due to the limited amount of data obtained during this outburst, only an upper limit of $<7 \times 10^{-13}$ Hz s$^{-1}$ could be obtained on the pulse-frequency derivative (Chakrabarty & Morgan 1998). Studies of the X-ray spectrum (Gilfanov et al. 1998; Heindl & Smith 1998; see also Gierlinski et al. 2002 and Poutanen & Gierlinski 2003) and the aperiodic rapid X-ray variability (Wijnands & van der Klis 1998b; see also Van Straaten et al. 2005) showed an object that, apart from its pulsations, is remarkably similar to other LMXBs with comparable luminosities (the atoll sources). There is apparent modulation of the X-ray intensity at the orbital period, with a broad minimum when the pulsar is behind the companion (Chakrabarty & Morgan 1998; Heindl & Smith 1998). Cui et al. (1998) and Ford (2000) reported on the harmonic content, energy dependency, and soft phase lag of the pulsations. The main result of those studies is that the low-energy pulsations lag the high-energy ones by as much as $\sim$200 $\mu$s ($\sim$8% of the pulsation period; see Cui et al. \[1998\], Ford \[2000\], and Poutanen & Gierlinski \[2003\] for possible explanations for these soft lags). Another interesting aspect is that the source first showed a steady decline in X-ray flux, which after  2 weeks suddenly accelerated (Gilfanov et al. 1998; Cui et al. 1998; Fig. \[fig:1808\_1998v2002\]). This behavior has been attributed to the fact that the source might have entered the ’propeller regime’ in which the accretion is centrifugally inhibited (Gilfanov et al. 1998). However, after the onset of the steep decline the pulsations could still be detected (Cui et al. 1998) making this interpretation doubtful. A week after the onset of this steep decline, the X-ray flux leveled off (Cui et al. 1998; Wang et al. 2001), but as no further [*RXTE*]{}/PCA observations were made, the X-ray behavior of the source at the end of the outburst remained unclear. The source might have displayed a similar long-term episode of low-luminosity activity as seen at the end of its 2000 and 2002 outbursts (see § \[subsection:2000outburst\] and § \[subsection:2002outburst\]). SAX J1808.4–3658 was not only detected and studied in X-rays but also in the optical, IR, and in radio bands. The optical/IR counterpart of SAX J1808.4–3658 (later named V4580 Sgr; Kazarovets et al. 2000) was first discovered by Roche et al. (1998) and subsequently confirmed by Giles et al. (1998). A detailed study of the optical behavior during this outburst was reported by Giles et al. (1999) and Wang et al. (2001). Both papers reported that the peak V magnitude of the source was $\sim$16.7 and the source decayed in brightness as the outburst progressed. The brightness of the source leveled off at around V $\sim$ 18.5 (I $\sim$ 17.9) about $\sim$2 weeks after the peak of the outburst. It stayed at this level for at least several weeks before it further decreased in brightness. This behavior suggests that the source was indeed still active for a long period after the main outburst. It was also reported (Giles et al. 1999) that the optical flux was modulated at the 2-hr orbital period of the system. Modeling the X-ray and optical emission from the system using an X-ray-heated accretion disk model yielded a Av of 0.68 and an inclination of cos i = 0.65 (Wang et al. 2001), resulting in a mass of the companion star of 0.05–0.10 solar masses. During some of the IR observations, the source was too bright to be consistent with emission from the disk or the companion star, even when considering X-ray heating. This IR excess might be due to synchrotron processes, likely related to an outflow or ejection of matter (Wang et al. 2001). Such an ejection event was also confirmed by the discovery of the radio counterpart (Gaensler et al. 1999). The source was detected with a 4.8 GHz flux of $\sim$0.8 mJy on 1998 April 27, but it was not detected at earlier or later epochs. The January 2000 outburst \[subsection:2000outburst\] ----------------------------------------------------- On January 21, 2000, SAX J1808.4–3658 was again detected (Wijnands et al. 2001) with the [*RXTE*]{}/PCA at a flux level of $\sim$10–15 mCrab (2–10 keV), i.e. about a tenth of the peak fluxes observed during the two previous outbursts. Using follow-up [*RXTE*]{}/PCA observations, it was found that the source exhibited low-level activity for several months (Wijnands et al. 2001). Due to solar constraints the source could not be observed before January 21 but likely a true outburst occurred before that date and we only observed the end stages of this outburst. This is supported by the very similar behavior of the source observed near the end of its 2002 October outburst (see § \[subsection:2002outburst\]; Fig. \[fig:lc\]). [c]{} During the 2000 outburst, SAX J1808.4–3658 was observed (using [*RXTE*]{}/PCA) on some occasions at luminosities of  $\sim$$10^{35}$ ergs s$^{-1}$, but on other occasions (a few days earlier or later) it had luminosities of $\sim$$10^{32}$ ergs s$^{-1}$ (as seen during [*BeppoSAX*]{} and [*XMM-Newton*]{} observations; Wijnands et al. 2002, Wijnands 2003; see Fig. \[fig:XMM\] left panel). This demonstrates that the source exhibited extreme luminosity swings (a factor of $>$1000) on timescales of days. During the [*RXTE*]{} observations, it was also found that on several occasions the source exhibited strong (up to 100 % r.m.s. amplitude) violent flaring behavior with a repetition frequency of about 1 Hz (Van der Klis et al. 2000; Fig. \[fig:1hz\]). During this episode of low-level activity, the pulsations at 401 Hz were also detected. [c]{} The source was again detected in optical, albeit at a lower brightness than during the 1998 outburst (Wachter & Hoard 2000). This is consistent with the lower X-ray activity seen for the source. The source was frequently observed during this outburst and preliminary results were presented by Wachter et al. (2000). The main results are presented in Figure \[fig:2000\_lc\] (reproduced with permission from Stefanie Wachter). The optical and X-ray brightness of the source are correlated at the end of the outburst, although one optical flare (around day 435–440 in Fig. \[fig:2000\_lc\]) was not accompanied by an X-ray flare. However, the optical and X-ray observations were not simultaneous, which means that a brief (around a few days) X-ray flare could have been missed. During the earlier stages of the outburst, the X-ray and the optical behavior of the source were not correlated (Fig. \[fig:2000\_lc\] lower panel): the source is highly variable in X-rays, but quite stable in optical with only low amplitude variations. This stable period in the optical is very similar to the episode of stable optical emission in the late stages of the 1998 outburst, suggesting this is typical behavior for this source. The October 2002 outburst \[subsection:2002outburst\] ----------------------------------------------------- In 2002 October, the fourth outburst of SAX J1808.4–3658 was detected (Markwardt et al. 2002b), immediately launching an extensive [*RXTE*]{}/PCA observing campaign. The main results are summarized below. [c]{} ### The X-ray light curve \[subsubsection:lightcurve\] The [*RXTE*]{}/PCA light curve for this outburst is shown in Figure \[fig:lc\] (see Fig. \[fig:1808\_asm\] for the ASM light curve). During the first few weeks, the source decayed steadily, until the rate of decline suddenly increased, in a manner similar to what was observed during the 1998 outburst (see § \[subsection:1998outburst\]). During both the 1998 and 2002 outbursts, the moment of acceleration of the decline occurred at about two weeks after the peak of the outburst. Approximately five days later the X-ray count rate rapidly increased again until it reached a peak of about a tenth of the outburst maximum. After that the source entered a state in which the count rate rapidly fluctuated on time scales of days to hours, very similar to the 2000 low-level activity (see § \[subsection:2000outburst\]). The 2002 outburst light curve is the most detailed one seen for this source and it exhibits all features seen during the previous three outbursts of the source (the initial decline, the increase in the decline rate, the long-term low-level activity), demonstrating that this behavior is typical for this source. ### The X-ray bursts and the burst oscillations \[subsubsection:bursts\] During the first five days of the outburst, four type-I X-ray bursts were detected. Burst oscillations were observed during the rise and decay of each burst, but not during the peak (Chakrabarty et al. 2003). The frequency in the burst tails was constant and identical to the spin frequency, while the oscillation in the burst rise showed evidence for a very rapid frequency drift of up to 5 Hz. This frequency behavior and the absence of oscillations at the peak of the bursts is similar to the burst oscillations seen in other, non-pulsating neutron star LMXBs, demonstrating that indeed the burst-oscillations occur at the neutron-star spin frequency in all sources. As a consequence, the spin frequency is now known for 18 LMXBs (12 burst-oscillations sources and 6 pulsars) with the highest spin frequency being 619 Hz. The sample of burst-oscillation sources was used to demonstrate that neutron stars in LMXBs spin well below the break-up frequency for neutron stars. This could suggest that the neutron stars are limited in their spin frequencies, possible due to the emission of gravitational radiation (Chakrabarty et al. 2003; Chakrabarty 2004). [c]{} ### The kHz QPOs \[subsubsection:khzqpos\] Wijnands et al. (2003) reported on the discovery of two simultaneous kHz QPOs during the peak of the outburst with frequencies of $\sim$700 and $\sim$500 Hz (Fig. \[fig:khzqpos\_1808\] top panel). This was the first detection of twin kHz QPOs in a source with a known spin-frequency. The frequency separation of those two kHz QPOs is only $\sim$200 Hz, significantly below the 401 Hz expected in the beat-frequency models proposed to explain the kHz QPOs. Therefore, those models are falsified by the discovery of kHz QPOs in SAX J1808.4–3658. The fact that the peak separation is approximately half the spin frequency suggests that the kHz QPOs are indeed connected to the neutron-star spin frequency, albeit in a way not predicted by any existing model at the time of the discovery. The lower-frequency kHz QPO was only seen during the peak of the outburst (October 16, 2002) but the higher-frequency kHz QPO could be traced throughout the main part of the outburst (Wijnands et al. 2003). In addition to the twin kHz QPOs, a third kHz QPO was found with frequencies ($\sim$410 Hz) just exceeding the pulse frequency (Fig. \[fig:khzqpos\_1808\] bottom panel; Wijnands et al. 2003). The nature of this QPO is unclear but it might be related to the side-band kHz QPO seen in several other sources (Jonker et al. 2000). Wijnands et al. (2003) pointed out that there appear to exist two classes of neutron-star LMXBs: the ’fast’ and the ’slow’ rotators. The fast rotators have spin frequencies $>$$\sim$400 Hz and the frequency separation between the kHz QPOs is roughly equal to half the spin frequency. In contrast, the slow rotators have spin frequencies below $<$$\sim$400 Hz and a frequency separation roughly equal to the spin frequency. These latest kHz QPO results have spurred new theoretical investigations into the nature of kHz QPO, involving spin induced resonance in the disk (e.g., Wijnands et al. 2003; Kluzniak et al. 2004; Lee et al. 2004; Lamb & Miller 2004; Kato 2004). [c]{} ### The low-frequency QPOs \[subsubsection:lfqpos\] During the peak of the outburst and in its subsequent decay, broad-noise and QPOs with frequencies between 10 and 80 Hz were detected in the power spectra (Fig. \[fig:lfqpos\]). Similar phenomena have been observed in other non-pulsating systems and are likely to be related to the noise components seen in SAX J1808.4–3658. Van Straaten et al. (2004, 2005) have studied the broad-band power spectra (including the noise components, the low-frequency QPOs, and the kHz QPOs) of SAX J1808.4–3658 in detail as well as the frequency correlations between the different power-spectral components. Interestingly, using those frequency correlations, van Straaten et al. (2004, 2005) suggested that the higher-frequency kHz QPO could also be identified during the 1998 outburst but at the lowest frequencies found so far in any kHz QPO source (down to $\sim$150 Hz). Previous work (Wijnands & van der Klis 1998b) on the aperiodic timing features of SAX J1808.4–3658 during its 1998 outburst had already found these features but they could not be identified as the higher-frequency kHz QPO due to their low frequency and broad character. Van Straaten et al. (2004, 2005) also compared the results of SAX J1808.4–3658 with those obtained for other non-pulsating neutron-star LMXBs. In those other sources, the frequencies of the variability components follow an universal scheme of correlations. The correlations observed for SAX J1808.4–3658 are similar but they show a shift in the frequencies of the kHz QPOs. It is unclear what physical mechanism(s) underlies this difference among sources (van Straaten et al. 2004, 2005). During the 1998 and 2002 outbursts of SAX J1808.4–3658, the source exhibited similar X-ray fluxes. However, at similar flux levels, the characteristic frequencies observed during the 1998 outburst are much lower (by a factor of$\sim$10) than during the 2002 outburst (van Straaten et al. 2005; see Fig. \[fig:1808\_1998v2002\]). Again it is unclear what causes this huge difference between the two outbursts but it might be related to the ’parallel track’ phenomena observed for the kHz QPOs in the non-pulsating neutron-star LMXBs (e.g., van der Klis 2000). -----     ----- ------      ------ ### The violent 1 Hz flaring \[subsubsection:1hzflaring\] Violent flaring was observed on many occasions at a $\sim$1 Hz repetition frequency during the late stages of the 2002 outburst (Fig. \[fig:1hz\]), similar to what had been observed during the 2000 outburst. This proves that also this violent flaring is a recurrent phenomenon and can likely be observed every time the source is in this prolonged low-level activity state. Preliminary results presented in Figure \[fig:1hz\] (right panels) show examples of power spectra obtained during the end stages of the 2002 outburst. During certain observations the 1 Hz QPO is rather narrow and its first overtone can be seen clearly (Fig. \[fig:1hz\] top right panel). During other observations, the 1 Hz QPO is much broader and its wings blend with the first overtone (Fig. \[fig:1hz\] middle two right panels). In addition to the 1 Hz QPO, other QPOs around 30–40 Hz are sometimes seen (see also van Straaten et al. 2005). It is unclear if this 30–40 Hz QPO is related to the low-frequency QPOs discussed in § \[subsubsection:lfqpos\] or if it is due to a different mechanism. During certain observations the 1 Hz QPO becomes very broad, turning into a band-limited noise component (Fig. \[fig:1hz\] bottom right panel). The mechanism behind these violent flares is not yet known and a detailed analysis of this phenomenon is in progress. ### The pulsations \[subsubsection:pulsations\] The pulsations could be detected at all flux levels with an amplitude of 3% – 10%. There was no evidence for a 200.5 Hz subharmonic in the data (upper limit of 0.38% of the signal at 401 Hz; Wijnands et al. 2003) confirming the interpretation of 401 Hz as the pulsar spin frequency. A detailed analysis and discussion of the coherent timing analysis will be presented by Morgan et al. (2005, in preparation). ### Observations at other wavelengths \[subsubsection:otherwavelengths\] Rupen et al. (2002b) reported the detection of SAX J1808.4–3658 at radio wavelengths. On October 16, 2002, they found a 0.44-mJy source at 8.5 GHz and a day later, the source was detected at 0.3 mJy. Monard (2002) reported that on October 16, 2002, the optical counterpart was detectable again at magnitudes similar to those observed at the peak of the 1998 outburst. SAX J1808.4–3658 in quiescence \[subsection:1808q\] --------------------------------------------------- In quiescence, SAX J1808.4–3658 has been observed on several occasions with the [*BeppoSAX*]{} and [*ASCA*]{} satellites (Stella et al. 2000; Dontani et al. 2000; Wijnands et al. 2002). The source was very dim in quiescence, with a luminosity close to or lower than $10^{32}$ ergs s$^{-1}$. Due to the low number of source photons detected, these luminosities had large errors and no information could be obtained on the spectral shape or possible variability in quiescence. Due to the limited angular resolution of [*BeppoSAX*]{}, doubts were raised as to whether the source detected by this satellite was truly SAX J1808.4–3658 or an unrelated field source (Wijnands et al. 2002). Campana et al. (2002) reported on a quiescent observation of the source performed with [*XMM-Newton*]{} which resolved this issue. They detected the source at a luminosity of $5 \times 10^{31}$ ergs s$^{-1}$ and found that the field around SAX J1808.4–3658 is rather crowded with weak sources. Two such sources are relatively close to SAX J1808.4–3658 and might have conceivably caused a systematic positional offset during the [*BeppoSAX*]{} observations. Despite this fact it is very likely that SAX J1808.4–3658 was indeed detected during those [*BeppoSAX*]{} observations. Using [*XMM-Newton*]{}, Campana et al. (2002) obtained enough photons to extract a quiescent X-ray spectrum, which was not dominated by the same thermal component seen in other quiescent neutron star transients; such a thermal component is thought to be due to the cooling of the neutron star in-between outbursts. However, the spectrum of SAX J1808.4–3658 was dominated by a power-law shaped component. The non-detection of the thermal component was used to argue that the neutron star was anomalously cool, possibly due to enhanced core cooling processes (Campana et al. 2002). It has been argued (Stella et al. 2000; Campana et al. 2002) that the propeller mechanism, which might explain (some of) the hard X-ray emission in quiescence, is not likely to be active since this mechanism is expected to stop operating at luminosities $<$$10^{33}$ ergs s$^{-1}$, because at those luminosities the source should turn on as a radio pulsar. Instead, it was proposed that the quiescent X-rays originate in the shock between the wind of a turned-on radio pulsar and the matter flowing out from the companion star (Stella et al. 2000; Campana et al. 2002). Di Salvo & Burderi (2003) suggested that the quiescent X-rays could also be due to direct dipole radiation from the radio pulsar. Using simple accretion disk physics and the quiescent luminosity found by Campana et al. (2002), they determined that the magnetic field strength of the neutron star in SAX J1808.4–3658 should be in a quite narrow range of $(1 - 5) \times 10^8$ Gauss. The quiescent optical counterpart of SAX J1808.4–3658 was studied by Homer et al. (2001). They reported that on August 10, 1999, the orbital modulation was still present in white light observations (estimated V magnitude of $\sim$20), with a semi-amplitude of $\sim$6%. It has the same phasing and approximately sinusoidal modulation as seen during outburst, and with photometric minimum when the pulsar is behind the companion star. During observations taken in July 2000 the quiescent counterpart was even fainter and no significant orbital modulation could be detected. Using these results, it has been suggested that the optical properties of SAX J1808.4–3658 in quiescence are evidence of an active radio pulsar (Burderi et al. 2003). Campana et al. (2004) reported on the first optical spectrum of this source during its quiescent state. They concluded that a very high irradiating luminosity, a factor of $\sim$100 larger than directly observed from the X-rays, must be present in the systems, which was suggested to be derived from a rotation-powered neutron star. If true, a pulsating radio source might be expected, but a search at 1.4 GHz could not detected the source (Burgay et al. 2003). This could be due to the effects of free-free absorption and searches at higher frequencies to limit these effects might still yield a pulsating radio source during the quiescent state of SAX J1808.4-3658. XTE J1751–305 \[section:1751\] ============================== The 2002 outburst \[subsection:1751\_2002outburst\] --------------------------------------------------- The second accretion-driven millisecond pulsar (XTE J1751–305) was discovered on April 3, 2002 (Markwardt et al. 2002a). Its spin frequency is 435 Hz (Fig. \[fig:pulsations\]) and the neutron star is in a very small binary with an orbital period of only 42 minutes. The timing analysis of the pulsations gave a minimum mass for the companion star of 0.013 solar mass and a pulse-frequency derivative of $<$$3 \times 10^{-13}$ Hz s$^{-1}$. Assuming that the mass transfer in this binary system was driven by gravitational radiation, the distance toward the source could be constrained to at least 7 kpc and the orbital inclination to 30$^\circ$–85$^\circ$, resulting in a companion mass of 0.013–0.035 solar masses, suggesting a heated helium dwarf (Markwardt et al. 2002a). [*Chandra*]{} briefly observed the source, resulting in an arcsecond position (Markwardt et al. 2002a). The source reached a peak luminosity of $>$$2 \times 10^{37}$ ergs s$^{-1}$, an order of magnitude brighter than the peak luminosity of SAX J1808.4–3658. However, the outburst was very short with an e-folding time of only $\sim$7 days (compared to $\sim$14 days for SAX J1808.4–3658; Fig. \[fig:lc\]) resulting in a low outburst fluence of only $\sim$$2.5 \times 10^{-3}$ ergs cm$^{-2}$ (Markwardt et al. 2002a). A potential re-flare was seen two weeks after the end of the outburst during which also a type-I X-ray burst was seen. Preliminary analysis of the burst indicated that the burst did not come from XTE J1751–305 but from another source in the field of view. This was later confirmed (In ’t Zand et al. 2003) and the burst likely originated from the bright X-ray transient in Terzan 6. It was also determined that the transient in Terzan 6 could not have produced the re-flare (In ’t Zand et al. 2003) suggesting that this re-flare could still have come from XTE J1751–305. However, van Straaten et al. (2005) suggested (based on a X-ray color study using [*RXTE*]{}/PCA observations) that this re-flare was emitted by one of the background sources and not by XTE J1751–305. Van Straaten et al. (2005) also investigated the aperiodic timing properties of the source (an example power spectrum is shown in Fig. \[fig:lfqpos\]) and the correlations between the characteristic frequencies of the observed power-spectral components. The frequency correlations were similar to those of the non-pulsating neutron-star LMXBs. In contrast with the results obtained for SAX J1808.4–3658 (§ \[subsubsection:lfqpos\]), no frequency shift was required for XTE J1751–305 to make the frequency correlations consistent with those of the non-pulsating sources. Using these correlations, van Straaten et al. (2005) suggested that the highest-frequency noise components in XTE J1751–305 are likely due to the same physical mechanisms as the kHz QPOs. They also investigated the correlations between the characteristic frequencies and the X-ray colors of the source and concluded that it did not behave like an atoll source. A previous outburst in June 1998 was discovered using archival [*RXTE*]{}/ASM data (Markwardt et al. 2002a), suggesting a tentative recurrence time of $\sim$3.8 years. Miller et al. (2003) reported on high spectral resolution data of the source obtained with [*XMM-Newton*]{} to search for line features in the X-ray spectrum. However, they only detected a continuum spectrum dominated by a hard power-law shaped component (power-law index of $\sim$1.44) but with a 17% contribution to the 0.5–10 keV flux from a soft thermal (black-body) component with temperature of $\sim$1 keV. Gierlinski & Poutanen (2004) studied in detail the X-ray spectrum of the source during its 2002 outburst using the archival [*RXTE*]{} and [*XMM-Newton*]{} observations. They find that XTE J1751–305 exhibited very similar behavior as SAX J1808.4–3658 during its 1998 outbursts. They also find that the pulse profile cannot be described by a simple sinusoid, but that a second harmonic is needed (the peak-to-peak amplitude of the fundamental was found to be 4.5% but of the second harmonic only 0.15%). Gierlinski & Poutanen (2004) also report that a clear energy dependency of the pulse profile was observed and that the higher energy photons arrive earlier than the softer (this ’soft lag’ reached $\sim$100 $\mu$s at about 10 keV, where it saturated). Searches for the optical and near-infrared counterparts were performed but no counterparts were found (Jonker et al. 2003), likely due to the high reddening toward the source. These non-detections did not constrain any models for the accretion disk or possible donor stars. XTE J1751–305 in quiescence \[subsection:1751q\] ------------------------------------------------ Recently, XTE J1751–305 was observed in quiescence using [*Chandra*]{} (Wijnands et al. 2005). Sadly, they could not detect the source in their $\sim$43 ksec observation, with 0.5–10 keV flux upper limits between 0.2 and $2.7 \times 10^{-14}$ ergs s$^{-1}$ cm$^{-2}$ depending on assumed spectral shape, resulting in 0.5–10 keV luminosity upper limits of $0.2 - 2 \times 10^{32}$ ($d$/ 8 kpc)$^2$ ergs s$^{-1}$, with $d$ the distance toward the source in kpc. Using simple accretion disk physics in combination with these luminosity upper limits, Wijnands et al. (2005) could constrain the magnetic field of the neutron star in this system to be less than $3 - 7 \times 10^8 {d \over {\rm 8~kpc}}$ Gauss (depending on assumed spectral shape of the quiescent spectrum). XTE J0929–314 \[section:0929\] ============================== The 2002 outburst \[subsection:0929\_2002outburst\] --------------------------------------------------- The third accretion-driven millisecond X-ray pulsar XTE J0929–314 had already been detected with the [*RXTE*]{}/ASM on April 13, 2002 (Remillard 2002) but was only found to be harboring a millisecond pulsar with a pulsations frequency of 185 Hz (Fig. \[fig:pulsations\]) on May 2nd when observations of the source were made using the [*RXTE*]{}/PCA (Remillard et al. 2002). Galloway et al. (2002) reported on the detection of the 44-min orbital period of the system which is remarkably similar to that of XTE J1751–305. A minimum mass of 0.008 solar mass was obtained for the companion star and a pulse-frequency derivative of $(-9.2 \pm 0.4) \times 10^{-14}$ Hz s$^{-1}$. Galloway et al. (2002) suggested that this spin down torque may arise from magnetic coupling to the accretion disk, a magneto-hydrodynamic wind, or gravitational radiation from the rapidly spinning neutron star. Assuming gravitational radiation as the driving force behind the mass transfer, Galloway et al. (2002) found a lower limit to the distance of 6 kpc. They also reported on the detection of a QPO at 1 Hz (Fig. \[fig:lfqpos\]). Full details of this QPO and the other aperiodic power-spectral components are presented by van Straaten et al. (2005). Just as they found for SAX J1808.4–3658, the frequency correlations for XTE J0929–314 were similar to those observed for the non-pulsating sources but with an offset in the frequencies of the highest-frequency components. These correlations allowed van Straaten et al. (2005) to identify those components as related to the kHz QPOs. Studying the correlated spectral and timing variability, they concluded that the behavior of XTE J0929–314 was consistent with that of an atoll source. Juett et al. (2003) obtained high resolution spectral data using the [*Chandra*]{} gratings. Again the spectrum is well fitted by a power-law plus a black body component, with a power-law index of 1.55 and a temperature of 0.65 keV. Similar to XTE J1751–305, no emission or absorption features were found. No orbital modulation of the X-ray flux was found implying an upper limit on the inclination of 85$^\circ$. Greenhill et al. (2002) reported the discovery of the optical counterpart of the system with a V magnitude of 18.8 on May 1st, 2002 (see also Cacella 2002). Castro-Tirado et al. (2002) obtained optical spectra of the source on May 6–8 in the range 350–800 nm and found emission lines from the C III - N III blend and H-alpha, which were superposed on a blue continuum. These optical properties are typical of X-ray transients during outburst. Rupen et al. (2002a) discovered the radio counterpart of the source using the VLA with 4.86 GHz flux of 0.3–0.4 mJy. XTE J0929–314 in quiescence \[subsection:0929q\] ------------------------------------------------ Recently, Wijnands et al. (2005) also observed XTE J0929–314 in its quiescent state with [*Chandra*]{}. For this source, they detected 22 source photons (in the energy range 0.3–8 keV) in $\sim$24.4 ksec of on-source time. This small number of photons detected did not allow for a detailed spectral analysis of the quiescent spectrum, but they could demonstrate that the spectrum is harder than a simple thermal emission (which might have been due to the cooling neutron star that has been heated during outbursts). Assuming a power-law spectral model for the time-averaged (averaged over the whole observation) X-ray spectrum, they obtained a power-law index of $\sim$1.8 and an unabsorbed X-ray flux of $\sim$$6 \times 10^{-15}$ ergs s$^{-1}$ cm$^{-2}$ (for the energy range 0.5–10 keV), resulting in a 0.5–10 keV X-ray luminosity of $\sim$$7\times 10^{31}$ ($d$/10 kpc)$^2$ ergs s$^{-1}$, with $d$ the distance in kpc. The thermal component usually seen in quiescent neutron star LMXBs could not be detected, with a maximum contribution to the 0.5–10 keV flux of $\sim$30%. Wijnands et al. (2005) also found that the quiescent count rate of XTE J0929–314 was variable at the 95% confidence level, but no conclusive evidence was found for associated spectral variability. The properties of XTE J0929–314 in its quiescent state are remarkably similar to that observed for SAX J1808.4–3658 (§ \[subsection:1808q\]) which might suggest that such behavior is common among accretion-driven millisecond X-ray pulsars. However, recent work on several other weak quiescent neutron-star X-ray binaries (e.g., Jonker et al. 2004a, b ; Tomsick et al. 2004), which do not exhibit pulsations during their X-ray outbursts, suggests that also such systems can resemble SAX J1808.4–3658 during quiescent (i.e., they could be almost as faint and hard as SAX J1808.4–3658 in quiescence; see Wijnands et al. 2005 for a in-dept discussion). Similar to what they did for XTE J1751–305, Wijnands et al. (2005) could constrain the neutron-star magnetic field strength in XTE J0929–314 to be $<$$3\times 10^9 {d\over {\rm 10~kpc}}$ Gauss. XTE J1807–294 \[section:1807\] ============================== [c]{} The fourth millisecond X-ray pulsar XTE J1807–294 with a frequency of 191 Hz, was discovered on February 21, 2003 (Markwardt et al. 2003a; Fig. \[fig:pulsations\]). The peak flux was only 58 mCrab (2–10 keV, measured on February 21, 2003). The orbital period was determined (Markwardt et al. 2003c) to be $\sim$40 minutes making it the shortest orbital period of all accretion-driven millisecond pulsars now known. Markwardt et al. (2003c) reported the best known position of the source based on a [*Chandra*]{} observation. Using the [*RXTE*]{}/PCA data, kHz QPOs have been detected for this system and it was found that the frequency separation between the two kHz QPOs was consistent with being equal to the neutron-star spin frequency (Markwardt et al. 2005 in preparation; Fig. \[fig:khz1807\]). This makes XTE J1807–294 consistent with the classification of Wijnands et al. (2003) of the neutron-star LMXBs into ’fast’ and ’slow’ rotators, with XTE J1807–294 a slow rotator. A detailed analysis of the correlations between the kHz QPOs and the low-frequency features (see Fig. \[fig:lfqpos\]) in this source will be reported by Linares et al. (2005, in preparation). The preliminary results of that analysis shows that also for XTE J1807–294 the frequency correlations are similar to those observed for the non-pulsating sources but with an offset in the frequencies of the highest-frequency components (similar to what was found for SAX J1808.4–3658 and possible XTE J0929–314; Van Straaten et al. 2005). Campana et al. (2003) reported on a [*XMM-Newton*]{} observation of this source taken on March 22, 2003. Assuming a distance of 8 kpc, the 0.5–10 keV luminosity during that observation was $2 \times 10^{36}$ ergs s$^{-1}$. They could detect the pulsations during this observation with a pulsed fraction of 5.8% in the 0.3–10 keV band (increasing with energy) and a nearly sinusoidal pulse profile. Furthermore, using the same data Kirsch et al. (2004; see also Kirsch & Kendziorra 2003) reported on the mass function of this system and found a minimal mass for the companion star of 0.007 $M_\odot$ when assuming a canonical neutron star mass of 1.4 $M_\odot$. The spectral data are well fit by a continuum model, assumed to be an absorbed Comptonisation model plus a soft component. The latter component only contributed 13% to the flux. Again no emission or absorption lines were found. No detections of the counterparts of the system at other wavelengths have been reported so far. XTE J1814–338 \[section:1814\] ============================== The fifth system (XTE J1814–338) was discovered on June 5, 2003 and has a pulse frequency of 314 Hz (Markwardt et al. 2003b; Fig. \[fig:pulsations\]), with an orbital period of 4.3 hr and a minimum companion mass of 0.15 solar mass (Markwardt et al. 2003d). This 4.3 hr orbital period makes it the widest binary system among the accretion-driven millisecond pulsars and also the one most similar to the general population of low-luminosity neutron star LMXBs (the atoll sources). XTE J1814–338 exhibited many type-I X-ray bursts, which showed burst oscillations with a frequency consistent with the neutron star spin frequency (Markwardt et al. 2003d, Strohmayer et al. 2003). A distance of $\sim$8 kpc was obtained from the only burst which likely reached the Eddington luminosity. The burst oscillations are strongly frequency- and phase-locked to the persistent pulsations (as was also seen for SAX J1808.4–3658; Chakrabarty et al. 2003) and two bursts showed evidence of a frequency decrease of a few tenths of a Hz during the onset of the burst, suggesting a spin down. Strohmayer et al. (2003) also reported on the detection of the first harmonic of the burst oscillations, which is the first time that this has been seen in any burst-oscillation source. This harmonic could arise from two hot-spots on the surface, but Strohmayer et al. (2003) suggested that if the burst oscillations arise from a single bright region, the strength of the harmonic would suggest that the burst emission is beamed (possibly due to a stronger magnetic field strength than in non-pulsating LMXBs). Bhattacharyya et al. (2004) used the non-sinusoidal burst oscillation light curves to constrain the parameters of the neutron star in XTE J1814–338; they obtained a dimensionless radius to mass ratio of $Rc^2/GM = 3.9 - 4.9$ for the neutron star in this source. They find that the bursting hot spot on the neutron-star surface remains always large, with an angular radius $>$$25^\circ$. Their study also suggest that the inclination of the source is greater than 50$^\circ$ and that the secondary companion is a hydrogen main sequence star that is significantly bloated (possibly due to X-ray heating). Wijnands & Homan (2003) analyzed the [*RXTE*]{}/PCA data of the source obtained between June 8 and 11, 2003. The overall shape of the 3–60 keV power spectrum is dominated by a strong broad band-limited noise component (Fig. \[fig:lfqpos\]), which could be fitted by a broken power-law model with a broad bump superimposed on it at frequencies above the break frequency. These characteristics make the power spectrum of XTE J1814–338 very similar to that observed in the non-pulsing low-luminosity neutron-star LMXBs (the atoll sources) when they are observed at relatively low X-ray luminosities (i.e., in the so-called island state). This is consistent with the hard power-law X-ray spectrum of the source reported by Markwardt et al. (2003d). This resemblance of XTE J1814–338 to the atoll sources was further strengthened (Wijnands & Homan 2003) by the fact that the source is consistent with the relation between the break frequency and the frequency of the bump found for atoll sources by Wijnands & van der Klis (1999). Van Straaten et al. (2005) performed an in-depth analysis of all publicly available [*RXTE*]{}/PCA data of XTE J1814–338 to study the power-spectral components and the correlations between their characteristic frequencies. Using those correlations and by comparing them to other sources, they could identify several components that are related to the kHz QPOs. They also found that the frequency correlations were identical to the non-pulsating sources with no need for a frequency shift. This is similar to what they found for XTE J1751–305 but different from SAX J1808.4–3658 and XTE J0929–314 (and for XTE J1807–294 as found by Linares et al. 2005, in preparation). The reason(s) for this difference between accreting millisecond X-ray pulsars is not know (see Van Straaten et al. 2005 for an extended discussion). From the correlations between the spectral and timing variability it was confirmed that the behavior of XTE J1814–338 was consistent with that of an atoll sources (van Straaten et al. 2005; see also Wijnands & Homan 2003). Wijnands & Reynolds (2003) reported that the position of XTE J1814–338 was consistent with the [*EXOSAT*]{} slew source EXMS B1810–337 which was detected on September 2nd, 1984. If this identification is correct, then its recurrence time can be inferred to be less than 19 years but more than 5 years (the time since the [*RXTE*]{}/PCA bulge scan observations started in February 1999), unless the recurrence time of the source varies significantly. Krauss et al. (2003) reported the best position of the source based on a [*Chandra*]{} observation and on the detection of the likely optical counterpart of the source (with magnitudes of B = 17.3 and R = 18.8 on June 6). Steeghs (2003) reported on optical spectroscopy of this possible counterpart, specifically on the discovery of prominent hydrogen and helium emission lines, confirming the connection between the optical source and XTE J1814–338. IGR J00291+5934 =============== Very recently, on December 2, 2004, the European Gamma-ray satellite [*INTEGRAL*]{} discovered a new X-ray transient named IGR J00291+5934 (Eckert et al. 2004). A day later, [*RXTE*]{} observed the source and it was found that this source harbors a 598.88 Hz accretion-driven millisecond X-ray pulsar (Markwardt et al. 2004a; see Galloway et al. 2005 for the analyze of all data obtained for this source). The pulsed amplitude was approximately 6% with no evidence for harmonics. The amplitude decreased with increasing photon energy and the soft photons arrive later than the hard ones (by up to 85 $\mu$s; Galloway et al. 2005) The X-ray spectrum could be fitted with an absorbed power-law model with photon index of 1.7 and a column density of $7\times 10^{21}$ cm$^{-2}$. In Figure \[fig:lfqpos\] the power spectrum between 0.001 and 10,000 Hz is show of the source, clearly showing significant aperiodic variability (see also Markwardt et al. 2004a). Interestingly, of all six accreting millisecond X-ray pulsars the break in the power spectrum is at the lowest frequency for IGR J00291+5934: during the peak of the outburst the break frequency for this source was $\sim$0.01 Hz compared to $>$0.1 Hz for the other sources at their outburst peaks. Markwardt et al. (2004b) used additional [*RXTE*]{} observations to determine that the orbital period of the system was $\sim$2.45 hours and they obtained a mass function of $(2.81 \pm 0.02) \times 10^{-5}$ M$_\odot$ (see also Galloway et al. 2005). For a neutron star mass of 1.4 M$_\odot$ this results in a lower limit on the mass for the companion star of 0.038 M$_\odot$, possible a brown dwarf (Galloway et al. 2005). The source reached its peak flux (29 mCrab; 2–10 keV) on December 3, 2004, after which the fluxes decreased in a linear way, until around December 11, 2004, when the rate of decline had increased slightly (Swank & Markwardt 2004). On December 14, 2004, [*Chandra*]{} performed a brief ($\sim$18 ksec) observation of the source using the ACIS-S/HETG combination. Nowak et al. (2004) reported that the source was at a flux level of $\sim$1 mCrab and its X-ray spectrum could be well-fitted by an absorbed power-law model with a column density of $\sim$$3\times 10^{21}$ cm$^{-2}$ and a photon index of $\sim$1.9. A possible iron line feature was also reported by the same authors. The flux value confirmed the steady decline as seen by Swank & Markwardt (2004) and soon after this the source could not be detected anymore with [*RXTE*]{}/PCA (see Galloway et al. 2005 for the full X-ray light curve of the source during this outburst). No X-ray bursts, dips, eclipses, or kHz QPOs (with upper limits of $\sim$1% rms) were found (Galloway et al. 2005). [c]{} An optical counterpart was proposed by Fox & Kulkarni (2004; see this reference for a finding chart) with an R magnitude during outburst of $\sim$17.4. The detection of broad emission lines of HeII and H$_\alpha$ from this tentative optical counterpart strongly support this identification (Roelofs et al. 2004; see also Filippenko et al. 2004). Pooley (2004) found a 1.1 mJy radio source (15 GHz) at a position consistent with that of the optical counterpart which likely is the radio counterpart of the source which faded during the outburst (Fender et al. 2004, who also detected the source at 5 GHz at a flux of $\sim$250 $\mu$Jy), although the decay did not seem to be very rapid (Rupen et al. 2004). Steeghs et al. (2004) reported the detection of a decaying infrared counterpart of the source. After the discovery of this new source, Remillard (2004) constructed a mission-long light curve using the [*RXTE*]{}/ASM data and found that in November 26–28, 1998, and in September 11–21, 2001, the source might have exhibited other outbursts (see Fig. \[fig:0029\_ASM\] for the [*RXTE*]{}/ASM light curve of the December 2004 and September 2001 outbursts of the source). If confirmed this would give a recurrence time for the outburst of approximately 3 years. In ’t Zand & Heise (2004) did not detected the source with the WFCs aboard [*BeppoSAX*]{} during a net exposure time of 2.9 Msec on the source. The [*RXTE*]{}/ASM detections reported by Remillard (2004) were not covered; the first WFCs data were obtained 16 days and 11 days after the two possible outbursts, respectively, and it is well possible that the source had by then decayed below the sensitivity limit of the WFCs. Theoretical work \[section:theory\] =================================== This chapter is intended to be an observational overview, thus I will not go into detail on the theoretical papers published on accretion-driven millisecond pulsars. Instead, I will briefly list some of those papers, most of which focus on SAX J1808.4–3658 since the other five systems have only been found recently (with IGR J00291+5934 discovered very recently). Since the discovery of SAX J1808.4–3658, several studies have tried to constrain the properties (i.e., radius, mass, magnetic field strength) of the neutron star in this system (Burderi & King 1998; Psaltis & Chakrabarty 1999; Bhattacharyya 2001), while others proposed that the compact object is not a neutron star at all, but instead a strange star (see, e.g., Li et al. 1999; Datta et al. 2000; Zdunik et al. 2000 and references there-in). Other studies focused on the evolutionary history of SAX J1808.4–3658 (Ergma & Antipova 1999) or on the nature of the companion star (Bildsten & Chakrabarty 2001, who suggested a brown dwarf companion star). Recently, Nelson & Rappaport (2003) investigated the evolutionary history of ultracompact binaries such as XTE J0929–314 and XTE J1751–305 (they focused on those two systems but likely their conclusions can also be applied to XTE J1807–294). Rappaport et al. (2004) investigated how accretion could occur in millisecond X-ray pulsars and they postulated that those systems can continue to accrete from a thin disk, even for accretion rates that place the magnetospheric radius well beyond the co-rotation radius. The discovery of the accretion-driven millisecond pulsars raises an important question: why are those systems different from other neutron star LMXBs for which no pulsations have been found. Cumming et al. (2001; see also Rai Choudhuri & Konar 2002; Cumming 2004) suggested that the low time-averaged accretion rate of SAX J1808.4–3658 might explain why this source is a pulsar. Although the remaining four pulsars were not know at the time of writing of that paper, the same arguments can be used for those systems: when the time-averaged accretion rate is sufficiently high, the neutron star magnetic field might be buried by the accreted matter and does not have time to dissipate through the accreted material. However, for the millisecond X-ray pulsars the time-averaged accretion rate is sufficiently low that the magnetic field dissipation can indeed happen, giving those systems a magnetic field still strong enough to disturb the flow of the accreted matter. However, more neutron-star LMXBs with low time-averaged accretion rate must be found and studied in detail to verify that they all indeed harbor a millisecond pulsar. If an exception is found, the screening model might only be part of the explanation and alternative ideas need to be explored (see, e.g., Titarchuk et al. 2002). Conclusion \[section:conclusions\] ================================== From this review it is clear that [*RXTE*]{} has played a vital role in the discovery and study of accretion-driven millisecond X-ray pulsars. The detailed studies performed with [*RXTE*]{} for those systems have yielded break-throughs in our understanding of kHz QPOs and burst oscillations. Furthermore, three of these accreting pulsars are in ultrashort binaries which will constrain evolutionary paths for this type of systems (e.g., see Nelson & Rappaport 2003). However, it is also clear that the six systems do not form a homogeneous group; their pulsation frequencies span the range between 185 Hz and 599 Hz (Fig. \[fig:pulsations\]), their orbital periods fall between 40 minutes and 4.3 hours, their X-ray light curves are very different (Fig. \[fig:lc\]), and also their aperiodic variability properties (Fig. \[fig:lfqpos\]). More, well studied outbursts of the currently known systems are needed as well as discoveries of additional systems. At the moment only [*RXTE*]{} is capable of performing the necessary timing observations. After [*RXTE*]{}, an instrument with similar or better capabilities is highly desirable for our understanding of accretion-driven millisecond pulsars and their connection with the non-pulsating neutron-star LMXBs. Acknowledgments {#acknowledgments .unnumbered} =============== I thank Stefanie Wachter for kindly providing the optical data used in Figure \[fig:2000\_lc\]. References ========== Bhattacharya, D., & van den Heuvel, E.P.J., 1991, [*Physics Reports*]{}, 203, 1–124 Bhattacharyya, S., 2001, [*ApJ*]{}, 554, L185–L188 Bhattacharyya, S. et al., 2004, [*ApJ*]{}, in press (astro-ph/0402534) Bildsten, L. & Chakrabarty, D., 2001, [*ApJ*]{}, 557, 292–296 Burderi, L. & King. A.R., 1998, [*ApJ*]{}, 505, L135–L137 Burderi, L. et al., 2003, [*A&A*]{}, 404, L43–L46 Burgay, M., et al., 2003, [*ApJ*]{}, 589, 902–910 Cacella, P., 2002, [*IAUC*]{}, 7893 Campana, S. et al., 2002, [*ApJ*]{}, 575, L15–L19 Campana, S. et al., 2003, [*ApJ*]{}, 594, L39–L42 Campana, S. et al., 2004, [*ApJ*]{}, 614, L49–L52 Castro-Tirado, A.J. et al., 2002, [*IAUC*]{}, 7895 Chakrabarty, D. 2004, To appear in [*Binary Radio Pulsars*]{}, F. A. Rasio & I. H. Stairs (eds.), ASP Conf. Ser. (astro-ph/0408004) Chakrabarty, D. & Morgan, E. H., 1998, [*Nature*]{}, 394, 346–348 Chakrabarty, D. et al., 2003, [*Nature*]{}, 424, 42–44 Cui, W. et al., 1998, [*ApJ*]{}, 504, L27–L30 Cumming, A. 2004, To appear in [*Binary Radio Pulsars*]{} F. A. Rasio & I. H. Stairs (eds.), ASP Conf. Ser. (astro-ph/0404518) Cumming, A. et al., 2001, [*ApJ*]{}, 557, 958–966 Datta, B. et al., 2000, [*A&A*]{}, 355, L19–L22 Di Salvo, T. & Burderi, L., 2003, [*A&A*]{}, 397, 723–727 Dotani, T. et al., 2000, [*ApJ*]{}, 543, L145–L148 Eckert, D. et al., 2004, [*ATEL*]{}, 352 Ergma, E. & Antipova, J., 1999, [*A&A*]{}, 343, L45–L48 Fender, R. et al., 2004, [*ATEL*]{}, 361 Filippenko, A. V. et al., 2004, [*ATEL*]{}, 366 Ford, E.C., 2000, [*ApJ*]{}, 535, L119–L122 Fox, D. B. & Kulkarni S. R., 2004, [*ATEL*]{}, 354 Gaensler, B.M. et al., 1999, [*ApJ*]{}, 522, L117–L119 Galloway, D.K. et al., 2002, [*ApJ*]{}, 576, L137–L140 Galloway, D.K. et al., 2005, ApJL, submitted (astro-ph/0501064) Gierlinski, M. & Poutanen, J., 2004, [*MNRAS*]{}, submitted (astro-ph/0411716) Gierlinski, M. et al., 2002, [*MNRAS*]{}, 331, 141–153 Giles, A.B. et al., 1998, [*IAUC*]{}, 6886 Giles, A.B. et al., 1999, [*MNRAS*]{}, 304, 47–51 Gilfanov, M. et al., 1998, [*A&A*]{}, 338, L83–L86 Greenhill, J.G. et al., 2002, [*IAUC*]{}, 7889 Heindl, W.A. & Smith, D.M., 1998, [*ApJ*]{}, 506, L35–L38 Heise, J. et al., 1999, [*ApL&C*]{}, 38, 297–300 Homer, L. et al., 2001, [*MNRAS*]{}, 325, 1471–1476 Jonker, P.G. et al., 2000, [*ApJ*]{}, 540, L29–L32 Jonker, P.G. et al., 2003, [*MNRAS*]{}, 344, 201–206 Jonker, P.G. et al., 2004a, [*MNRAS*]{}, 349, 94–98 Jonker, P.G. et al., 2004b, [*MNRAS*]{}, 354, 666–674 Juett, A.M. et al., 2003, [*ApJ*]{}, 587, 754–760 In ’t Zand, J. & Heise, J., 2004, [*ATEL*]{}, 362 In ’t Zand, J.J.M. et al., 1998, [*A&A*]{}, 331, L25–L28 In ’t Zand, J.J.M. et al., 2001, [*A&A*]{}, 372, 916–921 In ’t Zand, J.J.M. et al., 2003, [*A&A*]{}, 409, 659–663 Kato, S., 2004, [*PASJ*]{}, 56, 905–922 Kazarovets, E.V. et al., 2000, [*IBVS*]{}, 4870 Kirsch, M.G.F. & Kendziorra, E., 2003, [*ATEL*]{}, 148 Kirsch, M. G. F. et al., 2004, [*A&A*]{}, 423, L9–L12 Kluzniak, W. et al., 2004, [*ApJ*]{}, 603, L89–L92 Krauss, M.I. et al., 2003, [*IAUC*]{}, 8154 Lamb, F.K. & Miller, M.C., 2004, [*ApJ Letters*]{}, submitted (astro-ph/0308179) Lee, W. H., et al., 2004, [*ApJ*]{}, 603, L93–L96 Li, X.-D. et al., 1999, [*Physical Review Letters*]{}, 83, 3776–3779 Markwardt, C.B. et al., 2002a, [*ApJ*]{}, 575, L21–L24 Markwardt, C.B. et al., 2002b, [*IAUC*]{}, 7993 Markwardt, C.B. et al., 2003a, [*IAUC*]{}, 8080 Markwardt, C.B. et al., 2003b, [*IAUC*]{}, 8144 Markwardt, C.B. et al., 2003c, [*ATEL*]{}, 127 Markwardt, C.B. et al., 2003d, [*ATEL*]{}, 164 Markwardt, C.B. et al., 2004a, [*ATEL*]{}, 353 Markwardt, C.B. et al., 2004b, [*ATEL*]{}, 360 Marshall, F.E., 1998, [*IAUC*]{}, 6876 Miller, J.M. et al., 2003, [*ApJ*]{}, 583, L99–L102 Monard, B., 2002, [*VSNet alert*]{}, 7550 Nelson, L. A. & Rappaport, S., 2003, [*ApJ*]{}, 598, 431–445 Nowak, M. A. et al., 2004, [*ATEL*]{}, 369 Pooley, G., 2004, [*ATEL*]{}, 355 Poutanen, J. & Gierlinski, M., 2003, [*MNRAS*]{}, 343, 1301–1311 Psaltis, D. & Chakrabarty, D., 1999, [*ApJ*]{}, 521, 332–340 Rai Choudhuri, A. & Konar, S., 2002, [*MNRAS*]{}, 332, 933–944 Rappaport, S. A. et al., 2004, [*ApJ*]{}, 606, 436–443 Remillard, R.A., 2002, [*IAUC*]{}, 7888 Remillard, R.A. et al., 2002, [*IAUC*]{}, 7893 Remillard, R., 2004, [*ATEL*]{}, 357 Revnivtsev, M.G., 2003, [*AstL*]{}, 29, 383–386 Roche, P. et al., 1998, [*IAUC*]{}, 6885 Roelofs, G. et al., 2004, [*ATEL*]{}, 356 Rupen, M.P. et al., 2002a, [*IAUC*]{}, 7893 Rupen, M.P. et al., 2002b, [*IAUC*]{}, 7997 Rupen, M. et al., 2004, [*ATEL*]{}, 364 Steeghs, D., 2003, [*IAUC*]{}, 8155 Steeghs, D. et al., 2004, [*ATEL*]{}, 363 Stella, L. et al., 2000, [*ApJ*]{}, 537, L115–L118 Strohmayer, T. & Bildsten, L., 2003, To appear in [*Compact Stellar X-ray sources*]{}, W.H.G. Lewin & M. van der Klis (eds.), Cambridge University Press (astro-ph/0301544) Strohmayer, T.E. et al., 1996, [*ApJ*]{}, 469, L9–L12 Strohmayer, T.E. et al., 2003, [*ApJ*]{}, 596, L67–L70 Swank, J.H. & Markwardt, C.B., 2004, [*ATEL*]{}, 365 Titarchuk, L. et al., 2002, [*ApJ*]{}, 576, L49–L52 Tomsick, J.A. et al., 2004, [*ApJ*]{}, 610, 933–940 Vaughan, B.A. et al., 1994, [*ApJ*]{}, 435, 362–371 Van der Klis, M., 2000, [*ARA&A*]{}, 38, 717–760 Van der Klis, M., 2004, To appear in [*Compact Stellar X-ray sources*]{}, W.H.G. Lewin & van der Klis (eds.), Cambridge University Press (astro-ph/0410551) Van der Klis, M. et al., 1996, [*ApJ*]{}, 469, L1–L4 Van der Klis, M. et al., 2000, [*IAUC*]{}, 7358 Van Straaten, S. et al., 2004, In [*The Restless High-Energy Universe*]{}, eds. E.P.J. van den Heuvel, J.J.M. in ’t Zand, & R.A.M.J. Wijers (Elsevier), [*Nuclear Physics B*]{}, 132, 664–667 Van Straaten, S. et al., 2005, [*ApJ*]{}, in press (astro-ph/0410505) Wachter, S. & Hoard, D.W., 2000, [*IAUC*]{}, 7363 Wachter, S. et al., 2000, [*HEAD*]{}, 32, 24.15 Wang, Z. et al., 2001, [*ApJ*]{}, 563, L61–L64 Wijnands, R., 2003, [*ApJ*]{}, 588, 425–429 Wijnands, R., 2004a, In [*The Restless High-Energy Universe*]{}, eds. E.P.J. van den Heuvel, J.J.M. in ’t Zand, & R.A.M.J. Wijers (Elsevier), [*Nuclear Physics B*]{}, 132, 496–505 Wijnands, R., 2004b, In [*X-ray Timing 2003: Rossi and Beyond*]{}, eds. P. Kaaret, F.K. Lamb, & J.H. Swank (Melville, NY), [*AIP*]{}, 714, 209–216 Wijnands, R. & van der Klis, M., 1998a, [*Nature*]{}, 394, 344–346 Wijnands, R. & van der Klis, M., 1998b, [*ApJ*]{}, 507, L63–L66 Wijnands, R. & van der Klis, M., 1999, [*ApJ*]{}, 514, 939–944 Wijnands, R. & Homan, J., 2003, [*ATEL*]{}, 165 Wijnands, R. & Reynolds, A., 2003, [*ATEL*]{}, 166 Wijnands, R. et al., 2001, [*ApJ*]{}, 560, 892–896 Wijnands, R. et al., 2002, [*ApJ*]{}, 571, 429–434 Wijnands, R. et al., 2003, [*Nature*]{}, 424, 44–47 Wijnands, R. et al., 2005, [*ApJ*]{}, in press (astro-ph/0406057) Zdunik, J.L. et al., 2000, [*A&A*]{}, 359, 143–147
{ "pile_set_name": "ArXiv" }
--- abstract: 'We study the modular symmetry in magnetized D-brane models on $T^2$. Non-Abelian flavor symmetry $D_4$ in the model with magnetic flux $M=2$ (in a certain unit) is a subgroup of the modular symmetry. We also study the modular symmetry in heterotic orbifold models. The $T^2/Z_4$ orbifold model has the same modular symmetry as the magnetized brane model with $M=2$, and its flavor symmetry $D_4$ is a subgroup of the modular symmetry.' author: - 'Tatsuo Kobayashi, Satoshi Nagamoto, Shintaro Takada, Shio Tamba, and Takuya H. Tatsuishi' title: | Modular symmetry and non-Abelian discrete flavor symmetries\ in string compactification --- Introduction ============ Non-Abelian discrete flavor symmetries play an important role in particle physics. In particular, many models with various finite groups have been studied in order to explain quark and lepton masses and their mixing angles. (See for review [@Altarelli:2010gt; @Ishimori:2010au; @King:2013eh].) Those symmetries may be useful for dark matter physics and multi-Higgs models. Superstring theory is a promising candidate for unified theory including gravity. It has been shown that some non-Abelian discrete flavor symmetries appear in superstring theory with certain compactifications. Heterotic string theory on toroidal $Z_N$ orbifolds can realize non-Abelian flavor symmetries, e.g. $D_4$, and $\Delta (54)$ [@Kobayashi:2006wq]. (See also [@Kobayashi:2004ya; @Ko:2007dz].)[^1] Furthermore, magnetized D-brane models within the framework of type II superstring theory can lead to similar flavor symmetries [@Abe:2009vi; @Abe:2009uz; @BerasaluceGonzalez:2012vb; @Marchesano:2013ega; @Abe:2014nla]. Intersecting D-brane models are T-dual to magnetized D-brane models. Then, one can realize the same aspects in intersecting D-brane models as in the magnetized ones.[^2] On the other hand, superstring theories on tori and orbifolds have the modular symmetry. Recently, behavior of zero-modes under modular transformation was studied in magnetized D-brane models in Ref. [@Kobayashi:2017dyu].(See also [@Cremades:2004wa].) Also, behavior of twisted sectors under modular transformation was alreadly studied in Ref. [@Lauer:1989ax; @Lerche:1989cs; @Ferrara:1989qb]. These modular transformations also act non-trivially on flavors and transform mutually flavors each other. The remarkable difference is that modular transformation also acts Yukawa couplings as well as higher order couplings, while those couplings are trivial singlets under the usual non-Abelian symmetries. The purpose of this paper is to study more how modular transformation is represented by zero-modes in magnetized D-brane models, and to discuss relations between modular transformation and non-Abelian flavor symmetries in magnetized D-brane models. Intersecting D-brane models have the same aspects as magnetized D-brane models, because they are T-dual to each other. Furthermore, intersecting D-brane models in type II superstring theory and heterotic string theory have similarities, e.g. in two-dimensional conformal field theory. Thus, here we study modular symmetry and non-Abelian discrete flavor symmetries in heterotic orbifold models, too. This paper is organized as follows. In section \[sec:magne-D\], we study the modular symmetries in magnetized D-brane models and the relation to the $D_4$ flavor symmetry. In section \[sec:Het\], we study the modular symmetries in heterotic orbifold models. Section \[sec:conclusion\] is conclusion and discussions. We give brief reviews on non-Abelian discrete flavor symmetries in magnetized D-brane models and heterotic orbifold models in Appendix \[sec:app-1\] and \[sec:app-2\], respectively. Modular transformation in magnetized D-brane models {#sec:magne-D} =================================================== In this section, we study modular transformation of zero-mode wavefunctions in magnetized D-brane models. Zero-mode wavefunction {#sec:zero-mode} ---------------------- Here, we give a brief review on zero-mode wavefunctions on torus with magnetic flux [@Cremades:2004wa]. For simplicity, we concentrate on $T^2$ with $U(1)$ magnetic flux. The complex coordinate on $T^2$ is denoted by $z= x^1+ \tau x^2$, where $\tau$ is the complex modular parameter, and $x^1$ and $x^2$ are real coordinates. The metric on $T^2$ is given by $$g_{\alpha \beta} = \left( \begin{array}{cc} g_{zz} & g_{z \bar{z}} \\ g_{\bar{z} z} & g_{\bar{z} \bar{z}} \end{array} \right) = (2\pi R)^2 \left( \begin{array}{cc} 0 & \frac{1}{2} \\ \frac{1}{2} & 0 \end{array} \right) .$$ We identify $z \sim z +1$ and $z \sim z + \tau$ on $T^2$. On $T^2$, we introduce the $U(1)$ magnetic flux $F$, $$F = i\frac{\pi M}{{\text{Im}\,\tau}} (dz \wedge d\bar{z}),$$ which corresponds to the vector potential, $$A(z) = \frac{\pi M}{{\text{Im}\,\tau}} {\text{Im}\,(\bar{z}dz)}.$$ Here we concentrate on vanishing Wilson lines. On the above background, we consider the zero-mode equation for the spinor field with the $U(1)$ charge $q=1$, $$i \slashb{D} \Psi = 0.$$ The spinor field on $T^2$ has two components, $$\Psi(z,\bar{z}) = \left( \begin{array}{c} \psi_+ \\ \psi_- \end{array} \right).$$ The magnetic flux should be quantized such that $M$ is integer. Either $\psi_+$ or $\psi_-$ has zero-modes exclusively for $M \neq 0$. For example, we set $M$ to be positive. Then, $\psi_+$ has $M$ zero-modes, while $\psi_-$ has no zero-mode. Hence, we can realize a chiral theory. Their zero-mode profiles are given by $$\psi^{j,M}(z) = \mathcal{N} e^{i\pi M z \frac{{\text{Im}\,z}}{{\text{Im}\,\tau}}} \cdot \vartheta \left[ \begin{array}{c} \frac{j}{M} \\ 0 \end{array} \right] \left( M z, M\tau \right),$$ with $j=0,1,\cdots, (M-1)$, where $\vartheta$ denotes the Jacobi theta function, $$\vartheta \left[ \begin{array}{c} a \\ b \end{array} \right] (\nu, \tau) = \sum_{l \in {\bf Z}} e^{\pi i (a+l)^2 \tau} e^{2 \pi i (a+l)(\nu+b)} .$$ Here, $\mathcal{N}$ denotes the normalization factor given by $$\label{eq:normalization} \mathcal{N} = \left( \frac{2{\text{Im}\,\tau} M}{\mathcal{A}^2} \right)^{1/4},$$ with $\mathcal{A}= 4 \pi^2 R^2 {\text{Im}\,\tau}$. The ground states of scalar fields also have the same profiles as $\psi^{j,M}$. Thus, the Yukawa coupling including one scalar and two spinor fields can be computed by using these zero-mode waverfunctions. Zero mode wavefunctions satisfy the following relation, $$\begin{aligned} \label{eq:psi-psi-psi} \psi^{i,M}\psi^{j,M} &=& \mathcal{A}^{-1/2} (2{\text{Im}\,\tau})^{1/4}\left(\frac{MN}{M+N} \right)^{1/4} \nonumber \\ & & \times \sum_m \psi^{i+j+Mm,M+N} \cdot \vartheta \left[ \begin{array}{c} \frac{Ni-Mj+MNm}{MN(M+N)} \\ 0 \end{array} \right] \left( 0, MN(M+N)\tau \right).\end{aligned}$$ By use of this relation, their Yukawa couplings are given by the wavefunction overlap integral, $$\begin{aligned} Y_{ijk} &=& y \int d^2 z \psi^{i,M}\psi^{j,N}(\psi^{k,M'})^{*} \nonumber \\ &=& y \left(\frac{2{\text{Im}\,\tau} }{ \mathcal{A}^2}\right)^{1/4} \sum_{m \in Z_{M'}} \delta_{k,i+j+Mm} \cdot \vartheta \left[ \begin{array}{c} \frac{Ni-Mj+MNm}{MNM'} \\ 0 \end{array} \right] \left( 0, MNM'\tau \right),\end{aligned}$$ where $y$ is constant. This Yukawa coupling vanishes for $M' \neq M+N$. Similarly, we can compute higher order couplings using the relation (\[eq:psi-psi-psi\]) [@Abe:2009dr]. In the above equation, the Kronecker delta $\delta_{k,i+j+Mm}$ implies the coupling selection rule. For $g = {\rm gcd}(M,N,M')$, non-vanishing Yukawa couplings appear only if $$\label{eq:selection} i+j=k \qquad ({\rm mod~} g).$$ Hence, we can definite $Z_g$ charges in these couplings [@Abe:2009vi]. Modular transformation ---------------------- Here, we study modular transformation. First we give a brief review on results of modular transformation [@Kobayashi:2017dyu]. (See also [@Cremades:2004wa].) Then, we will study more in details. The $T^2$ is constructed by ${\mathbb R}^2/\Lambda$, and the lattice $\Lambda$ is spanned by the vectors $(\alpha_1,\alpha_2)$, where $\alpha_1 = 2\pi R$ and $\alpha_2 = 2 \pi R \tau$. However, the same lattice can be described by another basis, $$\label{eq:SL2Z} \left( \begin{array}{c} \alpha'_2 \\ \alpha'_1 \end{array} \right) =\left( \begin{array}{cc} a & b \\ c & d \end{array} \right) \left( \begin{array}{c} \alpha_2 \\ \alpha_1 \end{array} \right) ,$$ where $a,b,c,d$ are integer with satisfying $ad-bc = 1$. That is $SL(2,Z)$ transformation. The modular parameter $\tau = \alpha_2/\alpha_1$ transforms as $$\tau \longrightarrow \frac{a\tau + b}{c \tau + d},$$ under (\[eq:SL2Z\]). This transformation includes two important generators, $S$ and $T$, $$\begin{aligned} & &S:\tau \longrightarrow -\frac{1}{\tau}, \\ & &T:\tau \longrightarrow \tau + 1.\end{aligned}$$ They satisfy $$S^2=\mathbb{I}, \qquad (ST)^3=\mathbb{I}.$$ On top of that, if we impose the algebraic relation, $$T^N=\mathbb{I},$$ that corresponds to the congruence subgroup of modular group, $\Gamma(N)$. For example, it is found that $\Gamma(2) \simeq S_3$, $\Gamma(3) \simeq A_4$, $\Gamma(4) \simeq S_4$, and $\Gamma(5) \simeq A_5$. Since the group $A_4$ is the symmetry of tetrahedron, it is often called the tetrahedral group $T=A_4$. Also, it may be useful to mention about $\Delta(3N^2) \simeq (Z_N \times Z_N) \rtimes Z_3$ and $\Delta(6N^2) \simeq (Z_N \times Z_N) \rtimes S_3$. We find that $S_3 \simeq \Delta(6)$, $A_4 \simeq \Delta(12)$, and $S_4 \simeq \Delta(24)$. Following [@Kobayashi:2017dyu], we restrict ourselves to even magnetic fluxes $M$ ($M>0$). Under $S$, the zero-mode wavefunctions transform as [@Cremades:2004wa; @Kobayashi:2017dyu] $$\label{eq:magne-S} \psi^{j,M} \rightarrow \frac{1}{\sqrt{M}}\sum_k e^{2\pi i jk/M} \psi^{k,M}.$$ On the other hand, the zero-mode wavefunctions transform as [@Kobayashi:2017dyu] $$\label{eq:magne-T} \psi^{j,M} \rightarrow e^{ \pi i j^2/M} \psi^{j,M},$$ under $T$. Generically, the $T$-transformation satisfies $$T^{2M} = \mathbb{I},$$ on the zero-modes, $\psi^{j,M}$. Furthermore, in Ref. [@Kobayashi:2017dyu] it is shown that $$(ST)^3 = e^{\pi i/4},$$ on the zero-modes, $\psi^{j,M}$. In what follows, we study more concretely. ### Magnetic flux $M=2$ Let us study the case with the magnetic flux $M=2$ concretely. There are two zero-modes, $\psi^{0,2}, \psi^{1,2}$. The $S$-transformation acts on these zero-modes as $$\left( \begin{array}{c} \psi^{0,2} \\ \psi^{1,2} \end{array} \right) \longrightarrow S_{(2)}\left( \begin{array}{c} \psi^{0,2} \\ \psi^{1,2} \end{array} \right), \qquad S_{(2)} = \frac{1}{\sqrt 2}\left( \begin{array}{cc} 1 & 1 \\ 1 & -1 \end{array} \right).$$ The $T$-transformation acts as $$\left( \begin{array}{c} \psi^{0,2} \\ \psi^{1,2} \end{array} \right) \longrightarrow T_{(2)}\left( \begin{array}{c} \psi^{0,2} \\ \psi^{1,2} \end{array} \right), \qquad T_{(2)} = \left( \begin{array}{cc} 1 & 0 \\ 0 & i \end{array} \right).$$ They satisfy the following algebraic relations, $$\label{eq:ST-M=2} S^2_{(2)}=\mathbb{I}, \qquad T^4_{(2)}=\mathbb{I}, \qquad (S_{(2)}T_{(2)})^3=e^{\pi i /4}\mathbb{I}.$$ They construct a closed algebra with the order 192, which we denote here by $G_{(2)}$. By such an algebra, modular transformation is represented by two zero-modes, $\psi^{0,2}, \psi^{1,2}$. We find that $(S_{(2)}T_{(2)})^3$ is a center. Indeed, there are eight center elements and their group is $Z_8$. Other diagonal elements correspond to $Z_4$, which is generated by $T_{(2)}$. Here, we denote $$a = (S_{(2)}T_{(2)})^3, \qquad a' = T_{(2)}.$$ The diagonal elements are represented by $a^ma'^n$, i.e. $Z_8 \times Z_4$. Here, we examine the right coset $Hg$ for $g \in G_{(2)}$, where $H$ is the above $Z_8 \times Z_4$, i.e. $H=\{ a^ma'^n \}$. There would be $6(=192/(8\times 4))$ cosets. Indeed, we obtain the following six cosets: $$H, \quad HS_{(2)}, \quad HS_{(2)}T^k_{(2)}, \quad HS_{(2)}T^2_{(2)}S_{(2)},$$ with $k=1,2,3$. By simple computations, we find $HS_{(2)}T^k_{(2)}S_{(2)} \sim HS_{(2)}T^{4-k}_{(2)}S_{(2)} $ and $HS_{(2)}T^2_{(2)}S_{(2)}T \sim HS_{(2)}T^2_{(2)}S_{(2)}$. Furthermore, we would make a (non-Abelian) subgroup with the order 6 by choosing properly six elements such that we pick one element up from each coset and their algebra is closed. The non-Abelian group with the order 6 is unique, i.e. $S_3$. For example, we may be able to obtain the $Z_3$ generator from $ HS_{(2)}T_{(2)}$ because $(S_{(2)} T_{(2)})^3 \in H$. That is, we define $$b=a^ma'^nS_{(2)}T_{(2)}.$$ Then, we require $b^3 = \mathbb{I}$. There are three solutions, $(m,n)=(3,2)$, (5,0) mod (8,4). Similarly, we can obtain the $Z_2$ generator e.g. form $HS_{(2)}T_{(2)}^2S_{(2)}$ because $(S_{(2)}T_{(2)}^2S_{(2)})^2 \in H$. Then, we define $$c=a^{m'}a'^{n'}S_{(2)}T_{(2)}^2S_{(2)}.$$ We find $c^2 = \mathbb{I}$ when $n'=-m'$ mod 4. On top of that, we require $(bc)^2 = \mathbb{I}$, and that leads to the conditions, $n=-m'-1$ mod 4 and $m=m' + 2$ mod 8. As a result, there are six solutions, $(m,n,m')=(3,2,1)$, (3,2,5), (5,0,3), (5,0,7) with $n'=-m'$ mod 4. For example, for $(m,n,m')=(3,2,5)$ we write $$b = \frac{1}{\sqrt 2}\left( \begin{array}{cc} \rho^3 & \rho^{-3} \\ \rho^{-1} & \rho^{-3} \end{array}\right), \qquad c = \left( \begin{array}{cc} 0 & \rho^{-3} \\ \rho^3 & 0 \end{array}\right).$$ The six elements of the subgroup are written explicitly, $$\begin{aligned} & & \left( \begin{array}{cc} 1 & 0 \\ 0 & 1 \end{array} \right), \qquad \frac{1}{\sqrt 2}\left( \begin{array}{cc} 1 & 1 \\ 1 & -1 \end{array}\right), \qquad \left( \begin{array}{cc} 0 & \rho^{-3} \\ \rho^3 & 0 \end{array}\right), \nonumber \\ & & \frac{1}{\sqrt 2}\left( \begin{array}{cc} -1 & i \\ -i & 1 \end{array} \right), \qquad \frac{1}{\sqrt 2}\left( \begin{array}{cc} \rho^3 & \rho^{-3} \\ \rho^{-1} & \rho^{-3} \end{array}\right), \qquad \frac{1}{\sqrt 2}\left( \begin{array}{cc} \rho^{-3} & \rho \\ \rho^3 & \rho^3 \end{array}\right), \end{aligned}$$ where $\rho = e^{2\pi i /8}$. They correspond to $S_3\simeq \Gamma(2) \simeq \Delta(6)$ because they satisfy the following algebraic relations, $$c^2 = b^3 = (bc)^2=\mathbb{I}.$$ Moreover, they satisfy the following algebraic relation with $Z_8\times Z_4$, $$b^{-1}ab^{1}= a,\qquad cac=a,\qquad b^{-1}a'b=a, \qquad ca'c^{-1} = a^2a'^3.$$ Thus, the algebra of $G_{(2)}$ is isomorphic to $(Z_8 \times Z_4) \rtimes S_3$. We have started by choosing $HS_{(2)}T_{(2)}^2S_{(2)}$ for a candidate of the $Z_2$ generator. We can obtain the same results by starting with $HS_{(2)}$ for a candidate of the $Z_2$ generator. ### Magnetic flux $M=4$ Similarly, we study the case with the magnetic flux $M=4$. There are four zero-modes, $\psi^{i,M}$ with $i=0,1,2,3$. The $S$ and $T$-transformations are represented by $\psi^{i,M}$, $$S_{(4)} = \frac12\left( \begin{array}{cccc} 1 & 1 & 1 & 1 \\ 1 & i & -1& -i \\ 1 & -1 & 1 & -1 \\ 1 &-i & -1 & i \end{array} \right), \qquad T_{(4)} = \left( \begin{array}{cccc} 1 & & & \\ & e^{\pi i/4} & & \\ & & -1 & \\ & & & e^{\pi i/4} \end{array} \right).$$ This is a reducible representation. In order to obtain irreducible representations, we use the flowing basis, $$\left( \begin{array}{c} \psi^{0.4} \\ \psi^{1,4}_{+} \\ \psi^{2,4} \end{array} \right) = \left( \begin{array}{c} \psi^{0.4} \\ \frac{1}{\sqrt{2}}(\psi^{1,4} + \psi^{3,4}) \\ \psi^{2,4} \end{array} \right), \qquad \psi^{1,4}_- = \frac{1}{\sqrt{2}}(\psi^{1,4} - \psi^{3,4}).$$ This is nothing but zero-modes on the $T^2/Z_2$ orbifold [@Abe:2008fi]. The former corresponds to $Z_2$ even states, while the latter corresponds to the $Z_2$ odd state. Note that $(ST)^3$ transforms the lattice basis $(\alpha_1,\alpha_2) \rightarrow (-\alpha_1,-\alpha_2)$. Thus, it is reasonable that the zero-modes on the $T^2/Z_2$ orbifold correspond to the irreducible representations. The $S$ and $T$-representations by the $Z_2$ odd zero-mode are quite simple, and these are represented by $$S_{(4-)} = i, \qquad T_{(4)-} = e^{\pi i/4}.$$ Their closed algebra is $Z_8$. On the other hand, the $S$ and $T$-transformations are represented by the $Z_2$ even zero-modes, $$S_{(4)+}=\frac12\left( \begin{array}{ccc} 1 & \sqrt2 & 1 \\ \sqrt 2 & 0 & -\sqrt 2 \\ 1 & -\sqrt 2 & 1 \end{array} \right), \qquad T_{(4)+}=\left( \begin{array}{ccc} 1 & & \\ & e^{\pi i/4} & \\ & & -1 \end{array}\right).$$ They satisfy the following algebraic relation, $$(S_{(4)+})^2= \mathbb{I},\qquad (T_{(4)+})^8=\mathbb{I}, \qquad (S_{(4)+}T_{(4)+})^3=e^{\pi i/4} \mathbb{I}.$$ We denote the closed algebra of $S_{(4)+}$ and $T_{(4)+}$ by $G_{(4)+}$. Its order is equal to 768, and it includes the center element $(S_{(4)+}T_{(4)+})^3$, i.e. $Z_8$. Other diagonal elements correspond to $Z_8$, which is generated by $T_{(4)+}$. Again, we denote $a=(S_{(4)+}T_{(4)+})^3$ and $a'=T_{(4)+}$, and the diagonal elements are written by $a^ma'^n$, i.e. $Z_8 \times Z_8$. Similar to the case with $M=2$, we examine the coset structure, $Hg$. Indeed, there are the following 12 cosets: $$H, \quad HS_{(4)+}, \quad HS^k_{(4)+}, \quad HS_{(4)+}T^\ell_{(4)+} S_{(4)+},$$ where $k=1,\cdots,7$ and $\ell = 2,4,6$. By simple computation, we find that $$HS_{(4)+}T^k_{(4)+} S_{(4)+} \sim HS_{(4)+}T^{8-k}_{(4)+}, \qquad HS_{(4)+}T^\ell_{(4)+} S_{(4)+}T \sim HS_{(4)+}T^{8-\ell}_{(4)+} S_{(4)+}$$ for $k=$ odd and $\ell=$ even. We make a subgroup with the order 12 by choosing properly 12 elements such that we pick one element up from each coset and their algebra is closed. The non-Abelian group with the order 12 are $D_6$, $Q_6$ and $A_4$. Among them, $A_4$ would be a good candidate. Indeed, we can obtain the $Z_3$ generator from $HS_{(4)+}T_{(4)+}$, gain. That is, we define $$t=a^ma'^nS_{(4)+}T_{(4)+}.$$ The solutions for $t^3=\mathbb{I}$ are obtained by $(m,n)=(1,4)$, (3,6), (5,0), and (7,2). We also define $$s=a^{m'}a'^{n'}S_{(4)+}T^4_{(4)+}S_{(4)+}.$$ The solutions for $s^2=\mathbb{I}$ are obtained by $(m',n')=(0,0)$, (0,4), (4,0) and (4,4). These two generators satisfy $(st)^3=\mathbb{I}$ if $(m',n')=(0,4)$, and (4,0), i.e. $$s=\left( \begin{array}{ccc} 0 & 0& \pm 1 \\ 0 & -1 & 0 \\ \pm 1 & 0 &0 \end{array}\right).$$ As a result, they satisfy $$s^2=t^3=(st)^3=\mathbb{I}.$$ That is the $A_4$ algebra. ### Large magnetic flux $M$ For larger magnetic fluxes, $S$ and $T$-transformations are represented by zero-modes $\psi^{j,M}$, but those are reducible representations. The irreducible representations are obtained in the $T^2/Z_2$ orbifold basis, $$\psi_{\pm}^{j,M} = \frac{1}{\sqrt 2}\left( \psi^{j,M} \pm \psi^{M-j,j} \right).$$ The representations of $T_{(M)}$ are simply obtained by $$T_{(M)+}\left( \begin{array}{c} \psi^{0,M} \\ \psi^{1,M}_+ \\ \vdots \\ \psi^{j,M}_+ \\ \vdots\\ \psi^{M/2,M} \end{array} \right) = \left( \begin{array}{cccccc} 1 & & & & & \\ & e^{\pi i/M} & & & & \\ & & \ddots & & & \\ & & & e^{\pi i j^2/M} & & \\ & & & & \ddots & \\ & & & & & e^{\pi i M/4} \end{array}\right) \left( \begin{array}{c} \psi^{0,M} \\ \psi^{1,M}_+ \\ \vdots \\ \psi^{j,M}_+ \\ \vdots\\ \psi^{M/2,M} \end{array} \right),$$ and $$T_{(M)-}\left( \begin{array}{c} \psi^{1,M}_- \\ \vdots \\ \psi^{j,M}_- \\ \vdots\\ \psi^{M/2-1,M}_- \end{array} \right) = \left( \begin{array}{ccccc} e^{\pi i/M} & & & & \\ & \ddots & & & \\ & & e^{\pi i j^2/M} & & \\ & & & \ddots & \\ & & & & e^{\pi (M/2-1)^2/M} \end{array}\right) \left( \begin{array}{c} \psi^{1,M}_- \\ \vdots \\ \psi^{j,M} \\ \vdots\\ \psi^{M/2-1,M}_- \end{array} \right).$$ Both correspond to $Z_{2M}$. On the other hand, the $S_{(M)\pm}$ transforms $$S_{(M)\pm}\psi^{j,M}_{\pm} = \frac{1}{\sqrt {2M}}\sum_k \left(e^{2 \pi jk/M}\pm e^{2 \pi i (M-j)k/M,M} \right)\psi^{k,M} .$$ This representation is also written by $$S_{(M)\pm}\psi^{j,M}_{\pm} = \frac{1}{\sqrt {2M}}\sum_k \left(e^{2 \pi (M-j)(M-k)/M}\pm e^{2 \pi i j(M-k)/M,M} \right)\psi^{M-k,M} .$$ Thus, the $S$-transformation is represented on the $T^2/Z_2$ orbifold basis by $$S_{(M)\pm}\psi^{j,M}_{\pm} = \frac{1}{\sqrt {M}}\sum_{k \leq M/2 } \left(e^{2 \pi jk/M}\pm e^{2 \pi i (M-j)k/M} \right)\psi^{k,M}_{\pm} .$$ These are written by $$\begin{aligned} S_{(M)+}\psi^{j,M}_{+} = \frac{2}{\sqrt {M}}\sum_{k \leq M/2 } \cos (2 \pi jk/M) \psi^{j,M}_{+}, \nonumber \\ S_{(M)-}\psi^{j,M}_{-} = \frac{2i}{\sqrt {M}}\sum_{k \leq M/2 } \sin (2 \pi jk/M) \psi^{j,M}_{-}. \end{aligned}$$ For example, for $M=6$, $S$ and $T$ are represented by $Z_2$ even zero-modes, $$S_{(6)+}\left( \begin{array}{c} \psi^{0,6} \\ \psi^{1,6}_+ \\ \psi^{2,6}_+ \\ \psi^{3,6} \end{array}\right) = \frac{1}{\sqrt 6}\left( \begin{array}{cccc} 1 &\sqrt 2 & \sqrt 2 & 1 \\ \sqrt 2 & 1 & -1 & -\sqrt 2 \\ \sqrt 2 & -1 & -1 & \sqrt 2 \\ 1& -\sqrt 2 & \sqrt 2 & -1 \end{array} \right) \left( \begin{array}{c} \psi^{0,6} \\ \psi^{1,6}_+ \\ \psi^{2,6}_+ \\ \psi^{3,6} \end{array}\right),$$ $$T_{(6)+}\left( \begin{array}{c} \psi^{0,6} \\ \psi^{1,6}_+ \\ \psi^{2,6}_+ \\ \psi^{3,6} \end{array}\right) = \left( \begin{array}{cccc} 1 & & & \\ & e^{\pi i /6} & & \\ & & e^{2 \pi i/3} & \\ & & & e^{^{3 \pi i /2}} \end{array} \right) \left( \begin{array}{c} \psi^{0,6} \\ \psi^{1,6}_+ \\ \psi^{2,6}_+ \\ \psi^{3,6} \end{array}\right),$$ while $S$ and $T$ are represented by $Z_2$ odd zero-mode, $$S_{(6)-}\left( \begin{array}{c} \psi^{1,6}_- \\ \psi^{2,6}_- \end{array}\right) = \frac{i}{ \sqrt 2}\left( \begin{array}{cc} 1 & 1 \\ 1 & -1 \end{array} \right) \left( \begin{array}{c} \psi^{1,6}_- \\ \psi^{2,6}_- \end{array}\right),$$ $$T_{(6)-}\left( \begin{array}{c} \psi^{1,6}_- \\ \psi^{2,6}_- \end{array}\right) = \left( \begin{array}{cc} e^{\pi i /6} & 0 \\ 0 & e^{2 \pi i/3} \end{array} \right) \left( \begin{array}{c} \psi^{1,6}_- \\ \psi^{2,6}_- \end{array}\right).$$ Non-Abelian discrete flavor symmetries {#sec:D4} -------------------------------------- In Ref.[@Abe:2009vi], it is shown that the models with $M=2$ as well as even magnetic fluxes have the $D_4$ flavor symmetry. See Appendix \[sec:app-1\]. One of the $Z_2$ elements in $D_4$ corresponds to $(T_{(2)})^2$ on the zero-modes, $\psi^{0,2}$ and $\psi^{1,2}$, i.e. $$Z = \left( \begin{array}{cc} 1 & 0 \\ 0 & -1 \end{array} \right) = (T_{(2)})^2.$$ In addition, the permutation $Z^C_2$ element in $D_4$ corresponds to $S_{(2)}T_{(2)}T_{(2)}S_{(2)}$, i.e. $$C = \left( \begin{array}{cc} 0 & 1 \\ 1 & 0 \end{array} \right) = S_{(2)}T_{(2)}T_{(2)}S_{(2)}.$$ Thus, the $D_4$ group, which includes the eight elements (\[eq:D4-8\]), is subgroup of $G_{(2)} \simeq (Z_8 \times Z_4) \rtimes S_3$. However, there is the difference between the modular symmetry and the $D_4$ flavor symmetry, which studied in Ref.[@Abe:2009vi]. The modular symmetry transforms the Yukawa couplings, while the Yukawa couplings are invariant under the $D_4$ flavor symmetry. In order to study this point, here we examine the Yukawa couplings among $\psi^{i,2}$, $\psi'^{j,2}$ and $\psi^{k,4}$. Both $\psi^{i,2}$ and $\psi'^{j,2}$ are $D_4$ doublets, and their tensor product ${\bf 2}\times {\bf 2}$ is expanded by $${\bf 2} \times {\bf 2} = {\bf 1}_{++} + {\bf 1}_{+-} + {\bf 1}_{-+} + {\bf 1}_{--}.$$ Thus, the products $\psi^{i,2} \psi'^{j,2}$ correspond to four singlets, $$\begin{aligned} & & {\bf 1}_{+\pm}: \psi^{0,2} \psi'^{0,2} \pm \psi^{1,2} \psi'^{1,2}, \qquad {\bf 1}_{-\pm}: \psi^{0,2} \psi'^{1,2} \pm \psi^{1,2} \psi'^{0,2}.\end{aligned}$$ On the other hand, by use of Eq.(\[eq:psi-psi-psi\]), the products $\psi^{i,2}\psi'^{j,2}$ are expanded by $\psi^{k,4}$. For example, we can expand as $$\begin{aligned} \psi^{0,2} \psi'^{0,2} \pm \psi^{1,2} \psi'^{1,2} &\sim & \left( Y^{(0)}( 16\tau ) + Y^{(8/16)}( 16\tau ) \pm \left( Y^{(4/16)}( 16\tau ) + Y^{(12/16)}( 16\tau ) \right) \right) \nonumber \\ & & \times \left( \psi^{0,4} \pm \psi^{2,4} \right)\end{aligned}$$ up to constant factors, where $$Y^{(j/M)}(M\tau) = \mathcal{N} \cdot \vartheta \left[ \begin{array}{c} \frac{j}{M} \\ 0 \end{array} \right] \left( 0, M\tau \right).$$ Note that $Y^{(j/M)}(M\tau) = Y^{(1-j/M)}(M\tau)$. It is found that $$\begin{aligned} & & (T_{(4)})^2 \left( \psi^{0,4} \pm \psi^{2,4} \right) = \left( \psi^{0,4} \pm \psi^{2,4} \right), \nonumber \\ & & (S_{(4)}T_{(4)}T_{(4)}S_{(4)}) \left( \psi^{0,4} \pm \psi^{2,4} \right) = \pm \left( \psi^{0,4} \pm \psi^{2,4} \right).\end{aligned}$$ Thus, the zero-modes $\psi^{0,4} \pm \psi^{2,4} $ are indeed $D_4$ singlets, ${\bf 1}_{+\pm}$ when we identify $(T_{(4)})^2 $ and $(S_{(4)}T_{(4)}T_{(4)}S_{(4)})$ as $Z_2$ and $Z^C_2$ of $D_4$. In this sense, the $D_4$ flavor symmetry is a subgroup of the modular symmetry. Also, it is found that the above Yukawa couplings, $Y^{(m/4)}(16\tau)$, with $m=0,1,2,3$ are invariant under $T^2$ and $STTS$ transformation. Similarly, we can expand $$\begin{aligned} \label{eq:01-13} \psi^{0,2} \psi'^{1,2} + \psi^{1,2} \psi'^{0,2} &\sim & \left( Y^{(2/16)}( 16\tau ) + Y^{(6/16)}( 16\tau ) \right) \times \left( \psi^{1,4} + \psi^{3,4} \right),\end{aligned}$$ up to constant factors. It is found that $$\begin{aligned} & & (T_{(2)})^2 \left( \psi^{0,2} \psi'^{1,2} + \psi^{1,2} \psi'^{0,2} \right) = - \left( \psi^{0,2} \psi'^{1,2} + \psi^{1,2} \psi'^{0,2} \right) .\end{aligned}$$ On the other hand, we obtain $$\begin{aligned} & & (T_{(4)})^2 \left( \psi^{1,4} + \psi^{3,4} \right) = i \left( \psi^{1,4} + \psi^{3,4} \right) .\end{aligned}$$ In addition, we find $$T^2: \left( Y^{(2/16)}( 16\tau ) + Y^{(6/16)}( 16\tau ) \right) \rightarrow i \left( Y^{(2/16)}( 16\tau ) + Y^{(6/16)}( 16\tau ) \right) .$$ Thus, the $T^2$ transformation is consistent between left and right hand sides in (\[eq:01-13\]). However, when we interpret $T^2$ as $Z_2$ of the $D_4$ flavor symmetry, we face inconsistency, because Yukawa couplings are not invariant and $ (\psi^{1,4} + \psi^{3,4} )$ has transformation behavior different from $(\psi^{0,2} \psi'^{1,2} + \psi^{1,2} \psi'^{0,2})$. We can make this consistent by defining $Z_2$ of the $D_4$ on $ (\psi^{1,4} + \psi^{3,4} )$ such that its transformation absorbs the phase of Yukawa couplings under $T^2$ transformation. Then, the mode $ (\psi^{1,4} + \psi^{3,4} )$ exactly corresponds to the $D_4$ singlet, ${\bf 1}_{-+}$. We find that $(\psi^{0,2} \psi'^{1,2} + \psi^{1,2} \psi'^{0,2})$ is invariant under $S_{(2)}T_{(2)}T_{(2)}S_{(2)}$, and $ (\psi^{1,4} + \psi^{3,4} )$ is also invariant under $S_{(4)}T_{(4)}T_{(4)}S_{(4)}$. That is consistent. Therefore, the $D_4$ flavor symmetry is a subgroup of the modular symmetry on $\psi^{j,2}$ $(j=0,1)$. However, when the model includes couplings to zero-modes with larger $M$, we have to modify their modular symmetries such that coupling constants are invariant under the flavor symmetry. Then, we can define the $D_4$ flavor symmetry. Here, we give a comment on the $T^2/Z_2$ orbifold. The $T^2/Z_2$ orbifold basis gives the irreducible representations of the modular symmetry. The $D_4$ flavor symmetry is defined through the modular symmetry, as above. That is the reason why the $D_4$ flavor symmetry remains on the $T^2/Z_2$ orbifold [@Marchesano:2013ega; @Abe:2014nla]. Heterotic orbifold models {#sec:Het} ========================= Intersecting D-brane models in type II superstring theory is T-dual to magnetized D-brane models. Thus, intersecting D-brane models also have the same behavior under modular transformation as magnetized D-brane models. Furthermore, intersecting D-brane models in type II superstring theory and heterotic string theory on orbifolds have similarities, e.g. in two-dimensional conformal field theory. For example, computations of 3-point couplings as well as $n$-point couplings are similar to each other. Here, we study modular symmetry in heterotic orbifold models. Using results in Ref. [@Lauer:1989ax; @Lerche:1989cs; @Ferrara:1989qb], we compare the modular symmetries in heterotic orbifold models with non-Abelian flavor symmetries and also the modular symmetries in the magnetized D-brane models, which have been derived in the previous section. Twisted sector -------------- Here, we give a brief review on heterotic string theory on orbifolds. The orbifold is the division of the torus $T^n$ by the $Z_N$ twist $\theta$, i.e. $T^n/Z_N$. Since the $T^n$ is constructed by ${\mathbb R}^n/\Lambda$, the $Z_N$ twist $\theta$ should be an automorphism of the lattice $\Lambda$. Here, we focus on two-dimensional orbifolds, $T^2/Z_N$. The six-dimensional orbifolds can be constructed by products of two-dimensional ones. All of the possible orbifolds are classified as $T^2/Z_N$ with $N=2,3,4,6$. On orbifolds, there are fixed points, which satisfy the following condition, $$\label{eq:fp} x^i = (\theta^n x)^i + \sum_k m_k \alpha_k^i,$$ where $x^i$ are real coordinates, $\alpha_k^i$ are two lattice vectors, and $m_k$ are integer for $i,k=1,2$. Thus, the fixed points can be represented by corresponding space group elements $(\theta^n, \sum_k m_k \alpha_k^i)$, or in short $(\theta^n, ( m_1,m_2 ))$. The heterotic string theory on orbifolds has localized modes at fixed points, and these are the so-called twisted strings. These twisted states can be labeled by use of fixed points, $\sigma_{\theta,(m_1,m_2)}$. All of the twisted states $\sigma_{\theta,(m_1,m_2)}$ have the same spectrum, if discrete Wilson lines vanish. Thus, the massless modes are degenerate by the number of fixed points. On the $T^2/Z_2$ orbifold, there are four fixed points, which are denoted by $(\theta,(0,0))$, $(\theta,(1,0))$, $(\theta,(0,1))$, $(\theta,(1,1))$. The corresponding twisted states are denoted by $\sigma_{\theta,(m,n)}$ for $m,n=0,1$. On the $T^2/Z_3$ orbifold, $\alpha_1$ and $\alpha_2$ correspond to the $SU(3)$ simple roots and they are identified each other by the $Z_3$ twist. Thus, three fixed points on the $T^2/Z_3$ orbifold are represented by the space group elements, $(\theta,m \alpha_1)$ for $m=0,1,2$, or in short $(\theta, m)$. The corresponding twisted states are denoted by $\sigma_{\theta,m}$ for $m=0,1,2$. Similarly, we can obtain the fixed points and twisted states on the $T^2/Z_4$, where $\alpha_1$ and $\alpha_2$ correspond to the $SO(4)$ simple roots and they are identified each other by the $Z_4$ twist. For the $Z_4$ twist $\theta$, two fixed points satisfy Eq.(\[eq:fp\]), and these can be represented by $(\theta, m\alpha_1)$ for $m=0,1$, or in short $(\theta, m)$. Then, the first twisted states are denoted by $\sigma_{\theta,m}$ for $m=0,1$. In addition, for $\theta^2$, there are four points, which satisfy Eq.(\[eq:fp\]), and these can denoted by $(\theta^2,(m,n))$ for $m,n=0,1$. Indeed, these correspond to the four fixed points on the $T^2/Z_2$ orbifold. Then, the second twisted states are denoted by $\sigma_{\theta^2,(m,n)}$ for $m,n=0,1$. However, the fixed points $(\theta^2,(1,0))$ and $(\theta^2,(0,1))$ transform each other under the $Z_4$ twist $\theta$. Thus, the $Z_4$ invariant states are written by [@Kobayashi:1990mc] $$\sigma_{\theta^2,(0,0)},\qquad \sigma_{\theta^2,+}, \qquad \sigma_{\theta^2,(1,1)},$$ while $\sigma_{\theta^2,-}$ transforms to $-\sigma_{\theta^2,-}$ under the $Z_4$ twist, where $$\sigma_{\theta^2,\pm} = \frac{1}{\sqrt 2} \left( \sigma_{\theta^2,(1,0)} \pm \sigma_{\theta^2,(0,1)} \right).$$ Similarly, we can obtain the fixed points on $T^2/Z_6$. There is a fixed point $(\theta,0)$ for the $Z_6$ twist $\theta$, and a single twisted state $\sigma_{\theta,0}$. The second twisted sector has three fixed points $(\theta^2,m)$ ($m=0,1,2$), which correspond to the three fixed points on the $T^2/Z_3$ orbifold. The two fixed points $(\theta^2,1)$ and $(\theta^2,2)$ transform each other by the $Z_6$ twist, while $(\theta^2,0)$ is invariant. Thus, we can write the $Z_6$-invariant $\theta^2$-twisted states by $$\sigma_{\theta^2,0},\qquad \sigma_{\theta^2,+},$$ while $\sigma_{\theta^2,-}$ transforms to $-\sigma_{\theta^2,-}$ under the $Z_6$ twist, where $$\sigma_{\theta^2,\pm} = \frac{1}{\sqrt 2} \left( \sigma_{\theta^2,1} \pm \sigma_{\theta^2,2} \right).$$ The third twisted sector has four fixed points, which correspond to the fixed points on $T^2/Z_2$, and the corresponding $\theta^3$ twisted states. Their linear combinations are $Z_6$ eigenstates similar to the second twisted states. Since the first twisted sector has the single fixed point and twisted state, the modular symmetry as well as non-Abelian discrete flavor symmetry is rather trivial. We do not discuss the $T^2/Z_6$ orbifold itself. Modular symmetry ---------------- In Ref. [@Lauer:1989ax], modular symmetry in heterotic string theory on orbifolds was studied in detail. Here we use those results. ### $T^2/Z_4$ orbifold The $S$ and $T$ transformations are represented by the first twisted sectors of $T^2/Z_4$ orbifold as [@Lauer:1989ax], $$\begin{aligned} & & \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \end{array} \right) \longrightarrow S_{Z_4}\left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \end{array} \right), \qquad S_{Z_4} = \frac{1}{\sqrt 2}\left( \begin{array}{cc} 1 & 1 \\ 1 & -1 \end{array} \right), \nonumber \\ & & \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \end{array} \right) \longrightarrow T_{Z_4}\left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \end{array} \right), \qquad T_{Z_4} = \left( \begin{array}{cc} 1 & 0 \\ 0 & i \end{array} \right).\end{aligned}$$ These are exactly the same as representations of $S_{(2)}$ and $T_{(2)}$ on two-zero modes, $\psi^{0,2}$ and $\psi^{1,2}$ in the magnetized model with magnetic flux $M=2$. Hence, the twisted sectors on the $T^2/Z_4$ orbifold has the same behavior of modular symmetry as the magnetized model with magnetic flux $M=2$. Indeed, the twisted sectors have the $D_4$ flavor symmetry and two twisted states, $\sigma_{\theta,0}$ and $\sigma_{\theta,1}$ correspond to the $D_4$ doublet [@Kobayashi:2006wq]. The whole flavor symmetry of the $T^2/Z_4$ orbifold model is slightly larger than $D_4$. (See Appendix \[sec:app-2\].) The $T^2/Z_4$ orbifold model has the $Z_4$ symmetry, which transforms the first twisted sector, $$\label{eq:Z4} \sigma_{\theta,m} \longrightarrow e^{\pi i/2} \sigma_{\theta,m},$$ for $m=0,1$ and the second twisted sector, $$\sigma_{\theta^2,(m,n)} \longrightarrow e^{\pi i} \sigma_{\theta^2,(m,n)},$$ for $m,n=0,1$. The above $Z_4$ transformation (\[eq:Z4\]) is nothing but $(S_{Z_4}T_{Z_4})^6$ as clearly seen from Eq. (\[eq:ST-M=2\]). Thus, the whole flavor symmetry originates from the modular symmetry. The second twisted sectors correspond to $D_4$ singlets, ${\bf 1}_{\pm 1, \pm}$ [@Kobayashi:2006wq] as $$\begin{aligned} {\bf 1}_{+\pm}: \sigma_{\theta^2,(0,0)} \pm \sigma_{\theta^2,(1,1)}, \qquad {\bf 1}_{-\pm}: \sigma_{\theta^2,\pm},\end{aligned}$$ up to coefficients. Compared with the results in section \[sec:D4\], the $D_4$ behavior of the second twisted states correspond to one of the zero-modes $\psi^{m,4}$ with magnetic flux $M=4$. Their correspondence can be written as $$\begin{aligned} & & \sigma_{\theta^2,(0,0)} \sim \psi^{0,4}, \qquad \sigma_{\theta^2,(1,1)} \sim \psi^{2,4}, \nonumber \\ & & \sigma_{\theta^2,(1,0)} \sim \psi^{1,4}, \qquad \sigma_{\theta^2,(1,0)} \sim \psi^{3,4}.\end{aligned}$$ The above correspondence can also been seen from the Yukawa couplings. By use of operator product expansion, we obtain the following relations [@Lauer:1989ax], $$\begin{aligned} & & \sigma_{\theta,0} \sigma_{\theta,0} \sim Y_{0,0} \left( \sigma_{\theta^2,(0,0)} + \sigma_{\theta^2,(1,1)}\right) , \nonumber \\ & & \sigma_{\theta,1} \sigma_{\theta,1} \sim Y_{1,1} \left( \sigma_{\theta^2,(0,0)} + \sigma_{\theta^2,(1,1)} \right), \\ & & \sigma_{\theta,0} \sigma_{\theta,1} + \sigma_{\theta,1} \sigma_{\theta,0}\sim Y_{0,1} \sigma_{\theta^2,+} \nonumber \end{aligned}$$ up to constants. The second twisted state $\sigma_{\theta^2,-}$ can not couple with the first twisted sectors. Using results in Ref. [@Lauer:1989ax], it is found that $$(T_{Z_4})^2 \left( \begin{array}{c} Y_{0,0} \\ Y_{1,1} \\ Y_{0,1} \end{array} \right) = \left( \begin{array}{ccc} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & -1 \end{array}\right) \left( \begin{array}{c} Y_{0,0} \\ Y_{1,1} \\ Y_{0,1} \end{array} \right) .$$ This is the same as behavior of the Yukawa couplings under $T^2$ studied in section \[sec:D4\]. ### $T^2/Z_2$ orbifold Here, let us study the $T^2/Z_2$ orbifold in a way to similar to the previous section on the $T^2/Z_4$. The $S$ transformation is represented by the four twisted states on the $T^2/Z_2$ orbifold [@Lauer:1989ax], $$\left( \begin{array}{c} \sigma_{\theta,(0,0)} \\ \sigma_{\theta,(0,1)} \\ \sigma_{\theta,(1,0)} \\ \sigma_{\theta,(1,1)} \end{array} \right) \longrightarrow S_{Z_2}\left( \begin{array}{c} \sigma_{\theta,(0,0)} \\ \sigma_{\theta,(0,1)} \\ \sigma_{\theta,(1,0)} \\ \sigma_{\theta,(1,1)} \end{array} \right), \qquad S_{Z_2} = \frac{1}{ 2}\left( \begin{array}{cccc} 1 & 1 & 1 & 1\\ 1 & -1 & 1 & -1 \\ 1 & 1 & -1 & -1 \\ 1 & -1 & -1 & 1 \end{array} \right).$$ Also the $T$ transformation is represented as $$\left( \begin{array}{c} \sigma_{\theta,(0,0)} \\ \sigma_{\theta,(0,1)} \\ \sigma_{\theta,(1,0)} \\ \sigma_{\theta,(1,1)} \end{array} \right) \longrightarrow T_{Z_2}\left( \begin{array}{c} \sigma_{\theta,(0,0)} \\ \sigma_{\theta,(0,1)} \\ \sigma_{\theta,(1,0)} \\ \sigma_{\theta,(1,1)} \end{array} \right), \qquad T_{Z_2} = \left( \begin{array}{cccc} 1 & 0 & 0 & 0\\ 0 & -1 & 0 & 0 \\ 0 & 0 & -1 & 0 \\ 0 & 0 & 0 & -1 \end{array} \right).$$ The representation $S_{Z_2}$ is similar to $S_{Z_4}$ and $S_{(2)}$. Indeed, we find that $S_{Z_2} = S_{(2)} \otimes S_{(2)}$. However, the representation $T_{Z_2}$ is different from $T_{Z_4}$ and $T_{(2)}$. The matrices $S_{Z_2}$ and $T_{Z_2}$ satisfy the following relations, $$(S_{Z_2})^2=(T_{Z_2})^2=(S_{Z_2}T_{Z_2})^6=\mathbb{I}.$$ These correspond to the $D_6$. Indeed, the order of closed algebra including $S_{Z_2}$ and $T_{Z_2}$ is equal to 12. At any rate, these matrices are reducible. We change the basis in order to obtain irreducible representations, $$\left( \begin{array}{c} \sigma_1 \\ \sigma_2 \\ \sigma_3 \\ \sigma_4 \end{array}\right)=\left( \begin{array}{cccc} 1 & 0 & 0& 0\\ 0 & \frac{1}{\sqrt 3} & \frac{1}{\sqrt 3} & \frac{1}{\sqrt 3} \\ 0 & \frac{1}{\sqrt 2} & \frac{-1}{\sqrt 3} & 0 \\ 0 & \frac{1}{\sqrt 6} & \frac{1}{\sqrt 6} & \frac{-2}{\sqrt 6} \end{array}\right) \left( \begin{array}{c} \sigma_{\theta,(0,0)} \\ \sigma_{\theta,(1,0)} \\ \sigma_{\theta,(0,1)} \\ \sigma_{\theta,(1,1)} \end{array}\right).$$ Then, $\sigma_1$ and $\sigma_2$ correspond to the $D_6$ doublet, while $\sigma_3$ and $\sigma_4$ correspond to the $D_6$ singlets. For example, $S_{Z_2}T_{Z_2}$ and $T_{Z_2}$ are represented by $$\begin{aligned} & & S_{Z_2}T_{Z_2} \left( \begin{array}{c} \sigma_1 \\ \sigma_2 \\ \sigma_3 \\ \sigma_4 \end{array}\right)=\left( \begin{array}{cccc} \cos (2\pi/6) & -\sin(2\pi/6) & 0 & 0 \\ \sin(2\pi /6) & \cos(2 \pi /6) & 0 & 0 \\ 0 & 0& 1 & 0 \\ 0 & 0& 0 & -1 \end{array}\right) \left( \begin{array}{c} \sigma_1 \\ \sigma_2 \\ \sigma_3 \\ \sigma_4 \end{array}\right), \nonumber \\ & & T_{Z_2} \left( \begin{array}{c} \sigma_1 \\ \sigma_2 \\ \sigma_3 \\ \sigma_4 \end{array}\right)=\left( \begin{array}{cccc} 1 & 0 & 0 & 0 \\ 0 & -1 & 0 & 0 \\ 0 & 0& -1 & 0 \\ 0 & 0& 0 & -1 \end{array}\right) \left( \begin{array}{c} \sigma_1 \\ \sigma_2 \\ \sigma_3 \\ \sigma_4 \end{array}\right).\end{aligned}$$ It is found that $\sigma_{3}$ and $\sigma_4$ correspond to ${\bf 1}_{--}$ and ${\bf 1}_{-+}$. The twisted sector on the $T^2/Z_2$ orbifold has the flavor symmetry $(D_4 \times D_4)/Z_2$. However, this flavor symmetry seems independent of the above $D_6$, because they do not include any common elements. The twisted sector on the $S^1/Z_2$ orbifold has the flavor symmetry $D_4$. The flavor symmetry of $T^2/Z_2$ orbifold is obtained as a kind of product, $D_4 \times D_4$, although two $D_4$ groups have a common $Z_2$ element. Thus, the flavor symmetry of $T^2/Z_2$ originates from the product of symmetries of the one-dimensional orbifold. On the other hand, the modular symmetry appears in two or more dimensions, but not in one dimension. Hence, these symmetries would be independent. When we include the above $D_6$ as low-energy effective field theory in addition to the flavor symmetry $(D_4 \times D_4)/Z_2$, low-energy effective field theory would have larger symmetry including $D_6$ and $(D_4 \times D_4)/Z_2$, although Yukawa couplings as well as higher order couplings transform non-trivially under $D_6$. ### $T^2/Z_3$ orbifold The $S$ and $T$ transformations are represented by the first twisted sectors of $T^2/Z_3$ orbifold as [@Lauer:1989ax], $$\begin{aligned} & & \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \\ \sigma_{\theta,2} \end{array} \right) \longrightarrow S_{Z_3}\left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \\ \sigma_{\theta,2} \end{array} \right), \qquad S_{Z_3} = \frac{1}{\sqrt 3}\left( \begin{array}{ccc} 1 & 1 & 1\\ 1 & e^{2\pi i/3} & e^{-2\pi i/3} \\ 1 & e^{-2\pi i/3} & e^{2\pi i/3} \end{array} \right), \nonumber \\ & & \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \\ \sigma_{\theta,2} \end{array} \right) \longrightarrow T_{Z_3}\left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \\ \sigma_{\theta,2} \end{array} \right), \qquad T_{Z_3} = \left( \begin{array}{ccc} 1 & 0 & 0\\ 0 & e^{2\pi i/3} & 0 \\ 0 & 0 & e^{2\pi i/3} \end{array} \right).\end{aligned}$$ These forms look similar to $S$ and $T$ transformations in magnetized models (\[eq:magne-S\]) and (\[eq:magne-T\]). Indeed, they correspond to submatrices of $S_{(6)}$ and $T_{(6)}$ in the magnetized models with the magnetic flux $M=6$. Alternatively, in Ref. [@Lerche:1989cs] the following $S$ and $T$ representations were studied[^3] $$\begin{aligned} & & S'_{Z_3} = -\frac{i}{\sqrt 3}\left( \begin{array}{ccc} 1 & 1 & 1\\ 1 & e^{2\pi i/3} & e^{-2\pi i/3} \\ 1 & e^{-2\pi i/3} & e^{2\pi i/3} \end{array} \right), \qquad T'_{Z_3} = \left( \begin{array}{ccc} e^{2\pi i/3} & 0 & 0\\ 0 & 1 & 0 \\ 0 & 0 & 1 \end{array} \right).\end{aligned}$$ At any rate, the above representations are reducible representations. Thus, we use the flowing basis, $$\left( \begin{array}{c} \sigma_+ \\ \sigma_0 \\ \sigma_- \end{array} \right),$$ where $\sigma_\pm = (\sigma_1 \pm \sigma_-)/\sqrt 2$. The $(\sigma_+,\sigma_0)$ is a doublet, while $\sigma_-$ is a singlet. The former corresponds to the $Z_6$ invariant states among the $\theta^2$ twisted sector on the $T^2/Z_6$ orbifold. Similarly, $\sigma_-$ is the $\theta^2$ twisted state, which transforms $\sigma_- \rightarrow -\sigma_-$ under the $Z_6$ twist. Alternatively, we can say that the doublet $(\sigma_+,\sigma_0)$ corresponds to $Z_2$ even states and the singlet $\sigma_-$ is the $Z_2$ odd states, where the $Z_2$ means the $\pi$ rotation of the lattice vectors, $(\alpha_1,\alpha_2) \rightarrow (-\alpha_1,-\alpha_2)$. This point is similar to the aspect in magnetized D-brane models, where irreducible representations correspond to the $T^2/Z_2$ orbifold basis. Also, note that the first twisted states of the $T^2/Z_4$ orbifold correspond already to the $Z_2$-invariant basis. For example, we represent $S'_{Z_3}$ and $T'_{Z_3}$ on the above basis [@Lerche:1989cs] , $$S'_{Z_3} = \frac{i}{\sqrt{3}}\left( \begin{array}{cc} 1 & \sqrt 2 \\ \sqrt 2 & -1 \end{array} \right), \qquad T'_{Z_3} = \left( \begin{array}{ccc} e^{2 \pi i/3} & 0 \\ 0 & 1 \end{array} \right) ,$$ on the doublet $(\sigma_+,\sigma_0)^T$, while $\sigma_-$ is the trivial singlet. Here, we define $$Z=\left( \begin{array}{cc} -1 & 0 \\ 0 & -1 \end{array} \right), \qquad \tilde T_{Z_3} = ZT'_{Z_3}.$$ Then, they satisfy the following algebraic relations [@Lerche:1989cs; @Ferrara:1989qb], $$(S'_{Z_3})^2=(\tilde T_{Z_3})^3=(S'_{Z_3}\tilde T_{Z_3})^3=Z,\qquad Z^2=\mathbb{I}.$$ This group is the so-called $T'$, which is the binary extension of $A_4 = T$. The non-Abelian discrete flavor symmetry on the $T^2/Z_3$ orbifold is $\Delta(54)$, and the three twsisted states correspond to the triplet of $\Delta(54)$. Thus, this modular symmetry seems independent of the $\Delta(54)$ flavor symmetry. Two representations are related as $$S'_{Z_3} = -i S_{Z_3}, \qquad T'_{Z_3} = e^{2\pi i /3}(T_{Z_3})^{-1}.$$ When we change phases of $S$, $T$ and $ST$, the group such as $(Z_N \times Z_M) \rtimes H$ in sections \[sec:magne-D\] \[sec:Het\] and would change to $(Z_{N'} \times Z_{M'}) \rtimes H$. Conclusion {#sec:conclusion} ========== We have studied the modular symmetries in magnetized D-brane models. Representations due to zero-modes on $T^2$ are reducible except the models with the magnetic flux $M=2$. Irreducible representations are provided by zero-modes on the $T^2/Z_2$, i.e. $Z_2$ even states and odd states. It is reasonable because $(ST)^3$ transforms the lattice vectors $(\alpha_1,\alpha_2)$ to $(-\alpha_1,-\alpha_2)$. The orders of modular groups are large, and in general, they include the $Z_8$ symmetry as the center. The $D_4$ flavor symmetry is a subgroup of the modular group, which is represented in the models with the magnetic flux $M=2$. The system including zero-modes with $M=2$, $M=4$ and larger even $M$, also includes the $D_4$ flavor symmetry, when we define transformations of couplings in a proper way. We have also studied the modular symmetries in hetetrotic orbifold models. The heterotic model on the $T^2/Z_4$ has exactly the same representation as the magnetized model with $M=2$, and the modular symmetry includes the $D_4$ flavor symmetry. The representation due to the twisted states on the $T^2/Z_3$ orbifold is reducible, similar to representations due to zero-modes in magnetized D-brane models on $T^2$. Their irreducible representations correspond to $Z_2$ even and odd states, similar to those in magnetized D-brane models. Thus, the $\Delta(54)$ flavor symmetry seems independent of the modular symmetry in the $T^2/Z_3$ orbifold models. Note that the first twisted states on the $T^2/Z_4$ are $Z_2$-invariant states. In this sense, we find that the modular symmetry is the symmetry on the $Z_2$ orbifold in both heterotic orbifold models and magnetized D-brane models. The symmetries, which remain under the $Z_2$ twist, can be realized as the modular symmetry. We have set vanishing Wilson lines. It would be interesting to extend our analysis to magnetized D-brane models with discrete Wilson lines on orbifolds [@Abe:2013bca]. It would be also interesting to extend our analysis on zero-modes to higher Kaluza-Klein modes [@Hamada:2012wj]. Four-dimensional low energy-effective field theory is modular invariant [@Ferrara:1989bc; @Ferrara:1990ei; @Cvetic:1991qm]. Anomalies of the modular symmetry were studied [@Derendinger:1991hq; @Ibanez:1992hc], and they have important aspects [@Ibanez:1991zv; @Kawabe:1994mj; @Kobayashi:2016ovu]. The non-Abelian flavor symmetries such as $D_4$ can be anomalous. (See for anomalies of non-Abelian discrete symmetries, e.g. [@Araki:2008ek; @Ishimori:2010au; @Talbert:2018nkq]. ) In certain models, the modular symmetries are related with the non-Abelian flavor symmetry $D_4$. It would be interesting to study their anomaly relations. We also give a comment on phenomenological application. Recently, the mixing angles in the lepton sector were studied in the models, whose flavor symmetries are congruence subgroups, $\Gamma(N)$ [@Feruglio:2017spp; @Kobayashi:2018vbk]. In those models, the couplings are non-trivial representations of $\Gamma(N)$ and modular functions. Our models show massless modes represent larger finite groups. It would be interesting to apply our results to derive realistic lepton mass matrices as well as quark mass matrices. Acknowledgments {#acknowledgments .unnumbered} =============== T. K. was is supported in part by MEXT KAKENHI Grant Number JP17H05395 and JSP KAKENHI Grant Number JP26247042. Non-Abelian discrete flavor symmetry in magnetized D-brane models {#sec:app-1} ================================================================= In this Appendix, we give a brief review on non-Abelian discrete flavor symmetries in magnetized D-brane models [@Abe:2009vi]. As mentioned in section \[sec:zero-mode\], the Yukawa coulings as well as higher order couplings have the coupling selection rule (\[eq:selection\]). That is, we can define $Z_g$ charges for zero-modes. Such $Z_g$ transformation is represented on $\psi^{i,M=g}$ by $$Z = \left( \begin{array}{ccccc} 1 & & & & \\ &\rho & & & \\ & & \rho^2& & \\ & & & \ddots & \\ & & & & \rho^{g-1} \end{array} \right),$$ where $\rho = e^{2 \pi i /g}$. Furthermore, their effective field theory has the following permutation symmetry, $$\psi^{i,g} \to \psi^{i+1, g},$$ and such permutation can be represented by $$\label{eq:nonableC} C =\left( \begin{array}{cccccc} 0 & 1 & 0 & 0 & \cdots & 0 \\ 0 & 0 & 1 & 0 & \cdots & 0 \\ & & & & \ddots & \\ 1 & 0 & 0 & 0 & \cdots & 0 \end{array} \right).$$ This is another $Z_g^C$ symmetry. However, these two generators do not commute each other, $$CZ = \rho ZC.$$ Thus, the flavor symmetry corresponds to the closed algebra including $Z$ and $C$. Its diagonal elements are given by $Z^mZ'^n$, i.e. $Z_g \times Z'_g$ where $$Z' = \left( \begin{array}{ccc} \rho & & \\ & \ddots & \\ & & \rho \end{array} \right),$$ and the full group corresponds to $(Z_g \times Z'_g) \rtimes Z_g^C$. Furthermore, the zero-modes $\psi^{i,M=gn}$ with the magnetic flux $M=gn$ also represent $(Z_g \times Z'_g) \rtimes Z_g^C$. The zero-modes, $\psi^{i,M=gn}$ have $Z_g$ charges (mod $g$). Under $C$, they transform as $$\psi^{i,M=gn} \to \psi^{i + n, M=gn} .$$ For example, the model with $g=2$ has the $D_4$ flavor symmetry. The zero-modes, $$\left( \begin{array}{c} \psi^{0,2} \\ \psi^{1,2} \end{array} \right),$$ correspond to the $D_4$ doublet ${\bf 2}$, where eight $D_4$ elements are represented by $$\label{eq:D4-8} \pm\left( \begin{array}{cc} 1 & 0 \\ 0 & 1 \end{array} \right), \qquad \pm \left( \begin{array}{cc} 0 & 1 \\ 1 & 0 \end{array} \right), \qquad \pm \left( \begin{array}{cc} 0 & 1 \\ -1 & 0 \end{array} \right), \qquad \pm \left( \begin{array}{cc} 1 & 0 \\ 0 & -1 \end{array} \right).$$ In addition, when the model has the zero-modes $\psi^{i,4}$ $(i=0,1,2,3)$, the zero-modes, $\psi^{0,4}$ and $\psi^{2,4}$ ( $\psi^{1,4}$ and $\psi^{3,4}$) transform each other under $C$, and they have $Z_2$ charge even (odd). Thus, $\psi^{0,4}\pm\psi^{2,4}$ correspond to ${\bf 1}_{+\pm}$ of $D_4$ representations, while $\psi^{1,4}\pm\psi^{3,4}$ correspond to ${\bf 1}_{-\pm}$. Furthermore, among the zero-modes $\psi^{i,6}$ $(i=0,1,2,3,4,5)$, the zero-modes $\psi^{i,6}$ and $\psi^{i+3,6}$ transform each other under $C$. Hence, three pairs of zero-modes, $$\left( \begin{array}{c} \psi^{0,6} \\ \psi^{3,6} \end{array} \right), \qquad \left( \begin{array}{c} \psi^{1,6} \\ \psi^{4,6} \end{array} \right), \qquad \left( \begin{array}{c} \psi^{2,6} \\ \psi^{5,6} \end{array} \right), $$ correspond to three $D_4$ doublets. These results are shown in Table \[tab:D4\]. Magnetic flux $M$ $D_4$ representations ------------------- ---------------------------------------------------------------- 2 ${\bf 2}$ 4 ${\bf 1}_{++}$, ${\bf 1}_{+-}$, ${\bf 1}_{-+}$, ${\bf 1}_{--}$ 6 3 $\times {\bf 2}$ : $D_4$ representation[]{data-label="tab:D4"} Non-Abelian discrete flavor symmetry in heterotic orbifold models {#sec:app-2} ================================================================= Here, we give a brief review on non-Abelian discrete flavor symmetries in heteotic orbifold models [@Kobayashi:2006wq]. The twisted string $x^i$ on the orbifold satisfy the following boundary condition: $$\label{eq:string-BC} x^i(\sigma = 2 \pi) = (\theta^n x(\sigma = 0))^i + \sum_k m_k \alpha_k^i,$$ similar to Eq. (\[eq:fp\]). Thus, the twisted string can be characterized by the space group element $g= (\theta^{n}, \sum_k m_k \alpha_k^i)$. The product of the two space group elements $(\theta^{n_1},v_1)$ and $(\theta^{n_2},v_2)$ is computed as $$(\theta^{n_1},v_1)(\theta^{n_2},v_2) = (\theta^{n_1}\theta^{n_2},v_1 + \theta^{n_1}v_2).$$ The space group element $g$ belongs to the same conjugacy class as $hgh^{-1}$, where $h$ is any space group element on the same orbifold. Now, let us consider the couplings among twisted strings corresponding to space group elements $(\theta^{n_k},v_k)$. Their couplings are allowed by the space group invariance if the following condition: $$\prod_k (\theta^{n_k},v_k) = (1,0),$$ is satisfied up to the conjugacy class. That includes the point group selection rule, $\prod_k \theta^{n_k} = 1$, which is the $Z_N$ invariance on the $Z_N$ orbifold. We can define discrete Abelian symmetries from the space group invariance as well as the point group invariance. These symmetries together with geometrical symmetries of orbifolds become non-Abelian discrete flavor symmetries in heterotic orbifold models. We show them explicitly on concrete orbifolds. $S^1/Z_2$ orbifold ------------------ The $S^1/Z_2$ orbifold has two fixed points, which are denoted by the space group elements, $(\theta,m\alpha)$ with $m=0,1$, where $\alpha$ is the lattice vector. In short, we denote them by $(\theta,m)$ and the corresponding twisted states are denoted by $\sigma_{(\theta,m)}$. These states transform $$\label{eq:Z2-twist} \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \end{array}\right) \longrightarrow \left( \begin{array}{cc} -1 & 0 \\ 0 & -1 \end{array} \right) \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \end{array}\right) ,$$ under the $Z_2$ twist. In addition, the space group invariance requires $\sum_k m_k =0$ (mod 2) for the couplings corresponding to the product of the space group elements $\prod_k (\theta,m_k)$ with $m_k=0,1$. Hence, we can define another $Z_2$ symmetry, under which $\sigma_{(\theta,0)}$ is even, while $\sigma_{(\theta,1)}$ is odd. That is, another $Z_2$ transformation can be written by $$\label{eq:Z2-space} \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \end{array}\right) \longrightarrow \left( \begin{array}{cc} 1 & 0 \\ 0 & -1 \end{array} \right) \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \end{array}\right) .$$ Furthermore, there is the geometrical permutation symmetry, which exchange two fixed points each other. Such a permutation is represented by $$\label{eq:Z2-perm} \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \end{array}\right) \longrightarrow \left( \begin{array}{cc} 0 & 1 \\ 1 & 0 \end{array} \right) \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \end{array}\right) .$$ The closed algebra including Eqs.(\[eq:Z2-twist\]), (\[eq:Z2-space\]) and (\[eq:Z2-perm\]) is $D_4 \simeq (Z_2 \times Z_2) \rtimes Z_2$. $T^2/Z_3$ orbifold ------------------ As shown in Section \[sec:Het\], the $T^2/Z_3$ orbifold has three fixed points denoted by $(\theta,m)$ with $m=0,1,2$, and the corresponding twisted states are denote by $\sigma_{(\theta,m)}$. The $Z_3$ twist transforms $$\label{eq:Z3-twist} \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \\ \sigma_{\theta,2} \end{array}\right) \longrightarrow \left( \begin{array}{ccc} e^{2\pi i/3} & 0 & 0 \\ 0 & e^{2 \pi i/3} & 0 \\ 0 & 0 & e^{2 \pi i/3} \end{array} \right) \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \\ \sigma_{\theta,2} \end{array}\right) .$$ The space group invariance requires $\sum_k m_k =0$ (mod 3) for the couplings corresponding to the product of the space group elements $\prod_k (\theta,m_k)$ with $m_k=0,1,2$. Then, we can define another $Z_3$ symmetry, under which $\sigma_{(\theta,m)}$ transform $$\label{eq:Z3-space} \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \\ \sigma_{\theta,2} \end{array}\right) \longrightarrow \left( \begin{array}{ccc} 1 & 0 & 0 \\ 0 & e^{2\pi i/3} & 0 \\ 0 & 0& e^{2\pi i/3} \end{array} \right) \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \\ \sigma_{\theta,2} \end{array}\right) .$$ There is also the permutation symmetry of the three fixed points, that is, $S_3$. Thus, the flavor symmetry is $\Delta(54) \simeq (Z_3 \times Z_3) \rtimes S_3$. $T^2/Z_4$ orbifold ------------------- As shown in Section \[sec:Het\], the $T^2/Z_4$ orbifold has two $\theta$ fixed points denoted by $(\theta,m)$ with $m=0,1$, and the corresponding twisted states are denote by $\sigma_{(\theta,m)}$. The $Z_4$ twist transforms $$\label{eq:Z4-twist} \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \end{array}\right) \longrightarrow \left( \begin{array}{cc} i & 0 \\ 0 & i \end{array} \right) \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \end{array}\right) .$$ The space group invariance requires $\sum_k m_k =0$ (mod 2) for the couplings corresponding to the product of the space group elements $\prod_k (\theta,m_k)$ with $m_k=0,1$. Then, we can define another $Z_2$ symmetry, under which $\sigma_{(\theta,m)}$ transform $$\label{eq:Z4-space} \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \\ \sigma_{\theta,2} \end{array}\right) \longrightarrow \left( \begin{array}{ccc} 1 & 0 \\ 0 & -1 \end{array} \right) \left( \begin{array}{c} \sigma_{\theta,0} \\ \sigma_{\theta,1} \end{array}\right) .$$ There is also the permutation symmetry of the two fixed points. Thus, the flavor symmetry is almost the same as one on the $S^1/Z_2$ orbifold. The difference is the $Z_4$ twist, although its squire is nothing but the $Z_2$ twist. Hence, the flavor symmetry can be written as $(D_4 \times Z_4)/Z_2$. $T^2/Z_2$ orbifold ------------------ As shown in Section \[sec:Het\], the $T^2/Z_4$ orbifold has two $\theta$ fixed points denoted by $(\theta,(m,n))$ with $m,n=0,1$, and the corresponding twisted states are denote by $\sigma_{\theta,(m,n)}$. The space group invariance requires $\sum_k m_k =\sum_j n_j=0$ (mod 2) for the couplings corresponding to the product of the space group elements $\prod_k (\theta,(m_k,n_j))$ with $m_k,n_j=0,1$. There are two independent permutation symmetries between $(\theta,(0,n))$ and $(\theta,(1,n))$, and $(\theta,(m,0))$ and $(\theta,(m,1))$. Thus, this structure seems be a product of two one-dimensional orbifolds, $S^1/Z_2$. However, the $Z_2$ twist is comment such as $\sigma_{\theta,(m,n)} \longrightarrow -\sigma_{\theta,(m,n)}$. Thus, the flavor symmetry can be written by $(D_4 \times D_4)/Z_2$. [99]{} G. Altarelli and F. Feruglio, Rev. Mod. Phys.  [**82**]{}, 2701 (2010) \[arXiv:1002.0211 \[hep-ph\]\]. H. Ishimori, T. Kobayashi, H. Ohki, Y. Shimizu, H. Okada and M. Tanimoto, Prog. Theor. Phys. Suppl.  [**183**]{}, 1 (2010) \[arXiv:1003.3552 \[hep-th\]\]; Lect. Notes Phys.  [**858**]{}, 1 (2012). S. F. King and C. Luhn, Rept. Prog. Phys.  [**76**]{}, 056201 (2013) \[arXiv:1301.1340 \[hep-ph\]\]. T. Kobayashi, H. P. Nilles, F. Ploger, S. Raby and M. Ratz, Nucl. Phys. B [**768**]{}, 135 (2007) \[hep-ph/0611020\]. T. Kobayashi, S. Raby and R. J. Zhang, Nucl. Phys. B [**704**]{}, 3 (2005) \[hep-ph/0409098\]. P. Ko, T. Kobayashi, J. h. Park and S. Raby, Phys. Rev. D [**76**]{}, 035005 (2007) Erratum: \[Phys. Rev. D [**76**]{}, 059901 (2007)\] \[arXiv:0704.2807 \[hep-ph\]\]. F. Beye, T. Kobayashi and S. Kuwakino, Phys. Lett. B [**736**]{}, 433 (2014) \[arXiv:1406.4660 \[hep-th\]\]. H. Abe, K. S. Choi, T. Kobayashi and H. Ohki, Nucl. Phys. B [**820**]{}, 317 (2009) \[arXiv:0904.2631 \[hep-ph\]\]. H. Abe, K. S. Choi, T. Kobayashi and H. Ohki, Phys. Rev. D [**80**]{}, 126006 (2009) \[arXiv:0907.5274 \[hep-th\]\]; Phys. Rev. D [**81**]{}, 126003 (2010) \[arXiv:1001.1788 \[hep-th\]\]. M. Berasaluce-Gonzalez, P. G. Camara, F. Marchesano, D. Regalado and A. M. Uranga, JHEP [**1209**]{}, 059 (2012) \[arXiv:1206.2383 \[hep-th\]\]. F. Marchesano, D. Regalado and L. Vazquez-Mercado, JHEP [**1309**]{}, 028 (2013) doi:10.1007/JHEP09(2013)028 \[arXiv:1306.1284 \[hep-th\]\]. H. Abe, T. Kobayashi, H. Ohki, K. Sumita and Y. Tatsuta, JHEP [**1406**]{}, 017 (2014) \[arXiv:1404.0137 \[hep-th\]\]. T. Higaki, N. Kitazawa, T. Kobayashi and K. j. Takahashi, Phys. Rev. D [**72**]{}, 086003 (2005) \[hep-th/0504019\]. T. Kobayashi and S. Nagamoto, Phys. Rev. D [**96**]{}, no. 9, 096011 (2017) \[arXiv:1709.09784 \[hep-th\]\]. D. Cremades, L. E. Ibanez and F. Marchesano, JHEP [**0405**]{}, 079 (2004) \[hep-th/0404229\]. J. Lauer, J. Mas and H. P. Nilles, Phys. Lett. B [**226**]{}, 251 (1989); Nucl. Phys. B [**351**]{}, 353 (1991). W. Lerche, D. Lust and N. P. Warner, Phys. Lett. B [**231**]{}, 417 (1989). S. Ferrara, .D. Lust and S. Theisen, Phys. Lett. B [**233**]{}, 147 (1989). H. Abe, K. S. Choi, T. Kobayashi and H. Ohki, JHEP [**0906**]{}, 080 (2009) \[arXiv:0903.3800 \[hep-th\]\]. H. Abe, T. Kobayashi and H. Ohki, JHEP [**0809**]{}, 043 (2008) \[arXiv:0806.4748 \[hep-th\]\]. T. Kobayashi and N. Ohtsubo, Phys. Lett. B [**245**]{}, 441 (1990); Int. J. Mod. Phys. A [**9**]{}, 87 (1994). T. H. Abe, Y. Fujimoto, T. Kobayashi, T. Miura, K. Nishiwaki and M. Sakamoto, JHEP [**1401**]{}, 065 (2014) \[arXiv:1309.4925 \[hep-th\]\]; Nucl. Phys. B [**890**]{}, 442 (2014) \[arXiv:1409.5421 \[hep-th\]\]. Y. Hamada and T. Kobayashi, Prog. Theor. Phys.  [**128**]{}, 903 (2012) \[arXiv:1207.6867 \[hep-th\]\]. S. Ferrara, D. Lust, A. D. Shapere and S. Theisen, Phys. Lett. B [**225**]{}, 363 (1989). S. Ferrara, N. Magnoli, T. R. Taylor and G. Veneziano, Phys. Lett. B [**245**]{}, 409 (1990). M. Cvetic, A. Font, L. E. Ibanez, D. Lust and F. Quevedo, Nucl. Phys. B [**361**]{}, 194 (1991). J. P. Derendinger, S. Ferrara, C. Kounnas and F. Zwirner, Nucl. Phys. B [**372**]{}, 145 (1992). L. E. Ibanez and D. Lust, Nucl. Phys. B [**382**]{}, 305 (1992) \[hep-th/9202046\]. L. E. Ibanez, D. Lust and G. G. Ross, Phys. Lett. B [**272**]{}, 251 (1991) \[hep-th/9109053\]. H. Kawabe, T. Kobayashi and N. Ohtsubo, Nucl. Phys. B [**434**]{}, 210 (1995) \[hep-ph/9405420\]. T. Kobayashi, S. Nagamoto and S. Uemura, PTEP [**2017**]{}, no. 2, 023B02 (2017) \[arXiv:1608.06129 \[hep-th\]\]. T. Araki, T. Kobayashi, J. Kubo, S. Ramos-Sanchez, M. Ratz and P. K. S. Vaudrevange, Nucl. Phys. B [**805**]{}, 124 (2008) \[arXiv:0805.0207 \[hep-th\]\]. J. Talbert, arXiv:1804.04237 \[hep-ph\]. F. Feruglio, arXiv:1706.08749 \[hep-ph\]. T. Kobayashi, K. Tanaka and T. H. Tatsuishi, arXiv:1803.10391 \[hep-ph\]. [^1]: In Ref.[@Beye:2014nxa], a relation between gauge symmetries and non-Abelian flavor symmetries is discussed at the enhancement point. [^2]: See also [@Higaki:2005ie]. [^3]: See also Ref. [@Ferrara:1989qb].
{ "pile_set_name": "ArXiv" }
--- abstract: 'The thermodynamics of solvable isotropic chains with arbitrary spins is addressed by the recently developed quantum transfer matrix (QTM) approach. The set of nonlinear equations which exactly characterize the free energy is derived by respecting the physical excitations at $T=0$, spinons and RSOS kinks. We argue the implication of the present formulation to spinon character formula of level $k=2S$ SU(2) WZWN model .' author: - 'J. Suzuki[^1]\' title: Spinons in Magnetic Chains of Arbitrary Spins at Finite Temperatures --- Introduction ============ The 1D spin systems have been providing problems of both physical and mathematical interests. Among them, there exists a family of solvable models of Heisenberg type with spin-$S$ [@Tak; @Bab]. This includes, for instance, $$\begin{aligned} {\cal H} &=& J \sum_{i=1}^L \{\vec{S}_i \vec{S}_{i+1}+ \frac{1}{4} \} \label{Hamilspinhalf} \\ {\cal H} &=& \frac{J}{4}\sum_{i=1}^L \{\vec{S}_i \vec{S}_{i+1}-(\vec{S}_i \vec{S}_{i+1})^2 +3 \} \label{Hamilspin1} {\nonumber\\}\end{aligned}$$ as $S=1/2, 1$, respectively. Ground state properties as well as low lying excitations have been elucidated by the powerful machinery of solvable models, the Bethe Ansatz Equation (BAE). It has been demonstrated in many contexts [@Affleck; @AGZ; @AlcMart] that the underlying field theory is level $k=2S$ SU(2) WZWN model [@Witten]. Although this 1D quantum model is equivalent to a 2D vertex model, physical excitations carry both natures of vertex models and restricted SOS models. This is firstly demonstrated in [@ReshetS] based on the $S-$ matrix argument. The space of states is identified in [@ITIJMN]. By decomposition of crystals, such double feature is made explicit in terms of “type of domain walls” and “type of domain”. There are independent justifications of this: The double feature in the spectral decomposition is shown by path space approach[@Nagoya] . See also [@HKKOTY] for the decomposition of space picture realized in Fermionic forms. Here we are interested in the finite temperature problem. Standard arguments employ the string hypothesis[@Bab; @BabTsv]. The excitation is described not in terms of “physical excitations” in the above sense, but in the “string basis”. There strings of arbitrary lengths are allowed, which results into infinitely many coupled integral equations among infinitely many unknown functions. The description successfully reproduces the expected specific heat anomaly. It may not be, however, best suitable for practical numerics. We revisit the problem via the recently developed commuting Quantum Transfer Matrix (QTM) approach [@KluZeit]-[@JSE8]. The formulation does not reply on the string hypothesis. Rather, it only relies on analyticity structures of the object called QTM[@MSPRB; @KluRSOS]. The problem of the combinatorial summation, i.e., evaluation of the partition function, then reduces to investigations on the analyticity of suitably chosen auxiliary functions. Up to now, two kinds of choices are adopted independently. \(A) The eigenvalue of the QTM as given by the Quantum Inverse Scattering Method consists of several terms. The auxiliary functions are chosen from combinations of products of these terms [@KluZeit; @JKStJ; @JKS2p; @JKSHub]. A convenient choice leads to a finite number of coupled non-linear integral equations for a finite number of unknown functions \(B) A set of auxiliary functions may be chosen from the fusion hierarchy among “generalized” QTMs [@KluRSOS; @JKSFusion; @KSS; @JSE8]. Generically, one obtains an infinite number of coupled non-linear integral equations for an infinite number of unknown functions. This can be shown to recover the conventional TBAs based on the string hypothesis. Of course, the new approach is entirely free of any assumption about excitations like string hypothesis. The spin-1 case is analyzed in the related problem, in the context of finite size corrections[@KBP]. There six functions are introduced in the spirit of (A). The structure of NLIE among them is much more involved in comparison to the spin-1/2 case, and seems to defy a simple-minded generalization to higher $S$ cases. In a sense the most subtle point in the QTM approach appears; one does not know a priori “better” set of auxiliary functions. Here, we adopt other choice of auxiliary functions, in particular for the spin 1 case the number of these functions is 3 in contrast to 6 as in [@KBP]. A simple idea of combining the two formulations (A) and (B) works well so that the generalization to arbitrary $S$ is possible. The adopted functions agree with the picture in [@ReshetS]. Roughly speaking, the fusion part (B) of the auxiliary functions is related to the RSOS piece of the excitation, while auxiliary functions from (A) correspond to spinons. We remark that the fusion hierarchy itself is not truncated, by brute force, into finite set. Instead, the spinon part makes the functional relations among them strictly closed. Thus we obtain $2S+1$ coupled integral equations for $2S+1$ unknown functions. Besides the practical advantage, it implies the universality in the description of thermodynamics of solvable quantum 1D chains, i.e., the description only in terms of objects which reduce to physical excitations in $T \rightarrow 0$. This has been already demonstrated for several models in highly correlated 1D electron systems including the supersymmetric $t-J$ model [@JKStJ], the supersymmetric extended Hubbard model [@JKS2p] and the Hubbard model [@JKSHub]. There the exact thermodynamics are formulated in terms of “spinons” and “holons”, although they lose sense at sufficiently high temperatures. The present study adds one successful example even in the fusion models and gives further supports on the above conjecture. This paper is organized as follows. In the next section, we define the main object in this approach, the quantum transfer matrix (QTM). A minimal information of the novel approach is sketched. Section 3 is devoted to a brief description of the fusion hierarchy of generalized QTMs. After these preparations, we introduce auxiliary functions and examine functional relations among them in Section 4. The analytic structure studied numerically leads to nonlinear integral equations as discussed in Section 5. Based on these equations the low temperature asymptotics is studied analytically in Section 6. The central charge of the level $k=2S$ SU(2) WZWN model is successfully recovered. We also present the numerical evaluation of specific heat of $S=1/2, 1, 3/2 $ models for wider ranges of temperatures. In Section 7, the implication of the present formulation to the spinon character formula of the WZWN model [@BLS; @Nagoya; @NakYam] is discussed. The summary of the paper is given in Section 8. QTM Formulation =============== The present QTM formulation originates from two ingredients: the equivalence theorem between 1D quantum and 2D classical systems [@MSPRB] on one side and the integrability structure on the other[@Baxbook]. Especially, the latter provides the way to introduce commuting QTMs which reduce the problem of combinatorial counting to that of analyticity of suitable auxiliary functions[@KluRSOS]. Such a strategy has been successfully applied to several interesting models [@KluRSOS]-[@JSE8]. We also mention earlier studies on thermodynamics [@Koma]-[@Mizuta] which essentially utilize only the former part of ideas. A classical analogue to solvable spin-$S$ XXX model is already found as a $2S+1$ state vertex model [@SAA]. The Boltzmann weights are identified with the matrix elements of $\widehat{\mathfrak{sl}_2}$ invariant $R^{\vee}$ matrix: $$\begin{aligned} R^{\vee}(u) &=&\sum_{j=0}^k \rho_{2k-2j}(u) P_{2k-2j} {\nonumber\\}\rho_{2k-2j} &=& \prod_{\ell=0}^{j-1} \frac{2(k-\ell)-u}{2(k-\ell)} \prod_{\ell=j}^{k-1} \frac{2(k-\ell)+u}{2(k-\ell)} \label{Rmat}\end{aligned}$$ where $k=2S$ and $P_j$ is the projector to $V_j$, the $j+1$ dimensional irreducible module of $sl_2$. We choose $\{-j/2, -j/2+1, \cdots, j/2 \} $ as basis for $V_j$. The spectral parameter $u$ represents the anisotropy of the vertex weights. The Yang-Baxter equation implies the commutation of row-to row transfer matrices for arbitrary spectral parameters $u$, $v$: with $${\cal T}^\beta_\alpha(u)=\sum_\mu\prod_{i=1}^L {\cal R}_{\alpha_i\beta_i}^{\mu_i\mu_{i+1}}(u), \label{transfer}$$ where $L$ denotes the real system size, $\alpha_i, \beta_i, \mu_i\in V_k$, ${\cal R}= P R^{\vee}$ and $P(x\otimes y) = y\otimes x$. ![ Graphical representation of the R matrix ( its element ${\cal R}_{\alpha \beta}^{\mu \mu'})$. []{data-label="Rmatfig"}](rmatrix.ps){width="20.00000%"} The Hamiltonian is obtained as the logarithmic derivative at $u=0$, $${\cal H}= J \frac{{\rm d}}{{\rm d}u} \ln {\cal T}(u){\Big|_{u=0}}. \label{logderi}$$ It is an easy exercise to verify (\[logderi\]) gives (\[Hamilspinhalf\]) and( \[Hamilspin1\]) for $S=1/2, 1$ respectively. This may be most easily done by representing $P_j$ in (\[Rmat\]) by $$P_j = \prod_{p=0, p\ne j}^k \frac{\vec{S} \vec{S} -x_p} {x_j-x_p},$$ and $x_j = 1/2 (\frac{j}{2}(\frac{j}{2}+1) - k(\frac{k}{2}+1))$. The Hamiltonian for general $S$ can be extracted similarly. This is the well-known expression of the equivalence between 1D quantum systems and 2D classical models. To utilize the equivalence in evaluating finite $T$ quantities, especially, free energy, we need to proceed further. Let us introduce Boltzmann weights $\widetilde{{\cal R}}$ ($\overline{{\cal R}}$ ) of models related to (\[Rmat\]) by clockwise (anticlockwise) $90^0$ rotations $$\widetilde{{\cal R}}_{\alpha\beta}^{\mu\nu}(v)= {{\cal R}}^{\beta\alpha}_{\mu\nu}(v) \qquad \overline{{\cal R}}_{\alpha\beta}^{\mu\nu}(v)= {{\cal R}}^{\alpha\beta}_{\nu\mu}(v).$$ ![ In the same spirit to Fig.\[Rmatfig\], R matrices, $ \widetilde{{\cal R}}$ and $\overline{{\cal R}}$ are depicted as above. []{data-label="rmatsp"}](rmatrices.ps){width="45.00000%"} The standard initial condition of the $R^{\vee}$ matrix and (\[logderi\]) lead to significant relations, $${\cal T}(u)={\cal T}_R\, {\rm e}^{u{\cal H}/J+{\cal O}(u^2)}, \qquad {\overline{\cal T}}(u)={\cal T}_L\, {\rm e}^{u{\cal H}/J+{\cal O}(u^2)}, \label{qtm-exp}$$ where ${\overline{\cal T}}$ is defined in analogy to (\[transfer\]) , replacing ${\cal R}$ by $\overline{{\cal R}}$. ${\cal T}_{R,L}$ are the right- and left-shift operators, respectively and they commute with the Hamiltonian. We are ready to apply Trotter formula; by substitution $$u=-J \beta/N, \label{u}$$ we find $$\Big({\cal T}(u)\,\overline{{\cal T}}(u)\Big)^{N/2}= {\rm e}^{-\beta{\cal H}+{\cal O}(1/N)}. \label{qtm-element}$$ where $\beta$ denotes the inverse temperature. $N$ is a large integer “Trotter” number, interpreted as a fictitious system size in a virtual direction. Thus the partition function of the quantum system (size $L$, inverse temperature $\beta$) $$Z=\lim_{N\to\infty}{\rm Tr} \Big({\cal T}(u)\,\overline{{\cal T}}(u)\Big)^{N/2}, \label{Z}$$ is identical to the partition function of an inhomogeneous 2$S$+1 vertex model with alternating rows on a virtual 2D lattice of size $L\times N$ . ![ Graphic representation of the partition function on a virtual two dimensional lattice of $N\times L$. The operators ${\cal T}, \overline{{\cal T}}$ transfer states from bottom to top, while ${\tilde {\cal T}}(u)$ and ${\cal T}_{\rm QTM}$ do from right to left. []{data-label="zmnpic"}](zmn.ps){width="40.00000%"} See (Fig .\[zmnpic\]). Although the above mapping is exact, the expression (\[Z\]) is not yet efficient. The eigenvalues of ${\cal T}(u)\,\overline{{\cal T}}(u)$ are almost degenerate. Hence it is still difficult task to evaluate the trace. The intriguing point in [@MSPRB] is to consider a transfer matrix ${\tilde {\cal T}}(u)$ propagating in the “horizontal” direction. This novel operator acting on $N$ sites, has gaps in the eigenvalues provided $T>0$. Here we adopt further sophisticate approach developed in [@KluRSOS]-[@JSE8]. Explicitly, we define QTM by $${\cal T}_{\rm QTM}(u,x)=\sum_\mu\prod_{i=1}^{N/2} {\cal R}_{\alpha_{2i-1}\,\beta_{2i-1}}^{\mu_{2i-1}\,\mu_{2i}}(u+ix) \,\widetilde{{\cal R}}_{\alpha_{2i}\,\beta_{2i}} ^{\mu_{2i}\,\mu_{2i+1}}(u-ix), \label{qtm}$$ which reduces to above-mentioned ${\tilde {\cal T}}(u)$ by putting $x=0$. ![ Graphic representation of ${\cal T}_{\rm QTM}(u,v)$. []{data-label="qtmpic"}](qtm.ps){width="45.00000%"} Fig .\[qtmpic\] represents QTM graphically . Though the introduction of the extra parameter $x$ seems to be redundant, there is a remarkable property; $$[{\cal T}_{\rm QTM}(u,x),{\cal T}_{\rm QTM}(u,x')]=0 \label{commute}$$ by fixing $u$. This originates from the fact that ${\cal R}$ and $\widetilde{{\cal R}}$ operators possess the same intertwiner. Thus for each $T$, one can associate a auxiliary complex plane $x$ to the partition function. Due to the gap in spectra, the free energy $f$ of 1D quantum spin chains is given only by the largest eigenvalue $\Lambda_{{\rm QTM}}(u,x)$, $$f=-\frac{1}{\beta}\lim_{L\to\infty}\frac{1}{L} \ln Z =-\frac{1}{\beta}\lim_{N\to\infty} \ln\Lambda_{{\rm QTM}}(u=\frac{-\beta J}{N},x=0). \label{free-energy}$$ This is the starting point of our analysis. The difficulty in evaluating (\[free-energy\]) lies in the $N$ dependence of the vertex weights. The numerical extrapolation through finite $N$ studies may be plagued by marginal perturbations[@KluZitt]. The prescription is to utilize the existence of complex plane $x$ for each $T$. The analytic properties of ${\cal T}_{\rm QTM}$ and suitably chosen auxiliary functions in the $x-$ plane make the evaluation possible and transparent. Before closing this section, we shall describe how to modify the above relations in the presence of an external magnetic field $H$, namely by inclusion of the Zeeman term $-2 H \sum_i S_i^z $ to Hamiltonian (\[logderi\]). This contribution is described by diagonal operator $D(H)$, $$\left ( \begin{array}{c} {\rm e}^{-2 S\beta H} \\ {\rm e}^{-(2S-1)\beta H} \\ \vdots \\ {\rm e}^{2S\beta H} \\ \end{array} \right ) \otimes \left( \begin{array}{c} {\rm e}^{-2S\beta H} \\ {\rm e}^{-(2S-1)\beta H} \\ \vdots \\ {\rm e}^{2S\beta H} \\ \end{array} \right) \otimes \cdots \otimes \left( \begin{array}{c} {\rm e}^{-2S\beta H} \\ {\rm e}^{-(2S-1)\beta H} \\ \vdots \\ {\rm e}^{2S\beta H} \\ \end{array} \right).$$ Thus one only has to insert this inside the trace (\[Z\]) $$Z=\lim_{N\to\infty}{\rm Tr} \Big({\cal T}(u)\,\overline{{\cal T}}(u)\Big)^{N/2} D(H). \label{ZD}$$ In the rotated frame, the effect of the insertion of $D(H)$ is translated to the boundary weight, $B(\mu_1)={\rm e}^{\beta \mu_1 H} $; $${\cal T}_{\rm QTM}(u,x)=\sum_\mu B(\mu_1) \prod_{i=1}^{N/2} {\cal R}_{\alpha_{2i-1}\,\beta_{2i-1}}^{\mu_{2i-1}\,\mu_{2i}}(u+ix) \,\widetilde{{\cal R}}_{\alpha_{2i}\,\beta_{2i}} ^{\mu_{2i}\,\mu_{2i+1}}(u-ix).$$ Fusion Hierarchy ================ We consider a hierarchy of quantum transfer matrices acting on $V_{k}^{\otimes_N}$. Let $T_j(u,x)$ be a member of the hierarchy with the auxiliary space $V_j$. ![ For $T_j(u,x)$, vertical arrows carry $V_k$ while horizontal (broken) arrow does $V_j$. []{data-label="qtmjp"}](qtmj.ps){width="45.00000%"} In other words, it is the transfer matrix of the vertex model of which spins $S(=k/2)$ are assigned to vertical edges and spins $j/2$ to horizontal edges (Fig.\[qtmjp\]) The quantity of our interest, $T_{\rm QTM}(u,x)$ coincides with $T_k(u,x)$ apart from over-all normalization, which is specified later. For brevity, we shall only give matrix elements of ${\cal R}(u)$ defining the most fundamental $T_1(u,x)$. $$\begin{aligned} {\cal R}^{\pm 1/2,\pm 1/2}_{\ell,\ell} (u)&=&u+1\pm 2 \ell , \qquad {\cal R}^{\ell'-\ell,\ell-\ell'}_{\ell',\ell} (u)= \sqrt{(k+2 +2{\rm min}(\ell,\ell')) ( k -2min (\ell, \ell') ) } \\ $$ where $\ell, \ell' \in \{-k/2 \cdots, k/2\} $ and $|\ell-\ell'| = 1$. Similar to (\[Rmat\]), the corresponding $R^{\vee}$ matrix has decomposition, $$R^{\vee}(u) = (u+k+1) P_{k+1} +(u-k-1) P_{k-1}. \label{Rsing}$$ ${\cal R}-$ matrix for $T_j(u,x)$ is obtained from the above elementary ${\cal R}(u)$ by $j-1$ times fusion in the auxiliary space. By the construction, arbitrary pairs in this hierarchy are commutative if they share the same $u$, $$[T_j(u,x),T_{j'}(u,x') ]=0.$$ This is a generalization of (\[commute\]). In the following, we fix $u$ for all QTMs and omit the dependency on $u$. Due to the consequential commutativity, one needs not to distinguish operators $T_j$ from their eigenvalues. Then the explicit eigenvalue of the most elementary transfer matrix $T_1(x)$ reads $$\begin{aligned} T_1(x) &=& \phi_{+}(x-(k-1) i) \phi_{-}(x- (k+1)i) e^{\beta H} \frac{Q(x+2 i)}{Q(x)} {\nonumber\\}& & +\phi_{-}(x+ (k-1) i)\phi_{+} (x+(k+1)i) e^{-\beta H} \frac{Q(x-2 i)}{Q(x)} {\nonumber\\}\phi_{\pm} (x) &:=& (x \pm i u) ^{N/2}, {\nonumber\\}Q(x) &:=& \prod_{j=1}^{m} (x-x_j) {\nonumber\\}\label{t1expr}\end{aligned}$$ where $x_j, (j=1,\cdots, m)$ denotes the solution to the Bethe ansatz equation: $$\frac{\phi_{-}(x_j+ (k-1) i)\phi_{+}(x_j+ (k+1) i)} {\phi_{+}(x_j- (k-1) i)\phi_{-}(x_j-(k+1) i)} = - e^{2\beta H} \frac{Q(x_j+2 i)}{Q(x_j-2 i)}.$$ The number of BAE roots, $m$, differs for different eigenstates generally, and $m=Nk/2$ for the largest eigenvalue case. By construction of the fusion hierarchy and from singularities of the intertwining operators (\[Rsing\]) by tentatively replacing $k \rightarrow j$, the following relation is valid: $$\begin{aligned} T_j(x) T_1(x-i(j+1))&= &T_{j+1}(x-i) + g_j(x) T_{j-1}(x+i) \\ \qquad g_j(x) &=& \phi_-(x-i(k+j+2)) \phi_-(x+i(k-j)) \phi_+(x-i(j+k)) \phi_+(x+i(k-j+2)).\end{aligned}$$ From this, one can prove the important functional relation ($T-$ system) by induction [@KP], $$\begin{aligned} T_p(x+i) T_p(x-i) &=& f_p(x)+T_{p-1}(x) T_{p+1}(x), \quad(p \ge 1) {\nonumber\\}f_p(x) &:=& \prod_{j=1}^p \prod_{\sigma=\pm} \phi_{\sigma} (x+i\sigma(p-k-2j+1)) \phi_{\sigma} (x+i\sigma(k-p+2j+1)) {\nonumber\\}\label{Tsys}\end{aligned}$$ and $T_0=1$. By substituting (\[t1expr\]) into (\[Tsys\]), we can successively obtain $T_p(x), (p \ge 2)$. Explicitly, $T_p(x)$ consists of a sum of $p+1$ terms, $$\begin{aligned} T_p(x) &:=& \sum_{\ell=1}^{p+1} \lambda_{\ell}^{(p)}(x) , {\nonumber\\}\lambda_{\ell}^{(p)}(x) &:=& e^{\beta H(p+2-2\ell)} \psi_{\ell}^{(p)}(x) \frac{ Q(x+i(p+1)) Q(x-i(p+1))}{Q(x+i(2\ell-p-1)) Q(x+i(2\ell-p-3)) }, {\nonumber\\}\psi_{\ell}^{(p)}(x) &:=& \prod_{j=1}^{p-\ell+1} \phi_-(x+i(p-k-2j)) \phi_+(x+i(p-k+2-2j)) {\nonumber\\}& & \times \prod_{j=1}^{\ell-1} \phi_-(x-i(p-k+2-2j)) \phi_+(x-i(p-k-2j)). {\nonumber\\}\label{tj}\end{aligned}$$ As is previously noted, $T_{k}(x)$ has a normalization trivially different from $T_{\rm QTM}(u,x)$ in the previous section: $$\begin{aligned} T_{\rm QTM}(u,x)&=& \frac{T_{k}(x)}{\prod_{p=1}^{k} \phi_0(2i p)} {\nonumber\\}\phi_0(x) &:=& x^{N/2} {\nonumber\\}\label{normalize}\end{aligned}$$ In the original problem of the spin-$S$ chain, only $T_{k}(x)$ is of our interest. The auxiliary $T_j$’s, however, make the evaluation of $T_{k}(x)$ transparent as is shown in the following. Auxiliary Functions and Functional Relations among them ======================================================== To explore the analyticity of the transfer matrix $T_{k}(x)$, we introduce $k+1$ auxiliary functions. The first $k-1$, functions, $\{ y_j(x) \}$ are already found in several literatures and have sound basis in the $sl_2$ fusion hierarchy. They are defined by [@KP; @KNS] $$y_j(x) :=\frac{T_{j-1}(x) T_{j+1}(x) }{f_j(x)}, \qquad j\ge 1.$$ The functional relations among them are sometimes referred to as $Y-$ system: $$\begin{aligned} y_j(x+i) y_j(x-i) &= &Y_{j-1}(x)Y_{j+1}(x), \qquad j \ge 1 {\nonumber\\}Y_j(x) &:=& 1+y_j(x) \label{Ysys}\end{aligned}$$ and $y_0(x):=0$ which is a consequence of eq.(\[Tsys\]). Note that the $Y-$ system is not truncated to a finite set in this case. The $k-1$-th equation, which characterizes $y_{k-1}(x)$ inevitably contains $y_{k}(x)$ in r.h.s., and so on. Thus another device is needed to construct a finite set of auxiliary functions satisfying a complete and closed set of functional relations. The remaining two functions $\mathfrak{b}(x), \bar{\mathfrak{b}}(x)$ and their “relatives” $\mathfrak{B}(x):=1+\mathfrak{b}(x), \bar{\mathfrak{B}}(x):=1+\bar{\mathfrak{b}}(x)$ play this role. We define them by ratios of $\lambda$’s in $T_{k}(x)$ as, $$\begin{aligned} \mathfrak{b}(x) &:=& \frac{\lambda_1^{(k)}(x+i)+\cdots+\lambda_k^{(k)}(x+i)} {\lambda_{k+1}^{(k)}(x+i)} {\nonumber\\}\bar{\mathfrak{b}}(x) &:=& \frac{\lambda_2^{(k)}(x-i)+\cdots+\lambda_{k+1}^{(k)}(x-i)} {\lambda_{1}^{(k)}(x-i)}. {\nonumber\\}\label{bdef}\end{aligned}$$ The following relations are direct consequences of the above definitions: $$\begin{aligned} \mathfrak{B}(x) \lambda_{k+1}^{(k)}(x+i) &=& e^{-k\beta H} \mathfrak{B}(x) \prod_{\sigma=\pm} \prod_{j=1}^k \phi_{\sigma} (x+(2j+\sigma)i) \frac{Q(x-ik)}{Q(x+ik)} = T_{k}(x+i) {\nonumber\\}\bar{\mathfrak{B}}(x) \lambda_1^{(k)}(x-i) &=& e^{k\beta H} \bar{\mathfrak{B}}(x) \prod_{\sigma=\pm} \prod_{j=1}^k \phi_{\sigma} (x-(2j-\sigma)i) \frac{Q(x+ik)}{Q(x-ik)} = T_{k}(x-i). {\nonumber\\}\label{BTrel}\end{aligned}$$ We have $k-1$ equations for $y_j, (j=1,\cdots, k-1)$ in terms of $Y_j(x), (j=1,\cdots, k-1), \mathfrak{B}(x)$ and $\bar{\mathfrak{B}}(x)$. The first $k-2$ equations are chosen directly from the $Y-$ system. In the $k-1$ th equation, we rewrite $Y_{k}(x)$ in the rhs of $Y-$ system ($j=k-1$ in (\[Ysys\])) by $\mathfrak{B}(x) \bar{\mathfrak{B}}(x)$, thanks to (\[BTrel\]) , the definitions of $y_k, Y_k$ and the functional relation (\[Tsys\]): $$y_{k-1}(x-i) y_{k-1}(x+i) = Y_{k-2}(x) \mathfrak{B}(x) \bar{\mathfrak{B}}(x). \label{YBrel}$$ Finally, equations for $\mathfrak{b}$’s in terms of $Y_j(x), (j=1,\cdots, k-1), \mathfrak{B}(x)$ and $\bar{\mathfrak{B}}(x)$ are to be found. By comparing explicit forms, one finds $$\begin{aligned} \mathfrak{b}(x) &=& e^{\beta(k+1) H} \prod_{\sigma=\pm} \frac{\phi_{\sigma}(x+i\sigma )} {\prod_{j=1}^k \phi_{\sigma}(x+ (2j+\sigma)i) } \frac{Q(x+i(k+2))}{Q(x-ik)} T_{k-1}(x) {\nonumber\\}\bar{\mathfrak{b}}(x) &=&e^{-\beta(k+1) H} \prod_{\sigma=\pm} \frac{\phi_{\sigma}(x+i\sigma )} {\prod_{j=1}^k \phi_{\sigma}(x- (2j-\sigma)i) } \frac{Q(x-i(k+2))}{Q(x+ik)} T_{k-1}(x). {\nonumber\\}\label{bTrel}\end{aligned}$$ Note that $T_{k-1}(x)$ is presented by $Y_{k-1}(x)$: $$T_{k-1}(x-i) T_{k-1}(x+i) = f_{k-1}(x) Y_{k-1}(x) \label{TYrel}$$ which originates directly from definitions of $y_{k-1}, Y_{k-1}$ and the functional relation (\[Tsys\]). In what follows, we analyze these functional relations via the Fourier transformation. One denotes $\widehat{dl}\mathfrak{b}[q]$ to mean the Fourier transformation of the logarithmic derivative of $\mathfrak{b}(x)$, $$\widehat{dl}\mathfrak{b}[q]:= \int_{-\infty}^{\infty} \frac{d \log \mathfrak{b}(x)}{dx} e^{iqx} dx,$$ and similarly for other functions. Under some assumptions on analytic properties of the auxiliary functions, the above functional relations can be transformed into algebraic equations in the Fourier space. Roughly speaking, one can solve $\widehat{dl}Q[q]$ functions in terms of $\widehat{dl}\mathfrak{B}[q]$ and $\widehat{dl}\bar{\mathfrak{B}}[q]$ by deleting $\widehat{dl}T_k[q]$ from algebraic equations originated from (\[BTrel\]). Similarly $\widehat{dl}T_{k-1}[q]$ is solved by $\widehat{dl}Y_{k-1}[q]$ from (\[TYrel\]). Substituting these results into Fourier transformations of logarithmic derivatives of (\[bTrel\]), one finds expressions of $\widehat{dl}\mathfrak{b}[q], \widehat{dl}\bar{\mathfrak{b}}[q]$ in terms of $\widehat{dl}\mathfrak{B}[q], \widehat{dl}\bar{\mathfrak{B}}[q]$ and $\widehat{dl}Y_{k-1}[q]$. After the inverse Fourier transforming and the integration over $x$, we shall find the desired finite set of equations. We will make the above mentioned analytic assumptions explicit and examine them in the next section. Before going into details, let us discuss the physical interpretation of the above functions. As is argued in [@ReshetS], the $S-$matrix of excitations in the spin-$S$ model factorizes into two pieces: spin-$\frac{1}{2}$ SU(2) $S-$matrix and RSOS $S-$matrix of $sl_2$ level $k=2S$ . This is consistent with the general expectation that the underlying field theory is the level-$k$ SU(2) WZWN model. The latter is known to “decompose” into Gaussian and $Z_{k}$ parafermionic field theories [@FZ]. One finds, see for instance [@ABF], several evidences for the equivalence between the $sl_2$ RSOS model in regime II and the $Z_{k}$ parafermion field theory in the scaling limit. In the present description, $\mathfrak{b}, \bar{\mathfrak{b}}$ are to be identified with up- and down- spinons. As we will see later, only they couple to the magnetic field directly. For the $S=1/2$ case, there are further direct evidences for this identification [@JSch]. On the other hand, $\{ y_j(x) \}$ are insensitive to the external field. We are led to identify $y_j(x)$ in our choice as the RSOS piece of the excitations. The RSOS model possesses a subset of the $Y-$ system (\[Ysys\]). The additional condition $y_{k}=0$ for the model leads to the truncated set of equations among $k-1$ $y$’s [@BR; @KP]. In the present problem, at sufficiently low temperatures, one observes, $|\mathfrak{b}|, |\bar{\mathfrak{b}}| \sim 0$ for $x<< \ln \beta$. Thus the substitution of $\mathfrak{B}=\bar{\mathfrak{B}}=1$ in (\[YBrel\]) might be legitimate in the vicinity of the origin. Then the resultant approximated $Y-$ system coincides with that of the RSOS model. In this sense, (\[YBrel\]) represents a gluing relation between spinon and RSOS parts of excitations. Nonlinear Integral Equations ============================ We derive the nonlinear integral equations among auxiliary functions introduced in the previous section. The crucial observation is, that all nontrivial zeros and singularities of these functions are determined by zeros of $Q(x)$ and $T_j(x), (j=1, \cdots, k)$. For the largest eigenvalue sector of $T_k(x)$, zeros of $Q(x)$ form so-called $k-$ strings. Imaginary parts of zeros are approximately located at $(k+1)- 2\ell, \quad \ell=1,\cdots, k$. For later use, we introduce notations, $$\Psi_1(x):= Q(x-ik), \qquad \Psi_2(x):=Q(x+ik). \label{defPsi}$$ Empirically, similar patterns are found for zeros of $T_j(x)$: they distribute approximately on lines, $\Im x= \pm( k+j-2 \ell), \ell=0, \cdots, j-1$. We assume that these observation from numerics with fixed $N$ is valid and that the deviations from lines are very small in the limit $N \rightarrow \infty$. Then one deduces the following ansatz on the strips where auxiliary functions are Analytic, NonZero and have Constant asymptotic behavior (ANZC). $$\begin{aligned} &&\mathfrak{b}(x), \mathfrak{B}(x) \qquad -1< \Im x \le 0, \\ &&\bar{\mathfrak{b}}(x), \bar{\mathfrak{B}}(x) \qquad 0 \le \Im x <1, \\ &&y_j(x), Y_j(x) (j=1, \cdots, k-1), \quad T_p(x), (p=1, \cdots, k) \qquad -1 \le \Im x \le 1 \\ &&\Psi_1(x), \qquad \Im x<0 \\ &&\Psi_2(x), \qquad \Im x>0\end{aligned}$$ We find it convenient to shift the definition of the arguments in $\mathfrak{b}, \mathfrak{B}, \bar{\mathfrak{b}}$ and $ \bar{\mathfrak{B}}$. To avoid confusion, these new functions are denoted as $\mathfrak{a}$ etc, $$\mathfrak{a}(x) :=\mathfrak{b}(x-i\gamma), \qquad \bar{\mathfrak{a}}(x) :=\bar{\mathfrak{b}}(x+i\gamma),$$ and similarly for capital functions. Here $0<\gamma<1/2$ is an arbitrary but fixed parameter. Note that this is equivalent to adopt small shifts in the definition of the integration contours for the Fourier transformation. Due to the ANZC properties of $\mathfrak{b}, \bar{\mathfrak{b}}$, such modifications are almost trivial in the Fourier space. Having identified ANZC strips, we revisit eqs(\[BTrel\]). Consider the integral, $$\int_{\cal C} \frac{d}{d z} \log T_k(z) e^{i q z} dz,$$ where ${\cal C}$ encircles the edges of “square” : $[-\infty-i, \infty-i]\cup[\infty-i,\infty+i ]\cup [\infty+i,-\infty+i ]\cup[-\infty+i, -\infty-i]$ in counterclockwise manner. Due to the ANZC property of $ \frac{d}{d z} \log T_k(z)$ inside ${\cal C}$, the following equation is valid from Cauchy’s theorem: $$0= \int^{\infty}_{-\infty} \frac{d}{d x} \log T_k(x-i) e^{iq (x-i)} dx -\int^{\infty}_{-\infty} \frac{d}{d x} \log T_k(x+i) e^{iq (x+i)} dx.$$ One substitutes eqs(\[BTrel\]), rewritten in terms of $\Psi_{1,2}(x), \mathfrak{A}(x), \bar{\mathfrak{A}}(x)$, into the above equation and derives identities among $\widehat{dl}\Psi_{1,2}[q], \widehat{dl}\mathfrak{A}[q], \widehat{dl}\bar{\mathfrak{A}}[q]$, $$\begin{aligned} \widehat{dl}\Psi_1[q<0] &=& 0\\ \widehat{dl}\Psi_1[q>0] &=& \frac{e^{(1-\gamma)q}}{2\cosh q} dl \bar{\mathfrak{A}}[q] - \frac{e^{-(1-\gamma)q}}{2\cosh q} dl\mathfrak{A}[q] \\ &+& \pi i N \frac{e^{-k q} \sinh kq \cosh(1+u)q}{\cosh q \sinh q} \\ \widehat{dl}\Psi_2[q<0] &=& -\frac{e^{(1-\gamma)q}}{2\cosh q} dl \bar{\mathfrak{A}}[q] + \frac{e^{-(1-\gamma)q}}{2\cosh q} dl\mathfrak{A}[q] \\ &-& \pi i N \frac{e^{k q} \sinh kq \cosh(1+u)q}{\cosh q \sinh q} \\ \widehat{dl}\Psi_2[q>0] &=& 0,\\ $$ and $\widehat{dl}\Psi_1[q=0]=-\widehat{dl}\Psi_2[q=0]=\pi N ki$. Similarly, one can derive an identity for $\widehat{dl}y_j[q]$’s and $\widehat{dl}Y_j[q]$’s from (\[Ysys\]), and $\widehat{dl}T_{k-1}[q]$ and $\widehat{dl}Y_{k-1}[q]$ from (\[TYrel\]). Substituting these relations into the original definitions of $\mathfrak{a}$ and $ \bar{\mathfrak{a}}$, we obtain $k+1$ algebraic relations in Fourier space. (Remember $\Psi_{1,2}(x)$ are related to $Q(x)$ by (\[defPsi\]). After taking the inverse Fourier transformation and integrating over $x$ once, we arrive at the $k+1$ coupled nonlinear integral equations: $$\left( \begin{array}{c} \log y_1(x) \\ \vdots \\ \log y_{k-1}(x) \\ \log \mathfrak{a}(x) \\ \log \bar{\mathfrak{a}}(x) \\ \end{array} \right ) = \left( \begin{array}{c} 0 \\ \vdots \\ 0 \\ \beta H+ d(u,x-i\gamma) \\ - \beta H+ d(u,x+i\gamma) \\ \end{array} \right) + \cal {K} * \left( \begin{array}{c} \log Y_1(x) \\ \vdots \\ \log Y_{k-1}(x) \\ \log \mathfrak{A}(x) \\ \log \bar{\mathfrak{A}}(x) \\ \end{array} \right ) \label{nlie}$$ where $({\cal K}*g)_i$ denotes the matrix convolution, $ \sum_j \int {\cal K}_{i,j}(x-y) (g(y))_j dy$ and the “driving” function $d(u,x)$ reads $$d(u,x) = \frac{N}{2} \int \frac{\sinh uq}{q \cosh q} e^{-iqx} dq.$$ The integration constants ($\pm\beta H$) are fixed by comparing asymptotic values ($|x| \rightarrow \infty$) of both sides. Explicitly the kernel matrix is given by $${\cal K}(x) := \left( \begin{array}{cccccccc} 0,& K(x),& 0,& \cdots,& 0,& 0, & 0,& 0 \\ K(x),& 0,& K(x),& \cdots,& 0,& 0, & & \\ 0,& K(x),& & & 0,& 0, & \vdots& \vdots \\ \vdots& & & & \vdots& \vdots& & \\ 0,& &\cdots& & 0,& K(x),& 0,& 0 \\ 0,& & & &K(x),& 0,& K(x+i\gamma),&K(x-i\gamma) \\ 0,& &\cdots& & 0,&K(x-i\gamma)& F(x),& -F(x+2i(1-\gamma)) \\ 0,& &\cdots& & 0,&K(x+i\gamma)& -F(x-2i(1-\gamma)),& F(x) \\ \end{array} \right) \label{ksym}$$ where $$\begin{aligned} K(x) &:=& \frac{1}{4 \cosh\pi x/2} \\ F(x) &:=& \frac{1}{2\pi} \int_{-\infty}^{\infty} \frac{e^{-|q|-iqx}}{2 \cosh q} dq. \\\end{aligned}$$ $T_k$, in terms of these auxiliary functions, can be derived similarly. Technically, we find it convenient to introduce, $$T_k^R(x) :=\frac{T_k(x)}{\prod_{p=1}^k \phi_{e(p)}(x-2i(k+1-p)) \phi_{e(p+1)}(x+2i(k+1-p)) } \label{renormT}$$ and $e(p)=+ (-)$ for $p=$ even (odd). Then the product of two equations in (\[BTrel\]) leads to a simple algebraic relation: $$T^R_k(x-i) T^R_k(x+i) = \begin{cases} \mathfrak{B}(x) \bar{\mathfrak{B}}(x) \text{ for $k$ even} \\ \mathfrak{B}(x) \bar{\mathfrak{B}}(x) \frac{\phi_-(x+i)\phi_+(x-i)}{\phi_-(x-i)\phi_+(x+i)} \text{ for $k$ odd}\\ \end{cases} \label{defTR}$$ From ANZC property of the both sides of (\[defTR\]) in appropriate strips, the logarithmic derivative of them reduces to a simple algebraic equation in Fourier space. Taking account of the normalization (\[normalize\]) and (\[renormT\]) , we have $$\begin{aligned} \log \Lambda_{QTM}(u,x) &=& \log \Lambda_{QTM}^{(0)}(u,x) {\nonumber\\}& & +\int \frac{\log \mathfrak{A}(y)}{4 \cosh \pi/2(x-y+i\gamma)}dy+ \int \frac{\log \bar{\mathfrak{A}}(y)}{4 \cosh \pi/2(x-y-i\gamma)} dy {\nonumber\\}\log \Lambda_{QTM}^{(0)}(u,x) &:=& - \delta_{k=1 (mod 2)} \frac{N}{2} \int e^{-iqx-|q|}\frac{\sinh uq}{q\cosh q} dq \label{QTMA1} \\ & & +\sum_{p=1}^k \log\{ \frac{\phi_{e(p)}(x-2i(k+1-p)) \phi_{e(p+1)}(x+2i(k+1-p))} {\phi_0(2ip)^2 } \}. {\nonumber\\}\label{QTMA2}\end{aligned}$$ Finally put $u=-\beta J/N$ and send $N \rightarrow \infty$ [*analytically*]{}. This merely amounts to replacements: $$\begin{aligned} d(u,x) &\rightarrow& D(x) = -\frac{\beta J \pi}{ 2 \cosh \pi/2 x} {\nonumber\\}\log \Lambda_{QTM}^{(0)}(u,x=0) &\rightarrow& -\beta e_0 = -\frac{\beta J}{2} \sum_{j=1}^k \frac{(-1)^{k-j}}{j} + \delta_{k=1 (mod 2)} \beta J \log 2. \label{Ninf}\end{aligned}$$ Note that $e_0$ coincides with the known ground state energy [@Bab] after a trivial shift which stems from the difference in the normalization of $R^{\vee}$. The $k+1$ coupled nonlinear integral equations and $\log \Lambda_{QTM}$ do not carry the fictitious parameter $N$ any longer. They efficiently describe the thermodynamics of the solvable spin-$S$ XXX model. For an illustration, the specific heat for $S=1/2,1,3/2$ is evaluated for a wide range of temperature and plotted below. ![specific heats for $S=$ 3/2, 1 and 1/2 from top to bottom](sp.ps){width="45.00000%"} Each curve is produced by 10-30 minutes CPU time calculation on a Micro Sparc Work Station. In the next section, we derive the low temperature properties using (\[nlie\]), (\[QTMA1\]) and (\[Ninf\]). Analytic Evaluation of the Low Temperature Asymptotics ====================================================== We consider $T\rightarrow 0$ for the vanishing magnetic field. In sufficiently low temperature regime, $\mathfrak{a}, \log{\mathfrak{A}} $ shows a sharp crossover behavior like a step function: $|\mathfrak{a}|, |\log{\mathfrak{A}}| \ll 1$ for $|x| < \frac{2}{\pi} \log \pi \beta J $ and $|\mathfrak{a}|,|\log{\mathfrak{A}}| =O(1)$ for $|x| > \frac{2}{\pi} \log \pi \beta J $. Thus the following scaling functions [@KBP] control the asymptotic behavior: $$\begin{aligned} la^{\pm}(\xi) &:=& \log \mathfrak{a} (\pm \frac{2}{\pi}(\xi+\log \pi \beta J)), \qquad lA^{\pm}(\xi):= \log \mathfrak{A} (\pm \frac{2}{\pi}(\xi+\log \pi \beta J)), {\nonumber\\}l\bar{a}^{\pm}(\xi) &:=& \log \bar{\mathfrak{a}} (\pm \frac{2}{\pi}(\xi+\log \pi \beta J)), \qquad l\bar{A}^{\pm}(\xi):= \log \bar{\mathfrak{A}} (\pm \frac{2}{\pi}(\xi+\log \pi \beta J)), {\nonumber\\}ly_p^{\pm}(\xi) &:=& \log y_p (\pm \frac{2}{\pi}(\xi+\log \pi \beta J)), \quad lY_p^{\pm}(\xi) := \log Y_p (\pm \frac{2}{\pi}(\xi+\log \pi \beta J)).\end{aligned}$$ In terms of these scaling functions, NLIE are expressed by, $$\begin{aligned} \left( \begin{array}{c} ly_1^{\pm}(\xi) \\ \vdots \\ ly_{k-1}^{\pm} (\xi) \\ la^{\pm}(\xi) \\ l\bar{a}^{\pm} (\xi) \\ \end{array} \right ) &= & \left( \begin{array}{c} 0 \\ \vdots \\ 0 \\ - e^{-\xi \pm i\gamma \pi/2} \\ - e^{-\xi \mp i\gamma \pi/2} \\ \end{array} \right) + \bar{\cal K} * \left( \begin{array}{c} l Y_1^{\pm}(\xi) \\ \vdots \\ l Y_{k-1}^{\pm}(\xi) \\ lA^{\pm}(\xi) \\ l{\bar{A}}^{\pm}(\xi) \\ \end{array} \right ) {\nonumber\\}\bar{\cal K}(\xi) &=& \frac{2}{\pi} {\cal K}(\frac{2 \xi}{\pi}) {\nonumber\\}\label{scalingnlie}\end{aligned}$$ Note that neglect of small corrections $\sim O(T)$ leads to the decoupling equations for $\pm$. The thermal contribution (the second and the third term in (\[QTMA1\])) to $\lim_{N \rightarrow \infty} \log \Lambda_{\rm QTM}(u=-\frac{\beta J}{N},x)$ reads, $$\begin{aligned} &&\frac{e^{\pi x /2 }}{\pi^2 \beta J} [ e^{i \gamma \pi/2 } \int e^{-\xi} lA^{+} d \xi + e^{-i \gamma \pi/2 } \int e^{-\xi} l\bar{A}^{+} d \xi ] {\nonumber\\}&& + \frac{e^{-\pi x/2 }}{\pi^2 \beta J} [ e^{-i \gamma \pi/2 } \int e^{-\xi} lA^{-} d \xi + e^{i \gamma \pi/2 } \int e^{-\xi} l\bar{A}^{-} d \xi ]. {\nonumber\\}\label{T0free}\end{aligned}$$ The crucial observation in [@KBP] is that one needs not solve (\[scalingnlie\]) to evaluate (\[T0free\]) provided that the kernel matrix function satisfies a symmetry, $ {\cal K}_{i,j}(x-y)={\cal K}_{j,i}(y-x) $. This property is valid in the present case. See (\[ksym\]). We define $F_{\pm}$ by $$\begin{aligned} F_{\pm}&:=& \int_{-\infty}^{\infty} \sum_{p=1}^{k-1} [ (\frac{d}{d\xi} ly_{p}^{\pm}) lY_{p}^{\pm} - (\frac{d}{d\xi} l Y_{p}^{\pm}) ly_{p}^{\pm} ] d\xi {\nonumber\\}& & + \int_{-\infty}^{\infty}[ (\frac{d }{d\xi} l a^{\pm})lA^{\pm} + (\frac{d }{d\xi}l\bar{ a}^{\pm})l\bar{A}^{\pm} - (\frac{d }{d\xi}l A^{\pm}) la^{\pm} - \frac{d}{d\xi} l\bar{ A}^{\pm})l\bar{a}^{\pm}] d\xi. \label{Fdef}\end{aligned}$$ Then the trick in [@KBP] is as follows. First, take the derivative of both sides of (\[scalingnlie\]) and multiply them by a row vector $$\Bigl( lY_1^{\pm}(\xi), \cdots, lY_{k-1}^{\pm}(\xi), lA^{\pm}(\xi), l\bar{A}^{\pm}(\xi) \Bigr ). \label{rvec}$$ We call the resultant equality (A). Second, multiply both sides of (\[scalingnlie\]) by the derivative of the row vector (\[rvec\]), which is referred to as (B). Finally, subtract both sides of (A) and (B), and integrate over $\xi$. Then the lhs of the resultant equality is nothing but $F_{\pm}$. Remarkably, the most complicated terms in the rhs, like $$\int d\xi d \xi' lY^{+}_i(\xi) \frac{d \bar{{\cal K}}_{i,j}(\xi-\xi')}{d\xi}lY^{+}_j(\xi')$$ and $$- \int d\xi d \xi' \frac{d}{d\xi} lY^+_j(\xi) \bar{{\cal K}}_{j,i}(\xi-\xi') lY^+_i(\xi')$$ cancel with each other. To be precise, the first integral can be converted step by step, $$\begin{aligned} &=&-\int d\xi d \xi' lY^+_i(\xi) \frac{d \bar{{\cal K}}_{i,j}(\xi-\xi')}{d\xi'}lY^+_j(\xi') \\ &=& -\int d\xi d \xi' lY^+_i(\xi) \frac{d \bar{{\cal K}}_{j,i}(\xi'-\xi)}{d\xi'}lY^+_j(\xi') \\ &=& \int d\xi d \xi' lY^+_i(\xi) \bar{{\cal K}}_{j,i}(\xi'-\xi) \frac{d}{d\xi'} lY^+_j(\xi') \\ &=& \int d\xi d \xi' \frac{d }{d\xi} lY^+_j(\xi) \bar{{\cal K}}_{j,i}(\xi-\xi') lY^+_i(\xi') \\ $$ where the symmetry of the kernel matrix, partial integration and the change of integration variables $\xi \leftrightarrow \xi'$ are used. Similar cancellation happens for other terms. and the following equality results, $$F_{\pm} = 2 \int [ e^{-\xi\pm i \gamma \pi/2} lA^{\pm} + e^{-\xi\mp i \gamma \pi/2} l\bar{A}^{\pm} ] d\xi.$$ The first thermal correction (\[T0free\]) is thus given by $$\frac{ e^{\pi x/2 } F_+}{2 \pi^2 \beta J} + \frac{ e^{-\pi x/2 } F_-}{2 \pi^2 \beta J}.$$ To evaluate $F_{\pm}$ explicitly, one rewrites the integration variable from $\xi$ to $a, \bar{a}, y_p$. For example, the first summation term in $F_{\pm}$ is transformed to $$\sum_{p=1}^{k-1} \int_{y_p^\pm (-\infty)}^{y_p^\pm (\infty)} dy ( \frac{ \log(1+y)}{y}-\frac{\log y}{1+y} ) =2\sum_{p=1}^{k-1} \{ L_+(y_p^\pm (\infty))- L_+(y_p^\pm (-\infty)) \},$$ and similarly for others. $L_+(x)$ is a dilogarithm function and is related to Rogers’ dilogarithm function $L(x)$ by $ L_+(x) = L(x/(1+x))$, $$\begin{aligned} L_+(x) &:=& \frac{1}{2}\int_0^x ( \frac{ \log(1+y)}{y}-\frac{\log y}{1+y} ) dy,{\nonumber\\}L(x) &:=& -\frac{1}{2} \int_0^x (\frac{\log(1-y)}{y}+\frac{\log y}{1-y}) dy {\nonumber\\}\label{defRogers}\end{aligned}$$ The asymptotic values of scaling functions are easily extracted. For $x \rightarrow \infty$, $a^{\pm}$ coincides with original $\mathfrak{b}$. Thus one derives the limiting value by its definition (\[bdef\]) in terms of $\lambda^{(k)}_p$’s. Similar for $y^{\pm}_p(\infty)$. For $x \rightarrow -\infty$, one should rather consult (\[scalingnlie\]). We send the argument $x \rightarrow - \infty$ in both side and solve the resultant algebraic equations. The results are summarized as, $$\begin{aligned} a^{\pm}(- \infty)&=&\bar{a}^{\pm}(- \infty)=0, \qquad a^{\pm}( \infty)=\bar{a}^{\pm}( \infty)=k, {\nonumber\\}y_p^{\pm} (-\infty) &=& \frac{\sin\frac{\pi p}{k+2} \sin\frac{\pi (p+2)}{k+2} } { \sin^2\frac{\pi }{k+2} } \quad 1\le p \le k-1, {\nonumber\\}y_p^{\pm} (\infty) &=& p(p+2) \quad 1\le p \le k-1. \label{asym}\end{aligned}$$ With these pieces of information, $F_{\pm}$ is now given by $$F_{+}=F_-= 2\sum_{p=1}^{k-1} [ L(\frac{p(p+2)}{(p+1)^2}) - L( \frac{\sin\frac{\pi p}{k+2} \sin\frac{\pi (p+2)}{k+2} } { \sin^2\frac{\pi(p+1) }{k+2} } )] + 4L(\frac{k}{1+k})$$ Finally we use three relations[@Kir]: $$\begin{aligned} L(1) &=& L(x)+L(1-x) = \frac{\pi^2}{6} , \quad x \in [0,1] {\nonumber\\}2 L(1) &=& 2 L(\frac{1}{n+1}) + \sum_{j=0}^{n-1} L(\frac{1}{(1+j)^2}),\quad n \in Z_{\ge 0}, {\nonumber\\}L(1) \frac{3 n}{n+2} &=& \sum_{j=0}^{n-1} L( \frac{\sin^2\frac{\pi }{n+2} } { \sin^2\frac{\pi(j+1) }{n+2} } ), \quad n \in Z_{\ge 0}\end{aligned}$$ which yield the neat result $$F_{\pm} = \frac{ \pi^2 k}{k+2}.$$ Thus we conclude the low temperature asymptotics of free energy $$f \sim e_0 - \frac{\pi}{6 v_s \beta^2} c(k), \qquad c(k)= 3k/(k+2), \label{centralc}$$ where $v_s= J \pi/2$ coincides with the known spin velocity[@Sogo] and $c(k)$ is nothing but the central charge of level-$k$ SU(2) WZWN model. This is the desired result from the WZWN description of massless quantum spin chains. We remark that the final part of the calculation is quite parallel to that in [@BabTsv] utilizing the same dilogarithm function identities. There the spin-$S$ XXZ model is discussed via the standard string approach at “root of unity” where the number of strings is truncated finitely from the beginning. Spinon Characters ================= The character formulae obviously depend on the base of the space. Recently, the quasi-particle representation has been attracted much attention in the context of long-range interacting model [@BLS; @BPS; @NakYam], spectral decomposition of path space in lattice models [@Nagoya], or in the statistical interacting picture of Bethe Ansatz solvable models [@SUNY931]-[@BerkMC96]. See also [@Fring; @Gaite] in the different view points. For the spin-1/2 case, it has been shown recently that the novel thermodynamics formulation yields a natural spinon character and such character formula is generalized to $\widehat{\mathfrak{sl}}(n)_{k=1}$ [@JSch]. Thus it is tempting to find analogues for $\widehat{\mathfrak{sl}}(n)_{k=2S}$. The results given in the previous sections provide the first step for the simplest $n=2$ case as discussed below. Remark that we consider “chiral-half”, such that the only $+$ part contributions (say $F_+$ ) in the previous section are taken into account. The character needs the description of all excited states. In the present contexts, this information might be encoded in the additional zeros of $\mathfrak{a}, \bar{\mathfrak{a} }, y$ and their capitals in their “physical strips”. Indeed, some low excitations are identified in such a way [@JKSFusion; @KSS], and corresponding excited state TBA are derived. Such an analysis is of considerable interest, however it requires extensive numerical efforts. We leave it as a future problem and make a short-cut detour here employing the strategy in [@KNSch]. The central charge is described by the dilogarithm function of which the integration contour is simple. On the other hand, one can define an analytically continued dilogarithm function $L_{\cal C}(z)$. This is established by adopting general contour ${\cal C}$ for the integration contour of the dilogarithm function. We then generalize a successful observation from specific examples; all excitation spectra, or effective central charge $c_{\rm eff}$, in the conformal limit shall be described by $ L_{\cal C}(z)$. Namely, the replacement of a simple contour in the integral representation of the dilogarithm function by complex one leads to an excited state. Regarding $c_{\rm eff}$ as a function of $L_{\cal C}(z)$, the summation of $q^{-c_{\rm eff}/24}$ over a certain set of ${\cal C}$ is thus expected to reproduce affine characters. (Readers should not confuse this formal variable $q$, the standard notation in this field, with the Fourier variable used in previous sections.) Let us be more precise. By ${\cal C}$ we denote a contour starting from $f_-$ and terminating at $f_+$, such that it crosses firstly $[1, \infty)$ $\eta_1 (\ne 0)$ times then crosses $(-\infty,0]$ $\xi_1$ times then again $\eta_2$ times w.r.t. $[1, \infty)$ and so on. The intersections are counted as $+1 (-1) $ if the contour goes across the cut $[1, \infty)$ in the counterclockwise (clockwise) manner and $(-\infty,0]$ in the clockwise (counterclockwise) manner. (Note the definition is slightly different from [@KNSch; @JSch].) We denote this by ${\cal C}[f_-,f_+|\{\xi_1, \xi_2, \cdots \}|\{\eta_1, \eta_2, \cdots\}]$. ![A contour ${\cal C}[f_-,f_+|\{2 \}|\{-1,1\}]$.[]{data-label="fig:figcv"}](curve.ps){width="45.00000%"} The set of contours are parameterized by ${\cal S} =\{ {\cal C}[f_-^{(p)},f_+^{(p)}|\xi^{(p)}_1, \cdots, | \eta^{(p)}_1, \cdots, ] \}$, where $f^{(p)}_{\pm} = y^+_p(\pm \infty)/Y^+_p(\pm \infty), (1\le p\le k-1)$, $f^{(k)}_+=\mathfrak{a}^+(\infty)/\mathfrak{A}^+(\infty)$, $f^{(k+1)}_+=\bar{\mathfrak{a}}^+(\infty)/\bar{\mathfrak{A}}^+(\infty)$ and $f^{(k)}_-=f^{(k+1)}_-=0$. In the absence of additional zeros of auxiliary functions in the “physical strip”, we have $$0=\log(f_+^{(p)})- \sum_{p'} \mathfrak{g}_{p,p'} \log(1-f_+^{(p')}), \label{asytba}$$ where the “statistical interaction” matrix $\mathfrak{g}$ is related to the zero mode of the Fourier transformation of the kernel matrix (\[ksym\]) by $$\mathfrak{g} =I - {\cal K}[q=0]. \label{defg}$$ This is the situation we have treated in previous sections, and (\[asytba\]) follows from (\[scalingnlie\]). One replaces $\log$ in (\[asytba\]) by an analytically continued one, ${\rm Log}_{\cal C} (z)$ in excited states: $$\pi D^{(p)}= {\rm Log}_{{\cal C}^{(p)}}(f_+^{(p)})- \sum_{p'} \mathfrak{g}_{p,p'}{\rm Log }_{{\cal C}^{(p')}}(1-f_+^{(p')}), \label{excTBA}$$ where ${\rm Log}_{{\cal C}^{(p)}}(f_+^{(p)})= \log (f_+^{(p)})-2\pi i \sum_{\ell} \xi^{(p)}_{\ell}$ and so on. Here $D^{(p)}$ is introduced for consistency of both sides, and is interpreted as “chemical potential” [@KNSch]. On the other hand, $D^{(p)}$ should originate from the zeros of auxiliary functions in the physical strips. One may be able to prove that such $D^{(p)}$ actually agree with one in (\[excTBA\]), in principle. Though such microscopic derivation is yet to be done, we assume the coincidence in the following. Now, the excitation spectrum is solely implemented in the effective central charge $c_{\rm eff}({\cal S})$ $$\begin{aligned} c_{\rm eff}({\cal S}) = \frac{6}{\pi^2} \sum_p \Bigl ( L_{{\cal C}^{(p)}}(f_-^{(p)}, f_+^{(p)}) -\frac{\pi i}{2} D^{(p)} {\rm Log}_{{\cal C}^{(p)}}(1-f_+^{(p)}) \Bigr ). \label{ceff0} \end{aligned}$$ Here $D^{(p)}$ term is included by hand, so as to match its interpretation as chemical potential [@KNSch]. $L_{\cal C}$ is given by, $$L_{\cal C}(f_-,f_+) = -\frac{1}{2} \int_{\cal C} ( \frac{{\rm Log}_{\cal C}(1-z)}{z}+ \frac{{\rm Log}_{\cal C}(z)}{1-z}) dz.$$ for a path ${\cal C}$ from $f_-$ to $f_+$. After straightforward manipulations, one finds [@KNSch], $$\begin{aligned} L_{{\cal C}^{(p)}}(f_-^{(p)}, f_+^{(p)})& =& L(f_+^{(p)})-L(f_-^{(p)}) -\pi i \sum_{\ell} \xi^{(p)}_{\ell} \log(1-f_+^{(p)})- \pi i \sum_{\ell} \eta^{(p)}_{\ell} \log(f_+^{(p)}) {\nonumber\\}&& + 2 \pi^2 (\sum_{\ell} \xi^{(p)}_{\ell}) (\sum_{\ell} \eta^{(p)}_{\ell} ) -4\pi^2 \sum_{\ell} \xi^{(a)}_{\ell} (\eta^{(a)}_1+\cdots +\eta^{(a)}_{\ell}) .{\nonumber\\}\label{Lexpand} \end{aligned}$$ Remember $L(z)$ is defined in (\[defRogers\]). The substitution of (\[Lexpand\]) in (\[ceff0\]), using (\[excTBA\]) and the definitions of ${\rm Log}_{{\cal C}^{(p)}}$ leads to a remarkable result: $c_{\rm eff}$ can only be written in terms of $c(k)$ (see (\[centralc\])), $\{ \xi_{\ell}^{(p)} \} $ and $\{ \eta_{\ell}^{(p)} \} $, $$\begin{aligned} c_{\rm eff}({\cal S}) &=& c(k) - 24 {\cal T}({\cal S}) , \qquad {\cal T}({\cal S}) = \frac{1}{2} \phantom{}^t \mathbf{n} \mathfrak{g}\mathbf{n} + \sum_{a=1}^{k+1} \sum_{\ell \ge 1} \xi^{(a)}_{\ell} (\eta^{(a)}_1+\cdots +\eta^{(a)}_{\ell}) {\nonumber\\}\phantom{}^t \mathbf{n} &=& (n^{(1)}, \cdots, n^{(k+1)}) \label{ceff}\end{aligned}$$ where $n^{(p)}= \sum_{\ell} \eta^{(p)}_{\ell}$ . Note that the explicit forms of $D^{(p)}$ are not needed in the above transformation. In the following, we shall argue that the summation of $q^{-c_{\rm eff}}$ over some subset ${\cal O}$ of all possible contours reproduce the character ${\rm ch}_j(z,q)$ of level $k$ WZWN model with spin-$j$ ($j=$ some fixed integer or half-integer). We present necessary conditions for such ${\cal O}$ below. - $\eta's$ and $\xi's$ are non-negative. Such path can be parameterized by $${\cal C}[f^{(p)}_-, f^{(p)}_+ | \xi_1^{(p)}, \cdots, \xi_{n^{(p)}}^{(p)} | \underbrace{1,\cdots, 1}_{n^{(p)}}]$$ and $\xi_{\ell}^{(p)}\ge 0, (p=1,\cdots,n^{(p)})$. - For $p=k, k+1$, we require $\frac{n^{(k)}+n^{(k-1)}}{2}-j \in {\bf Z}_{\ge 0}$ in addition. - For $p \le k-1$, $\xi_{n^{(p)}}^{(p)}\ge 1$. Graphically, one can associate a Young diagram $YD^{(p)}$ to a set of winding numbers: $\{\xi^{(p)}_{\ell} \}, \{\eta^{(p)}_{\ell} \}$ (or a Young diagram with tail for some cases in $p=k,k+1$). First draw a line of length 1 downwards. Next draw a line of length $\xi^{(p)}_1$ to the left. Then again draw a line of length 1 downwards. We continue this procedure $n^{(p)}$ times. Finally draw a horizontal line from the starting point to the left and also draw a vertical line from the end point upwards. ![Young diagrams $YD^{(p)}$ of $n^{(p)}=3$ corresponding to a set of winding numbers: $\{\xi \}=\{1,1,1 \}, \{ \eta \}=\{1,1,2\}$ (left), $\{\xi \}=\{1,1,1 \}, \{ \eta \}=\{0,1,1\}$ (middle) and $\{\xi \}=\{1,1,1 \}, \{ \eta \}=\{1,1,0\}$ (right). Arrows are put for a guide to eyes. The most right diagram is termed “with tail” in the text. []{data-label="fig:fig2ab"}](ygexp.ps){width="45.00000%"} See the Figure \[fig:fig2ab\]. Obviously, the number of boxes in the diagram is equal to the second term in ${\cal T}({\cal S})$. We allow for contours which are isomorphic to a set of Young diagrams $YD^{(p)}, (p=1, \cdots k-1)$ such that - $n^{(k-2p+1)}=$ odd, $n^{(k-2p)}=$ even. - By the definition, the depth of $YD^{(p)}$ is $n^{(p)}$. The width is restricted by the maximum value $w^{(p)}_{\rm max} $ which is determined by depths of “adjacent” diagrams: $$w^{(p)}_{\rm max}= 1/2(n^{(p-1)}+n^{(p+1)}-2 n^{(p)} +\delta_{2j,k-p}),\qquad(p=1,\cdots,k-2)$$ for the fixed $j$ and $n^{(0)}=0$ . The case $p=k-1$ is exceptional, $w^{(k-1)}_{\rm max} = 1/2(n^{(k-2)}+n^{(k)}+n^{(k+1)}-2n^{(k-1)}+ \delta_{j,1/2})$. Under the above restrictions on ${\cal O}$, we find $$\begin{aligned} q^{-\triangle(j) +j/2} {\rm ch}_j(z,q) &=& \sum_{ {\cal S} \in {\cal O} } q^{-c_{\rm eff}({\cal S})/24} z^{(n^{(k)} - n^{(k+1)})/2} {\nonumber\\}&=& \sum_{n^{(k)}, n^{(k+1)} \ge 0} q^{-(n^{(k)}+n^{(k+1)})^2/ 4} \Psi_{A_k}^{n^{(k)}+n^{(k+1)}}(u_j;q) {\cal S}_{n^{(k)}, n^{(k+1)}}(z;q).\end{aligned}$$ ${\cal S}_{M,N}(z;q)$ stands for contributions from $(M,N)$ spinons; $${\cal S}_{M,N}(z;q) =\frac{1}{(q)_M (q)_N} z^{(M-N)/2}$$ resulting from summations over “nodes” $\xi^{(p)}_{\ell}\ge 0, \ell=1, \cdots, n^{(p)}$ and $p=k,k+1$. $A_k$ is the Cartan matrix for $sl_{k+1}$ and $\Psi_{A_k}$ denotes $$\Psi_{A_k}^{m_1} (u_j;q) = \sum_{m_2, m_3, \cdots, m_k} q^{1/4 m. A_k m} \prod_{i=2}^{k} \begin{bmatrix} \frac{1}{2}((2-A_k).m +u_j)_i \\ m_i \\ \end{bmatrix}$$ and $(u_j)_i= \delta_{i,2j+1}$. The summations are taken over odd (even) positive integers for $m_{\rm even (odd)}$. Note we redefine $n^{(k-\ell)}=m_{\ell+1}, \,\, 1 \le \ell \le k-1$. The appearance of the Gaussian q-binomial usually originates from combinatorics on the truncated Bratteli diagram and is the reminiscence of the RSOS model. Here the origin is also simple. It comes from the restriction on the width of Young diagrams. We denote the number of boxes in a Young diagram $YD^{(p)}$ by $b^{(p)}$. For fixed $\{ n^{(a)} \}$, ${\cal T(S) }$ assumes the same value for diagrams having identical $b^{(p)}$. This multiplicity is given by $p(n^{(p)}, w^{(p)}_{\rm max}, b^{(p)}) =$ the number of partition of $b^{(p)}$ into at most $ w^{(p)}_{\rm max}$ part, each $\le n^{(p)}$. Thanks to the generating relation, $$\sum_{b^{(p)}} p(n^{(p)},w^{(p)}_{\rm max},b^{(p)}) q^{b^{(p)}} = \begin{bmatrix} n^{(p)}+w^{(p)}_{\rm max}\\ n^{(p)} \end{bmatrix} =\begin{bmatrix} (n^{(p-1)}+n^{(p+1)}+\delta_{p,k-2j})/2\\ n^{(p)} \end{bmatrix}$$ we obtain the product of Gaussian q-binomial equivalent to the one in $\Psi_{A_k}^{m_1} (u;q) $. Our result implies that the Hilbert space of level $k$ SU(2) WZWN model is isomorphic to a part of homotopy space of Rogers’ dilogarithm function. We believe that this provides a novel interesting view point in the prototype of CFT and deserves further examinations. Summary and Discussion ====================== In this report, we formulate a novel description of the thermodynamics of solvable spin-$S$ $XXX$ models. The suitable choices of auxiliary functions yield a natural generalization of the strategy in [@KBP]. The nonlinear integral equations close finitely, which clearly differs from the string formulation[@Bab] and is obviously efficient in numerics. The resultant formulation has an interpretation in terms of physical excitations, spinons and RSOS kinks. This has been demonstrated by the calculation of the low-temperature specific heat as well as the spinon character formula. The latter, however, is derived under several assumptions on the possible homotopy class of $L_{\cal C}$. Certainly these constraints on winding numbers have “microscopic” origins in patterns of zeros of $y-$ (or $T$) functions. This has been actually demonstrated for few cases: the superintegrable 3 state chiral Potts model [@KM] and the simplest case of $sl_2$ RSOS models with open boundaries [@Melb]. We hope to report on extensive numerical investigations on zeros in the present context in the near future. Finally we mention spinon pictures arisen from different view points [@Schoutensp; @BSlevel1; @Hannover]. where explicit “spinon” bases are constructed by vertex operators. Thus the meaning of “spinon” is more transparent in comparison to the present approach. Less obvious is their concrete relation to eigenstates of spin Hamiltonians. The method also involves an uncontrolled approximation, a truncation procedure, in evaluating the partition function as well as “one body” distribution functions. This contrasts to the present approach which involves no approximation. In a sense they are complementary, and their relations are to be explored. Acknowledgments {#acknowledgments .unnumbered} =============== The author thanks A Kl[ü]{}mper for discussions and critical reading of the manuscript. He thanks A Fujii for discussions and helpful advice on numerical calculations. Thanks are also due to M Jimbo for explanation of his paper and T Miwa for discussions and his interest in this paper. [10]{} Takhtajian L, Phys. Lett. [**A 87**]{} (1982) 479. Babujian H M, Nucl. Phys. B215 (1983) 317. Affleck I, Nucl. Phys. [**B265**]{} \[FS\] (1986) 409, Phys. Rev. Lett.[**56**]{}(1986) 746. Affleck I, Gepner D, Schulz H J and Ziman T, J. Phys. [**A22**]{} (1989) 511. Alcaraz F C and Martins M J, J. Phys. [**A 21**]{} (1988) L381, ibid. 4397. Witten E, Comm. Math. Phys.[**92**]{} (1984) 455, and references there in. Reshetikhin N Yu, J. Phys. [**A24**]{} (1991) 3299. Idzumi M, Tokihiro T, Iohara K, Jimbo M, Miwa T and Nakashima T Int J Mod Phys [**A8**]{}(1993) 1479. Arakawa T, Nakanishi T Ohshima K and Tsuchiya A, Comm. Math. Phys.[**181**]{} (1996) 157. Hatayama G, Kirillov A N, Kuniba A, Okado M, Takagi T and Yamada Y, Nucl. Phys. [**B536**]{} (1998) 575. Babujian H M and Tsvelick A M, Nucl. Phys. [**B265**]{} (1986) 24. Kl[ü]{}mper A, Ann. Physik [**1** ]{}(1992) 540. Klümper A, Z. Phys. [**B 91**]{} (1993) 507. J[ü]{}ttner G, Kl[ü]{}mper A and Suzuki J, Nucl. Phys. [**B487**]{} (1997) 650. J[ü]{}ttner G, Kl[ü]{}mper A and Suzuki J, J. Phys. A. [**30**]{} (1997) 1881. J[ü]{}ttner G, Kl[ü]{}mper A and Suzuki J, Nucl. Phys. [**B512**]{}(1998) 581. J[ü]{}ttner G, Kl[ü]{}mper A and Suzuki J, Nucl. Phys. [**B 522**]{}(1998) 471 (cond-mat/9711310). Kuniba A, Sakai K and Suzuki J, Nucl Phys [**B525**]{} (1998) 597-626 (math-QA/9803056). Kl[ü]{}mper A, Euro. Phys J [**B5**]{} (1998) 677. Suzuki J, Nucl Phys [**B528**]{} (1998) 683-700 (cond-mat/9805241). Suzuki M, Phys. Rev. [**B 31**]{} (1985) 2957 Kl[ü]{}mper A and Batchelor M T, J. Phys. [**A 23**]{} (1990) L189. Kl[ü]{}mper A, Batchelor M T and Pearce P A J. Phys. [**A 24**]{} (1991) 3111 Bouwknegt P, Ludwig A W W and Schoutens K, Phys. Lett. [**B359**]{} (1995) 304. Baxter R J, “Exactly Solved Model in Statistical Mechanics”, (Academic Press, London 1982) Koma T, Prog. Theor. Phys. [**78**]{} (1987) 1213. Suzuki M and Inoue M, Prog. Theor. Phys. [**78**]{} (1987) 787. Suzuki J, Akutsu Y and Wadati M, J. Phys. Soc. Japan [**59**]{} (1990) 2667. Suzuki J, Nagao T and Wadati M, Int. J. Mod. Phys. [**B6**]{} (1992) 1119-1180. Takahashi M, Phys. Rev. [**B 43**]{} (1991) 5788. Destri C and de Vega H J, Phys. Rev. Lett. [**69**]{} (1992) 2313, Nucl. Phys. [**B438**]{} (1995) 413. Mizuta H, Nagao T and Wadati M, J. Phys. Soc. Japan [**63**]{} (1994) 3951. Sogo K, Akutsu Y and Abe T, Prog. Theoret. Phys. [**70**]{} (1983) 730, ibid, 739. Zamolodchikov A B and Fateev V A, Sov. Phys. JETP [**62**]{} (1985) 215. Andrews G, Baxter R J and Forrester P, J. Stat. Phys. [**35**]{} (1984) 193. Kl[ü]{}mper A and Pearce P A , Physica [**A183**]{} (1992) 304. Kuniba A, Nakanishi T and Suzuki J, Int.J.Mod.Phys [**A9**]{} (1994) 5215, ibid 5267 Bazhanov V V and Reshetikhin N Yu, Int.J.Mod.Phys [**A4**]{} (1989) 115. See, e.g., Kirillov A N, Prog. Theoret. Phys. Suppl. [**118**]{} (1995) 61, and references therein. Sogo K, Phys. Lett. [**A 104**]{} (1984) 51. Bernard D, Pasquier V and Serben D, Nucl. Phys. [**B428**]{} (1994) 612. Nakayashiki A and Yamada Y, Comm. Math. Phys.[**178** ]{} (1996) 179. Kedem R, Klassen T R, McCoy B M and Melzer E, Phys.Lett [**B304**]{} (1993) 263, Phys.Lett [**B307**]{}(1993) 68. Kedem R, McCoy B M and Melzer E, “The sum of Rogers, Schur and Ramanujian and the Bose-Fermi correspondence in 1+1 dimensional quantum field theory” (hepth-9304056) in “ Recent Progress in Statistical Mechanics and Quantum Field Theory”, ed. by P. Bouwknegt et al.(World Scientific 1995) Melzer E, Lett. Math. Phys. [**31**]{} (1994) 233. Berkovich A, Nucl. Phys. [**B431**]{} (1994) 315. Berkovich A and McCoy B M , Lett Math Phys [**37**]{} (1996) 49. Bytsko A and Fring A, Nucl. Phys. [**B521**]{} (1998) 573, and Nucl. Phys. [**B532**]{} (1998) 588. Gaite J, Nucl. Phys. [**B525**]{} (1998) 627. Kuniba A, Nakanishi T and Suzuki J, Mod. Phys. Lett. [**A8**]{} (1993) 1649-1659. Suzuki J, J Phys A [**31**]{} (1998) 6887-6896 (cond-mat/9805242) Kedem R and McCoy B M, J. Stat. Phys. [**71**]{} (1993) 865. O’Brien D L and Pearce P A and Warnaar S O, Nucl. Phys. [**B501**]{} (1997) 773. Schoutens K, “ Exclusion statistics for non-abelian quantum Hall states” (cond-mat/9803169). Bouwknegt P and Schoutens K, “Exclusion Statistics in Conformal Field Theory – generalized fermions and spinons for level-1 WZW theories” (hep-th/9810113). Frahm H and Stahlsmeier M, “Spinon statistics in integrable spin-S Heisenberg chains” (cond-mat/9803381). [^1]: e-mail: [email protected]
{ "pile_set_name": "ArXiv" }
--- abstract: 'In this paper, we consider spatial-division multiple-access (SDMA) systems with one base station with multiple antennae and a number of single antenna mobiles under noisy limited CSIT feedback. We propose a robust noisy limited feedback design for SDMA systems. The solution consists of a real-time robust SDMA precoding, user selection and rate adaptation as well as an offline feedback index assignment algorithm. The index assignment problem is cast into a Traveling Sales Man problem (TSP). Based on the specific structure of the feedback constellation and the precoder, we derive a low complex but asymptotically optimal solution. Simulation results show that the proposed framework has significant goodput gain compared to the traditional naive designs under noisy limited feedback channel. Furthermore, we show that the average system goodput scales as $\mathcal{O}(\frac{n_T(1-\epsilon)}{n_T-1}(C_{fb}-\log_2(N_n)))$ and $\mathcal{O}(n_T\cdot\log_2P)$ in the interference limited regime ($C_{fb} < (n_T-1)\log_2P+\log_2N_n$) and noise-limited regime respectively. Hence, despite the noisy feedback channel, the average SDMA system goodput grows with the number of feedback bits in the interference limited regime while in noise limited regime increases linearly with the number of transmit antenna and the forward channel SNR ($\log_2P$).' author: - | Tianyu Wu, Vincent K. N. Lau\ [email protected], [email protected]\ Dept of EEE, Hong Kong University of Science and Technology [^1] title: 'Design and Analysis of Multi-User SDMA Systems with Noisy Limited CSIT Feedback' --- Introduction {#sect:intro} ============ It is widely known that spatial-division multiple-access(SDMA) is an important technique to enhance the throughput of multi-user wireless systems due to spatial multiplexing. However, SDMA system requires channel state information at transmitter(CSIT). In FDD systems, only a limited number of bits(e.g. 6 bits for WiMAX [@Wimax:05]) can be allocated to carry the CSIT feedback, namely the [*limited feedback*]{}. In [@Wang:07], the authors consider multiuser MISO system with a limited total feedback bits constraint and proposed a codebook design algorithm and a CSIT decomposition algorithm. The authors of [@Kaibin:07] analyzed the asymptotic performance of a per user unitary and rate control design. Other works like [@Murch:04]-[@Goldsmith:06] studied transmit beamforming using different criteria and methods. However, in all these works, the focus was to study the quantization effects on the CSIT under noiseless feedback[^2]. In practice, the CSIT feedback may not be error-free due to the feedback channel noise. Unlike the forward channel where heavy error correction coding can be applied to the time-interleaved payload, the limited CSIT feedback has to be received at the transmitter with minimum latency and hence, time interleaving is not possible. Furthermore, in most systems, the number of bits available for feedback is very limited (such as 6 bits in WiMAX) and hence, it will be more effective to utilize all the limited bits to carry the CSIT rather than wasting some bits to protect the CSIT feedback. The issue of noisy feedback is considered in [@Murthy:08; @Zhang:08; @Lim:09; @Housfater:09]. For example, in [@Murthy:08; @Zhang:08; @Lim:09], the authors analyzed the effect of noisy feedback on the point-to-point MISO system and broadcast channel [@Housfater:09]. However, the authors did not incorporate the noisy feedback into the algorithm design. In [@Jafarkhani:07; @Tianyu:09], the authors design a channel optimized quantizer for point-to-point MISO link to incorporate the noisy feedback link. However, extension to SDMA system is not-trivial. As illustrated in [@Hanzo:07], the sensitivity of noisy feedback is much higher in SDMA systems and it is critical to take into consideration of noisy feedback in the robust limited feedback for SDMA systems. When we have noisy CSIT feedback, there may be significant performance degradation because an erroneous CSIT feedback will make the base station selecting a wrong precoder for the user, which not only decreases the received signal to noise ratio (SNR) of the user but also increases the interference from other scheduled users since their assigned precoders are no longer orthogonal to the target user. As we shall illustrate in the paper, adopting a [*naive design approach*]{} (design the limited feedback codebook assuming error-free feedback and testing its performance in the noisy limited feedback situation) will result in very poor SDMA performance. In order to obtain a feedback-error-resilient design, there are several first-order technical challenges to be addressed: - [**[Robust SDMA precoding & user selection:]{}**]{} Due to the noisy limited feedback, the CSIT index received at the base station may not be the same as that sent by the mobiles. As such, the selected precoder may not match the actual CSI at all, resulting in additional spatial interference among the selected SDMA users. - [**[Robust rate adaptation:]{}**]{} To achieve a high system goodput advantage, robust rate adaptation is needed to control packet errors due to channel outage. - [**[Robust Index Assignment:]{}**]{} Index assignment refers to the mapping of the CSIT feedback indices with the precoder entries in the codebook. With noisy feedback, CSIT index assignment plays an important role on the robustness performance of the SDMA systems. - [**[Performance Analysis:]{}**]{} Beside robust limited feedback designs, it is important to have closed-form performance results to obtain useful design insights such as the sensitivity of CSIT errors in SDMA systems. In this paper, we propose a robust noisy limited feedback design for SDMA systems. The solution consists of a real-time robust SDMA precoding, user selection and rate adaptation as well as an offline feedback index assignment algorithm. We formulate the robust index assignment problem into a Traveling Salesman Problem (TSP) and derive a low complex but asymptotically optimal solution. Simulation results show that the proposed framework has significant goodput gain compared to the traditional naive uncoded design (SDMA design assuming error-free and uncoded feedback) and naive coded design (SDMA design assuming error-free feedback but the limited CSIT feedback bits are protected by FEC) under noisy limited feedback channel. Furthermore, we show that despite the noisy feedback channels, the average system goodput of the proposed robust SDMA design scales as $\mathcal{O}(\frac{n_T(1-\epsilon)}{n_T-1}(C_{fb}-\log_2(N_n)))$ for interference limited scenario ($C_{fb} < (n_T-1)\log_2P+\log_2N_n$) and $\mathcal{O}(n_T\cdot\log_2P)$ for noise limited scenario where $n_T$ is the transmit antenna number, $\epsilon$ is the target outage probability, $C_{fb}$ is the number of feedback bits, $P$ is the transmission power and $N_n$ is a constant. We find that in interference limited scenario, the average system goodput increase linearly with the number of feedback bits for fixed feedback SER and converge to a constant number for fixed feedback SNR. In the noise limited scenario, the average system goodput increase linearly with the number of transmit antenna and the forward channel SNR ($\log_2P$). System Model {#sect:sys-model} ============ In this paper, we shall adopt the following convention. $\mathbf{X}$ denotes a matrix and $\mathbf{x}$ denotes a vector. $\mathbf{X}^\dag$ denotes matrix hermitian. Forward MIMO Fading Channel Model {#sect:model-forward} --------------------------------- In this paper, we consider a multi-user system with a base station having $n_T$ transmit antennas simultaneously transmitting to $n_T$ one antenna active users from a total of K users. We shall focus on the case when $K>n_T$ so that user scheduling in addition to precoder adaptation is important. The base station separates $n_T$ data streams to the active users by precoding. Each active user *k* is assigned a $n_T\times 1$ precoding vector $\mathbf{w}_k$. The precoder $\{\mathbf{w}_k\}_{k=1}^{n_T}$ are a set of unitary orthogonal vectors selected from a codebook of multiple sets of unitary orthogonal vectors. Let $x_k$ be the transmitted symbol of user *k* with $E[|x_k|^2]=1$ and $y_k$ denote the received symbol of user *k*. The forward channel is modeled as: $$y_k=\sqrt{\frac{P}{n_T}}\mathbf{h}_k^\dag\sum_{i\in \mathcal{A}}\mathbf{w}_ix_i+z_k \label{eqn:forward-channel-model}$$ where $P$ is the transmission power[^3], $\mathbf{h}_k$ is the $n_T \times 1$ complex channel state vector of the $k^{th}$ user, $\mathcal{A}$ is the active user set and $z_k$ is the additive white Gaussian noise with zero-mean and unit variance. We assume that the transmit antennas and users are sufficient separated so that the channel fading between different users and different antenna are modeled as independent and identically distributed (i.i.d.) complex Gaussian process with zero-mean and unit variance[^4]. We consider slow fading channels where the fading is quasi-static within a scheduling time slot for each user. This is a realistic assumption for pedestrian mobility (  5km/hr) as the packet duration is of the order of 500ns (such as Wi-Fi and B3G). Due to the quasi static fading and noisy limited CSI feedback, there is uncertainty on a user’s instantaneous mutual information (a function of the instantaneous CSI of all $K$ users) at the transmitter. Hence, there exists potential packet errors (despite the use of powerful channel coding) due to channel outage when the transmitted data rate of user $\emph{k}$ exceeds its instantaneous mutual information. Limited Feedback Processing at the Mobiles {#sect:Quantization-MS} ------------------------------------------ In this paper, we consider FDD system and assume the CSI is estimated at each user (CSIR) perfectly and fed back to the base station through a feedback channel with a limited feedback capacity constraint $C_{fb}$ bits per fading block per user. The CSIR of user *k*, $\mathbf{h}_k$, consists of two parts: channel gain $\|\mathbf{h}_k\|=\sqrt{\sum_{i=1}^{n_T} {h_k^i}^2}$ and channel shape $\tilde{\mathbf{h}}_k=\mathbf{h}_k/\|\mathbf{h}_k\|$. As will be explained in section \[sect:feedback-strategy\], the average system goodput is dominated by channel shape and channel gain has little influence on it especially in high SNR scenario[^5]. Hence in this paper, we shall focus on utilizing all the $C_{fb}$ feedback bits on the feedback of channel shape $\tilde{\mathbf{h}}_k$ and do not feed back channel gain $\|\mathbf{h}_k\|$. We assume the $K$ mobile stations quantize the channel shape of the local CSIR with a codebook consists of multiple sets of orthonormal vectors: $$\mathcal{F}=\bigcup_{m=1}^M \mathcal{V}^{(m)} \label{eqn:codebook}$$ where $\mathcal{F}$ denotes the quantization codebook with a cardinality $|\mathcal{F}|=N$, $\mathcal{V}^{(m)}$ is the $m^{th}$ orthonormal set in the codebook and $M$ is the number of orthonormal sets which is given by $M=\frac{N}{n_T}$. The $M$ orthonormal bases of $\mathcal{F}$ are generated randomly and independently similar to [@Zyczkowski:94][^6]. Define the distortion function between two $n_T\times 1$ vectors $\mathbf{v}_1,\mathbf{v}_2$ as: $$d(\mathbf{v}_1,\mathbf{v}_2)=1-|\mathbf{v}_1^\dag\mathbf{v}_2|^2=\sin^2(\angle(\mathbf{v}_1,\mathbf{v}_2)) \label{eqn:distortion}$$ At the $\emph{k}-{th}$ mobile, the quantized channel shape $\widehat{\widetilde{\mathbf{h}}}_k $ can be expressed as: $$\widehat{\widetilde{\mathbf{h}}}_k = \arg\min_{\mathbf{v}\in \mathcal{F}} d(\mathbf{v},\widetilde{\mathbf{h}}_k) \label{eqn:quantized_h}$$ In order to reduce the total feedback overhead for all $K$ users, user *k* shall decide whether to feed back its CSIT or not based on the criteria: $$d(\widetilde{\mathbf{h}}_k,\widehat{\widetilde{\mathbf{h}}}_k)<\delta; \ \ \ \ \ |\mathbf{h}_k\|^2>g_{th} \label{eqn:ch-criteria}$$ where $\delta$ is the threshold for the distortion from actual channel shape to the quantized channel shape and $g_{th}$ is the threshold for the channel gain. The motivation of (\[eqn:ch-criteria\]) is allocating the feedback bits to the users with smaller CSI quantization error so as to reduce the potential spatial interference among the SDMA streams. $\delta$ is a system parameter and can be selected offline. When the number of total user $K$ is large, $\delta$ can be selected as a very small number to decrease the quantization error. CSIT Feedback Channel Model {#sect:model-feedback} --------------------------- Unlike most of the previous literature where the limited feedback channel is assumed to be noiseless, we are interested in the more realistic case where the feedback channel may be noisy. Note that since the CSI feedback has to be delivered in a timely manner, effective FEC coding over many CSI feedbacks is not possible and hence, the feedback error cannot be ignored in practice. The CSIT of user *k* is quantized locally and encoded into $C_{fb}$ bits. The set of CSI indices sent at each of the K mobile stations $\{\mathcal{I}^{(MS)}_k\}$ and the corresponding CSI indices received at the base station $\{\mathcal{I}^{(BS)}_k\}$ both have cardinality of $N=2^{C_{fb}}$. Assume the modulation symbol in the CSIT feedback channel has constellation $\mathcal{M}$ and the corresponding mapping from the CSIT index $\mathcal{I}^{(MS)}_k$ to the constellation point $\mathbf{m}\in\mathcal{M}$ is given by the 1-1 [*index mapping function*]{} $\mathcal{M}=\xi(\mathcal{I}^{(MS)}_k)$. The probabilistic relationship between the CSIT feedback symbol sent $\xi(\mathcal{I}^{(MS)}_k)$ and the CSIT feedback symbol received by the transmitter $\xi(\mathcal{I}^{(BS)}_k)$ can be characterized by the [*feedback channel transition matrix*]{} $\mathbf{P}_{ch}=\{P_{m_l,m_k}^{ch}\}$, where: $$\begin{aligned} P_{m_l,m_k}^{ch}=\Pr\left[\xi(\mathcal{I}^{(BS)}_k)=m_l|\xi(\mathcal{I}^{(MS)}_k)=m_k\right], m_l,m_k\in\mathcal{M}. \label{eqn:transfer}\end{aligned}$$ Hence, the [*reliability*]{} of the CSIT feedback channel is characterized by the CSIT feedback channel transition matrix $\mathbf{P}_{ch}$. Note that $\mathbf{P}_{ch}$ depends on the average feedback SNR, feedback constellation and so on and can be offline evaluated analytically or numerically through simulations. Given $\mathbf{P}_{ch}$ and the index mapping rule $\xi$, the stochastic relationship between the CSIT index sent by the mobile station $\mathcal{I}^{(MS)}_k$ and the CSIT index received by the base station $\mathcal{I}^{(BS)}_k$ is characterized by the [*CSIT index transition probability*]{} $\mathbf{P}_{CSIT}=\{P_{m_l,m_k}^{CSIT}\}$ given by: $$P_{ij}^{CSIT}= \Pr\left[\mathcal{I}^{(BS)}_k=j|\mathcal{I}^{(MS)}_k=i\right]=P_{\xi(i),\xi(j)}^{ch}, i,j\in[1,N]. \label{eqn:feedback-model2}$$ Suppose the condition in (\[eqn:ch-criteria\]) is satisfied for the *k*-th mobile, the index of $\widehat{\widetilde{\mathbf{h}}}_k$, $\mathcal{I}^{(MS)}_k$, is then mapped to a constellation point $m_k$ using an index mapping function $m_k=\xi(\mathcal{I}^{(MS)}_k)$ and feed back to the base station. Due to the noisy feedback channel, selection of the index mapping function $\xi$ becomes important and will affect the robustness of the SDMA system. Base Station Processing: SDMA Precoding, User Scheduling and Rate Adaptation {#sect:feedback-strategy} ============================================================================ In this section, we shall discuss the base station processing based on the limited feedback sent from the K mobiles over a noisy feedback channel with [*index transition probability*]{} $\mathbf{P}_{CSIT}$. Specifically, we shall discuss the SDMA precoding, user selection and rate adaptation. System Goodput {#sect:system-goodput} -------------- Consider a full multiplexing system where base station schedules $n_T$ active users for SDMA. Define $\mathcal{A}$ as the active user set and $\mathbf{w}_k$ as the precoder for a scheduled user *k*. We can write the instantaneous goodput (b/s/Hz successfully received by user *k*) $\rho_k$ for user $\emph{k}$ as: $$\rho_k = r_k(\mathcal{I}^{(BS)}_k)\cdot 1[r_k(\mathcal{I}^{(BS)}_k) < C_k(\mathbf{h}_k, \mathbf{w}_k)] \label{eqn:goodput-general}$$ where $r_k$ is the data rate of the packet of user *k* and is a function of received CSIT at base station, $C_k$ is user $\emph{k}$’s instantaneous mutual information and $1(A)$ is an indicator function which is 1 if the event $A$ is true and 0 otherwise. User $\emph{k}$’s instantaneous mutual information $C_k$ is a function of its channel state $\mathbf{h}_k$ as well as the assigned precoder $\mathbf{w}_k$. Specifically, $C_k$ can be written as: $$C_k(\mathbf{h}_k, \mathbf{w}_k) = \log_2(1+\frac{\frac{P}{n_T}|\mathbf{h}^\dag_k\mathbf{w}_k|^2}{1+\sum_{j\neq k,j\in \mathcal{A}}\frac{P}{n_T}|\mathbf{h}^\dag_k\mathbf{w}_j|^2}) \label{eqn:inst-mutual}$$ We define $\theta = \angle(\widetilde{\mathbf{h}}_k,\mathbf{w}_k)$ as the angle between the actual channel shape and the assigned precoder for user *k*. Similarly, define $\phi = \angle(\widetilde{\mathbf{h}}_k,\mathbf{v}_{\mathcal{I}^{(MS)}_k})$ and $\varphi(\mathbf{v}_i,\mathbf{v}_j) = \angle(\mathbf{v}_{i},\mathbf{v}_{j})$ as the angle between the actual channel shape and the quantized vector of user *k* and the angle between two $n_T \times 1$ vectors $\mathbf{v}_i, \mathbf{v}_j$. Consider high effective SNR asymptotic scenario where $\frac{P}{n_T}\|\mathbf{h}_k\|^2$ is sufficiently large. We can further simplify (\[eqn:inst-mutual\]) to: $$\begin{aligned} C_k &\approx& \log_2(1+\frac{\cos^2 \theta}{\sum_{j\neq k,j\in \mathcal{A}}\cos^2\varphi(\widetilde{\mathbf{h}}_k,\mathbf{w}_j)}) \label{eqn:inst-mutual-high-SNR}\end{aligned}$$ As shown in Figure \[fig:approx-capacity\], the approximation is quite good for moderate to high SNR. Remark 1: While the approximation in (\[eqn:inst-mutual-high-SNR\]) fails when $\sum_{j\neq k,j\in \mathcal{A}}\cos^2\varphi(\widetilde{\mathbf{h}}_k,\mathbf{w}_j)\rightarrow 0$, it will not affect our design and analysis because for a practical target PER (e.g. $10^{-2}$), when $\sum_{j\neq k,j\in \mathcal{A}}\cos^2\varphi(\widetilde{\mathbf{h}}_k,\mathbf{w}_j)\rightarrow 0$, both $C_k$ in (\[eqn:inst-mutual\]) and (\[eqn:inst-mutual-high-SNR\]) will be large enough and no outage will occur. In other words, the case $\sum_{j\neq k,j\in \mathcal{A}}\cos^2\varphi(\widetilde{\mathbf{h}}_k,\mathbf{w}_j)\rightarrow 0$ will not be the performance bottleneck and there is no loss of generality to focus on the bottleneck case when $\sum_{j\neq k,j\in \mathcal{A}}\cos^2\varphi(\widetilde{\mathbf{h}}_k,\mathbf{w}_j)$ is not close to 0. Remark 2: From equation (\[eqn:inst-mutual-high-SNR\]), we can see that at high SNR, the instantaneous mutual information does not depend on the channel gain $\|\mathbf{h}_k\|$ and hence the average system goodput at high SNR is dominated by channel shape. The average goodput for a scheduled user *k* can be expressed as: $$\begin{aligned} \overline{\mathcal{\rho}}_k &=& \mathbb{E}_{\mathcal{I}^{(BS)}_k}\mathbb{E}_{\mathbf{h}_k}[\mathcal{\rho}_k]\nonumber \\&=& \mathbb{E}_{\mathcal{I}^{(BS)}_k}[r_k(\mathcal{I}^{(BS)}_k) \cdot Pr(r_k<C_k|\mathcal{I}^{(BS)}_k)] \label{eqn:avg-goodput}\end{aligned}$$ In practice, there is a target PER requirement $\epsilon$ associated with different application streams (e.g. $\epsilon = 0.01$ for voice applications), which is expressed as: $$1-Pr(r_k(\mathcal{I}^{(BS)}_k)<C_k|\mathcal{I}^{(BS)}_k) = \epsilon \label{eqn:outage-constraint}$$ The average system goodput for all scheduled users can be written as: $$\begin{aligned} \mathcal{G}&=&\sum_{k\in \mathcal{A}}\overline{\rho}_k = n_T \cdot (1-\epsilon) \cdot \mathbb{E}_{\mathcal{I}^{(BS)}_k}[r_k(\mathcal{I}^{(BS)}_k)] \label{eqn:avg-goodput-whole}\end{aligned}$$ SDMA Precoding and User Selection {#sect:precoding-scheduling} --------------------------------- Generally speaking, based on the CSIT feedback, the active user set $\mathcal{A}$ and corresponding precoder $\mathbf{w}_k$ shall be jointly optimized to maximize the overall average system goodput in (\[eqn:avg-goodput-whole\]). Assume that the number of total user K is large enough such that the base station can always fully schedule $n_T$ active users. A natural algorithm includes an exhaustive search over all the combinations of $n_T$ users out of a pool of K users and jointly optimize the $n_T$ precoders for the selected users to maximize the average system goodput. However, exhaustive search has exponential complexity in terms of number of users and is not practical as an online algorithm especially for large number of users. Therefore we shall adopt a simple orthogonal scheduling and precoding algorithm. At base station, take the received CSIT $\mathbf{v}_{\mathcal{I}^{(BS)}_k}$ as the estimation of $\widetilde{\mathbf{h}}_k$, the term $\cos^2\theta$ in the signal power term in (\[eqn:inst-mutual-high-SNR\]) is maximized if $\mathbf{w}_k=\mathbf{v}_{\mathcal{I}^{(BS)}_k}$. Similarly, the interference term $\sum_{j\neq k,j\in \mathcal{A}}\cos^2\varphi(\widetilde{\mathbf{h}}_k,\mathbf{w}_j)$ is minimized by choosing $\mathbf{w}_j\perp\mathbf{v}_{\mathcal{I}^{(BS)}_k}, \ \ \forall\ j\neq k,j\in \mathcal{A}$. It can be easily shown that the above two equations can be satisfied simultaneously with: $$\mathbf{v}_{\mathcal{I}^{(BS)}_k}\perp \mathbf{v}_{\mathcal{I}^{(BS)}_j}, \ \ \forall j\neq k, j,k\in\mathcal{A}. \label{eqn:schedule}$$ In other words, the received CSIT of $n_T$ scheduled users $\{\mathbf{v}_{\mathcal{I}^{(BS)}_k}\}$ form an orthonormal set $\mathcal{V}^{(m)}$ in the codebook $\mathcal{F}$. In fact, SDMA with orthogonal precoding is also called *per unitary basis stream user and rate control* (PU2RC) [@PU2RC] and has been widely used in the standards such as 3GPP-LTE\[3\]. The main feature of PU2RC is that it could accommodate limited CSIT feedback in a natural way. For instance, the multiuser precoders are selected from a codebook of multiple orthonormal bases. The importance of PU2RC for the next-generation wireless communication motivates the investigation of its performance in this paper. In this paper, we consider a simplified PU2RC system where scheduled users have single data streams, which are separated by orthogonal precoders. The precoding and user scheduling strategy is briefly summarized as follows: - *Step 1: search $N/n_T$ groups of orthonormal sets $\mathcal{V}^{(m)}$ in $\mathcal{F}$ for one set $\mathcal{V}^{(m)*}$ in which each vector is received by the base station as a CSIT feedback for at least one user. If there exists multiple satisfying sets, randomly pick one.* - *Step 2: For an orthonormal vector $\mathbf{v}_k\in\mathcal{V}^{(m)*}$, randomly select one user from the group of users with received CSIT $\mathbf{v}_k$ as the scheduled user and set $\mathbf{v}_k$ as its precoder.* - *Step 3: Repeat Step 2 for all the vectors in $\mathcal{V}^{(m)*}$* Robust Rate Adaptation {#sect:rate-adaptation} ---------------------- With the proposed precoding and user scheduling strategy, $C_k$ in (\[eqn:inst-mutual-high-SNR\]) is given by the following lemma: At high downlink SNR, when the $n_T$ scheduled users are using $n_T$ orthonormal precoders to transmit, the mutual information of a scheduled user is given by: $$C_k = -2\log_2(\sin\theta). \label{eqn:inst-mutual-3}$$ \[lem:mutual\_info\] Please refer to Appendix A for details. $r_k$ can be calculated from the requirement of the conditional PER target in (\[eqn:outage-constraint\]), which is given by: $$P_{out} = Pr(-2\log_2(\sin\theta)<r_k|\mathcal{I}_k^{(BS)}) = \epsilon \label{eqn:rate-adaptation}$$ Yet, one critical challenge in solving for $r_k$ in (\[eqn:rate-adaptation\]) is the knowledge of the CDF of $\sin\theta$ conditioned on the CSIT $\mathcal{I}^{(BS)}_k$. This is in contrast with the conventional approach of maximizing the ergodic capacity in which only the first order moment of the random variable $\sin\theta$ is needed. In the following lemma, we shall give a tight upper bound on the conditional CDF of $\sin\theta$, which is critical to solving for a closed-form rate adaptation solution. The conditional PER $P_{out}$ of the forward channel (conditioned on the limited CSIT feedback received at the BS) is given by: $$\begin{aligned} P_{out}&\leq& (1-\frac{(2^{-\frac{r_k}{2}}-\sin\varphi_{\mathcal{I}^{(BS)}_k,i^*})^{2(n_T-1)}}{\delta^{(n_T-1)}}) \cdot P^{CSIT}_{i^*,\mathcal{I}^{(BS)}_k}\nonumber \\& & + \sum_{j\in \overline{Ns^{\epsilon}}_{\mathcal{I}_k^{(BS)}}}P^{CSIT}_{j,\mathcal{I}_k^{(BS)}} \label{eqn:outage-upper-bound}\end{aligned}$$ where $i^*= \arg\max_{i\in Ns^{\epsilon}(\mathcal{I}^{(BS)}_k)}\sin\varphi_{\mathcal{I}^{(BS)}_k,i}$ and $Ns^{\epsilon}(\mathcal{I}^{(BS)}_k)$ is the set of neighboring codewords of $\mathcal{I}^{(BS)}_k$ satisfying: $\sum_{j\in Ns^{\epsilon}(\mathcal{I}^{(BS)}_k)}P^{CSIT}_{j,\mathcal{I}^{(BS)}_k}\geq 1-\epsilon$ and $\sum_{j\in Ns^{\epsilon}(\mathcal{I}^{(BS)}_k)}P^{CSIT}_{j,\mathcal{I}^{(BS)}_k}-P^{CSIT}_{i^*,\mathcal{I}^{(BS)}_k}< 1-\epsilon.$ \[lem:lemma1\] Please refer to Appendix B for details. Note that in practice, $i^*$ and $Ns^{\epsilon}(\mathcal{I}^{(BS)}_k)$ can be offline precalculated from the channel transition matrix $P^{CSIT}$ and the distortion between the codewords $\sin\varphi_{i,j}$. For example, suppose our codebook is given by $\mathbf{F}=\{\mathbf{v_1,v_2,v_3,v_4}\}$. Assume $\epsilon = 0.1$ and $P^{CSIT}_{11}=0.70, \sin\angle(\mathbf{v_1,v_1})=0; P^{CSIT}_{21}=0.10,\sin\angle(\mathbf{v_1,v_2})=0.5; P^{CSIT}_{31}=0.11, \sin\angle(\mathbf{v_1,v_3})=0.4; P^{CSIT}_{41}=0.09, \sin\angle(\mathbf{v_1,v_4})=1$, then $Ns^{0.1}(1)=\{\mathbf{v_2,v_3}\}$ and $i^*=3$. Using Lemma \[lem:lemma1\], and define $\epsilon_{res}=\epsilon-\sum_{j\in \overline{Ns^{\epsilon}}_{\mathcal{I}_k^{(BS)}}}P^{CSIT}_{j,\mathcal{I}_k^{(BS)}}$, the transmission rate of user *k* is given by: $$r_k = -2\log_2\left\{\delta^{\frac{1}{2}}(1-\frac{\epsilon_{res}}{P^{CSIT}_{i^*,\mathcal{I}^{(BS)}_k}})^{\frac{1}{2(n_T-1)}}+\sin\varphi_{\mathcal{I}^{(BS)}_k,i^*}\right\}. \label{eqn:rk}$$ Robust Index Assignment {#sect:index-assigment} ======================= Index mapping algorithm is important when there is noise on feedback channel. In this section, we shall optimize the index assignment function $\xi$ to maximize the system goodput in (\[eqn:avg-goodput-whole\]). This is equivalent to minimize the feedback distortion between the channel shape $\widetilde{\mathbf{h}}_k$ and corresponding precoder of user *k* $\mathbf{v}_k^{(BS)}$, which is $\sin^2\theta$. This feedback distortion is contributed by two parts: distortion from quantization $\sin^2(\phi)$ and distortion from feedback error $\sin^2(\varphi(\mathbf{v}_k^{(BS)},\mathbf{v}_k^{(MS)}))$. Based on: $$\varphi_{\mathcal{I}^{(BS)}_k,\mathcal{I}^{(MS)}_k}-\phi\leq\theta\leq \varphi_{\mathcal{I}^{(BS)}_k,\mathcal{I}^{(MS)}_k}+\phi\label{eqn:theta-phi-varphi}$$ and $\sin^2\phi<\delta$ where $\delta$ is chosen to be a small value to avoid excessive spatial interference among the SDMA streams. As a result,we shall omit quantization distortion $\sin^2\phi$ and focus on minimizing the distortion introduced by feedback error $d(\mathbf{v}_k^{(BS)},\mathbf{v}_k^{(MS)})=\sin^2(\varphi(\mathbf{v}_k^{(BS)},\mathbf{v}_k^{(MS)}))$. The average distortion is given by: $$\mathbb{E}(d)=\sum_{i=1}^N \sum_{j=1}^N Pr(\mathbf{v}_i) \cdot P^{ch}_{\xi(i),\xi(j)} \cdot d(\mathbf{v}_i,\mathbf{v}_j) \label{eqn:expected-distortion}$$ Searching for optimal index mapping function $\xi$ can be summarized into the following problem: Find an optimal index assignment function to minimize the average distortion introduced by feedback error: $$\xi^*=\arg\min_{\xi}\sum_{i=1}^N \sum_{j=1}^N Pr(\mathbf{v}_i) \cdot P^{ch}_{\xi(i),\xi(j)} \cdot d(\mathbf{v}_i,\mathbf{v}_j) \label{eqn:feedback-distortion}$$ where $Pr(\mathbf{v}_i)$ is the probability that $\mathbf{v}_i$ is the quantization output for $\mathbf{h}_k$ and $d(\mathbf{v}_i,\mathbf{v}_j)$ is the distortion between two codeword given in (\[eqn:distortion\]).\[Prob:index-assignment\] Note that the insight of the above formulation is that a good index mapping function shall map 2 precoders $\mathbf{v}_1,\mathbf{v}_{2} \in \mathcal{F}$ with smaller distortion $d(\mathbf{v}_1,\mathbf{v}_{2})$ to the constellation points $m_1,m_2$ with larger transition probability $P^{ch}_{m_1,m_2}$. In general, finding the optimal mapping $\xi(.)$ involves combinatorial search. When the number of feedback bits is small, the computation complexity of exhaustive search is still acceptable. However, when the number of transmit and receive antennas gets larger and more feedback bits are required, the exhaustive searching time will increase double exponentially with $C_{fb}$. This motivates the study on the low-complexity solution of the problem. Consider a special case when the CSIT feedback index is modulated by one $N$-PSK symbol. When feedback error occurs, the erroneous symbol is likely to be one of the adjacent neighbors of the feedback $N$-PSK symbol, which is referred as the nearest constellation error. The average distortion introduced by feedback error in (\[eqn:feedback-distortion\]) can be simplified to: $$\begin{aligned} D(\xi)&=& P_e \cdot \{ (\frac{Pr(\mathbf{v}_{\xi^{-1}(1)})d(\mathbf{v}_{\xi^{-1}(1)},\mathbf{v}_{\xi^{-1}(2)})}{2}+\nonumber \\ & & \frac{Pr(\mathbf{v}_{\xi^{-1}(2)})d(\mathbf{v}_{\xi^{-1}(2)},\mathbf{v}_{\xi^{-1}(1)})}{2})+ \cdots +\nonumber \\ & &(\frac{Pr(\mathbf{v}_{\xi^{-1}(N)})d(\mathbf{v}_{\xi^{-1}(N)},\mathbf{v}_{\xi^{-1}(1)})}{2}+ \nonumber \\ & & \frac{Pr(\mathbf{v}_{\xi^{-1}(1)})d(\mathbf{v}_{\xi^{-1}(1)},\mathbf{v}_{\xi^{-1}(1)})}{2})\} \label{eqn:distortion-adjacent-error}\end{aligned}$$ where $P_e$ denotes the symbol error rate (SER) of the feedback channel and $D(\xi)=\mathbb{E}(d)$ is the average distortion. Assume that $\{Nd_1, \cdots, Nd_N\}$ are $N$ virtual cities, and the distance between the virtual cities $Nd_i$ and $Nd_j$ is given by: $$Dis(Nd_i,Nd_j)=P_e \frac{Pr(\mathbf{v}_{i})d(\mathbf{v}_i,\mathbf{v}_j)+Pr(\mathbf{v}_{j})d(\mathbf{v}_j,\mathbf{v}_i)}{2} \label{eqn:distance-CSIT}$$ Equation (\[eqn:distortion-adjacent-error\]) can be expressed in terms of distance between [*virtual cities*]{} as follows. $$\begin{aligned} D(\xi) &=& Dis(Nd_{\xi^{-1}(N)},Nd_{\xi^{-1}(1)})+\nonumber \\ & &Dis(Nd_{\xi^{-1}(1)},Nd_{\xi^{-1}(2)})+ \cdots +\nonumber \\ & &Dis(Nd_{\xi^{-1}(N-1)},Nd_{\xi^{-1}(N)}) \label{eqn:distance-TSP}\end{aligned}$$ Hence, the optimization metric is equivalent to the total distance of a Hamiltonian cycle [@Jensen:00]. From equation (\[eqn:distance-TSP\]), the index mapping problem in Problem \[Prob:index-assignment\] is equivalent to searching shortest path in a Hamiltonian cycle and this can be cast into a [*traveling salesman problem*]{} (TSP). This is summarized below. Given a number of cities $\{Nd_1, Nd_2, \cdots, Nd_N\}$, and the costs of traveling from any city to any other city $\{Dis( Nd_i, Nd_j)\}$, what is the round-trip route $[\xi^{-1}(1),\xi^{-1}(2),\cdots,\xi^{-1}(N)]$ that visits each city exactly once and then returns to the starting city to minimize the total distance (\[eqn:distance-TSP\]). \[Prob:tsp\] TSP is found to be an NP-hard (nondeterministic polynomial time) problem and yet, there are a number of efficient searching algorithms for the TSP such as the cutting-plane method [@David:01] and genetic algorithm[@PL:99]. In this paper, we propose a simple construction algorithm, namely the *Circled Nearest Neighbor Algorithm*(CNNA). The CNNA algorithm is described below: - *Step 1: Start the TSP travel from a randomly selected node $Nd_i$.* - *Step 2: Go to the nearest unvisited node from $Nd_i$. If there exists more than one such nodes, we select the one with smallest sum distortion to the previously visited nodes. For example, if $\mathcal{S}$ is the set of nodes already visited, and $\mathcal{N}_i$ is the set of unvisited neighboring nodes of $Nd_i$, we shall select $Nd_j$ via the criteria: $Nd_j=\arg\min_{Nd_j\in\mathcal{N}_i}\sum_{Nd_k\in\mathcal{S}}dis(Nd_k,Nd_j)$* - *Step 3: Repeat Step 2 till all the nodes are visited. Then go back to the start node.* Suppose we require that all feedback error results in the “nearest neighboring precoder” and we would like to find an index assignment such that this can be realized for “2-nearest neighbor error feedback channel”. To do that, we assume all the neighboring regions of codewords are “equi-probable” and isotropically distributed on the surface of a unitary hypersphere ($\widetilde{\mathbf{h}}_k$-space). For the given topology of the partition region, the above algorithm of index assignment will result in a circled pattern as illustrated in figure \[fig:TSP-ball\] and is similar to the following analogy. Suppose we start from north pole of the Earth and travel around the world along the latitudes to the south pole. When we finish traveling along a latitude, we go down to the next one until we reach the south pole. After south pole is arrived, we go back to the north pole directly. Define $d_{min}$ as the minimum distortion between two precoder, we have the following lemma: For N-PSK constellation with nearest-constellation error approximation, the index mapping solution given by the CNNA algorithm $\xi^*$ is asymptotically optimal for sufficiently large N. i.e. $$\lim_{N\rightarrow +\infty}\frac{D(\xi^*)}{N}=d_{min}. \label{eqn:index-optimal}$$ \[lem:index-assignment-optimal\] We provide a sketch of proof due to page limit. With the proposed algorithm, all error events shall result in nearest neighbor codeword errors (for 2-nearest constellation feedback channel) except for the codewords serving as the starting and ending nodes. Hence we have $$D=N\cdot d_{min} + c \label{eqn:D-optimal}$$ where $c$ is a constant. When $N$ is sufficiently large, the average distortion of the travel is $\lim_{N\rightarrow +\infty}\frac{D}{N}=d_{min}$. Hence the proposed algorithm is asymptotically optimal. Performance Analysis {#sect:performance-analysis} ==================== In this section, we shall focus on obtaining the asymptotic goodput performance under noisy limited feedback. The transmission rate for a scheduled user in (\[eqn:rk\]) can be bounded by: $-2\log_2(\delta^\frac{1}{2}+\sin\varphi_{\mathcal{I}^{(BS)}_k,i^*}) \leq r_k \leq -2\log_2(\sin\varphi_{\mathcal{I}^{(BS)}_k,i^*})$. In fact, the upper bound and lower bound of $r_k$ above are both very tight when $\delta$ is small. Asymptotically, the two bounds will meet each other as the number of feedback bits $C_{fb}$ goes to infinity since $\delta$ will approach to 0 and $\sin\varphi_{\mathcal{I}^{(BS)}_k,i^*}$ is the dominant factor to $r_k$. Since we are interested in the first-order analysis, the data rate $r_k$ can be taken as: $$r_k = \mathcal{O}(-2\log_2(\sin\varphi_{\mathcal{I}^{(BS)}_k,i^*})) \label{eqn:rk-approx}$$ Substitute (\[eqn:rk-approx\]) into (\[eqn:avg-goodput-whole\]), the average system goodput is given by: $$\begin{aligned} \mathcal{G}&=& \mathcal{O}(-2n_T(1-\epsilon)\log_2(\mathbb{E}_{\mathcal{I}^{(BS)}_k}[\sin\varphi_{\mathcal{I}^{(BS)}_k,i^*}])) \label{eqn:avg-goodput-2}\end{aligned}$$ From (\[eqn:avg-goodput-2\]), we can see that system goodput performance depends on the worst case distortion $\sin\varphi_{\mathcal{I}^{(BS)}_k,i^*}$ of an index assignment function $\xi$. Define $N_n^\epsilon(i)$ as the set of neighboring points of a constellation point $i$ (including itself) with $\sum_{j\in N_n^\epsilon(i)} P_{ch}(i,j)=1-\epsilon$. The cardinality of the set is $N_n=|N_n^\epsilon(i)|$. Note that $N_n^\epsilon(i)$ is the $N_n$ largest terms in the $i_th$ row of $P_{ch}$.[^7] As a result, both $P_e$ and $N_n$ are two first order parameters to characterize the quality of feedback channel. We have the following lemma: For sufficiently large N, we have: $$\mathbb{E}_{\mathcal{I}^{(BS)}_k}(\sin\varphi_{\mathcal{I}^{(BS)}_k,i^*})\geq (\frac{N_n}{N})^{\frac{1}{2(n_T-1)}} \label{eqn:istar-distortion}$$ \[lem:lemma2\] Please refer to Appendix C for details. Numerical results show that with optimal or near optimal index assignment, the lower bound is quite tight as illustrated in figure \[fig:approx-TSP\]. Take the lower bound in (\[eqn:istar-distortion\]) as an approximation of $\mathbb{E}_{\mathcal{I}^{(BS)}_k}(\sin\varphi_{\mathcal{I}^{(BS)}_k,i^*})$ and substitute into (\[eqn:avg-goodput-2\]), we have the following theorem: When the number of feedback bits $C_{fb}$ satisfies $C_{fb}<(n_T-1)\log_2P+\log_2N_n$, the system is dominated by interference. For sufficiently large transmission power P and quantization codebook size N, the average system goodput is given by: $$\begin{aligned} \mathcal{G}&=& \mathcal{O}(\frac{n_T(1-\epsilon)}{n_T-1}(C_{fb}-\log_2(N_n))) \label{eqn:avg-goodput-inter-limited}\end{aligned}$$ \[lem:theorem-1\] In interference dominant system with sufficiently large transmission power P and quantization codebook size N: - [With fixed SER $P_e$ on the feedback channel, $N_n$ is a finite constant and the average system goodput is given by:]{} $$\begin{aligned} \mathcal{G}&=& \mathcal{O}(\frac{n_T(1-\epsilon)}{n_T-1}C_{fb}) \label{eqn:avg-goodput-inter-limited-fix-SER}\end{aligned}$$ - [With fixed feedback SNR, $N_n$ scales with $N$ as $N_n=c\cdot2^{C_{fb}}$ where c is a constant and the average system goodput is given by:]{} $$\begin{aligned} \mathcal{G}&=& \mathcal{O}(\frac{n_T(1-\epsilon)}{n_T-1}) \label{eqn:avg-goodput-inter-limited-fix-SNR}\end{aligned}$$ \[cor:cor1\] The goodput order of growth results in (\[eqn:avg-goodput-inter-limited-fix-SER\]) and (\[eqn:avg-goodput-inter-limited-fix-SNR\]) are also verified against simulations in Figure \[fig:goodput-cfb-fxser\] and Figure \[fig:goodput-cfb-fxsnr\]. On the other hand, when the number of feedback bits is sufficiently large, the SDMA system will operate in the noise-dominated regime. When the number of feedback bits $C_{fb}$ satisfies $C_{fb}>(n_T-1)\log_2P+\log_2N_n$, the system is noise-dominated. For sufficiently large transmission power P and quantization codebook size N, the system goodput is given by: $$\begin{aligned} \mathcal{G}&=& \mathcal{O}(n_T\cdot\log_2P)\label{eqn:avg-goodput-noise-limited}\end{aligned}$$ \[lem:theorem-2\] Please refer to Appendix D for details. The above result also reduces to that for SDMA system with noiseless feedback when $N_n=\mathcal{O}(1)$ . Note that $N_n = \mathcal{O}(1)$ actually corresponds to an asymptotically noiseless feedback channel. We shall note that $C_{fb}>(n_T-1)\log_2P+\log_2N_n$ may not be able to be satisfied by simply increasing $C_{fb}$ since $C_{fb}$ will be canceled when $N_n=c\cdot2^{C_{fb}}$ (e.g. constant feedback SNR). This indicates that one could not enhance the CSIT quality by increasing $C_{fb}$ if the feedback SNR is kept constant. Results and Discussions {#sect:results} ======================= In this section, we study the performance of the proposed robust SDMA system under noisy limited feedback. We compare the performance of proposed design with two naive designs under the same feedback cost. In the uncoded naive design, the SDMA system is designed as if the limited CSIT feedback were noiseless and the limited CSIT feedback bits are uncoded. In the coded naive design, the SDMA system is similar to the uncoded naive design except that the limited CSIT feedback bits are protected by hamming code. In the simulations, we consider an SDMA system with $n_T = 4$ forward SNR 20dB and $K=100$. We set the thresholds $\delta=0.1$ and $g_{th}=2$. System Performance with respect to the Number of Feedback Bits $C_{fb}$ {#sect:results-cfb} ----------------------------------------------------------------------- Figure \[fig:goodput-cfb-fxser\] illustrates the average goodput versus the number of feedback bits of the SDMA system with a fixed feedback SER of 0.2[^8]. There are significant goodput gain compared with both the naive uncoded and coded designs. The order-of-growth expressions in Theorem \[lem:theorem-1\] and Corollary \[cor:cor1\] are also verified. Figure \[fig:goodput-cfb-fxsnr\] illustrates the average goodput versus the modulation level per feedback symbol with a fixed average feedback SNR of 10dB and a fixed number of feedback symbols. There are also significant goodput gain in the proposed scheme compared with both the naive uncoded and coded designs. In addition, there is a tradeoff relationship in the feedback constellation level for the baseline systems. With lower feedback constellation level (such as BPSK), the feedback is more robust but the average goodput performance is limited by the resolution in the CSI feedback. On the other hand, for large feedback constellation, the average goodput performance of the reference baselines are poor because the performance is limited by the feedback error. System Performance with respect to Feedback Quality {#sect:results-fb-quality} --------------------------------------------------- The feedback quality can be specified by feedback SER and feedback SNR. Figure \[fig:goodput-ser\] shows the average system goodput versus SNR with different feedback SER and fixed feedback bits $C_{fb}=8$. It is shown that with the proposed design, the system goodput decrease much slower with the increasing of the feedback SER. In figure \[fig:goodput-snr\], we show the average system goodput versus feedback SNR with fixed feedback bits $C_{fb}=6$. With proposed design the system goodput increases much faster with the increasing of the feedback SNR. Conclusion {#sect:summary} ========== In this paper, we proposed a robust noisy limited feedback design with a joint user scheduling and precoder scheme as well as rate adaptation and robust index assignment optimization algorithms. We convert the index assignment optimization problem to a *Traveling Salesman Problem*(TSP). Simulation results show that the proposed framework has significant goodput gain compared to the uncoded and coded naive designs. Furthermore, we show that despite the noisy feedback, the average system goodput scales as $\mathcal{O}(\frac{n_T(1-\epsilon)}{n_T-1}(C_{fb}-\log_2(N_n)))$ and $\mathcal{O}(n_T\cdot\log_2P)$ in the interference limited regime ($C_{fb} < (n_T-1)\log_2P+\log_2N_n$) and noise-limited regime respectively. Appendix-A: Proof of Lemma \[lem:mutual\_info\] {#appendix-a-proof-of-lemma-lemmutual_info .unnumbered} =============================================== From equation (\[eqn:inst-mutual\]), the instantaneous mutual information can be written as: $$\begin{aligned} C_k(\mathbf{h}_k, \mathbf{w}_k) &\approx& \log_2(1+\frac{\cos^2 \theta}{\sum_{j\neq k,j\in \mathcal{A}}\cos^2\varphi(\widetilde{\mathbf{h}}_k,\mathbf{w}_j)}) \nonumber \\ &=&-2\log_2(\sin\theta) \label{eqn:inst-mutual-deduction}\end{aligned}$$ (\[eqn:inst-mutual-deduction\]) is because all the scheduled precoders $\mathbf{w}_j$ forms an orthogonal bases of the $n_T$ dimensional space and $(\cos{\varphi(\widetilde{\mathbf{h}}_k,\mathbf{w}_{j1})},\cdots,\cos{\varphi(\widetilde{\mathbf{h}}_k,\mathbf{w}_{jn_T})})$ is a vector on the unit sphere of the $n_T$ dimensional space. where $\mathbf{w}_{j1}$ to $\mathbf{w}_{jn_T}$ are $n_T$ scheduled orthonomal precoders. Hence we have: $$\cos^2\theta + \sum_{j\neq k, j\in \mathcal{A}}\cos^2{\varphi(\widetilde{\mathbf{h}}_k,\mathbf{w}_{j})})=1. \label{eqn:theta-relation}$$ Appendix-B: Proof of Lemma \[lem:lemma1\] {#appendix-b-proof-of-lemma-lemlemma1 .unnumbered} ========================================= Given received CSIT $\mathcal{I}^{(BS)}_k$, the outage probability can be written as: $$\begin{aligned} P_{out}&\approx& \frac{Pr(2^{-\frac{r_k}{2}}-\sin\varphi_{\mathcal{I}^{(BS)}_k,i^*}<\sin\phi\leq \sqrt{\delta})}{Pr(\sin\phi<\sqrt{\delta})} \cdot P^{CSIT}_{i^*,\mathcal{I}^{(BS)}_k} \nonumber \\ & &+ \sum_{j\in \overline{Ns^{\epsilon}}_{\mathcal{I}_k^{(BS)}}}P^{CSIT}_{j,\mathcal{I}_k^{(BS)}} \label{eqn:outage-upper-bound-append}\end{aligned}$$ where $i^*= \arg\max_{i\in Ns^{\epsilon}(\mathcal{I}^{(BS)}_k)}\sin\varphi_{\mathcal{I}^{(BS)}_k,i}$ and $Ns^{\epsilon}(\mathcal{I}^{(BS)}_k)$ is the set of neighboring codewords of $\mathcal{I}^{(BS)}_k$ satisfying: $\sum_{j\in Ns^{\epsilon}(\mathcal{I}^{(BS)}_k)}P^{CSIT}_{j,\mathcal{I}^{(BS)}_k}\geq 1-\epsilon$ and $\sum_{j\in Ns^{\epsilon}(\mathcal{I}^{(BS)}_k)}P^{CSIT}_{j,\mathcal{I}^{(BS)}_k}-P^{CSIT}_{i^*,\mathcal{I}^{(BS)}_k}< 1-\epsilon$. From [@Dai:08], we have the CDF for $\sin\phi$: $$Pr(\sin\phi<x) = x^{2(n_T-1)} \label{eqn:distribution-sinphi}$$ Subscribing (\[eqn:distribution-sinphi\]) into (\[eqn:outage-upper-bound-append\]), we can simplify the outage probability as: $$\begin{aligned} P_{out}&\leq& (1-\frac{(2^{-\frac{r_k}{2}}-\sin\varphi_{\mathcal{I}^{(BS)}_k,i^*})^{2(n_T-1)}}{\delta^{(n_T-1)}}) \cdot P^{CSIT}_{i^*,\mathcal{I}^{(BS)}_k} \nonumber \\ & &+ \sum_{j\in \overline{Ns^{\epsilon}}_{\mathcal{I}_k^{(BS)}}}P^{CSIT}_{j,\mathcal{I}_k^{(BS)}}. \nonumber\end{aligned}$$ Appendix-C: Proof of Lemma \[lem:lemma2\] {#appendix-c-proof-of-lemma-lemlemma2 .unnumbered} ========================================= The asymptotic expression for $\mathbb{E}_{\mathcal{I}^{(BS)}_k}(\sin\varphi_{\mathcal{I}^{(BS)}_k,i^*})$ is deduced assuming all the channel shape quantized to the elements in $Ns^{\epsilon}(\mathcal{I}_k^{(BS)})$ forms a neighboring region of $\mathbf{v}_{\mathcal{I}_k^{(BS)}}$ defined as a hypersphere: $$\mathcal{C}_b(\mathbf{v}_{\mathcal{I}_k^{(BS)}})=\{\widetilde{\mathbf{h}}_k:\sin\angle(\widetilde{\mathbf{h}}_k,\mathbf{v}_{\mathcal{I}_k^{(BS)}})\leq r_{\mathcal{C}_b}\} \label{eqn:partition-ball}$$ where $r_{\mathcal{C}_b}$ is the radius of the hypersphere and $\mathbf{v}_{i^*}$ lies at the edge of the hypersphere. Since $Pr(\sin\angle(\widetilde{\mathbf{h}}_k,\mathbf{v}_i)<x)=x^{2(n_T-1)}$, and the size of the hypersphere is: $$Pr(\sin\angle(\widetilde{\mathbf{h}}_k,\mathbf{v}_{\mathcal{I}_k^{(BS)}})\leq r_{\mathcal{C}_b}) = \frac{|Ns^{\epsilon}(\mathcal{I}_k^{(BS)})|}{N}\geq \frac{N_n}{N} \label{eqn:size-ball}$$ where $|Ns^{\epsilon}(\mathcal{I}_k^{(BS)})|$ is the cardinality of $Ns^{\epsilon}(\mathcal{I}_k^{(BS)})$. Therefore we have $r_{\mathcal{C}_b} \geq (\frac{N_n}{N})^{\frac{1}{2(n_T-1)}}$ Since $\mathbf{v}_{i^*}$ lies at the edge of the hypersphere, we have: $$\mathbb{E}_{\mathcal{I}^{(BS)}_k}(\sin\varphi_{\mathcal{I}^{(BS)}_k,i^*})\geq (\frac{N_n}{N})^{\frac{1}{2(n_T-1)}}.\nonumber$$ Appendix-D: Proof of Theorem \[lem:theorem-2\] {#appendix-d-proof-of-theorem-lemtheorem-2 .unnumbered} =============================================== Consider the situation with noise power larger than interference power $1>\mathbb{E}[\frac{P}{n_T}\|\mathbf{h}_k\|^2\sin^2\theta]$ as noise dominant scenario. With optimal or near optimal index assignment, we have $\mathbb{E}[\sin^2\theta]\leq (\frac{N_n}{N})^{\frac{1}{n_T-1}}$. Therefore, we require $P\cdot(\frac{N_n}{N})^{\frac{1}{n_T-1}}<1$ which gives $C_{fb}>(n_T-1)\log_2P+\log_2N_n $. The instantaneous mutual information of a user in (\[eqn:inst-mutual\]) can be simplified into $C_k \approx \log_2(1+\frac{P}{n_T}\|h_k\|^2\cos^2\theta)$. As the channel gain $\|\mathbf{h}_k\|^2$ scale with $n_T$, without loss of generality, we can also select $g_{th}$ in (\[eqn:ch-criteria\]) at order $O(n_T)$. Therefore the transmission rate for user *k* has the order $r_k \sim O(\log_2n_T)$ hence the system goodput $\mathcal{G}$ has the order: $$\begin{aligned} \mathcal{G}&\sim& O(n_T\cdot\log_2P).\label{eqn:avg-goodput-noise-limited}\end{aligned}$$ [1]{} C. Wang, R. D. Murch, “MU-MISO Transmission with Limited Feedback,” *IEEE Transactions on Wireless Communications,* vol. 6, no.11, pp. 3907-3913, Nov, 2007. K. Huang, J.G. Andrews, R.W. Heath Jr, “Performance of Orthogonal Beamforming for SDMA with Limited Feedback,” *IEEE Transactions on Vehicular Technology,* vol. 58, no. 1, pp. 152-164, Jan, 2009. “IEEE 802.16e amendment: Physical and medium access control layers for combined fixed and mobile operation in licensed bands,” *IEEE Standard 802.16,* 2005. L.U. Choi and R. Murch, “A transmit preprocessing technique for multiuser MIMO systems using a decomposition approach,” *IEEE Transactions on Wireless Communications,* vol. 3, no. 1, pp. 20-24, 2004. T. Yoo and A. Goldsmith, “On the optimality of multiantenna broadcast scheduling using zero-forcing beamforming,” *IEEE Journal on Sel. Areas in Communications,* vol. 24, pp. 528-541, Mar, 2006. C.R.Murthy, J. Zheng, B.D.Rao, “Performance of Quantized Equal Gain Transmission With Noisy Feedback Channels,” *IEEE Transactions on Signal Processing,* vol. 56, no. 6, pp. 2451-2460, Jun, 2008. J. Zhang, J.S. Lehnert, “Throughput-Optimal Precoding and Rate Allocation for MISO Systems With Noisy Feedback Channels,” *IEEE Transactions on Information Theory,* vol. 54, no. 5, pp. 2139-2155, May, 2008. A.S. Housfater, T.J. Lim, “Noisy Feedback Linear Precoding: A Bayesian Cramer-Rao Bound,” *ISIT 2009, Seou, Korea*, pp. 1689-1693, Jul, 2009. A.S. Housfater, T.J. Lim, “Throughput of Precoded Broadcast Transmission with Noisy Feedback,” *ISIT 2009, Seou, Korea*, pp. 398-402, Jul, 2009. F.Etemadi, S.Ekbatani, H.Jafarkhani, “Joint Source-Channel Coding for Quasi-Static Fading Channels with Noisy Quantized Feedback,” *IEEE Information Theory Workshop, 2007. ITW ’07.,* pp. 277-282, Sep, 2007. T.Y. Wu, V. Lau, “ Robust Precoder Adaptation for MIMO Links With Noisy Limited Feedback,” *IEEE Transactions on Information Theory,* vol. 55, no. 4, pp. 1640-1649, Apr, 2009. D. Yang, L. Yang, L. Hanzo, “ Performance of SDMA Systems Using Transmitter Preprocessing Based on Noisy Feedback of Vector-Quantized Channel Impulse responses,” *VTC2007 Spring,* pp. 22-25, Apr, 2007. J.C. Roh, B.D. Rao, “ MIMO Spatial Multiplexing Systems with Limited Feedback,” *ICC 2005,* vol. 2, pp. 777-782, May, 2005. D. Love, R.W. Heath, Jr, “ Limited Feedback Unitary Precoding for Spatial Multiplexing Systems,” *IEEE Transactions on Information Theory,* vol. 51, no. 8, pp. 2967-2976, Aug, 2005. K. Zyczkowski and M. Kus, “Random unitary matrices,” *J. Phys.,* vol. A27, pp. 4235-4245, Jun 1994. Samsung Electronics, “Downlink MIMO for EUTRA,” in *3GPP TSG RAN WG1 44/R1-060335*. J.B.Jensen and G.Gutin, “Digraphs: Theory, Algorithms and Applications,” *Springer-Verlag,* London,Oct, 2000. D. Appletgate, R. Bixby, V. Chvatal, W. Cook, “TSP cuts which do not conform to the template paradigm,” *Computational Combinatorial Optimization, Optimal or Provably Near-Optimal Solutions,* LNCS,vol. 2241, pp. 261-304, 2000. P. Larranaga, C.M.H. Kuijpers, R.H. Murga, I. Inza1 and S. Dizdarevic, “Genetic Algorithms for the Travelling Salesman Problem: A Review of Representations and Operators,” *Artificial Intelligence Review,* pp. 129-170, April, 1999. W. Dai, Y. Liu, B. Rider, “Quantization Bounds on Grassmann Manifolds and Applications to MIMO Communications,” *IEEE Transactions on Information Theory,* vol. 54, No. 3, pp. 1108-1123, Mar, 2008. [T.Y. Wu]{} received B.Eng degree from Tsinghua University in 2001 and Ph.D degree from Hong Kong University of Science and Technology in 2009 respectively. He is currently a research engineer in Huawei Technologies Co.Ltd. His research interests include limited feedback, interference management, and cognitive radio networks. [Vincent K.N. Lau]{} obtained B.Eng (Distinction 1st Hons) from the University of Hong Kong (1989-1992) and Ph.D. from the Cambridge University (1995-1997). He joined the Bell Labs - Lucent Technologies as member of technical staff from 1997-2003 and the Department of ECE, Hong Kong University of Science and Technology (HKUST) as Associate Professor afterwards. His current research focus includes robust cross layer scheduling for MIMO/OFDM wireless systems with imperfect channel state information, communication theory with limited feedback as well as delay-sensitive cross layer optimizations. He is currently an associate editor of IEEE Transactions on Wireless Communications, IEEE JSAC, EUARSIP Wireless Communications and Networking. [^1]: This paper is supported by RGC funding 615606. [^2]: The CSIT feedback index is always received correctly at the transmitter. [^3]: In this paper we assume equal power allocation since power allocation will only bring marginal performance gain under high SNR. This is also assumed in [@Kaibin:07; @Rao:05; @Love:05] etc. [^4]: In this paper we assume the large scale fading parameters (path loss and shadowing) between a base station and all the users in the cell are the same. [^5]: When there is only 1 active user in a cell, the system will degrade to a MISO system and this claim will be invalid. However, in MISO systems, interference is no longer a severe problem and there exists a lot of optimizing schemes. In this paper we will focus on strict SDMA system where there are more than 1 user served by the BS. [^6]: Grassmannian codebook is not a good choice here. Maximizing the minimum distortion among M bases is not equivalent to maximizing the minimum distortion among all the vectors in the codebook. [^7]: Example: For 8PSK with 10dB feedback SNR we have $N_n=3$ when $\epsilon=0.03$. [^8]: This corresponds to a feedback SNR of 10 dB when $C_{fb}=3$ and 8PSK is adopted.
{ "pile_set_name": "ArXiv" }
--- author: - 'P. Mertsch,' - 'M. Ahlers' bibliography: - 'qualinani.bib' title: 'Cosmic Ray Small-Scale Anisotropies in Quasi-Linear Theory' --- Introduction ============ The arrival directions of cosmic rays (CRs) are highly isotropic. Usually, this is explained as a consequence of pitch-angle scattering between CRs and turbulent magnetic fields. If the large-scale distribution of CR sources results in a spatial gradient, quasi-linear theory [@1966ApJ...146..480J; @1966PhFl....9.2377K; @1967PhFl...10.2620H; @1970ApJ...162.1049H; @Jokipii1972] predicts a small dipole anisotropy. Yet, observations show fluctuations on smaller scales, down to $10^{\circ}$ degrees. These small-scale anisotropies are conveniently quantified by the angular-power spectrum of the relative intensity of cosmic rays or, equivalently, by the phase-space density $f({\ensuremath{\bm{r}}}_\oplus,{\ensuremath{\bm{p}}},t)$ inferred by an observer at position ${\ensuremath{\bm{r}}}_\oplus$ and time $t$. In the following, we will study the power spectrum per unit square of the phase-space volume defined by $$\label{eq:Cl} C_{\ell}(t) \equiv \frac{1}{4 \pi} \! \int \! \! {\ensuremath{\mathrm{d}}}{\ensuremath{\widehat{\bm{p}}}}_A \int \! \! {\ensuremath{\mathrm{d}}}{\ensuremath{\widehat{\bm{p}}}}_B P_{\ell} ( {\ensuremath{\widehat{\bm{p}}}}_A {\!\cdotOLD\!}{\ensuremath{\widehat{\bm{p}}}}_B) f_A f_B\,,$$ where we use the abbreviation ${\ensuremath{\widehat{\bm{p}}}} = {\ensuremath{\bm{p}}} / |{\ensuremath{\bm{p}}}|$ and $f_A = f({\ensuremath{\bm{r}}}_\oplus,{\ensuremath{\bm{p}}}_A,t)$, [*etc.*]{} Small-scale anisotropies are *not* present in the usual quasi-linear theory with uniform pitch-angle scattering. (See however Ref. [@Giacinti:2016tld]). For a recent review on observations and interpretations of the small-scale anisotropies see Ref. [@Ahlers:2016rox] One of the arguably most attractive explanations of the small-scale anisotropies is that they are due to magnetic turbulence itself [@Giacinti:2011mz; @Ahlers:2013ima; @Ahlers:2015dwa]. Standard quasi-linear theory only predicts the ensemble-averaged phase-space density ${\langle f \rangle}$ and we can therefore only predict the angular power spectrum $C^{\text{std}}_{\ell}$ obtained from Eq.  through $f_A f_B \to {\langle f_A \rangle}{\langle f_B \rangle}$. Under the commonly used assumptions $C^{\text{std}}_{\ell} \sim 0$ for $\ell \geq 2$. (See again Ref. [@Giacinti:2016tld] for modifications to this simple picture.) However, it is easy to see that in the ensemble-average the angular power spectrum ${\langle C_{\ell} \rangle}$ can have small-scale power, i.e. ${\langle C_{\ell} \rangle} > C^{\text{std}}_{\ell}$, if ${\langle f_Af_B \rangle} > {\langle f_A \rangle}{\langle f_B \rangle}$. In other words, if there are correlations between the fluxes of CRs arriving under an angle (with $\ell$ the orbital quantum number corresponding to this angle $\theta$) then the average angular power spectrum ${\langle C_{\ell} \rangle}$, computed from the ensemble average of the product of phase-space densities, will be larger than the standard angular power spectrum $C^{\text{std}}_{\ell}$, computed from the product of ensemble-averaged phase-space densities. Therefore, correlations lead to small-scale anisotropies. These correlations are to be expected if particles propagate through a turbulent magnetic field: Particles arriving under an angle $\theta$ will have experienced similar fields for a certain amount of time before observation. It can be motivated [@Ahlers:2013ima] that this time is of the order $\tau_{\text{sc}} / (\ell (\ell + 1))$ where $\tau_{\text{sc}}$ is the scattering time. It is therefore ultimately the spatial correlations of the turbulent magnetic field that are reflected in the angular correlations of CR arrival directions. In the following we will predict the angular power spectrum Eq.  in an extended quasi-linear theory, taking into account the angular correlation between phase-space densities. We will consider the case with an isotropic turbulence tensor and without regular magnetic field. In this configuration, the unperturbed trajectories are straight lines, thus particles are propagating ballistically. The remainder of this paper is organised as follows: In Sec. \[sec:single\], we present a formalism describing the evolution in a random magnetic field of the ensemble-averaged cosmic ray phase-space density ${\langle f \rangle}$ from time $t_0$ to time $t$ by the propagator ${U^{}_{t, t_0}}$. We extend on this in Sec. \[sec:pair\] to treat the correlated evolution of the ensemble-average of a product of phase-space densities by a pair propagator. Evaluating the lowest order terms of the ensuing perturbative series, we formulate an ordinary differential equation for the ensemble-averaged angular power spectrum ${\langle C_{\ell} \rangle}$ and present an analytical expression for its steady-state. We fix the only free parameter of this model by a comparison with test particle simulations in Sec. \[sec:validation\]. In Sec. \[sec:results\], we show the predicted angular power spectrum and compare to observations from HAWC and IceCube. We summarise and conclude in Sec. \[sec:summary\]. Single-Particle Propagator {#sec:single} ========================== In the following, we will make use of a diagrammatic formalism for solving stochastic differential equations, as used for instance in propagation of waves through random media. Here, we briefly review this formalism to fix our notation. We refer the interested reader to Refs. [@Frisch:1968aa; @1977JPlPh..18...49P] for details. For simplicity, we will assume relativistic cosmic rays and work in natural units, $c=1$. The problem of propagation of (relativistic) charged particles through a static regular and turbulent magnetic field $\overline{{\ensuremath{\bm{B}}}}$ and ${\ensuremath{\bm{\delta B}}}({\ensuremath{\bm{r}}})$ can be formulated using Liouville’s equation for the phase-space density $f = f({\ensuremath{\bm{r}}}, {\ensuremath{\bm{p}}}, t)$, $$\partial_t f + {\ensuremath{\widehat{\bm{p}}}} {\!\cdotOLD\!}{\ensuremath{\bm{\nabla}}} f + {\mathcal{L}}{} f = -{\delta\mathcal{L}}{}(t) f\, , \label{eqn:Liouville}$$ with the deterministic and stochastic Liouville operators $$\begin{aligned} {\mathcal{L}}{} = - i {\ensuremath{\bm{\Omega}}} {\!\cdotOLD\!}{\ensuremath{\bm{L}}} \quad \text{and} \quad {\delta\mathcal{L}}{} = - i {\ensuremath{\bm{\omega}}}({\ensuremath{\bm{r}}}) {\!\cdotOLD\!}{\ensuremath{\bm{L}}} \, , \label{eqn:def_Liouville_operators}\end{aligned}$$ where ${\ensuremath{\bm{\Omega}}} = q \overline{{\ensuremath{\bm{B}}}} / p_0$ and ${\ensuremath{\bm{\omega}}}({\ensuremath{\bm{r}}}) = q {\ensuremath{\bm{\delta B}}}({\ensuremath{\bm{r}}}) / p_0$ are the (relativistic) gyrovectors of the regular and turbulent field, respectively, and $L_i \equiv-i \epsilon_{ijk}p_j\partial_{p_k}$ are angular momentum operators. In the following, we will assume that the spatial dependence of the phase-space density can be approximated by the first two terms of a Taylor expansion, $$f({\ensuremath{\bm{r}}}, {\ensuremath{\bm{p}}},t) \simeq {f_\oplus({\ensuremath{\bm{p}}}, t)} + ({\ensuremath{\bm{r}}} - {\ensuremath{\bm{r}}}_\oplus) {\!\cdotOLD\!}\nabla\overline{f} \, . \label{eqn:gradient_ansatz}$$ where $f_\oplus({\ensuremath{\bm{p}}}, t) \equiv f({\ensuremath{\bm{r}}}_\oplus,{\ensuremath{\bm{p}}}, t)$ denotes the local phase-space density and $\overline{f}$ is the local angular-averaged phase-space distribution. With this ansatz, the Liouville equation  evaluates to $$\partial_t f_\oplus + {\mathcal{L}}{} f_\oplus + \delta \mathcal{L}(t) f_\oplus \simeq - {\ensuremath{\widehat{\bm{p}}}} {\!\cdotOLD\!}\nabla\overline{f} \, . \label{eqn:Liouville_with_gradient}$$ Here, the stochastic Liouville operator depends on time, as the turbulent magnetic field is evaluated along the particle trajectory. Eq.  can be formally solved as $$f_\oplus({\ensuremath{\bm{p}}}, t) \simeq {U^{}_{t, t_0}} f_\oplus({\ensuremath{\bm{p}}}, t_0) - \int_{t_0}^t {\ensuremath{\mathrm{d}}}t' {U^{}_{t, t'}} {\ensuremath{\widehat{\bm{p}}}} {\!\cdotOLD\!}\nabla\overline{f} = {U^{}_{t, t_0}} f_\oplus({\ensuremath{\bm{p}}}, t_0) + \Delta {\ensuremath{\bm{r}}}(t_0){\!\cdotOLD\!}\nabla\overline{f} \, , \label{eqn:formal_solution}$$ with $\Delta {\ensuremath{\bm{r}}}(t_0) \equiv {\ensuremath{\bm{r}}}(t_0)-{\ensuremath{\bm{r}}}_\oplus$ and the aid of the time-evolution operator (also called propagator), written using the time-ordered (“latest–to–left”) exponential, $$\begin{aligned} {U^{}_{t, t_0}} &= \mathcal{T} \exp \left[ - \int_{t_0}^t {\ensuremath{\mathrm{d}}}t' \, ({\mathcal{L}}{} + {\delta\mathcal{L}}{} (t')) \right] = {U^{(0)}_{t, t_0}} \mathcal{T} \! \exp \left[ - \!\! \int_{t_0}^t {\ensuremath{\mathrm{d}}}t' \left( {U^{(0)}_{t', t_0}} \right)^{-1} {\delta\mathcal{L}}{}(t') {U^{(0)}_{t', t_0}} \right] . \label{eqn:time_ordered_exponential}\end{aligned}$$ Here, ${U^{(0)}_{t, t_0}}$ denotes the free propagator, $${U^{(0)}_{t, t_0}} = \exp \left[ - (t - t_0) {\mathcal{L}}{} \right] \, .$$ What complicates the solution of Eq.  is the stochastic nature of ${\delta\mathcal{L}}(t')$. One can only hope to predict moments of the propagator, its first moment being the expectation value. In the Gaussian limit, the expectation value of the propagator, ${\langle {U^{}_{t, t_0}} \rangle}$ contains only two-point functions of ${\delta\mathcal{L}}(t')$, $${\langle {\delta\mathcal{L}}{}(t_n) {\delta\mathcal{L}}{}(t_{n-1}) \mathellipsis {\delta\mathcal{L}}{}(t_1) \rangle} = {\langle {\delta\mathcal{L}}{}(t_n) {\delta\mathcal{L}}{}(t_{n-1}) \rangle} \mathellipsis {\langle {\delta\mathcal{L}}{}(t_1) {\delta\mathcal{L}}{}(t_0) \rangle} + \text{permut.} \, ,$$ for even $n$ and vanishes identically for odd $n$. The expansion of Eq.  becomes algebraically complex very quickly. It can be diagrammatically written in a more economic form, $$\vcenter{\hbox{ \includegraphics[scale=1,trim={0.5cm 0.6cm 0.5cm 0.7cm}, clip=false]{figures/single1_wide} }}$$ Here, solid lines represent free single-particle propagators ${U^{(0)}_{t'', t'}}$, dots correspond to insertions of ${\delta\mathcal{L}}{}(t')$ and dotted lines connecting such dots represent the expectation value of the two ${\delta\mathcal{L}}$’s that it connects. All intermediate time variables are integrated over. All connected diagrams can be resummed into the so-called mass operator, $$\vcenter{\hbox{ \includegraphics[scale=1,trim={0.7cm 0.5cm 0.8cm 0.7cm}, clip=false]{figures/single2} }}$$ such that the series for the propagator takes the simple form $$\vcenter{\hbox{ \includegraphics[scale=1,trim={0.2cm 0.7cm 0.1cm 0.8cm}, clip=false]{figures/single3} }}$$ The mass operator is difficult to evaluate at all orders, but approximating it with its lowest order term results in the so-called Bourret approximation to the single particle propagator, $$\vcenter{\hbox{ \includegraphics[scale=1,trim={0 0.8cm 0 0.7cm}, clip=false]{figures/single4} }} \label{eqn:Bourret_single1}$$ This series can now be resummed. In the simple case of vanishing regular magnetic field (${\ensuremath{\bm{\Omega}}} = {\ensuremath{\bm{0}}}$) the unperturbed trajectories are just straight lines and one finds [@Casse:2001be] $${\langle {U^{}_{t, t_0}} \rangle} \simeq {\overline{\overline{ {U^{}_{t, t_0}} }} \,}\equiv {U^{(0)}_{t, t_0}} e^{- (t-t_0) \nu{\ensuremath{\bm{L}}}^2/ 2} \, . \label{eqn:Bourret_single2}$$ The parameter $\nu$ contains integrals over the two-point functions of the turbulent field $\omega({\ensuremath{\bm{r}}})$. Pair Propagator {#sec:pair} =============== The small-scale anisotropies are a consequence of the fact that the trajectories of a pair of CRs are correlated for a (finite) amount of time before observation. Therefore, we need to consider the ensemble average of products of phase-space densities when computing the angular power spectrum. Note that in standard quasi-linear theory, we compute the ensemble average of single phase-space densities and are therefore missing the correlations between pairs of CR particles. In the following, we will use the abbreviations $f_A(t) \equiv f_\oplus({\ensuremath{\bm{p}}}_A,t)$, [*etc*]{}. From Eq. , we find for the ensemble average of the product of phase-space densities, $$\begin{aligned} {\langle f_A(t) f_B^*(t) \rangle} \simeq & {\langle {U^{A}_{t, t_0}} {U^{B*}_{t, t_0}} \rangle} {\langle f_A(t_0) f_B^*(t_0) \rangle} + {\langle (\Delta {\ensuremath{\bm{r}}}_A(t_0){\!\cdotOLD\!}\nabla\overline{f}) {U^{B*}_{t, t_0}} \rangle} {\langle f^*_B(t_0) \rangle} \nonumber \\ & + {\langle (\Delta {\ensuremath{\bm{r}}}_B(t_0){\!\cdotOLD\!}\nabla\overline{f}^*) {U^{A}_{t, t_0}} \rangle} {\langle f_A(t_0) \rangle} + {\langle (\Delta {\ensuremath{\bm{r}}}_A(t_0){\!\cdotOLD\!}\nabla\overline{f})(\Delta {\ensuremath{\bm{r}}}_B(t_0){\!\cdotOLD\!}\nabla\overline{f}^*) \rangle}\,. \label{eqn:formal_solution_product}\end{aligned}$$ In our previous analysis [@Ahlers:2015dwa] we identified the last term on the right-hand-side of Eq.  as the term that determines the asymptotic behavior of the power-spectrum for large look-back times, $t-t_0\gg\nu^{-1}$. Note that we have assumed that correlations between the propagators and the initial state $f_\oplus({\ensuremath{\bm{p}}},t_0)$ can be ignored. In the following, we will take a different approach and aim to establish a differential equation for the angular power spectrum in quasi-linear theory based on Eq. . This differential equation will describe the temporal evolution of the angular power spectrum *locally*, that is at one position and we will thus consider the infinitesimal limit $\Delta T \to 0$. By virtue of the Bethe-Salpeter equation [@1951PhRv...84.1232S], the double propagator can be expanded into a perturbative series. This series has a diagrammatic representation, somewhat similar to Feynman diagrams employed in quantum field theory, $$\vcenter{\hbox{ \includegraphics[scale=0.98,trim={0.6cm 0.4cm 0.2cm 0.6cm}, clip=false]{figures/double1_wide.pdf} }} \label{eqn:diagrams}$$ If the dashed lines are connecting the Liouville operators ${\delta\mathcal{L}}{}$ of two different particles A and B, then this can be considered an interaction between particles A and B mediated by the correlation structure of the turbulent magnetic field. It is the repeated action of these “interactions” that is inducing the correlations between particles A and B. While Eq.  allows computing the angular power spectrum anytime after preparing the initial state, $f({\ensuremath{\widehat{\bm{p}}}}, t_0)$, evaluating or even resumming all diagrams of Eq.  in all generality seems challenging at the very least. Instead, we seek to approximate the identity (\[eqn:formal\_solution\_product\]) by the stationary solution of a differential equation with respect to a small step in look-back time $\Delta T \equiv t-t_0$, $$\begin{aligned} \frac{1 - {\langle {U^{A}_{t, t_0}} {U^{B*}_{t, t_0}} \rangle}}{\Delta T} {\langle f_A(t_0)f^*_{B}(t_0) \rangle} &\simeq (\overline{f} - 3 {\ensuremath{\widehat{\bm{p}}}}_A {\!\cdotOLD\!}{\bf K} {\!\cdotOLD\!}\nabla\overline{f}) \left(\frac{\Delta{\ensuremath{\bm{r}}}_B}{\Delta T} {\!\cdotOLD\!}\nabla\overline{f}\right) \nonumber \\ & + \left(\frac{\Delta{\ensuremath{\bm{r}}}_A}{\Delta T} {\!\cdotOLD\!}\nabla\overline{f}\right) (\overline{f} - 3 {\ensuremath{\widehat{\bm{p}}}}_B {\!\cdotOLD\!}{\bf K} {\!\cdotOLD\!}\nabla\overline{f}) + \mathcal{O}(\Delta T)\,,\end{aligned}$$ where we applied the quasi-stationary solution of the diffusion equation ${\langle f_A(t_0) \rangle} \simeq \overline{f}-3 {\ensuremath{\widehat{\bm{p}}}}_A {\!\cdotOLD\!}{\bf K} {\!\cdotOLD\!}\nabla\overline{f}$. In this limit $\Delta T \to 0$, we can approximate $\Delta{\ensuremath{\bm{r}}}/\Delta T \simeq - {\ensuremath{\widehat{\bm{p}}}}$. This allows writing down an ordinary differential equation for the angular power spectrum $C_{\ell}$, $$A_{\ell \ell_0} C_{\ell_0}(t) \simeq \frac{8 \pi}{9} K \left| \nabla\overline{f} \right|^2 \delta_{\ell 1} \, , \label{eqn:ODE}$$ where we assume isotropic diffusion $K_{ij} = K\delta_{ij}$ and define the transition matrix $$A_{\ell \ell_0}(t) = \lim_{t_0 \to t} \frac{\delta_{\ell \ell_0} - M_{\ell \ell_0}(t, t_0)}{t - t_0}\,, \label{eqn:A}$$ and where $$\begin{aligned} M_{\ell \ell_0}(t, {t_0}) & = \frac{1}{4 \pi} \int {\ensuremath{\mathrm{d}}}{\ensuremath{\widehat{\bm{p}}}}_A \int {\ensuremath{\mathrm{d}}}{\ensuremath{\widehat{\bm{p}}}}_B P_{\ell} ({\ensuremath{\widehat{\bm{p}}}}_A {\!\cdotOLD\!}{\ensuremath{\widehat{\bm{p}}}}_B) {\langle {U^{A}_{t, t_0}} {U^{B*}_{t, t_0}} \rangle} \frac{2 \ell_0 + 1}{4 \pi} P_{\ell_0} ({\ensuremath{\widehat{\bm{p}}}}_A {\!\cdotOLD\!}{\ensuremath{\widehat{\bm{p}}}}_B) \, . \label{eqn:mixing_matrix}\end{aligned}$$ Once we have computed $M_{\ell \ell_0}(t, {t_0})$, it is easy to find the steady-state angular power spectrum $C^{\text{stdy}}$ by solving $$A_{\ell \ell_0} C^{\text{stdy}}_{\ell_0}(t) = \frac{8 \pi}{9} K \left| \nabla\overline{f} \right|^2 \delta_{\ell 1} \, . \label{eqn:ODE_stdy}$$ In evaluating ${\langle {U^{A}_{t, t_0}} {U^{B*}_{t, t_0}} \rangle}$, we confine ourselves to considering the leading and next-to-leading order terms, that is the first line of Eq. . We label the contributions of those diagrams to the double propagator as follows, $$\begin{aligned} &\vcenter{\hbox{ \includegraphics[scale=1,trim={0.5cm 0cm 1cm 0.5cm}, clip=false]{figures/term0.pdf} }} \, , \vcenter{\hbox{ \includegraphics[scale=1,trim={0.5cm 0.3cm 0.7cm 0.7cm}, clip=false]{figures/term1a.pdf} }} \, , \\ &\vcenter{\hbox{ \includegraphics[scale=1,trim={0.5cm 0.1cm 0.7cm 0.5cm}, clip=false]{figures/term1b.pdf} }} \, , \vcenter{\hbox{ \includegraphics[scale=1,trim={0.5cm 0cm 0.5cm 0.5cm}, clip=false]{figures/term1c.pdf} }} \, .\end{aligned}$$ The leading order term, the free double propagator, $$\left( {\langle {U^{A}_{t, t_0}} {U^{B*}_{t, t_0}} \rangle} \right)^{(0)} = 1 \, ,$$ is trivial since the free single propagator is trivial, , in the limit of vanishing regular magnetic field (${\ensuremath{\bm{\Omega}}} = {\ensuremath{\bm{0}}}$). The contribution of $\left( {\langle {U^{A}_{t, t_0}} {U^{B*}_{t, t_0}} \rangle} \right)^{(0)}$ to $M_{\ell \ell_0}(t, {t_0})$ thus takes the simple form $$\begin{aligned} M_{\ell \ell_0}^{(0)} =& \delta_{\ell \ell_0} \, . \label{eqn:M0}\end{aligned}$$ The first next-to-leading order term (1a) can be computed assuming ballistic trajectories, and an isotropic turbulence tensor, $$\begin{aligned} {\langle \tilde{\omega}_i({\ensuremath{\bm{k}}}) \tilde{\omega}^*_j({\ensuremath{\bm{k}}}') \rangle} = \frac{g(k)}{k^2} \left( \delta_{ij} - \widehat{k}_i \widehat{k}_j \right) \delta({\ensuremath{\bm{k}}} - {\ensuremath{\bm{k}}}') \, .\end{aligned}$$ With the help of a plane wave expansion we find $$\begin{aligned} \left( {\langle {U^{A}_{t, t_0}} {U^{B*}_{t, t_0}} \rangle} \right)^{(1a)} &= \int_{t_0}^t {\ensuremath{\mathrm{d}}}{t_{2}} \int_{t_0}^{t_2} {\ensuremath{\mathrm{d}}}{t_{1}} {U^{A(0)}_{t, {t_{2}}}} {\langle \delta\mathcal{L}^A_{{t_{2}}} {U^{A(0)}_{{t_{2}}, {t_{1}}}} \delta\mathcal{L}^A_{{t_{1}}} \rangle} {U^{A(0)}_{{t_{1}}, t_0}} {U^{B*(0)}_{t, t_0}} \nonumber \\ &= -\sum_{\ell_A} (2 \ell_A + 1) {\imath}^{\ell_A} \Lambda_{\ell_A}(t-t_0) \int {\ensuremath{\mathrm{d}}}\widehat{k} P_{\ell_A}( {\ensuremath{\widehat{\bm{k}}}} {\!\cdotOLD\!}{\ensuremath{\widehat{\bm{p}}}}_A ) \left( \delta_{ij} - \widehat{k}_i \widehat{k}_j \right) L_i^A L_j^A\,, \label{eqn:double1a}\end{aligned}$$ where $P_{\ell}(\cdotOLD)$ denotes the Legendre polynomial of degree $\ell$ and we introduce the quantity $$\begin{aligned} \Lambda_{\ell_A}(\Delta T) &= \int_{0}^{\Delta T} {\ensuremath{\mathrm{d}}}T \int_{0}^{T} {\ensuremath{\mathrm{d}}}\tau \int {\ensuremath{\mathrm{d}}}k \, g(k) j_{\ell_A}(k \tau) \, ,\end{aligned}$$ where $ j_{\ell}(\cdotOLD)$ is the spherical Bessel function of the first kind. Computing the contribution of $\left( {\langle {U^{A}_{t, t_0}} {U^{B*}_{t, t_0}} \rangle} \right)^{(1a)} $ to the mixing matrix $M_{\ell \ell_0}(t, {t_0})$ via Eq.  we find $$\begin{aligned} M_{\ell \ell_0}^{(1a)} =& - \frac{8 \pi}{3} \delta_{\ell \ell_0} \! \left( \! \Lambda_0(\Delta T) - \frac{1}{2} \Lambda_2(\Delta T) \! \right) \! \ell (\ell+1) \, . \label{eqn:M1a}\end{aligned}$$ There is evidence that the energy spectral density of interstellar turbulent magnetic fields follows a power law in wavenumber, with theoretically motivated values of the spectral index of $-7/2$ or $-11/3$ [@Elmegreen:2004wj]. Here, we will adopt a band-limited white noise spectrum, that is $g(k) = {g_{\text{out}}}{}$ if ${k_{\text{out}}}{} \leq k < {k_{\text{in}}}{}$ and $0$ otherwise. Below we will show that for this spectrum, the limit $t_0 \to t$ of Eq.  only exists if we simultaneously let ${k_{\text{in}}}{} \to \infty$, while keeping $\alpha \equiv {k_{\text{in}}}{} \Delta T$ finite. Physically this means that as we let $\Delta T \to 0$, we need to also extend the turbulence spectrum to arbitrarily small scales such that particles can experience changes in the magnetic field during the time $\Delta T$. The parameter $\alpha$ encodes how many wavelengths of the smallest modes the particle traverses in the time interval $\Delta T$. We consider this to be a free parameter and determine it by comparing with numerical simulations and observational data below. We then find $$\begin{aligned} M_{\ell \ell_0}^{(1a)} = - \pi \frac{{g_{\text{out}}}{}}{{k_{\text{out}}}{}} r \delta_{\ell \ell_0} \Big\{ & - \frac{1}{\alpha} \cos \alpha + 2 \alpha \, {}_2 F_3\left( \frac{1}{2}, \frac{1}{2}; \frac{3}{2}, \frac{3}{2}, \frac{3}{2}; - \left( \frac{\alpha}{2} \right)^2 \right) \nonumber \\ & + \alpha^2 \sin \alpha - \text{Si} [\alpha] \Big\} \ell (\ell+1) \, . \label{eqn:M1a_limit}\end{aligned}$$ For ${k_{\text{out}}}{} \Delta T \ll 1$, $M_{\ell \ell_0}^{(1a)} / \Delta T$ is a function of $\alpha$ only. Specifically, $$\frac{M_{\ell \ell_0}^{(1a)}}{\Delta T} \propto \left\{ \begin{array}{l l} \alpha & \text{for } \alpha \ll 1 \, , \\ \ln \alpha & \text{for } 1 \ll \alpha \ll {k_{\text{in}}}{}/{k_{\text{out}}}{} \, , \\ \text{const.} & \text{for } \alpha \gg {k_{\text{in}}}{}/{k_{\text{out}}}{} \, . \end{array} \right.$$ These cases correspond, respectively, to the particle travelling a distance less than $1/{k_{\text{in}}}{}$, between $1/{k_{\text{in}}}{}$ and $1/{k_{\text{out}}}{}$ and more than $1/{k_{\text{out}}}{}$ in the time $\Delta T$. The first interacting contribution (1c) is $$\int_{t_0}^t {\ensuremath{\mathrm{d}}}{t_{1}} \int_{t_0}^t {\ensuremath{\mathrm{d}}}{\tau_{1}} {U^{(0)A}_{t, {t_{1}}}} {U^{(0)B*}_{t, {\tau_{1}}}} {\langle \delta\mathcal{L}^A_{{t_{1}}} \delta\mathcal{L}^{B*}_{{\tau_{1}}} \rangle} {U^{(0)A}_{{t_{1}}, t_0}} {U^{(0)B*}_{{\tau_{1}}, t_0}} \, .$$ We use Eq. , Fourier transform the correlation function and, again assuming ballistic trajectories, , perform a free wave expansion for the exponential factors $\exp[{\imath}{\ensuremath{\bm{k}}} {\!\cdotOLD\!}{\ensuremath{\bm{r}}}(t) ]$. This leads to $$\begin{aligned} & \int_{t_0}^t {\ensuremath{\mathrm{d}}}{t_{1}} \int_{t_0}^t {\ensuremath{\mathrm{d}}}{\tau_{1}} {U^{(0)A}_{t, {t_{1}}}} {U^{(0)B*}_{t, {\tau_{1}}}} {\langle \delta\mathcal{L}^A_{{t_{1}}} \delta\mathcal{L}^{B*}_{{\tau_{1}}} \rangle} {U^{(0)A}_{{t_{1}}, t_0}} {U^{(0)B*}_{{\tau_{1}}, t_0}} \nonumber \\ & = (4 \pi)^2 \int {\ensuremath{\mathrm{d}}}\widehat{k} \int {\ensuremath{\mathrm{d}}}k \, {\langle \tilde{\omega}_i({\ensuremath{\bm{k}}}) \tilde{\omega}^*_j({\ensuremath{\bm{k}}}') \rangle} \sum_{\substack{\ell_A,m_A\\\ell_B,m_B}} {\imath}^{-\ell_A + \ell_B} \int_0^{k \Delta T} {\ensuremath{\mathrm{d}}}{t_{1}}' j_{\ell_A}({t_{1}}') \int_0^{k \Delta T} {\ensuremath{\mathrm{d}}}{\tau_{1}}' j_{\ell_B}({\tau_{1}}') \nonumber \\ & \times Y_{\ell_A m_A}({\ensuremath{\widehat{\bm{k}}}}) Y^*_{\ell_B m_B}({\ensuremath{\widehat{\bm{k}}}}) Y^*_{\ell_A m_A}({\ensuremath{\widehat{\bm{p}}}}_A)Y_{\ell_B m_B}({\ensuremath{\widehat{\bm{p}}}}_B) L^A_i L^{B*}_j \, .\end{aligned}$$ In the following, we split $(\delta_{ij} - \widehat{k}_i \widehat{k}_j)$ into a monopole and a quadrupole contribution (in ${\ensuremath{\widehat{\bm{k}}}}$), $$\left( \delta_{ij} - \widehat{k}_i \widehat{k}_j \right) = \frac{2}{3} \delta_{ij} + \left( \frac{1}{3} \delta_{ij} - \widehat{k}_i \widehat{k}_j \right) \, .$$ leading to $M_{\ell \ell_0}^{(1c,0)}$ and $M_{\ell \ell_0}^{(1c,2)}$, respectively. We find $$\begin{aligned} M_{\ell \ell_0}^{(1c,0)} &= \frac{8 \pi}{3} \sum_{\ell_A} (2 \ell_A + 1) \kappa_{\ell_A \ell_A}(t-t_0) {\left( \begin{array}{ccc} \ell & \ell_A & \ell_0 \\ 0 & 0 & 0 \end{array} \right)}^2 (2 \ell_0 + 1) \ell_0 (\ell_0 + 1) \, , \label{eqn:M1c0}\end{aligned}$$ where $( {\!\cdotOLD\!})$ denotes the Wigner $3j$-symbol and with the triple integral $$\begin{aligned} \kappa_{\ell_A \ell_B} (\Delta T) &\equiv \int_{{k_{\text{out}}}{}}^{{k_{\text{in}}}{}} \mathrm{d} k \frac{g(k)}{k^2} h_{\ell_A} \left( k, \Delta T \right) h_{\ell_B} \left( k, \Delta T \right) \nonumber \\ &= \frac{{g_{\text{out}}}{}}{{k_{\text{out}}}{}} \int\displaylimits_{1}^{{k_{\text{in}}}{}/{k_{\text{out}}}{}} \mathrm{d} k' \, k'^{-2} \!\!\!\! \int\displaylimits_0^{k' {k_{\text{out}}}{} \Delta T} \!\!\!\!\mathrm{d} {t_{1}}' j_{\ell_A}({t_{1}}') \!\!\!\! \int\displaylimits_0^{k' {k_{\text{out}}}{} \Delta T} \!\!\!\! \mathrm{d} {\tau_{1}}' j_{\ell_B}({\tau_{1}}') \, .\end{aligned}$$ The quadrupole contribution requires significantly more algebra, but eventually reads $$\begin{aligned} M_{\ell,\ell_0}^{(1c,2)} & = \frac{4 \pi}{3} \frac{\ell_0 (\ell_0+1) (2 \ell_0 + 1)}{{\left( \begin{array}{ccc} \ell_0 & 2 & \ell_0 \\ 0 & 0 & 0 \end{array} \right)}} (-1)^{\ell_0} \sum_{\ell_A,\ell_B} {\imath}^{\ell_A + \ell_B} \kappa_{\ell_A \ell_B} (t-t_0) (2 \ell_A + 1)(2 \ell_B + 1) \nonumber \\ & \times {\left\{ \begin{array}{ccc} 2 & \ell_0 & \ell_0 \\ \ell & \ell_A & \ell_B \end{array} \right\}} {\left( \begin{array}{ccc} 2 & \ell_A & \ell_B \\ 0 & 0 & 0 \end{array} \right)} {\left( \begin{array}{ccc} \ell & \ell_A & \ell_0 \\ 0 & 0 & 0 \end{array} \right)} {\left( \begin{array}{ccc} \ell & \ell_B & \ell_0 \\ 0 & 0 & 0 \end{array} \right)}\,. \label{eqn:M1c2}\end{aligned}$$ The curly brackets in Eq.  denote the Wigner 6$j$-symbol. Also $M_{\ell \ell_0}^{(1c)} / \Delta T$ shows some simple dependence on $\alpha$, $$\frac{M_{\ell \ell_0}^{(1c)}}{\Delta T} \propto \left\{ \begin{array}{l l} \alpha^{2 + 2 |\ell - \ell_0|} & \text{for } \alpha \ll 1 \, , \\ \alpha & \text{for } 1 \ll \alpha \ll {k_{\text{in}}}{}/{k_{\text{out}}}{} \, . \end{array} \right.$$ Eventually, we compute $M_{\ell \ell_0} = M_{\ell \ell_0}^{(0)} + 2 M_{\ell \ell_0}^{(1a)} + M_{\ell \ell_0}^{(1c,0)} + M_{\ell \ell_0}^{(1c,2)}$ (see Eqs. (\[eqn:M0\]), (\[eqn:M1a\_limit\]), (\[eqn:M1c0\]) and (\[eqn:M1c2\])), determine $A_{\ell \ell_0}$ from Eq.  and find the steady-state angular power spectrum from Eq. . Validation {#sec:validation} ========== In order to validate the results of our analytical computation we now compare to numerical results following the method in Ref. [@Ahlers:2015dwa]. The power spectrum can be derived from the last term of Eq.  in the limit $\Delta T \to \infty$. We have back-tracked test particles in isotropic turbulent magnetic fields with band-limited white-noise spectrum. We have not assumed any regular component. Specifically, we have chosen ${k_{\text{out}}}{} r_g = 10^{-3}$ and ${k_{\text{in}}}{} r_g = 10^2$, with $r_g$ the particles gyroradius in the RMS turbulent field. This guarantees that there is a broad enough range of wavenumbers to be in resonance with ($r_g k_{\text{res}} \approx 1$, $k_{\text{res}}$ being the resonant wavenumber) while satisfying the requirement ${k_{\text{in}}}{}/{k_{\text{out}}}{} \gg 1$. The numerical backtracking results in a set of trajectories that converge at ${\ensuremath{\bm{r}}}_\oplus$. Thanks to Liouville’s theorem, we can use this to compute the angular distribution at position ${\ensuremath{\bm{r}}}_\oplus$ and time $t$ by assuming a certain phase-space density at time $t_0$. To make the connection with our analytical approach, we adopt the same gradient dependence as in Eq. . For each of 100 random realisations of the turbulent magnetic field, we compute the angular power spectrum from the phase-space density at position ${\ensuremath{\bm{r}}}_\oplus$ and time $t$ and finally compute the ensemble averaged angular power spectrum. ![Validation of the analytical method by comparison with numerical result. The data points show the angular power spectrum determined in test particle simulations for three different backtracking times $t_{\text{back}}$ after subtraction of the estimated shot noise. The shot noise levels due to the finite number of trajectories is indicated by the horizontal dashed lines. For comparison, the lines show the results of our analytical approach with different values of the free parameter $\alpha = {k_{\text{in}}}{} \Delta T$. Fixing this free parameter to ${k_{\text{in}}}{} \Delta T \approx 50$ results in excellent agreement with the test particle simulations.[]{data-label="fig1"}](figures/validation_k0_0.0010_k1_100_nModes_100_nside_64.pdf) In Fig. \[fig1\], we show this ensemble averaged angular power spectrum for three different backtracking times $t_{\text{back}}$. It can be seen that the angular power spectrum converges to an asymptotic form for $\Omega t_{\text{back}} \gtrsim 10$ where $\Omega$ is the gyro frequency in the RMS turbulent field. (See also Ref. [@Ahlers:2015dwa].) Comparing with the analytical results allows fixing the free parameter, $\alpha = {k_{\text{in}}}{} \Delta T$, for which we otherwise only have the constraint $\alpha \gg 1$. It appears that $\alpha \approx 50$ gives excellent agreement between analytical and numerical results. Results {#sec:results} ======= ![The angular power spectrum computed in quasi-linear theory for different values of ${k_{\text{in}}}{} \Delta T$, without (upper panel) and with (lower panel) adding the noise contribution. For comparison, we also show the observations by HAWC [@Abeysekara:2018qho] and the IceCube-HAWC combined data [@TheHAWC:2017uyf] with the shot noise level estimated for the latter.[]{data-label="fig2"}](figures/results_stdy_wo_noise.pdf "fig:")\ ![The angular power spectrum computed in quasi-linear theory for different values of ${k_{\text{in}}}{} \Delta T$, without (upper panel) and with (lower panel) adding the noise contribution. For comparison, we also show the observations by HAWC [@Abeysekara:2018qho] and the IceCube-HAWC combined data [@TheHAWC:2017uyf] with the shot noise level estimated for the latter.[]{data-label="fig2"}](figures/results_stdy_w_noise.pdf "fig:") While the band-limited white-noise spectrum serves only as an approximation for the magnetic turbulence in our local environment, it is nevertheless instructive to compare our model predictions to the power spectrum observed by HAWC and IceCube [@Abeysekara:2018qho; @TheHAWC:2017uyf]. In Fig. \[fig2\] we show the steady-state angular power spectrum $C^{\text{stdy}}$ derived by numerically solving Eq. . In the upper panel, we have fixed the gradient to $K \left|{\nabla\overline{f}}/{\overline{f}}\right|^2 = 10^{-4} {k_{\text{out}}}{}$ and show the result for a range of $\alpha = {k_{\text{in}}}{} \Delta T$. It can be seen that with increasing values of ${k_{\text{in}}}{} \Delta T$, the normalisation of the angular power spectrum is decreasing and the power spectrum tends to fall off much faster. We compare our model predictions to the angular power spectra inferred from HAWC data [@Abeysekara:2018qho] and the combined IceCube-HAWC data [@TheHAWC:2017uyf]. Note that we have not accounted for the shot noise necessarily present in the data or for cross talk between multipole moments due to IceCube’s limited field of view, see Ref. [@Ahlers:2016rox]. In the right panel of Fig. \[fig2\], we do account for the effect of shot noise by adding a constant noise power of $\mathcal{N}\simeq1.5 \times 10^9$. This is reproducing the data from the combined analysis of HAWC and IceCube data [@TheHAWC:2017uyf] which is dominated by shot noise above $\ell \gtrsim 10$. Overall, with $\alpha = {k_{\text{in}}}{} \Delta T \simeq 50$, as suggested by the numerical simulations, see Sec. \[sec:validation\], we find good agreement with the data, again for $K \left|{\nabla\overline{f}}/{\overline{f}}\right|^2 = 10^{-4} {k_{\text{out}}}{}$. Summary and Conclusion {#sec:summary} ====================== We have presented a computation of the angular power spectrum of CR small-scale anisotropies, based on the idea that the small-scale anisotropies are a consequence of cosmic ray streaming in the local configuration of the turbulent magnetic field. This model is based on a formal solution to the evolution equations for pairs of CR particles, and expressed as a steady-state solution of the ensemble-averaged products of their phase-space densities. We have evaluated this solution in a perturbative approach which can be represented by a series of diagrams. Considering only the contributions from the lowest order terms, we have formulated an ordinary differential equation for the angular power spectrum and solved for its steady state. We have assumed throughout the absence of a regular magnetic field such that the unperturbed trajectories are straight lines. In order to formulate the ordinary differential equation, we also needed to adopt a band-limited white noise power spectrum for the turbulent magnetic field. This introduced two free parameters, the inverse of the smallest turbulent scale, ${k_{\text{in}}}{}$, and the smallest time-interval considered, $\Delta T$, but the steady-state angular power spectrum only depends on their combination $\alpha = {k_{\text{in}}}{} \Delta T$. By comparing to numerical test particle simulations we have found a value of $\alpha \simeq 50$ to be appropriate. With this value, we find good agreement between our model predictions and the measurements by HAWC and IceCube. The most obvious limitations of the present model are the unrealistic power spectrum that needs to be adopted and the dependence on the parameter $\alpha$ that needed to be fixed with the help of numerical simulations. We are convinced, however, that both are artefacts introduced by the fact that we treat unperturbed trajectories as straight lines. While we have motivated this by the assumed absence of a regular magnetic field, it is true that even in that case particles will experience an average regular field, set by the largest scales on which there is significant power. In the future, it would therefore be desirable to consider a regular background magnetic field and unperturbed helical trajectories. This will also introduce resonance effects between the particles’ gyroradii and the turbulent wavelengths which are also absent due to the assumed straight-line trajectories. We stress that such resonances and the form or the turbulent power spectrum are ultimately responsible for the energy-dependencies of the pitch-angle scattering rate and of the spatial diffusion coefficients. Observationally, this would broaden the range of predictions of our model. Given that the HAWC collaboration has already started presenting angular power spectra for different energy bins [@Abeysekara:2018qho], this avenue seems most promising. Acknowledgments {#acknowledgments .unnumbered} =============== This work was supported by Danmarks Grundforskningsfond under grant no. 1041811001. MA also acknowledges support from <span style="font-variant:small-caps;">Villum Fonden</span> (project no. 18994).
{ "pile_set_name": "ArXiv" }
--- abstract: 'Intense extreme-ultraviolet (XUV) pulses enable the investigation of XUV-induced nonlinear processes and are a prerequisite for the development of attosecond pump - attosecond probe experiments. While highly nonlinear processes in the XUV range have been studied at free-electron lasers (FELs), high-harmonic generation (HHG) has allowed the investigation of low-order nonlinear processes. Here we suggest a concept to optimize the HHG intensity, which surprisingly requires a scaling of the experimental parameters that differs substantially from optimizing the HHG pulse energy. As a result, we are able to study highly nonlinear processes in the XUV range using a driving laser with a modest ($\approx 10$ mJ) pulse energy. We demonstrate our approach by ionizing Ar atoms up to Ar$^{5+}$, requiring the absorption of at least 10 XUV photons.' author: - Björn Senfftleben - Martin Kretschmar - Andreas Hoffmann - Mario Sauppe - Johannes Tümmler - Ingo Will - Tamás Nagy - 'Marc J. J. Vrakking' - Daniela Rupp - Bernd Schütte title: 'Highly nonlinear ionization of atoms induced by intense high-harmonic pulses' --- Three key features of intense XUV pulses from FEL and HHG sources have opened up new possibilities for a range of fields from life science to material science and fundamental physics: (i) Intense XUV pulses provide the possibility to perform pump-probe experiments, in which a first XUV pump pulse initiates dynamics in an atom or a molecule, and these dynamics are probed by a second, time-delayed XUV probe pulse. This capability has been extensively used with femtosecond time resolution at FELs (see e.g. Ref. [@jiang10]) and has been implemented with attosecond time resolution using HHG sources [@tzallas11; @takahashi13]. (ii) Intense, coherent XUV pulses have enabled single-shot coherent diffractive imaging (CDI) of isolated nanotargets with a resolution in the tens of nanometers range. While these experiments are predominantly performed at FELs [@bogan08], CDI on He nanodroplets using an intense HHG source was recently reported [@rupp17]. (iii) High XUV intensities enable the study of nonlinear optics in this spectral range. Examples include multiple ionization of atoms [@motomura09] and XUV-driven four-wave mixing schemes [@foglia18]. Advantages of HHG sources compared to FELs include their smaller size and lower costs resulting in easier access to these sources. Another important point is that two-color XUV-optical pump-probe experiments, which are one of the preferred applications of ultrashort XUV pulses, are challenging at FELs due to timing jitter between the optical and XUV beams [@savelyev17]. Such experiments are routinely performed using HHG sources, with a time resolution extending into the attosecond domain, given the fact that attosecond pulses based on HHG have been generated for more than 10 years [@krausz09]. The most common scheme used to generate intense HHG pulses requires powerful near-infrared (NIR) pulses that are loosely focused into a gas medium to increase the generation volume [@takahashi13; @schutte14a; @manschwetus16; @nayak18; @bergues18]. It has been shown that the HHG pulse energy increases with increasing NIR focal length [@boutu11]. In contrast, we demonstrate in this paper that for a given laboratory size the XUV intensity on target can be optimized by using an NIR focusing element with a relatively short focal length. This enables the generation of a smaller XUV beam waist radius in the experiment and a higher XUV beamline transmission. As a result, we achieve an XUV intensity of $7 \times 10^{14}$ W/cm$^2$, allowing us to study highly nonlinear processes, which we demonstrate in this paper by ionizing neutral Ar atoms up to Ar$^{5+}$. Investigations on HHG scaling have shown that the HHG conversion efficiency may be preserved when varying the laser power, as long as all relevant parameters such as the focusing geometry and the gas pressure and length of the HHG medium are scaled appropriately [@heyl16]. One can further assume that the NIR intensity used to optimize the HHG conversion efficiency is nearly independent of the chosen experimental conditions, given that the NIR intensity is dictated by phase-matching conditions that are largely determined by the choice of laser wavelength and target gas. As a result, the XUV intensity obtained directly behind the generation medium is to a good approximation independent of the NIR focusing geometry. Therefore, in order to optimize the peak XUV intensity in the experiment, we apply the following strategy: (1) We choose a beamline geometry leading to a large demagnification $D=w_{XUV,source}/w_{XUV,focus}$ (where $w_{XUV,source}$ is the XUV beam waist radius at the HHG source and $w_{XUV,focus}$ is the XUV beam waist radius after focusing). (2) The HHG beamline transmission is optimized by using as few XUV optical elements (filters, mirrors) as possible, as each of these elements introduces high losses. To study how the demagnification of the XUV beam radius depends on the distance between the NIR focusing element and the focal plane $d_{NIR}$, we have performed calculations based on Gaussian beams in the paraxial approximation. In these calculations, we have assumed that our laboratory allows a maximum distance of 18 m between the last NIR mirror preceding the HHG source and the mirror that focuses the XUV radiation onto the target. Furthermore, an XUV mirror with a focal length of $f_{XUV}=75$ mm was assumed, which is the shortest focal length that can be used in combination with our velocity-map imaging spectrometer (VMIS). The results of the calculations are shown in Fig. \[raytracing\]. The XUV beam waist radius in the experiment (blue curve in Fig. \[raytracing\](a)) increases rapidly with increasing $d_{NIR}$, both due to the increasing XUV source size (as a result of the increasing NIR focus) and the decreasing demagnification. The demagnification calculated by Gaussian optics (solid orange curve) is in good approximation equal to $D\approx d/f_{XUV}$ for short focal lengths (dotted orange curve), where $d$ is the distance from the HHG source to the XUV mirror. The two curves deviate for long NIR focal lengths where the XUV beam has a very low divergence. In the limit of a collimated XUV beam, the demagnification would increase with $d_{NIR}$ because of the increasing XUV source size. In our experimental setup, we chose $d\approx 13$ m. This results in a demagnification of 173, independent of whether we assume Gaussian optics or demagnification according to simple geometrical optics. The global HHG scaling laws presented in Ref. [@heyl16] show that a constant HHG conversion efficiency can be obtained following the change of one parameter (e.g. the NIR pulse energy), if all other relevant parameters (including the NIR focal length and the NIR beam size before focusing) are changed accordingly. In particular, the NIR intensity needs to be kept constant both before focusing (to avoid damage of the focusing optics) and at the focus (for optimal phase matching). This results in an expected linear increase of the XUV pulse energy as a function of the NIR focal length [@heyl16], as shown in Fig. \[raytracing\](b) (blue curve). The expected XUV intensity (orange curve) shows a contrary behavior: It increases with decreasing $f_{NIR}$, since the smaller XUV focal spot size more than compensates for the lower XUV energy that is obtained. According to these scaling laws, the use of a very short NIR focal length would be most advantageous. However, under such conditions the divergence of the XUV beam is substantial, and restrictions due to finite XUV filter and mirror apertures (which cannot be made arbitrarily large) have to be taken into account. Therefore, we have performed first experiments using $d\approx 13$m. Besides offering a large demagnification, a further advantage of this beamline geometry is the occurrence of a comparably low NIR intensity on the Al filter that is used to block the NIR light co-propagating with the XUV beam. This removes the need for additional Si / SiC [@takahashi04] or antireflective coated mirrors [@gustafsson07], which are often used in high-power HHG applications to attenuate the NIR intensity. Our experimental setup is shown in Fig. \[setup\](a). An optical parametric chirped pulse amplification (OPCPA) system that was developed in-house is used to drive HHG. This system consists of a three-stage optical parametric amplifier, pumped by a home-built frequency-doubled Yb:YAG thin disk laser that operates at 100 Hz repetition rate. Ultrashort seed pulses from a commercial Yb:KGW front-end are amplified to pulses with an energy of up to 30 mJ and a duration of 8 fs in a spectral window from 700 nm to 1025 nm [@kretschmar19]. A telescope consisting of a concave ($f=1.5$ m) and a convex mirror ($f=-1$ m) is used to focus the NIR pulses. This configuration has the advantages that astigmatism can be compensated and that the distance of the focal plane from the last curved mirror can be easily adjusted. In our setup the f-number of the NIR beam is $\approx 400$, and the distance between the last NIR mirror and the NIR focal plane is about 5 m. A motorized iris is placed behind the last NIR mirror to optimize the HHG yield, and we measured an NIR driving energy of 11 mJ behind this iris (for an aperture of 10 mm). A 30 cm long gas cell is placed 10 cm in front of the focus, and is statically filled with Xe using a backing pressure of 7 mbar. We have performed gas dynamics simulations that indicate a reduced pressure of 1.1 mbar to be present within the cell. A 100 nm thick Al filter is placed 12 m downstream from the HHG cell and is used to block the co-propagating NIR light. The XUV pulses are focused into the interaction zone of the VMIS using a spherical B$_4$C coated mirror with a focal length of 75 mm. Ar atoms are injected into the instrument using a piezoelectric valve. The central part of the atomic beam is selected by a molecular beam skimmer with an orifice diameter of 0.5 mm. Ion distributions are recorded in a spatially-resolved manner along the XUV propagation axis using the spatial imaging mode of the VMIS [@eppink97]. In addition, the HHG beam diameter can be measured by moving a planar grating into the beam path and steering the zeroth diffraction order onto a microchannel plate (MCP) / phosphor screen assembly that is placed 12.5 m downstream from the HHG cell. An XUV pulse energy of 56 nJ was measured behind 2 Al filters using an XUV photodiode (AXUV100G) placed 12.5 m downstream from the HHG cell. We estimated an XUV pulse energy of 0.7 $\mu$J directly behind the gas cell. Considering the transmission of a new Al filter ($\approx 40$ $\%$) that was used for the experiment (note that different Al filters have different transmissions, because the transmission decreases over time due to oxidization and / or deposition of material) and the reflectivity of the XUV focusing mirror ($\approx 25$ $\%$), the XUV pulse energy in the VMIS experiment is expected to be $\approx 70$ nJ. We have further estimated the XUV beam waist radius in the VMIS as 1.3 $\mu$m (see below). The nonlinearity of the HHG process is expected to lead to the production of XUV pulses that are substantially shorter than the duration of the driver pulse. Indeed, previous measurements have reported HHG pulse durations which are shorter than the fundamental pulse duration by factors ranging from 1.6 to 2.9 [@mauritsson04; @nabekawa05]. Accordingly, assuming an XUV pulse duration of 4 fs, the XUV peak intensity is estimated as $7\times 10^{14}~$W/cm$^2$. We note that the actual XUV peak intensity might be even higher due to the attosecond pulse train structure of HHG. To demonstrate and investigate the utility of the focused XUV pulses for XUV nonlinear optics, CDI and XUV pump - XUV probe experiments, we have studied multiphoton ionization of Ar atoms, and observed the generation of ions with charge states up to Ar$^{5+}$. The spatial ion distributions along the XUV propagation direction (Fig. \[multiphoton-ionization\]) are peaked at the XUV focal plane and become narrower for more highly charged ions, which is a signature of the higher orders of nonlinearity leading to the production of these ions. We note that the absence of sidebands in the electron spectrum shown in Fig. \[setup\](b) demonstrates that the residual NIR light did not play a role in the ionization of Ar. The spatially resolved measurements allow us to determine the relative contributions of the different ion species in the focal plane (Fig. \[multiphoton-ionization\](e)), showing significant contributions of Ar$^{2+}$ (27.5 $\%$), Ar$^{3+}$ (19.5 $\%$) and Ar$^{4+}$ (3 $\%$) ions as well as a low contribution of Ar$^{5+}$ (0.05 $\%$) ions. These results are comparable to previously obtained FEL results [@motomura09]. We have used the longitudinal Ar$^{2+}$ and Ar$^{3+}$ ion distributions (solid curves in Fig. \[multiphoton-ionization\](f)) to estimate the Rayleigh length $z_R$ of the focused XUV beam. For this measurement, we have inserted a second Al filter with a transmission of 16 $\%$ to avoid saturation of the individual ion yields. The transversely integrated ion yield scales with $I^n(z) \times w^2(z) \propto I^n/I= I^{n-1}$, where $I(z)\propto 1/w^2(z)$ is the XUV intensity as a function of the distance $z$ from the focal plane, $n$ is the order of nonlinearity, and $w(z)=w_0\sqrt{1+z^2/z_R^2}$ is the beam radius, with $w_0$ being the beam waist radius. Accordingly, the ion distributions are proportional to $(1+z^2/z_R^2)^{-(n-1)}$. As indicated by measurements of the intensity dependence of the ion yields (see below), we assumed that the generation of Ar$^{2+}$ requires the absorption of 2 photons, while the generation of Ar$^{3+}$ requires the absorption of 4 photons. Fits using these orders of nonlinearity are shown as dotted curves in Fig. \[multiphoton-ionization\](f), giving $z_R=49$ $\mu$m for the fit of the Ar$^{2+}$ and $z_R=58$ $\mu$m for the fit of the Ar$^{3+}$ result. Therefore, we conclude that $z_R=54\pm 8$ $\mu$m. The XUV beam radius on the MCP / phosphor screen was measured to be 1.7 mm, indicating an XUV beam radius $w_{XUV,mirror}=1.8$ mm on the XUV focusing mirror due to the divergence of the XUV beam. The beam waist radius is estimated according to $w_0=w_{XUV,mirror}\times (1+f_{XUV}^2/z_R^2)^{-1/2}\approx w_{XUV,mirror} \times z_R/f_{XUV}$, giving a value of $1.3\pm 0.2$ $\mu$m. Fig. \[intensity-dependence\] shows the intensity-dependent Ar$^{2+}$, Ar$^{3+}$ and Ar$^{4+}$ ion yields, from which we can extract the orders of nonlinearity of the different ion species. We find $n=1.8\pm 0.2$ for Ar$^{2+}$, $n=3.9 \pm 0.2$ for Ar$^{3+}$ and $n=8.2 \pm 0.9$ for Ar$^{4+}$, indicating that about 8 XUV photons are absorbed in the latter case. While we have not recorded intensity-dependent data for Ar$^{5+}$ due to the low signal, we estimate that 3-4 additional photons are required to generate Ar$^{5+}$ from the Ar$^{4+}$ ground state, which has an ionization potential of 74.8 eV [@biemont99]. This suggests that at least 10 HHG photons are absorbed by a single atom. We note that the production of ions up to Ar$^{4+}$ was observed in a recent experiment using HHG, in which XUV pulses with energies of up to 230 $\mu$J were reported [@nayak18]. Our findings demonstrate that we achieve a higher degree of ionization, although our XUV pulse energy is more than two orders of magnitude lower than the pulse energy reported in Ref. [@nayak18]. Our results therefore open a path for the investigation of highly nonlinear processes in the XUV range using HHG sources that are driven by lasers with moderately high pulse energies, which makes it straightforward to perform these experiments using lasers at kHz repetition rates in the future. In summary, we have demonstrated the generation of XUV pulses by HHG reaching intensities up to $7\times 10^{14}$ W/cm$^2$ on target using a novel scaling scheme. These XUV pulses enabled highly nonlinear ionization of Ar atoms up to Ar$^{5+}$, following the absorption of at least 10 XUV photons. Our results provide excellent opportunities for XUV pump - XUV probe experiments and for the single-shot CDI of isolated nanotargets. **Funding.** Leibniz Grant No. SAW/2017/MBI4. **Acknowledgment.** We kindly acknowledge the technical support by M. Krause, C. Reiter, W. Krüger and R. Peslin. **Disclosures.** The authors declare no conflict of interest. [22]{}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 [****,  ()](\doibase 10.1103/PhysRevLett.105.263002) [****,  ()](\doibase https://doi.org/10.1038/nphys2033) [****,  ()](https://www.nature.com/articles/ncomms3691) [****,  ()](\doibase 10.1021/nl072728k) [****, ()](https://www.nature.com/articles/s41467-017-00287-z) [****,  ()](\doibase 10.1088/0953-4075/42/22/221003) [****,  ()](\doibase 10.1103/PhysRevLett.120.263901) @noop [****,  ()]{} [****, ()](\doibase 10.1103/RevModPhys.81.163) [****,  ()](\doibase 10.1103/PhysRevLett.112.073003) [****,  ()](\doibase 10.1103/PhysRevA.93.061402) [****,  ()](\doibase 10.1103/PhysRevA.98.023426) [****,  ()](\doibase 10.1364/OPTICA.5.000237) [****,  ()](\doibase 10.1103/PhysRevA.84.053819) [****,  ()](\doibase 10.1088/1361-6455/50/1/013001) [****,  ()](\doibase 10.1364/OL.29.000507) [****,  ()](\doibase 10.1364/OL.32.001353) @noop [ ]{} [****,  ()](\doibase https://doi.org/10.1063/1.1148310) [****,  ()](\doibase 10.1103/PhysRevA.70.021801) [****,  ()](\doibase 10.1103/PhysRevLett.94.043001) [****,  ()](\doibase https://doi.org/10.1006/adnd.1998.0803)
{ "pile_set_name": "ArXiv" }
--- abstract: | Within the description of stochastic differential equations it is argued that the existence of Boltzmann-Gibbs type distribution in economy is independent of the time reversal symmetry in econodynamics. Both power law and exponential distributions can be accommodated by it. The demonstration is based on a mathematical structure discovered during a study in gene regulatory network dynamics. Further possible analogy between equilibrium economy and thermodynamics is explored.\ [ ]{}\ PACS numbers:\ 89.65.Gh Economics; econophysics, financial markets, business and management;\ 05.10.Gg Stochastic analysis methods (Fokker-Planck, Langevin, etc)\ 05.20.-y Classical statistical mechanics.\ 87.23.Ge Dynamics of social systems address: ' Department of Mechanical Engineering, University of Washington, Seattle, WA 98195, USA ' author: - 'P. Ao' date: 'June 10, (2005)' title: ' Boltzmann-Gibbs Distribution of Fortune and Broken Time-Reversible Symmetry in Econodynamics' --- Introduction ============ Carefully and extensive analysis of real economic and financial data have revealed various exponential and power-law distributions regarding to money, income, wealth, and other economic quantities in ancient and modern social societies [@ms; @bp; @dy; @bm]. A remarkable analogy between the economic system and a thermodynamical system has been revealed in a recent study [@dy]. Using a detailed microdynamical model with time reversal symmetry, it was demonstrated that a Boltzmann-Gibbs type distribution exists in economic systems. Indeed, ample empirical analysis support such suggestion [@dy; @sy]. Nevertheless, different microdynamical models lead to apparently different distributions [@bm; @dy]. Those distributions are supported by empirical data, too. The nature of such difference may reveal the difference in corresponding economic structure. For example, this difference has been tentatively attributed to the role played by time reversal symmetry in microdynamical models [@dy]. Following the tradition of synthesizing interdisciplinary knowledge, such as from biology, physics, and finance [@farmer], in this letter we argue that irrespective of the time reversal symmetry the Boltzmann-Gibbs type distribution always exists. A broader theoretical base is thus provided. The demonstration is performed within the framework of stochastic differential equations and is based on a novel mathematical structure discovered during a recent study in gene regulatory network dynamics [@zhu]. In the light of thermodynamical description, possible explanations for the origin of the difference in various empirical distributions are proposed. Boltzmann-Gibbs distribution in finance ======================================= Stochastic differential equations or Langevin equations and their corresponding Fokker-Planck equations have been shown to be a useful modelling tool in economy [@bs; @heston; @ms; @bp]. One of the best examples is the Black-Scholes formula in option pricing theory [@bs]. This kind of mathematical formulation provides a direct connections between the microdynamics and the stationary state and has been used to generate various distribution laws [@ms; @bp; @bm; @dy; @sy]. Specifically, the stochastic differential equation may take the following form: $$\label{standard} \dot{{\bf q}} = {\bf f}({\bf q}) + N_I({\bf q}) \xi(t) \; ,$$ where ${\bf f}$ and ${\bf q}$ are $n$-dimensional vectors and ${\bf f}$ a nonlinear function of the state variable ${\bf q}$. The state variable is the quantity to specify the economic system. It may be the money, the income, or other suitable indices. The noise ${\bf \xi}$ is a standard Gaussian white noise with $l$ independent components: $\langle \xi_i \rangle = 0$, $\langle \xi_i(t) \xi_j (t')\rangle= 2 \delta_{ij} \delta (t-t')$, and $i,j=1, 2, ..., l$. In Eq.(1) we explicitly factorize out the pure noise source and the state variable dependent part for the convenient of later description. The specification of the noise in Eq.(\[standard\]) is through the $n\times n$ diffusion matrix $D({\bf q})$ by the following matrix equation $$N_I({\bf q}) N_I^\tau ({\bf q}) = \epsilon \; D({\bf q}) \; ,$$ where $N_I$ is an $n\times l$ matrix, $N_I^\tau$ is its the transpose, and $\epsilon$ a nonnegative numerical constant to keep tract of the noise. It plays the role of temperature in thermodynamics. According to Eq.(2) the $n\times n$ diffusion matrix $D$ is both symmetric and nonnegative. For the dynamics of state variable ${\bf q}$, all what needed from the noise is the diffusion matrix $D$. Hence, it is not necessary to require the dimension of the noise vector $\xi$ be the same as that of the state vector ${\bf q}$ and to require more specific knowledge of $n\times l$ matrices $\{ N_I \}$ beyond Eq.(2). It is known that even in situations that Eq.(\[standard\]) is not an exact description, and perhaps it would never be in a rigorous sense in economy, it may still serve as the first approximation for further modelling [@ms; @bp]. Indeed, it has been empirically verified to be a rather accurate description in economy [@bs; @heston; @rs; @loffredo]. Because the energy function or Hamiltonian has played a dominant role in equilibrium physics processes, the crucial question is whether or not a similar quantity exists in a more general setting. In the following we present an argument leading to the positive answer. There exists several ways to deal with the stochastic equations equation in the form of Eq.(1) and (2). The most commonly used are those of Ito and Stratonovich methods [@vankampen; @ms; @bp]. However, with those methods the connection between the existence of energy function like quantity in Eq.(1) and the stationary distribution is not clear when the time reversal symmetry is broken [@vankampen]. The difficulty for finding such potential function can be illustrated by the fact that usually $D^{-1}({\bf q}) {\bf f}({\bf q})$ cannot be written as the gradient of a scalar function [@vankampen] in the absence of detailed balance condition or in the broken time reversal symmetry. This will become precise as we proceed. During a recent study of the robustness of the genetic switch in a living organism [@zhu], it was discovered that Eq.(1) can be transformed into the following form, $$\label{normal} [ A({\bf q}) + C({\bf q})] \dot{{\bf q}} = \partial_{\bf q} \phi({\bf q}) + N_{II}({\bf q})\xi(t) \; ,$$ where the noise $\xi(t)$ is from the same source as that in Eq.(1). Here we tentatively name $A({\bf q})$ the adaptation matrix, $C({\bf q})$ the conservation matrix, and the scalar function $\phi({\bf q})$ the fortune function. The gradient operation in state space is denoted by $\partial_{\bf q}$. The adaptation matrix $A({\bf q})$ is defined through the following matrix equation $$N_{II}({\bf q}) N_{II}^\tau ({\bf q})= \epsilon \; A({\bf q}) \; ,$$ which guarantees that $A$ is both symmetric and nonnegative. The $n\times n$ conservation matrix $C$ is antisymmetric. We define $$A({\bf q}) + C({\bf q}) = 1/[ D({\bf q}) + Q({\bf q})] \equiv M({\bf q}) \; .$$ with the $n\times n$ matrix $M$ is the solution of following two matrix equations [@ao04]. The first equation is the potential condition $$\partial_{\bf q} \times [ M({\bf q}) {\bf f}({\bf q}) ] = 0 \; ,$$ which gives $n(n-1)/2$ conditions \[the wedge product for two arbitrary vectors ${\bf v}_1$ and ${\bf v}_2$ in n-dimension: $ [{\bf v}_1 \times {\bf v}_{2}]_{ij} = {v_1}_i {v_2}_j - {v_1}_j {v_2}_i % = \partial^2/\partial {q_i} \partial {q_j} - % \partial^2/\partial {q_j} \partial {q_i} \; \; , \; i,j = 1,2, ... , n $ \]. The second equation is the generalized Einstein relation between the adaptation and diffusion matrices in the presence of conservation matrix $$M({\bf q}) D({\bf q}) M^{\tau}({\bf q}) = \frac{1}{2} [M({\bf q}) + M^{\tau}({\bf q}) ] \; .$$ which gives $n(n+1)/2$ conditions [@ao04]. The fortune function $\phi({\bf q})$ is connected to the deterministic force ${\bf f}({\bf q})$ by $$\partial_{\bf q} \phi({\bf q}) = M({\bf q}) {\bf f}({\bf q}) \; .$$ For simplicity we will assume $\det(A) \neq 0$ in the rest of the letter. Hence $\det(M) \neq 0$ [@kat]. Thus, the adaptation matrix $A$, the conservation matrix $Q$ and the fortune function $\phi$ in Eq.(3) and (4) can be completely determined by Eq.(1) and (2). The breakdown of detailed balance condition or the time reversal symmetry is represented by the finiteness of the conservation matrix $$C({\bf q}) \neq 0 \; ,$$ or equivalently $Q \neq 0$. The usefulness of the formulation of Eq.(3) and (4) is already manifested in the successful solution of outstanding stable puzzle in gene regulatory dynamics [@zhu] and in solving two fundamental controversies in population genetics [@ao05]. A few remarks on Eq.(3) are in order. In the light of classical mechanics in physics, Eq.(3) is in precisely the form of Langevin equation. The fortune function $\phi$ corresponds to the potential function but opposite in sign to reflect the fact that in economy there is a tendency to seek the peak or maximum of fortune. The adaptive matrix $A$ plays the role of friction. It represents adaptive dynamics and is the dynamical mechanism to seek the nearby fortune peak. The conservation matrix $C$ plays the role analogous to a magnetic field. Its dynamics is similar to that of the Lorentz force, hence conserves the fortune. As in classical mechanics, the finiteness of the conservation matrix $C$ breaks the time reversal symmetry. It was heuristically argued [@ao04] and rigorous demonstrated [@yin] that the stationary distribution $\rho({\bf q})$ in the state space is, if exists, $$\label{bg-dis} \rho({\bf q}) \propto \exp\left( {\phi({\bf q}) \over{\epsilon} } \right) \; .$$ Therefore, the fortune function $\phi$ acquires both the dynamical meaning through Eq.(\[normal\]) and the steady state meaning through Eq.(\[bg-dis\]). Specifically, in the so-called zero-mass limit to differentiate from Ito and Stratonovich methods, the Fokker-Planck equation for the probability distribution $\rho({\bf q},t)$ takes the form [@yin] $$\label{fp-eq} {\partial_t \rho({\bf q},t) } = \partial_{{\bf q}}^{\tau} M^{-1}({\bf q}) [\epsilon \partial_{\bf q} - \partial_{{\bf q}} \phi({\bf q})] \rho({\bf q},t) \; .$$ Here $\partial_t$ is a derivative with respect to time and $\partial_{\bf q}$ represents the gradient operation in state space. We note that Eq.(\[bg-dis\]) is a stationary solution to Eq.(\[fp-eq\]) even it may not be normalizable, that is, even when the partition function ${\cal Z} = \int d^{n}{\bf q} \; \rho({\bf q})$ is ill-defined. Again, we emphasize that no time reversal symmetry is assumed in reaching this result. This completes our demonstration on the existence of the Boltzmann-Gibbs distribution in economy. Using $ M^{-1}({\bf q})= D({\bf q}) + Q({\bf q})$ and $ {\bf f}({\bf q}) = [D({\bf q}) + Q({\bf q})] \partial_{{\bf q}} \phi({\bf q})$, Eq.(\[fp-eq\]) can be rewritten in a more suggestive form [@yin] $$\label{fp-eq2} {\partial_t \rho({\bf q},t) } = \partial_{{\bf q}}^{\tau} [\epsilon D({\bf q})\partial_{\bf q} + \epsilon (\partial_{{\bf q}}^{\tau} Q({\bf q})) - {\bf f}({\bf q})] \rho({\bf q},t) \; .$$ It is clear that in the presence of time reversal symmetry, [*i.e.*]{} $Q = 0$, one can directly read the fortune function $\phi$ from above form of Fokker-Planck equation as $ \partial_{{\bf q}} \phi({\bf q}) = D^{-1}({\bf q}) {\bf f}({\bf q})$. For the sake of completeness, we list the Fokker-Placnk equations corresponding to Ito and Stratonovich treatments of Eq.(1) and (2) [@vankampen]: $$\label{ito} \partial_t \rho_{I}({\bf q},t) = \sum_{i=1}^{n}\partial_{ q_i} \left[ - f_i({\bf q}) + \epsilon \sum_{j=1}^n \partial_{ q_j} D_{i,j}({\bf q}) \right] \rho_{I}({\bf q}, t) \; , {\ } {\ } {\ } (Ito)$$ and $$\label{str} \partial_t \rho_{S}({\bf q},t) = \sum_{i=1}^n \partial_{ q_i} \left[ - f_i({\bf q}) + \sum_{j=1}^n\sum_{k=1}^l {N_{I}}_{ik}({\bf q}) \partial_{ q_j} {N_I}_{jk}({\bf q}) \right] \rho_S({\bf q}, t) \; . {\ } {\ } {\ } (Str)$$ The connection of fortune function with both dynamical (Eq.(1) and (2)) and stationary state is indeed not clear in above two equations. Nevertheless, it has been shown [@yin] that there are corresponding fortune functions, adaptation and conservation matrices to Eq.(\[ito\]) and (\[str\]). We point out here that when the matrix $N_I$ is independent of state variable, Eq.(\[ito\]) and (\[str\]) are the same but may still differ from Eq.(\[fp-eq\]), because the gradient of the antisymmetric matrix $Q({\bf q})$ may not be zero. This last property shows that the time reversal symmetry is indeed important. Two Examples ============ The Fokker-Planck equation used by Silva and Yakovenko [@sy] has the form: $$\label{sy-eq} {\partial_t \rho_{sy}({ q},t) } = \partial_{ q} [ a({ q}) + \partial_{ q} b({ q})] \rho_{sy}({ q},t) \; ,$$ and Fokker-Planck equation used by Bouchaud and Mezard [@bm] has the form $$\label{bm-eq} {\partial_t \rho_{bm}({ q},t) } = \partial_{ q} [(J (q-1) + \sigma^2 q + \sigma^2 q \partial_{ q} q ] \rho_{bm}({ q},t) \; .$$ They are all in one dimension. We immediately conclude that the conservation matrix $C$, equivalently $Q$, is zero, because there is no conservation matrix in one dimension. In accordance with the definition in the present letter, which is consistent with that in nonequilibrium processes [@vankampen], the dynamics described by above two equations can be effectively classified as time reversal symmetric. Rewriting them in symmetric form with respect to the derivative of state variable $q$ as in Eq.(\[fp-eq2\]), we have $$\label{sy-eq} {\partial_t \rho_{sy}({ q},t) } = \partial_{ q} [ a({ q}) + (\partial_{ q} b({ q})) + b({ q}) \partial_{ q}] \rho_{sy}({ q},t) \; ,$$ and $$\label{bm-eq} {\partial_t \rho_{bm}({ q},t) } = \partial_{ q} [(J (q-1) + \sigma^2 q + \sigma^2 q + \sigma^2 q \partial_{ q} ] \rho_{bm}({ q},t) \; .$$ The corresponding wealth functions can be immediate read out as $$\begin{aligned} \phi_{sy}({ q}) & = & - \int_{q_0}^q d q' { a(q') + (\partial_{q'} b(q')) \over{ b(q') } } \nonumber \\ & = & - \frac{a}{b} (q-q_0) \; , \; if \; a, b = constant \\ \phi_{bm}({ q}) & = & - \int_{q_0}^{q} d q' {(J (q'-1) + 2 \sigma^2 q' \over { \sigma^2 {q'}^2 } } \nonumber \\ & = & - {J \over{\sigma^2 }} {1 \over{q}} - \left( 2 + {J \over{\sigma^2}} \right) \ln q + {J \over{\sigma^2 }} {1 \over{q_0}} + \left( 2 + {J \over{ \sigma^2}} \right) \ln q_0\end{aligned}$$ They are exactly what found in Ref.\[[@sy]\] and \[[@bm]\]: the first one corresponds to an exponential distribution and the second one a power law distribution according to the Boltzmann-Gibbs distribution Eq.(\[bg-dis\]). Ensembles and state variables ============================= Having defined a precise meaning of time reversal symmetric and have demonstrated that the Boltzmann-Gibbs distribution even in the absence of time reversal symmetry, we explore further connection between econodynamics and statistical physics. There are two general types of situations which would generate different distributions in statistical physics and thermodynamics. The first one is to link to constraints on the system under various conditions. In statistical physics such constraints are described by various ensembles and free energies. For example, there are canonical and grand-canonical ensembles. There are Gibbs and Helmholtz free energies, entropy, enthalpy, etc. Those ensembles have their characteristic distributions. It would be interesting to know the corresponding situations in economy. Even with a given constraint, the form of distribution depends on the choice of state variable. For example, for ideal gas model, the distributions are different if views from the kinetic energy and from velocity. Hence, there is a question of appropriate state variable for a given situation, with which the physics becomes particular transparent. It would be interesting to know what be the appropriate variables to describe an economic system. Within this context, the difference between what discovered by Dragulecu and Yakovenko [@dy] and Bouchaud and Mezard [@bm] is perhaps more due to the difference in choices of state variables, because it seems they are describing the same situation of same system under same constraints. A remark on terminology is in order. It was demonstrated above that regardless of the time reversal symmetry the Boltzmann-Gibbs distribution, Eq.(\[bg-dis\]), exists. The fortune function $\phi$ has an additional dynamical meaning defined in Eq.(\[normal\]). Both exponential and power law distribution can be represented by Eq.(\[bg-dis\]). In fact, it is well known that power law distributions exist in statistical physics. A nontrivial example is the Kosterlitz-Thouless transition [@kt]. Thus it does not appear appropriate to call the power law distribution non-Boltzmann-Gibbs distribution. Such a terminology confusion was already noticed before [@ls]. In the view of the dominant role of entropy in Kosterlitz-Thouless transition [@kt], the ubiquitous existence of power law distribution in economy may suggest that the entropy effect is rather important in econodynamics. This may corroborate with the suggestion of “superthermal” in economy [@sy]. Conclusions =========== In this letter we demonstrate that the existence of Botlzmann-Gibbs distribution in finance is independent of time reversal symmetry. Both power law and exponential distributions are within its description. In analogous to similar situation in statistical physics, the differences among those distributions discovered empirically in economy are likely the result of different choices of state variables to describe the same system in econodynamics. [ ]{} This work was supported in part by USA NIH grant under HG002894. [99]{} R.N. Mantegna and H.E. Stanley, An Introduction to Econophysics: correlation and complexity in finace, Cambridge University Press, Cambridge, 2000. J.-P. Bouchaud and M. Potters, Theory of Finanical Risks, 2nd edition, Cambridge University Press, Cambridge, 2003. J.-P. Bouchaud and M. Mezard, Physica [**A282**]{} (2000) 536. A. Dragulescu and V.M. Yakovenko, Euro. Phys. J. [**B17**]{} (2000) 723. A.C. Silva and V.M. Yakovenko, Europhys. Lett. [**69**]{} (2005) 304. J.D. Farmer, Industrial and Corporate Change [**11**]{} (2002) 895. X.-M. Zhu, L. Yin, L. Hood, and P. Ao, Func. Integr. Genomics [**4**]{} (2004) 188;\ J. Bioinf. Comput. Biology [**2**]{} (2004) 785. F. Black and M. Scholes, J. Finance [**27**]{} (1972) 399;\ J. Political Economy [**81**]{} (1973) 637. S.L. Heston, Rev. Finanical Studies [**6**]{} (1993) 327. P. Richmond and L. Sabatelli, Physica [**A336**]{} (2004) 27. M.I. Loffredo, Int. J. Mod. Phys. [**B18**]{} (2004) 705. N.G. van Kampen, [*Stochastic Processes in Physics and Chemistry*]{}. Elsevier, Amsteredam, 1992. P. Ao, J. Phys. [**A37**]{} (2004) L25. C. Kwon, P. Ao, and D.J. Thouless, submitted to PNAS. available upon request. P. Ao, Phys. Life Rev. [**2**]{} (2005) 117. L. Yin and P. Ao, submitted to PRL. available upon request. J.M. Kosterlitz and D.J. Thouless, J. Phys. [**C6**]{} (1973) 1181. M. Levy and S. Solomon, Int. J. Mod. Phys. [**C7**]{} (1996) 595.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We study a nearest-neighbor hopping model on the Cayley tree under the smooth boundary condition with the modulation function $f_s=\sin^2[\pi s/(2M+1)]$, where $s$ is a distance from the central site, and $M$ is the number of shells on the tree. As a result of this smoothing, the particle density in the ground state becomes nearly uniform in the bulk region even when $M$ is relatively small. We compare the calculated particle density at the center with exact result on the Bethe lattice, and they show a good agreement. The calculated bond energy at the center also agrees with that on the Bethe lattice.' address: ' Department of Physics, Graduate School of Science, Kobe University, Kobe 657-8501, Japan ' author: - title: Bulk Property on Cayley Tree with Smooth Boundary Condition --- Introduction ============ The Cayley tree is a loopless graph where the coordination number $z$ on each site is the same except at the boundary. It is known that boundary effects tend to be dominant on the Cayley tree, since the number of the boundary sites on the tree increases exponentially with respect to the diameter $M$ of the tree. In some cases, the boundary effect remains even in the thermodynamic limit $M\rightarrow \infty$, as is known in the case of ferromagnetic Ising model[@Runnels; @Eggarter; @Matsuda; @Morita; @Muller-Hartmann; @Baxter]; a thermodynamic quantity of the whole system is dominated by those area near the boundary. On the other hand, an expectation value of a local quantity at the center of the tree can be different from the averaged value for the entire system. Such a “bulk" property, which appears deeply inside the system, can be described by the Bethe lattice, which can be regarded as an infinite Cayley tree without boundary. In this article, we consider a way of obtaining the “bulk" property of observables on the Cayley tree. For this purpose we try to suppress the boundary effect. It is well known that the strength of the boundary effects is dependent on boundary conditions. The most simple one is the open boundary condition (OBC), where the boundary effect is relatively conspicuous. A typical example is the Friedel oscillation caused by the open boundary. It is possible to suppress such a sudden boundary effect by imposing the periodic boundary condition (PBC). However, PBC on the Cayley tree is non-trivial. A solution is to introduce so called the smooth boundary condition (SBC)[@sbc1; @sbc2], where the local energy scale varies smoothly from the maximum at the center to zero at the boundary. In one dimension, a class of SBC called “sine-square deformation" (SSD)[@ssd] efficiently reduces the boundary effect. For a variety of one-dimensional systems, it has been confirmed that the boundary effect in the ground state vanishes entirely under SSD[@ssd; @ssd-HN; @ssd-K1; @ssd-GDLN; @ssd-K2; @ssd-SH]. The SSD has been applied to quantum entanglement analysis on spin chains[@Hikihara] and the ground state study of a simple string theory[@Tada]. The validity of SSD for two-dimensional system is also confirmed[@ssd-MKH; @Hotta1; @Nishimoto]. Recently, a method of obtaining bulk quantities by means of SSD, the method which is called as the grand canonical analysis, is proposed[@Hotta1; @Hotta2] and applied to kagome Heisenberg antiferromagnet[@Nishimoto]. An advantage of this grand canonical method is that one can easily access bulk quantities from numerical observables of finite-size systems. In this paper, we compare bulk properties of the nearest-neighbor hopping model on the Bethe lattice, with observables on the Cayley tree, performing a numerical calculation with the aid of the grand canonical analysis. This paper is organized as follows. In Sec. \[sec:TB\], we introduce the nearest-neighbor hopping model on the Cayley tree. We shall see that the Friedel oscillation disturbs the bulk quantity under OBC. In Sec. \[sec:SBC\], we impose SBC to the model. The obtained results are compared with the bulk quantities on the Bethe lattice. We conclude the obtained result in Sec. \[sec:last\]. Nearest-Neighbor Hopping Model on Cayley Tree {#sec:TB} ============================================= We first consider how to label sites on the Cayley tree. We introduce a site-centered tree with coordination number $z$, where there are $M$ shells. Figure \[tree\] shows the case $z=3$ and $M=3$. All the interior sites have $z$ neighboring sites, but those boundary sites have only one. Regarding the center of the lattice as the 0-th shell, each site on the $s$-th shell can be labelled by a set of $s+1$ integers $$\begin{aligned} R=\{ r_0\: r_1\ldots r_s \}, \label{labeling}\end{aligned}$$ where $s=0,1,\ldots,M$, and where $r_i$ takes values as follows $$\begin{aligned} r_i = \begin{cases} 1 & (i=0)\\ 1,\ldots, z & (i=1)\\ 1,\ldots, z-1 & (i=2,\ldots,M) \end{cases} .\end{aligned}$$ ![A Cayley tree with coordination number with $z=3$. The number of shell $M$ is equal to 3. All sites are labeled according to the index rule in Eq. .[]{data-label="tree"}](fig1.eps){width="40.00000%"} Next we introduce a spinless nearest-neighbor hopping model on the Cayley tree. We introduce fermionic creation and annihilation operator, respectively, $c^{\dag}_{R}$ and $c^{}_{R}$ on each site labeled by $R$ in Eq. . The hopping between neighboring sites $R$ and $R'$ is expressed as $$\begin{aligned} h(R, R') = -t (c^{\dag}_{R} c^{}_{R'} + c^{\dag}_{R'} c^{}_{R}),\end{aligned}$$ where $t$ corresponds to the hopping amplitude, and the on-site potential is given by $$\begin{aligned} g(R) = -\mu c_{R}^{\dag} c^{}_{R},\end{aligned}$$ where $\mu$ denotes the chemical potential. By means of these notations, the nearest-neighbor hopping Hamiltonian on the tree is given by $$\begin{aligned} H =\sum_{s=0}^{M-1} \sum_{r_0\ldots r_{s+1}} h(\{ r_0\ldots r_s \}, \{ r_0\ldots r_{s+1} \}) +\sum_{s=0}^{M} \sum_{r_0\ldots r_s} g(\{ r_0\ldots r_s \}). \label{H_OBC}\end{aligned}$$ For example, in the case of $z=3,M=2$, the Hamiltonian is written as $$\begin{aligned} H = & h(\{1\}, \{11\}) + h(\{1\}, \{12\}) + h(\{1\}, \{13\}) + g(\{1\}) \notag\\ & + h(\{11\}, \{111\}) + h(\{11\}, \{112\}) + g(\{11\}) \notag\\ & + h(\{12\}, \{121\}) + h(\{12\}, \{122\}) + g(\{12\}) \notag\\ & + h(\{13\}, \{131\}) + h(\{13\}, \{132\}) + g(\{13\}) \notag\\ & + g(\{111\}) + g(\{112\}) + g(\{121\}) + g(\{122\}) + g(\{131\}) + g(\{132\}) .\end{aligned}$$ Since the hopping is terminated at the boundary, the open boundary condition (OBC) is imposed naturally. Now we shall see the behavior of local observables in the ground state under OBC. Since the Hamiltonian contains no interactions between particles, we can obtain the one-particle eigenstate from the following eigenvalue relation $$\begin{aligned} H \sum_{R} \psi_i(R) c^{\dag}_{R} {\lvert 0 \rangle} = E_i \sum_{R} \psi_i(R) c^{\dag}_{R} {\lvert 0 \rangle}, \label{eigen}\end{aligned}$$ where $E_i$ is the $i$-th eigenvalue, and where $\psi_i(R)$ is the corresponding one-particle wave function. We assume the increasing order for $E_i$. By means of $\psi_i(R)$, the site occupation is expressed as $$\begin{aligned} n_R \equiv {\langlec^{\dag}_R c^{}_R \rangle} = \sum_{i(E_i \leq 0)} \psi_i^* (R) \psi_i (R) \label{density}\end{aligned}$$ under the grand canonical ensembles. Similarly, the bond energy between neighboring sites $R$ and $R'$ can be obtained as $$\begin{aligned} \epsilon^{}_{R,R'} & \equiv -t {\langlec^{\dag}_{R} c^{}_{R'} + c^{\dag}_{R'} c_{R} \rangle} \notag \\ & = -t\sum_{i(E_i \leq 0)} \Bigl[ \psi_i^* (R) \psi_i (R') + \psi_i^* (R') \psi_i (R) \Bigr]. \label{bond_energy}\end{aligned}$$ The nearest-neighbor hopping model on the tree has many degenerate energy levels due to the symmetry of the tree. If and only if the chemical potential $\mu$ coincides with one of such energy levels, there exist a ground states degeneracy. Throught this article, we avoid such a coincidence, and we choose $\mu$ so that the ground state is determined uniquely. In this situation, the particle density $n_{R}$ in Eq.  and the bond energy $\epsilon^{}_{R,R'}$ in Eq.  are constant within each shell. Hence, they are the functions of $s$, the shell index. We thus use simpler forms $$\begin{aligned} n_s &\equiv n_{R_s},\\ \epsilon^{}_{s} & \equiv \epsilon^{}_{R_{s}, R_{s+1}},\end{aligned}$$ where $R_s$ denotes $\{ r_0\: r_1 \ldots r_s \}$. Since the total number of sites in the system increases exponentially with respect to $M$, it is not straight forward to solve the eigenvalue relation Eq.  numerically when $M$ is relatively large. This problem can be overcome by the block-diagonalization process[@Ogawa; @Lepetit]. All the blocks in the Hamiltonian matrix are symmetric tridiagonal matricies, whose size is $(M+1) \times (M+1)$ at most, so that one can easily diagonalize each. Note that this block-diagonalization procedure can be applied also for those cases where Hamiltonian terms are position dependent, if $t$ and $\mu$ are the same for all the bonds and sites in the same shell. We shall compare the calculated density $n_s$ under OBC with that on the Bethe lattice[@Economou; @Thorpe; @Mahan; @Eckstein; @Kollar], which is given by $$\begin{aligned} n (\mu) = \frac{1}{2\pi} \left[ z\theta_{\mu} - (z-2) \arctan \left( \frac{z}{z-2} \tan \theta_{\mu} \right) \right], \label{exact-n}\end{aligned}$$ where the relation $\mu = -2|t|\sqrt{z-1}\cos \theta_{\mu}$ is satisfied. We also compare the bond energy $\epsilon^{}_{s}$ with that of the Bethe lattice $$\begin{aligned} \epsilon (\mu) = -\frac{|t|}{\pi} \left[ \sqrt{4(z-1)- (\mu /t)^2} - (z-2) \arctan \left( \frac{\sqrt{4(z-1)-(\mu /t)^2}}{z-2} \right) \right]. \label{exact-e}\end{aligned}$$ In the following numerical calculations, we consider $t$ as the unit of energy. Figure \[ns\_obc\] shows $n_s$ when $z=3, M=20$ and $\mu=0.5$. The solid line shows the corresponding bulk value in Eq. . We observe that $n_s$ oscillates around this bulk value and the oscillation amplitude is not negligible even at the center. The similar fluctuation is observed in the bond energy as shown in Fig. \[es\_obc\]. Smooth Boundary Condition {#sec:SBC} ========================= In this section, we consider the nearest-neighbor hopping model on the Cayley tree under SBC. We start from the simplest case $z=2$, i.e., one-dimensional one. Since there are only sites that are labeled such as $\{ 111\ldots 1 \}$ and $\{ 121\dots 1\}$ on the tree, the Hamiltonian is reduced as $$\begin{aligned} H = \sum_{s=0}^{M-1} \sum_{r_1=1}^{2} h(\{ \underbrace{1\: r_1\: 1\ldots 1}_{s+1} \}, \{ \underbrace{1\: r_1\: 1 \ldots 1}_{s+2} \}) +g(\{1\}) + \sum_{s=1}^M \sum_{r_1=1}^{2} g(\{ \underbrace{1\: r_1\: 1\ldots 1}_{s+1} \}).\end{aligned}$$ In the SBC scheme, local energy scale is smoothly modulated by a function $f_s$, where $s$ is the distance from the center of the system. The modulation function $f_s$ is chosen to be smooth with respect to $s$ and vary from the maximum value near the center to zero at the edge. Under such a setup, the Hamiltonian with SBC is written by $$\begin{aligned} H_{\mathrm{SBC}} = \sum_{s=0}^{M-1} f_{s+\frac{1}{2}} \sum_{r_1=1}^{2} h(\{ \underbrace{1\: r_1\: 1\ldots 1}_{s+1} \}, \{ \underbrace{1\: r_1\: 1 \ldots 1}_{s+2} \}) \notag\\ + f_{0}\: g(\{1\}) + \sum_{s=1}^M f_{s} \sum_{r_1=1}^{2} g(\{ \underbrace{1\: r_1\: 1\ldots 1}_{s+1} \}). \label{deformed_z2}\end{aligned}$$ Now the generalization to those cases $z\geq 3$ is straight forward. The genelized Hamiltonian can be rewritten as $$\begin{aligned} H_{\mathrm{SBC}} = \sum_{s=0}^{M-1} f_{s+\frac{1}{2}} \sum_{r_0\ldots r_{s+1}} h(\{ r_0\ldots r_s \}, \{ r_0\ldots r_{s+1} \}) +\sum_{s=0}^{M} f_s \sum_{r_0\ldots r_s} g(\{ r_0\ldots r_s \}).\end{aligned}$$ As a choice of the smoothing function $f_s$, we choose $$\begin{aligned} f_s = \sin^2 \left[ \frac{\pi}{2M+1} \left( s+M+\frac{1}{2} \right) \right], \label{ssd-function}\end{aligned}$$ where the functional form is shown in Fig. \[fig-ssd\]. The deformation with this $f_s$ is called as sine-square deformation (SSD). ![The modulation function $f_s$ in Eq. .[]{data-label="fig-ssd"}](fig3.eps){width="0.5\linewidth"} Figure. \[ns\_ssd\] shows the particle density $n_s$ under SSD with $z=3$, $M=20$, and $\mu=0.5$, where the parameter set is equivalent to those in Fig. \[ns\_obc\]. In the bulk region ($s\sim 1$), the oscillation is suppressed conspicuously, and the converged value well agrees with the bulk value on the Bethe lattice. The diffrence lies within $O(10^{-4})$ near the center as shown in the inset. SSD also reduces the oscillation in the bond energy $\epsilon_s$ as shown in Fig. \[es\_ssd\]. The calculated value approaches the bulk one toward the center. The difference converge to the order of $10^{-4}$ as shown in the inset. Now we estimate the bulk density of the system by the ground canonical analysis applied to the calculated data. As seen in Fig. \[ns\_ssd\], the oscillation around the central site of the system is small. Thus we may consider $n_0$ as the candidate for the bulk density. We compare the obtained density $n_0$ with that on the Bethe lattice $n(\mu)$, which is given in Eq. . Figure \[n0\] shows the $\mu$ dependence in $n_0$. The curves of $n_0$, calculated for $M=10$ and 20, well coincides with $n(\mu)$, where the typical discrepancy is at most of the order of $10^{-3}$. We also compare $\epsilon_0$ with the bond energy on the Bethe lattice $\epsilon (\mu)$ in Eq. , in the same manner. The coincidence is observed as shown in Fig. \[e0\]. The difference between $\epsilon_0$ and $\epsilon (\mu)$ lies within the order of $10^{-3}$. Conclusion {#sec:last} ========== We have applied SSD to the nearest-neighbor hopping model on the Cayley tree. As an effect of the boundary condition, the particle density become relatively uniform in the bulk region, deep inside the tree. The $\mu$ dependence of the particle density at the center of the system coincide with that on the Bethe lattice even for a small $M$, the number of shells. We also observe such coincidence in the bond energy. These fact suggest that SSD enables us to extract bulk quantities of various local observables for the Cayley tree in an efficient way. A future problem is to examine other deformations such as sinusoidal[@ssd-GDLN; @Hikihara], exponential[@exp1; @exp2] and hyperbolic[@hyp1; @hyp2; @hyp3] deformation. Acknowledgement {#acknowledgement .unnumbered} =============== The author would like to thank T. Nishino for valuable comments. [99]{} L. K. Runnels, J. Math. Phys. **8**, 2081 (1967). T. P. Eggarter, Phys. Rev. B. **9**, 2989 (1974). H. Matsura, Prog. Theor. Phys. **51**, 1053 (1974). E. Müller-Hartmann and J. Zittartz, Phys. Rev. Lett. **33**, 893 (1974). T. Morita and T. Horiguchi, Prog. Theor. Phys. **54**, 982 (1975). R. Baxter, *Exactly solved models in statistical mechanics* (Academic Press, London 1982). M. Vekić and S. R. White, Phys. Rev. Lett. **71**, 4283 (1993). M. Vekić and S. R. White, Phys. Rev. B **53**, 14552 (1996). A. Gendiar, R. Krcmar, and T. Nishino, Prog. Theor. Phys. **122**, 953 (2009); **123**, 393 (2010). T. Hikihara and T. Nishino, Phys. Rev. B **83**, 060414(R) (2011). H. Katsura, J. Phys. A: Math. Thoer. **44**, 252001 (2011). A. Gendiar, M. Daniška, Y. Lee, and T. Nishino, Phys. Rev. A. **83**, 052118 (2011). N. Shibata and C. Hotta, Phys. Rev. B **84**, 115116 (2011). H. Katsura, J. Phys. A: Math. Thoer. **45**, 115003 (2012). T. Hikihara and T. Suzuki, Phys. Rev. A **87**, 042337 (2013). T. Tada, arXiv:1404.6343. I. Maruyama, H. Katsura, and T. Hikihara, Phys. Rev. B **84**, 165132 (2011). C. Hotta, S. Nishimoto, and N. Shibata, Phys. Rev. B **87**, 115128 (2013). S. Nishimoto, N. Shibata, and C. Hotta, Nat. Commun. **4**, 2287 (2013). C. Hotta and N. Shibata, Phys. Rev. B **86**, 041108(R) (2012). T. Ogawa, Prog. Theor. Phys **54**, 1028 (1975). M.-B. Lepetit, M. Cousy, and G. M. Pastor, Eur. Phys. J. B **13**, 421 (2000). E. N. Economou, *Green’s Functions in Quantum Physics* (Springer-Verlag, Berlin, 1979). M. F. Thorpe, In *Excitations in Disordered Systems*, ed. M. F. Thorpe (Plenum, New York, 1981), pp. 85-107. G. D. Mahan, Phys. Rev. B **63**, 15110 (2001). M. Eckstein, M. Kollar, K. Byczuk, and D. Vollhardt, Phys. Rev. B **71**, 235119 (2005). M. Kollar, et. al., Ann. Phys. **14**, 642 (2005). K. Okunishi and T. Nishino, Phys. Rev. B **82**, 144409 (2010). G. Ramiírez, J. Rodríguez-Laguna, and G. Sierra, J. Stat. Mech. **2014**, P10004 (2014). H. Ueda and T. Nishino, J. Phys. Soc. Jpn. **78**, 014001 (2008). H. Ueda, H. Nakano, K. Kusakabe, and T. Nishino, arXiv:0812.4513. H. Ueda, H. Nakano, K. Kusakabe, and T. Nishino, Prog. Thoer. Phys. **124**, 389 (2010).
{ "pile_set_name": "ArXiv" }
--- abstract: | The UN Sustainable Development Goals allude to the importance of infrastructure quality in three of its seventeen goals. However, monitoring infrastructure quality in developing regions remains prohibitively expensive and impedes efforts to measure progress toward these goals. To this end, we investigate the use of widely available remote sensing data for the prediction of infrastructure quality in Africa. We train a convolutional neural network to predict ground truth labels from the Afrobarometer Round 6 survey using Landsat 8 and Sentinel 1 satellite imagery. Our best models predict infrastructure quality with AUROC scores of 0.881 on Electricity, 0.862 on Sewerage, 0.739 on Piped Water, and 0.786 on Roads using Landsat 8. These performances are significantly better than models that leverage OpenStreetMap or nighttime light intensity on the same tasks. We also demonstrate that our trained model can accurately make predictions in an unseen country after fine-tuning on a small sample of images. Furthermore, the model can be deployed in regions with limited samples to predict infrastructure outcomes with higher performance than nearest neighbor spatial interpolation. author: - Barak Oshri - Annie Hu - Peter Adelson - Xiao Chen - Pascaline Dupas - Jeremy Weinstein - Marshall Burke - David Lobell - Stefano Ermon bibliography: - 'egbib.bib' - 'liter\_review\_ref.bib' title: Infrastructure Quality Assessment in Africa using Satellite Imagery and Deep Learning --- &lt;ccs2012&gt; &lt;concept&gt; &lt;concept\_id&gt;10010147.10010257.10010293.10010294&lt;/concept\_id&gt; &lt;concept\_desc&gt;Computing methodologies Neural networks&lt;/concept\_desc&gt; &lt;concept\_significance&gt;500&lt;/concept\_significance&gt; &lt;/concept&gt; &lt;/ccs2012&gt; &lt;ccs2012&gt; &lt;concept&gt; &lt;concept\_id&gt;10010147.10010257.10010293.10010294&lt;/concept\_id&gt; &lt;concept\_desc&gt;Computing methodologies Neural networks&lt;/concept\_desc&gt; &lt;concept\_significance&gt;500&lt;/concept\_significance&gt; &lt;/concept&gt; &lt;concept&gt; &lt;concept\_id&gt;10010405.10010455.10010460&lt;/concept\_id&gt; &lt;concept\_desc&gt;Applied computing Economics&lt;/concept\_desc&gt; &lt;concept\_significance&gt;300&lt;/concept\_significance&gt; &lt;/concept&gt; &lt;/ccs2012&gt;
{ "pile_set_name": "ArXiv" }
--- abstract: 'We show show that the singularities of the Fresnel surface for Maxwells equation on an anisotrpic material can be accounted from purely topological considerations. The importance of these singularities is that they explain the phenomenon of conical refraction predicted by Hamilton. We show how to desingularise the Fresnel surface, which will allow us to use Morse theory to find lower bounds for the number of critical wave velocities inside the material under consideration. Finally, we propose a program to generalise the results obtained to the general case of hyperbolic differential operators on differentiable bundles.' author: - | Carlos Valero\ Universidad Autonoma Metropolitana (UAM)\ Unidad Cuajimalpa\ Mexico City, Mexico\ \ bibliography: - 'myBib.bib' title: 'Maxwells Equations, The Euler index and Morse theory' --- \#1[\^[2]{}\#1]{} \#1[\_[0]{}\^[2]{}\#1]{} \#1[\#1]{} \#1\#2 \#1 Introduction ============ One of the most interesting problems in geometrical optics in the nineteenth century was the phenomenon of double refraction, in which a ray of light entering certain crystals is refracted into two rays. Later on, Hamilton discovered that if the ray of light enters a biaxial crystal in certain directions (known as the optical axis) then the ray of light must be refracted into a cone of rays. This phenomenon, now known as conical refraction, was confirmed a year later by Lloyd, who investigated aragonite at Hamilton’s suggestion. In mathematical terms, conical refraction is explained by the existence of conical singularities in the Fresnel surface associated to the crystal. The Fresnel surface is constructed as the set of allowed wave speeds imposed by Maxwell’s equations. In this paper we will show that the singularities of the Fresnel surface can be accounted for topological considerations only. Furthermore, we apply Morse theory to establish a connection between the number of these singularities and the critical speeds of wave propagation within the crystal. Inspired by this result, we propose a program to study this problem in the general context of hyperbolic differential equations on manifolds. Recall that Maxwell’s equations in a medium with dielectric tensor $\epsilon\in\sym{\RR^{3}}$ are given by (see [@kn:born pg. 678]) $$\begin{aligned} \nabla\times H-\dero t(\epsilon E) & = & 0,\\ \nabla\times E+\dero tH & = & 0,\\ \nabla\cdot(\epsilon E) & = & 0,\\ \nabla\cdot H & = & 0,\end{aligned}$$ where we have assumed the speed of light to be equal to one. These equations describe the behaviour of the electric and magnetic fields $E$ and $H$, in the medium under consideration. From now on we will always assume that the eigenvalues of $\epsilon$ are all strictly positive real numbers. If for constant vectors $E_{0},H_{0},\xi\in\RR^{3}$ and $\tau\in\RR$ we try to find planar wave solutions of the form $$\begin{aligned} E(x,t) & = & E_{0}\exp(i(<\text{\ensuremath{\xi}},x>-\tau t)),\\ H(x,t) & = & H_{0}\exp(i(<\xi,x>-\tau t)),\end{aligned}$$ we are lead to the equations $$Q_{\epsilon}(\xi)\left(\begin{array}{c} E_{0}\\ H_{0} \end{array}\right)=0,<\xi,\epsilon E_{0}>=0,<\xi,H_{0}>=0,\label{eq:PolarisationMaxwell}$$ where $$Q_{\epsilon}(\xi,\tau)=\left(\left(\begin{array}{cc} \epsilon & 0\\ 0 & I \end{array}\right)\tau+\left(\begin{array}{cc} 0 & P(\xi)\\ -P(\xi) & 0 \end{array}\right)\right)\label{eq:SymbolEpsilon}$$ and $P(\xi)$ is the anti-symmetric matrix such that $P(\xi)v=\xi\times v$ for any $v\in\RR^{3}$. For the first of the equations \[eq:PolarisationMaxwell\] to hold we need that $$\det(Q_{\epsilon}(\xi,\tau))=0.\label{eq:EikonalEquation}$$ We can assume without loss of generality that the principal axes of $\epsilon$ are aligned with the $x$,$y$ and $z$ axes in $\RR^{3}$, so that we can write $$\epsilon=\left(\begin{array}{ccc} \epsilon_{1} & 0 & 0\\ 0 & \epsilon_{2} & 0\\ 0 & 0 & \epsilon_{3} \end{array}\right).$$ A simple calculation shows that if we let $q_{\epsilon}(\xi,\tau)=\det(Q_{\epsilon}(\xi,\tau))/\tau^{2}$ (assuming $\tau\not=0)$ then we have that $$\begin{aligned} q_{\epsilon}(\xi,\tau) & = & (\xi_{1}^{2}+\xi_{2}^{2}+\xi_{3}^{3})(\epsilon_{1}\xi_{1}^{2}+\epsilon_{2}\xi_{2}^{2}+\epsilon_{3}\xi_{3}^{3})+\\ & + & (\epsilon_{1}\epsilon_{2}(\xi_{1}^{2}+\xi_{2}^{2})+\epsilon_{2}\epsilon_{3}(\xi_{2}^{2}+\xi_{3}^{2})+\epsilon_{1}\epsilon_{3}(\xi_{1}^{2}+\xi_{3}^{2}))\tau^{2}\\ & + & \epsilon_{1}\epsilon_{2}\epsilon_{3}\tau^{4}\end{aligned}$$ For $\xi\not=0$ and $\tau\not=0$ we can visualise condition \[eq:EikonalEquation\] if we consider the Fresnel surface $$\FH_{\epsilon}=\{(\tau\xi/||\xi||^{2})\in\RR^{3}|q_{\epsilon}(\text{\ensuremath{\xi},\ensuremath{\tau})}=0\}.\label{eq:FresnelSurface}$$ Physically, the Fresnel surface is the space of the allowable *phase velocities*. We can visualise $\FH_{\epsilon}$ by considering the following cases: 1. If $\epsilon_{1}=\epsilon_{2}=\epsilon_{3}\not=0$ then $$q_{\epsilon}(\xi,\tau)=\epsilon_{1}(\xi_{1}^{2}+\xi_{2}^{2}+\xi_{3}^{2}-\epsilon_{1}\tau^{2}),$$ and hence $\FH_{\epsilon}$ is a sphere of radius $\epsilon_{1}^{-1/2}$. 2. If exactly two if the $\epsilon_{i}$s are equal then $\FH_{\epsilon}$ consists of two smooth surfaces that intersect tangentially (see Figure \[fig:Uniaxial\]). 3. If all of $\epsilon_{i}$s are different from each other, then $\FH_{\epsilon}$ consists two singular surfaces (each having 4 singularities) intersecting at these singular points (see Figure \[fig:Biaxial\]). The existence of singularities in case $3$ accounts for the phenomena of conical refraction described by Hamilton (see [@kn:born pg. 681]) in which ray of light splits into a cone upon entering at certain directions of a biaxial crystal (e.g aragonite). ![\[fig:Uniaxial\]Fresnel surface for $\epsilon_{1}=\epsilon_{3}=3$ and $\epsilon_{2}=15$](3-15-3) ![\[fig:Biaxial\]Fresnel Surface for $\epsilon_{1}=1,\epsilon_{2}=2$ and $\epsilon_{3}=15$](1-2-15) Maxwell Equations and Symmetric Tensors on the Sphere ===================================================== In this section we will show that, for the case when the eigenvalues of the dielectric tensor are all different, the singularities of the associated Fresnel surface can be accounted purely from topological considerations. Throughout this section we will assume $F$ to be an orientable Riemannian real vector bundle of rank $2$ over an orientable closed surface $X$, and we will denote the metric in $F$ by $<,>_{F}$. We will let $\sym F$ stand for the bundle of symmetric endomorphism of $F$, i.e for $p\in X$ we have that $A\in\sym{F_{p}}$ iff $A$ is a linear map from $F_{p}$ to $F_{p}$ such that $$<Av,w>_{F}=<v,Aw>_{F}\hbox{\, for\, all\,\,}v,w\in F_{p},$$ For a section of $s$ of $\sym F$ we can consider its eigenvalue functions $\lambda_{s,1},\lambda_{s,2}:X\rightarrow\RR$, i.e for $p\in X$ we will let $\lambda_{s,1}(p)$ and $\lambda_{s,1}(p)$ be the two eigenvalues of $s(p)$ for all $p\in X$. From now on we will assume that we have chosen $\lambda_{s,1}$ and $\lambda_{s,2}$ so that $\lambda_{s,1}\leq\lambda_{s,2}$. Given a symmetric matrix $\epsilon\in\sym{\RR^{3}}$ we define $s_{\epsilon}:S^{2}\rightarrow\sym{(\tangent S^{2})}$ by letting $$<s_{\epsilon}(p)v,w>=<\epsilon v,w>\hbox{\,\, for\, all\,\,\,}p\in S^{k-1}\hbox{\,\, and\,\,}v,w\in\tangent_{p}S^{2}.$$ Observe that we are identifying $\tangent_{p}S^{2}$ with the subspace of $\RR^{3}$ consisting of all the vectors orthogonal to $p$, and that the metric we are using on $\tangent_{p}S^{2}$ is that induced from the ambient space $\RR^{3}$. To simplify notation we will let $$\lambda_{\epsilon,i}=\lambda_{s_{\epsilon},i}\hbox{\,\, for\,\,}i=1,2.$$ The following result expresses the Fresnel surface $\FH_{\epsilon}$ in terms of $\lambda_{\epsilon,1}$ and $\lambda_{\epsilon,2}$. If $\FH_{\epsilon}$ is the Fresnel surface associated with the dielectric tensor $\epsilon\in\sym{\RR^{3}}$, then we have that $$\FH_{\epsilon}=\bigcup_{i=1}^{2}\{\lambda_{\epsilon,i}^{-1/2}(\xi)\xi|\xi\in S^{2}\}.$$ As we have seen, we must have that $$Q_{\epsilon}(\xi,\tau)\left(\begin{array}{c} E_{0}\\ H_{0} \end{array}\right)=0,$$ where $Q_{\epsilon}$ is given by \[eq:SymbolEpsilon\]. The above formula is the same as the formulas $$\begin{aligned} \tau\epsilon E_{0}+\xi\times H_{0} & = & 0,\\ \tau H_{0}-\xi\times E_{0} & = & 0.\end{aligned}$$ From these last two equations we obtain that $$(\tau/||\xi||)^{2}\epsilon E_{0}-(\xi/||\xi||)\times((\xi/||\xi||\times E_{0})=0.$$ Since $\xi/||\xi||$ is a unitary vector, we have that the second term in the left hand side of the above equation is just the negative of the projection of $E_{0}$ onto the orthogonal space to $\xi$, and hence $$(\tau/||\xi||)^{2}<\epsilon E_{0},v>+<E_{0},v>=0\hbox{\,\, for\, all\,\,}v\hbox{\, shuch\, that\,}<v,\xi>=0.$$ The above equation means that $$(\sigma_{\epsilon}(\xi/||\xi||)+(||\xi||/\tau)^{2}I)(\pi_{\xi}E_{0})=0,$$ where $\pi_{\xi}:\RR^{3}\rightarrow\tangent_{\xi/||\xi||}S^{2}$ is projection onto the orthogonal complement of $\xi$, and $I$ is the identity morphism of the bundle $\sym{(\tangent S^{2})}$. Hence, we must have that $$\tau=||\xi||\lambda_{\epsilon,i}^{-1/2}(\xi/||\xi||)$$ where $\lambda_{i}:S^{2}\rightarrow\RR$ are the eigenvalue functions of $\sigma_{\epsilon}$. From this formula and the definition of $\FH_{\epsilon}$ (see \[eq:FresnelSurface\]) we obtain the desired result. Consider the case were $\epsilon$ is diagonal with eigenvalues $\epsilon_{1}=1,\epsilon_{2}=2$ and $\epsilon_{3}=15$, as shown in Figure \[fig:Biaxial\]. The surface at the bottom right corresponds to $\lambda_{\text{\ensuremath{\epsilon},1 }}$ and the one to the left $\lambda_{\epsilon,2}$. Observe that that these surfaces join at the points $\xi\in S^{2}$ where $\lambda_{\epsilon,1}(\xi)=\lambda_{\epsilon,2}(\xi)$, and exactly at these points are where the singularities of $\FH_{\epsilon}$ occur. The previous example motivates the importance of the following definition. If $s$ is a section of $\sym F$ we will say that $p\in X$ is a *multiple point* iff $\lambda_{s,1}(p)=\lambda_{s,2}(p)$, and we will denote the set of such points by $\mult_{s}$. Let $\symz F$ stand for traceless elements in $\sym F$. For a given section $s$ of $\sym F$ we can construct its traceless part (which is a section of $\symz F$) by letting $$s_{0}=s-\frac{1}{2}\tr(s)I_{F},$$ where $I_{F}$ is the identity section in $\sym F$. \[prop:MasZeros\]The set $\mult_{s}$ coincides with the set of zeros of the section $s_{0}$. Furthermore, if $s_{0}$ is transversal to the zero section of $\symz F$ then the eigenvalue function $\lambda_{s,1}$ and $\lambda_{s,2}$ have conic singularities on $\mult_{s}$. It is easy to see that for a section $s$ of $\sym F$ we have that $$\begin{aligned} \lambda_{s,1}(p) & = & \frac{1}{2}\tr(s(p))-||s_{0}(p)||\\ \lambda_{s,2}(p) & = & \frac{1}{2}\tr(s(p))+||s_{0}(p)||.\end{aligned}$$ where the norm used for an element $A\in\symz F_{p}$ is given by $||A||=\frac{1}{2}\tr(A^{2})$. From this it follows that $\lambda_{s,1}(p)=\lambda_{s,2}(p)$ iff $s_{0}(p)=0$. If $s_{0}$ is transversal to the zero section of $\symz F$ then near a zero of $s_{0}$ the norm function behaves like $(x,y)\map(x^{2}+y^{2})^{1/2}$, and hence we have conical singularities at these points. We are now ready to prove that four singularities of $\FH_{\epsilon}$ can be accounted from purely topological considerations. We will need the following result first. \[lem:S0FtoFF\]The bundle $\symz F$ is isomorphic to $F\otimes_{\CC}F$, as a real vector bundle. Since $F$ has a metric, we can consider it as an $\SO(2)$ bundle. For $$R=\left(\begin{array}{cc} \cos(\theta) & -\sin(\theta)\\ \sin(\theta) & \cos(\theta) \end{array}\right)\in\SO(2)$$ and $$A=\left(\begin{array}{cc} a & b\\ b & -a \end{array}\right)\in\symz{\RR^{2}}$$ we have that $$RAR^{T}=\left(\begin{array}{cc} p & q\\ q & -p \end{array}\right),$$ where $$\left(\begin{array}{c} p\\ q \end{array}\right)=\left(\begin{array}{cc} \cos(2\theta) & -\sin(2\theta)\\ \sin(2\theta) & \cos(2\theta) \end{array}\right)\left(\begin{array}{c} a\\ b \end{array}\right).$$ Hence, the transition functions of the bundle $\symz F$ are the same as those of the bundle $F\otimes_{\CC}F$. If $s$ is a section of $\sym F$ such that $s_{0}$ is transversal to the zero section of $\symz F$, then the cardinality of $\mult_{s}$ is at least $2\left|\int_{X}e(F)\right|$. From Proposition \[prop:MasZeros\] we have that the $\mult_{s}=Z_{0}$, where $Z_{0}$ is the set of zeros of $s_{0}$. We have that $$\int_{X}e(\symz F)=\sum_{p\in Z_{0}}i_{s_{0}}(p),$$ where $i_{s_{0}}(p)$ is the index of $s_{0}$ at $p$. From the transversality assumption we have that $i(p)=\pm1$, and from Lemma \[lem:S0FtoFF\] we have that $e(\symz F)=e(F\otimes_{\CC}F)=2e(F)$. We conclude that $$|Z_{0}|\geq2\left|\int e(F)\right|,$$ where $|Z_{0}|$ is the cardinality of $Z_{0}$. If $X$ has genus $g$ and $s:X\rightarrow\sym{(\tangent X)}$ is such that $s_{0}$ is transversal to the zero section of $\symz{(\tangent X)},$ then cardinality of $\mult_{s}$ is at least $4-4g$. Apply the above proposition and the formula $$\int_{X}e(\tangent X)=2-2g.$$ We know that the singularities of $\FH_{\epsilon}$ occur at the multiplicity set $\mult_{\epsilon}$ of the section $s_{\epsilon}$. And the from the above Corollary we know that the cardinality of $\mult_{\epsilon}$ must be greater or equal than $4$. This is accordance with the known fact that $\FH_{\epsilon}$ has exactly four singularities, but our argument has been purely topological. In this discussion we have implicitly assumed that $(s_{\epsilon})_{0}$ is transversal to the zero section of $\sym{(\tangent S^{2}})$. We can explicitly find the points in $\mult_{\epsilon}$ and verify that this is actually the case. \[prop:s\_epsilon\_0\_transversal\]If the the eigenvalues of $\epsilon\in\sym{\RR^{3}}$ are all different from each other, then the traceless part of $s_{\epsilon}$ is transversal to the zero section of $\symz{(\tangent S^{2})}$. We can assume that $\epsilon$ is diagonal with diagonal elements $0<\epsilon_{1}<\epsilon_{2}<\epsilon_{3}$. The zero set of $(s_{\epsilon})_{0}$ does not contain neither $(0,0,1)$ nor $(0,0,-1)$. We can use polar coordinates $$p(\theta,\varphi)=(\cos(\theta)\cos(\varphi),\sin(\theta)\cos(\varphi),\sin(\varphi))\hbox{\,\, for\,\,}0\leq\theta<2\pi,-\pi/2<\varphi<\pi/2,$$ and let $$\begin{aligned} u(\theta,\varphi) & = & (-\sin(\theta),\cos(\theta),0)\\ v(\theta,\varphi) & = & (-\cos(\text{\ensuremath{\theta})\ensuremath{\sin}(\ensuremath{\varphi}),-\ensuremath{\sin}(\ensuremath{\theta})\ensuremath{\sin}(\ensuremath{\theta}),\ensuremath{\cos}(\ensuremath{\varphi})). }\end{aligned}$$ The triple $(p,u,v)$ forms and orthonormal frame, so that locally we can write $$s_{\epsilon}=\left(\begin{array}{cc} <\epsilon u,u> & <\epsilon u,v>\\ <\epsilon v,u> & <\epsilon v,v> \end{array}\right)$$ If we identify $\symz{\RR^{2}}$ with $\RR^{2}$ by letting $$(a,b)\sim\left(\begin{array}{cc} a & b\\ b & -a \end{array}\right)$$ then we can write $$(s_{\epsilon})_{0}=\frac{1}{2}\left(<\epsilon u,u>-<\epsilon v,v>,2<\epsilon u,v>\right)$$ where $$\begin{aligned} 2<\epsilon u,v> & = & 2(\epsilon_{1}-\epsilon_{2})\cos(\theta)\sin(\theta)\sin(\varphi),\\ <\epsilon u,u>-<\epsilon v,v> & = & (\epsilon_{2}-\epsilon_{1}\sin^{2}(\varphi))\cos^{2}(\theta)+\\ & + & (\epsilon_{1}-\epsilon_{2}\sin^{2}(\varphi))\sin^{2}(\theta)-\epsilon_{3}\cos^{2}(\varphi).\end{aligned}$$ From this formula it is easy to see that the zeros of $(s_{\epsilon})_{0}$ occur at the points $(0,\varphi_{m}),(0,\varphi_{m}+\pi/2),(\pi,\varphi_{m})$ and $(0,\varphi_{m}+\pi/2)$, where $$\varphi_{m}=\arctan\left(\frac{\epsilon_{3}-\epsilon_{2}}{\epsilon_{2}-\epsilon_{1}}\right)^{\frac{1}{2}}.$$ A simple calculations shows that $$D(s_{\epsilon})_{0}(0,\varphi_{m})=(\epsilon_{2}-\epsilon_{1})^{3/2}(\epsilon_{3}-\epsilon_{2})(\epsilon_{3}-\epsilon_{1})^{-1/2}>0,$$ and hence $(s_{\epsilon})_{0}$ is transversal to the zero section of $\sym{(\tangent S^{2})}$ at that point. Due to the symmetry of the problem, the same condition holds at the other points. Desingularising the Fresnel Surface and Morse Theory ==================================================== We are interested in finding the critical points of the functions $\lambda_{\epsilon,1}$ and $\lambda_{\epsilon,2}$ described in the previous section. One reason why this problem is interesting is that the maximum and minimum are critical points of these functions, and they correspond to directions of maximal and minimal wave velocities. We want use a topological argument to find lower bounds to the number and types of critical points. To do this, we can apply Morse theory. Recall that if $f:X\rightarrow\RR$ is a smooth function and $p\in X$ is a critical point of $f$ (i.e $df(p)=0$) then we can define its hessian $d^{2}f(p)$, which is symmetric bilinear form in $\tangent_{p}X$. A critical point $p\in X$ of $f$ is said to be non-degenerate iff $d^{2}f(p)$ is non-degenerate, and in this case we define the index of $p$ as the dimension of the maximal subspace of $\tangent_{p}X$ at which $d^{2}f(p)$ is negative definite. Morse inequalities assert that if $f$ has only non-degenerate critical points the we have that (see [@kn:morse pg. 29]) $$\begin{aligned} C_{i}(f) & \geq & \dim(H^{i}(X,\RR))\\ \sum_{i=0}^{\dim(X)}(-1)^{i}\dim(H^{i}(X,\RR)) & = & \sum_{i=0}^{\dim(X)}(-1)^{i}C_{i}(f),\end{aligned}$$ where $C_{i}(f)$ are the number of critical points of $f$ of index $i$, and $H^{i}(X,\RR)$ is the cohomology group of $X$ with real coefficients. This formulas allow us to estimate the number of critical points of a given index in terms of topological invariants of $X$. We have seen in the previous section both $\lambda_{\epsilon,1}$ and $\lambda_{\epsilon,2}$ are non-smooth at $\mult_{\epsilon}$, so we cannot apply Morse Theory directly to them. To solve this problem we desingularise these functions as follows. For a section $s:X\rightarrow\sym F$ we define $$\EE_{s}=\bigcup_{p\in X}\{l_{x}\in\proj F|l_{p}\hbox{\, is\, spanned\, by\, an\, eigenvector\, of\,\,}s(p)\},$$ where $\proj F$ is the proyectivisation of $F$. We will refer to $\EE_{s}$ as the *space of eigenlines* of $s$. Over this space we can define the eigenvalue function of $s$ by $$\lambda_{s}(l_{p})=\hbox{eigenvalue\, of\,\,}s(p)\hbox{\, corresponding\, to\,\,}l_{p}.$$ It turns out that if $s_{0}$ is transversal to the zero section of $\symz F$ then we have that $\EE_{s}$ is a smooth submanifold of $\proj F$, and $\lambda_{s}$ is a smooth function (see [@kn:mocs Theorem 5]). Furthermore, the space $\EE_{s}$ is topologically equivalent to the space obtained by the following procedure (see [@kn:mocs Proposition 6]) 1. Let $X_{0}$ be obtained by removing small open disks, centred at the points of $\mult_{s}$, from $X.$ 2. The space $\EE_{s}$ is homeomorphic the the space obtained by joining two disjoint copies of $X_{0}$ along cylinders of the form $S^{1}\times(-\delta,\delta)$ (see Figure \[fig:Space of Eigenlines\]). ![\[fig:Space of Eigenlines\]The space $\EE_{s}$](blowup) If we restrict the projection map $\pi:\proj F\rightarrow X$ to $\EE_{s}$ we obtain a projection map $\pi_{s}:\EE_{s}\rightarrow X$ with the following properties (see [@kn:mocs Theorem 5]) 1. We have that $\pi_{s}^{-1}(X-\mult_{s})$ consists of two connected component $X_{1}$ and $X_{2}$ such that $\pi_{s}:X_{i}\rightarrow X-\mult_{s}$ is a diffeomorphism and $\lambda_{s,i}\circ\pi_{s}=\lambda_{s}$. 2. For any $p\in\mult_{s}$ we have that $\pi_{s}^{-1}(p)$ is diffeomorphic to a circle and $\lambda_{s}$ can have $0,1$ or $2$ critical points at $\pi^{-1}(p)$. Furthermore, if any of these critical points is non-degenerate it must have index equal to one. To simplify notation we define $$\EE_{\epsilon}=\EE_{s_{\epsilon}},\pi_{\epsilon}=\pi_{s_{\epsilon}}\hbox{\, and\,}\lambda_{\epsilon}=\lambda_{s_{\epsilon}}.$$ As an application of the above results, we have the following. If $\epsilon\in\sym{\RR^{3}}$ has eigenvalues $0<\epsilon_{1}<\epsilon_{2}<\epsilon_{3}$ then then space $\EE_{\epsilon}$ is a smooth surface of genus $3$ and $\lambda_{\epsilon}:\EE_{\epsilon}\rightarrow\RR$ is a smooth function having no critical points at $\pi_{\epsilon}^{-1}(\mult_{\epsilon}).$ The spaces $\EE_{\epsilon}$ is obtained joining two disjoint copies of $S^{2}-D_{1}\cup D_{2}\cup D_{3}\cup D_{4}$ along cylinders of the form $S^{1}\times(-\delta,\delta)$, where the $D_{i}$s are small open disks centred at the four points in $\mult_{\epsilon}$ (see Figure \[fig:Space of Eigenlines\]). From this it follows that $\EE_{\epsilon}$ is a surface of genus $3$. To check that $\lambda_{\epsilon}$ has no singularities in $\pi_{\epsilon}^{-1}(\mult_{\sigma})$ we write a explicit formula for $\lambda_{\epsilon}$ and check that this is actually the case for $0<\epsilon_{1}<\epsilon_{2}<\epsilon_{3}$. Consider polar coordinates on $S^{2}-\{(0,0,1),(0,0,-1)$ given by $$p(\theta,\varphi)=(\cos(\theta)\cos(\varphi),\sin(\theta)\cos(\varphi),\sin(\varphi))$$ and the orthonormal frame $\{u(\theta,\varphi),v(\theta,\varphi)\}$ in $\tangent(S^{2}-\{(0,0,1),(0,0,-1))$ given by $$\begin{aligned} u(\theta,\varphi) & = & (-\sin(\theta),\cos(\theta),0)\\ v(\theta,\varphi) & = & (-\cos(\text{\ensuremath{\theta})\ensuremath{\sin}(\ensuremath{\varphi}),-\ensuremath{\sin}(\ensuremath{\theta})\ensuremath{\sin}(\ensuremath{\theta}),\ensuremath{\cos}(\ensuremath{\varphi})). }\end{aligned}$$ If we let $\alpha$ be the angle of a line $l\in\proj{(\tangent S^{2})}$ with respect to $u(\theta,\varphi)$ then we can use $(\theta,\varphi,\alpha)$ as coordinates for $\proj{(\tangent S^{2})}$. In these coordinates we have that (see [@kn:mocs proof of Theorem 5 ] $$\begin{aligned} \EE_{\epsilon} & = & f_{\epsilon}^{-1}(0)\\ \lambda_{\epsilon} & = & g_{\epsilon}|\EE_{\epsilon}\end{aligned}$$ where $$\begin{aligned} f_{\epsilon}(\theta,\varphi,\alpha) & = & (\epsilon_{2}-\epsilon_{1})\cos(2\alpha)\cos(\text{\ensuremath{\theta})\ensuremath{\sin}(\ensuremath{\theta})\ensuremath{\sin}(}\varphi)\\ & + & \frac{1}{2}\sin(2\alpha)(\cos^{2}(\theta)(\epsilon_{2}-\epsilon_{1}\sin^{2}(\varphi))\\ & + & \frac{1}{2}\sin(2\alpha)(\sin^{2}(\theta)(\epsilon_{1}-\epsilon_{2}\sin^{2}(\varphi))-\epsilon_{3}\cos^{2}(\varphi))\\ g_{\epsilon}(\theta,\varphi,\alpha) & = & -\sin(2\alpha)\sin(\theta)\sin(\varphi)\\ & + & \frac{1}{2}(\epsilon_{1}\sin^{2}(\theta)+\epsilon_{2}\cos^{2}(\theta)+\epsilon_{3}\cos^{2}(\varphi)))\\ & + & \frac{1}{2}(\epsilon_{1}\cos^{2}(\theta)+\epsilon_{2}\sin^{2}(\theta))\sin^{2}(\varphi)\\ & + & \frac{1}{2}\cos(2\alpha)(\cos^{2}(\theta)(\epsilon_{2}-\epsilon_{1}\sin^{2}(\varphi))\\ & + & \frac{2}{3}\cos(2\alpha)(\sin^{2}(\theta)(\epsilon_{1}-\epsilon_{2}\sin^{2}(\varphi)-\epsilon_{\text{3}}\cos^{2}(\varphi))\end{aligned}$$ From this, we obtain that $$\begin{aligned} \nabla f_{\epsilon}(0,\varphi_{m},\alpha) & = & \left(\frac{(\epsilon_{1}-\epsilon_{2})(\epsilon_{3}-\epsilon_{2})^{1/2}}{(\epsilon_{3}-\epsilon_{1})^{1/2}}\cos(2\alpha),(\epsilon_{2}-\epsilon_{1})^{1/2}(\epsilon_{3}-\epsilon_{2})^{1/2}\sin(2\alpha),0\right)\\ \nabla g_{\epsilon}(0,\varphi_{m},\alpha) & = & \left(\frac{(\epsilon_{3}-\epsilon_{2})^{1/2}}{(\epsilon_{3}-\epsilon_{1})^{1/2}}\sin(2\alpha),(\epsilon_{2}-\epsilon_{1})^{1/2}(\epsilon_{3}-\epsilon_{2})^{1/2}(1-\cos(2\alpha),0\right).\end{aligned}$$ where $(0,\varphi_{m})$ corresponds to one of the zeros of $(s_{\epsilon})_{0}$ as in the proof of Proposition \[prop:s\_epsilon\_0\_transversal\] (by symmetry the other zeros of $(s_{\epsilon})_{0}$ are dealt in the same way). The critical points of $\lambda_{\epsilon}$ correspond to the $\alpha$s at which $\nabla f_{\epsilon}(0,\varphi_{m},\alpha)=\nabla g_{\epsilon}(0,\varphi_{m},\alpha)$. By using the above formulas we have that this equality can not hold if all the $\epsilon_{i}$s are different from each other. If $\epsilon\in\sym{\RR^{3}}$ has eigenvalues $0<\epsilon_{1}<\epsilon_{2}<\epsilon_{3}$ then we have that $C_{i}(\lambda_{\text{\ensuremath{\epsilon}}})=C_{i}(\lambda_{\epsilon,1})+C_{i}(\lambda_{\epsilon,2})$ for $i=0,1,2$. By the previous proposition we have that $\lambda_{\epsilon}$ has no critical points on $\pi_{\epsilon}^{-1}(\mult_{\epsilon})$. But we know that on each of the two connected components of $\EE_{\epsilon}-\pi_{\epsilon}^{-1}(\mult_{\epsilon})$ the function $\lambda_{\epsilon}$ coincides (up to diffeomorphism) with $\lambda_{\epsilon,1}$or $\lambda_{\epsilon,2}$. Applying Morse inequalities to $\lambda_{\epsilon}$ and using the above results we obtain that $$\begin{aligned} C_{0}(\lambda_{\epsilon,1})+C_{0}(\lambda_{\epsilon,2}) & \geq & 1\\ C_{1}(\lambda_{\epsilon,1})+C_{1}(\lambda_{\epsilon,2}) & \geq & 3\\ C_{2}(\lambda_{\epsilon,1})+C_{2}(\lambda_{\epsilon,2}) & \geq & 1\\ \sum_{i=0}^{2}C_{0}(\lambda_{\epsilon,1})+C_{0}(\lambda_{\epsilon,2}) & = & -4\end{aligned}$$ The above holds under the assumption that $\lambda_{\epsilon}$ is a Morse function, but it can easily be seen that is the case if the eigenvalues of $\epsilon$ are all different from each other. Generalisations - Future Work ============================= In this section we propose a program to generalise our results to general hyperbolic differential operators on bundles. Let $E$ and $F$ be vector bundles of rank $m$ and $k$ over a Riemannian manifold $X$ of dimension $n$. Consider a linear differential operator of degree $d\in\ZZ^{+}$ of the form $$L=\sum_{i=0}^{d}\left(L_{i}\circ\frac{\partial^{d-i}}{\partial t^{d-i}}\right),\label{eq:MainDiffOperator}$$ where $L_{i}:C^{\infty}(E)\rightarrow C^{\infty}(F)$ is a differential operator of order $i$. Observe that $L$ acts on time dependent sections of $E$. We are interested in *high frequency solutions* of the equation $Lu=0$, i.e in solutions to the asymptotic partial differential equation $$L\left(u_{0}e^{is\varphi}\right)=0\hbox{\, as\,\,}s\mapsto\infty,$$ where $u_{0}$ is a section of $E$ and $\varphi$ is a smooth real valued function in $X\times\RR$ (known in the physics literature as the *phase function*). The above asymptotic equation leads to the the equation (see [@kn:guillemin pg. 31]) $$\left(\sum_{i=0}^{d}\sigma_{i}(d_{x}\varphi(x,t))\left(\frac{\partial\varphi}{\partial t}(x,t)\right)^{d-i}\right)u_{0}(x)=0,\label{eq:PolarisationEquation}$$ where $\sigma_{i}:\cotangent X\rightarrow\hbox{Hom}(E,F)$ is the *principal symbol* of $L_{i}$. If we choose local coordinates for $X$ and local trivialisation of $E$ and $F$ over an open set $U\subset X,$ we can write $$L_{i}=\sum_{|\alpha|\leq i}A_{i,\alpha}\frac{\partial^{\alpha}}{\partial x^{\alpha}}$$ where $$\alpha=(\alpha_{1},\ldots,\alpha_{n}),|\alpha|=\alpha_{1}+\cdots+\alpha_{n}\hbox{\, and\,\,\,}\frac{\partial^{\alpha}}{\partial x^{\alpha}}=\frac{\partial^{\alpha_{1}+\cdots+\alpha_{n}}}{(\partial x_{1})^{\alpha_{1}}\cdots(\partial x_{n})^{\alpha_{n}}}$$ and for every $\alpha$ and $i$ we have that $A_{i,\alpha}$ is a function from $U$ to the $k\times m$ matrices with coefficients in $\RR$. The local expression for $\sigma_{i}$ is then given by $$\sigma_{i}(x,\xi)=\sum_{|\alpha|=i}A_{i,\alpha}(x)\xi^{\alpha}\hbox{\,\, where\,\,\,}\xi=\xi_{1}^{\alpha_{1}}\xi_{2}^{\alpha_{2}}\cdots\xi_{n}^{\alpha_{n}}.$$ We see that over the fibre variables of $\cotangent X$ the symbol $\sigma_{i}$ is an homogenous polynomial of degree $i$ whose coefficients are real $k\times m$ matrices. We define the *Fresnel hypersurface* associated to $L$ by $$\FH_{L}=\left\{ (x,-\tau\xi)\in\cotangent X\left|(x,\xi)\in S(\cotangent X)\hbox{\, and\,}\ker\left(\sum_{i=0}^{d}\sigma_{i}(x,\xi)\tau^{d-i}\right)\not=0\right.\right\} ,\label{eq:characteristicSet}$$ We will assume that the following conditions hold 1. We have that $F$ is a Riemannian real vector bundle and $E=F$ . For example, if we were modelling small oscillations of an elastic surface $X$ then we could assume that $E=F=\tangent X$. 2. For all $(x,\xi)\in\cotangent X$ and $1\leq i\leq d$ we have that $\sigma_{i}(x,\xi)$ is a symmetric operator with respect to a metric in $F$, i.e $$<\sigma_{i}(x,\xi)v,w>_{F}=<v,\sigma_{i}(x,\xi)w>_{F}\hbox{\,\, for\, all\,}v,w\in F_{x}.$$ It turns out that two conditions above hold in many cases of physical interest, and in particular those physical problems that arise from variational principles. Under the above assumptions we can write $$\FH_{L}=\{(x,-\tau\xi)\in S(\cotangent X)|p_{L}(x,\xi,\tau)=0\},$$ where $S(\cotangent X)$ are the unit vectors in $\cotangent X$ and $$p_{L}(x,\xi,\tau)=\det\left(\sum_{i=0}^{d}\sigma_{i}(x,\xi)\tau^{d-i}\right).$$ A polynomial $p=p(\tau)$ is said to be *hyperbolic* if it has as many real roots as its degree (counting multiplicities) and it is said to be *strictly hyperbolic* if all these roots are different from each other. For a fixed $(x,\xi)\in S(\cotangent X)$ let $$p_{L,x,\text{\ensuremath{\xi}}}(\tau)=p_{L}(x,\xi,\tau).$$ The operator $L$ is said to be hyperbolic if $p_{L,x,\xi}$ is hyperbolic for all $(x,\xi)\not=0$ in $\cotangent X$. If $L$ is hyperbolic we have functions $\lambda_{i}:\cotangent X\rightarrow\RR$ for $1\leq i\leq\dim(F)\cdot\hbox{deg}(L)$ given as as the roots of $p_{L,x,\xi}$, and we can write $$\FH_{L}=\bigcup_{i=1}^{d}\{-\lambda_{i}(\xi)\xi|\xi\in S(\cotangent X)\}.$$ We define $\mult_{L}$ as the set of points in $S(\cotangent X)$ at which two of more or the $\lambda_{i}$s coincide. The singularities of $\FH_{L}$ occur at points on the set $\mult_{L}$. Inspired by our work on Maxwells equations, we pose the following problems. 1. *Find topological obstructions to the condition* $\mult_{L}=\emptyset$. This will provide obstructions to the existence of strictly hyperbolic differential operators on $F$. It is expected that the characteristic classes of $F$ should be involved in the answer to this problem. 2. *Desingularisation of* $\FH_{L}$. As in the case of Maxwells equation, we would like to find a smooth space $\EE_{L}$ and a smooth function $\lambda_{L}:\EE_{L}\rightarrow\RR$ that contains all the information of the $\lambda_{L,i}$s. We would then expect to be able to apply Morse theory to $\lambda_{L}$ to obtain similar results as in the case of Maxwells equations. In relation with Problem 1 above we mention the work in the articles [@kn:braam; @kn:hormander; @kn:john; @kn:khesin; @kn:lax; @kn:nuij], which contain some local and global results related to the multiplicity of eigenvalues of symbols of differential and pseudo-differential operators. Regarding Problem 2, we mention that the non-smoothness of $\lambda_{L,i}$s at $\mult_{L}$ explain phenomena like wave transformation (see [@kn:arnold pg. 223]) in which longitudinal waves transforms into transversal ones (or viceversa).
{ "pile_set_name": "ArXiv" }
--- abstract: 'The objective of this paper is to study the monoid of all partial transformations of a finite set that preserve a uniform partition. In addition to proving that this monoid is a quotient of a wreath product with respect to a congruence relation, we show that it is generated by 5 generators, we compute its order and determine a presentation on a minimal generating set.' address: - 'Università degli Studi di Cagliari, Dipartimento di Matematica e Informatica, Via Ospedale, 72 - 09124 Cagliari, Italy; e-mail: [email protected]' - 'Departamento de Matemática, Faculdade de Ciências e Tecnologia, Universidade Nova de Lisboa, Monte da Caparica, 2829-516 Caparica, Portugal; also: Centro de Álgebra da Universidade de Lisboa, Av. Prof. Gama Pinto 2, 1649-003 Lisboa, Portugal; e-mail: [email protected]' - 'Centro de Álgebra da Universidade de Lisboa, Av. Prof. Gama Pinto 2, 1649-003 Lisboa, Portugal; e-mail: [email protected]' author: - 'Serena Cicalò, Vítor H. Fernandes and Csaba Schneider' title: Partial transformation monoids preserving a uniform partition --- Introduction {#intro} ============ The main objective of this paper is a study of the monoid of all partial transformations of a finite set that preserve a uniform equivalence relation. Given a set $\Omega$, a [partial transformation]{} on $\Omega$ is a map $\alpha:\Gamma\rightarrow \Omega$ where $\Gamma\subseteq \Omega$ allowing the case that $\Gamma=\emptyset$. The set of partial transformations of $\Omega$ is a monoid under the operation of composition. Given an equivalence relation $E$ on $\Omega$, the set of partial transformations that preserve $E$ is closed under composition, and hence it is a submonoid. In this paper we will investigate monoids of partial transformations preserving a uniform equivalence relation $E$; that is, an equivalence relation with the property that the equivalence classes have the same size. For $n{\geqslant}1$, the symbols $\PT_n$ and $\T_n$ denote the monoid of all partial transformations and the monoid of all transformations of the set $\underline n=\{1,\ldots,n\}$. The wreath product of two monoids $T$ and $S$ is denoted by $T\wr S$ (see Section \[notations\] for the definition). The rank of a monoid is the cardinality of a least-order generating set. The main results of this paper are collected in the following theorem. \[main\] Let $\Omega$ be a finite set, let $E$ be a uniform equivalence relation on $\Omega$ with $m$ equivalence classes each of which has size $n$, with $n, m{\geqslant}2$, and let $\PT_E$ be the monoid of all partial transformations on $\Omega$ that preserve $E$. Then 1. $|\PT_E|=(m(n+1)^n-m+1)^m$; 2. $\rank \PT_E=5$; 3. $\PT_E\cong (\PT_n\wr\T_m)/R$ where $R$ is a congruence relation on $\PT_n\wr\T_m$ generated by a single pair of elements. Parts (i) and (iii) of Theorem \[main\] are proved in Section \[wreathsec\], while part (ii) is verified in Section \[sectrank\]. In addition to proving the assertions in Theorem \[main\], we determine an explicit generating set for $\PT_E$ with 5 generators (Corollary \[5gen\]) and a presentation for $\PT_E$ satisfied by this generating set (Corollary \[relcor\]). The rank of a semigroup is an important invariant and the determination of the ranks of semigroups and monoids has been a focus of research in semigroup theory. Let $\Omega$ be a finite set with at least 2 elements. It is well-known that the full symmetric group of $\Omega$ is generated by 2 generators. (In the language of semigroup theory one might say that the rank of the symmetric group is 2, but, as the rank has a different meaning in permutation group theory, we avoid expressing this result this way.) Further, the monoid of all transformations and the monoid of all partial transformations of $\Omega$ have ranks $3$ and $4$, respectively; see [@GO Theorems 3.1.3 and 3.1.5]. The survey [@Fernandes:2002] collects these results and similar ones for other classes of transformation monoids, in particular, for monoids of order-preserving transformations and for some of their extensions. Let $\T_E$ denote the stabilizer of a non-trivial uniform equivalence relation $E$ in the monoid of transformations of a finite set $\Omega$. Huisheng [@Huisheng:2005a] proved that $\T_E$ is generated by its group of invertible transformations and two additional elements. In other words, the relative rank of $\T_E$ with respect to its group of units is 2. The group of units of $\T_E$ is isomorphic to a wreath product of two full symmetric groups and Huisheng observed that such a group can be generated by 4 elements. This led him to conclude that $\rank\T_E{\leqslant}6$. Later, Araújo and the third author improved this result by showing that the wreath product of two symmetric groups can always be generated by two elements, and hence $\rank\T_E=4$. Some natural submonoids of $\T_E$, such as those of the order-preserving or orientation-preserving transformations, have already been considered in , and their ranks were determined by the second author and Quinteiro . The wreath product construction in the class of monoids is one of the main tools of this paper. The full stabilizer in the symmetric group of a uniform equivalence relation is a wreath product of two smaller symmetric groups and the analogous result holds in the full transformation monoid (see ). One would expect that the monoid $\PT_E$ of all partial transformations preserving a uniform equivalence relation $E$ is a wreath product of two partial transformation monoids, but this is not the case. Nevertheless, $\PT_E$ can still be described in terms of wreath products. More precisely, $\PT_E$ is isomorphic to a quotient of the wreath product $\PT_n \wr \T_m$ where $n$ is the size of an equivalence class and $m$ is the number of classes of $E$. This quotient is taken modulo a well-described congruence relation $R$ (see Lemma \[l21\] for details). As claimed by Theorem \[main\](iii), the congruence $R$ is generated by a single pair of elements. The last part of the paper is concerned with determining a presentation of the monoid $\PT_E$ defined in the previous paragraph. Determining presentations of important semigroups, monoids, and groups has been a mainstream research topic since the end of the 19-th century. For instance, a presentation for the symmetric group $\S_n$ was given by Moore [@Moore:1897] in 1897. For the full transformation monoid $\T_n$, a presentation was given by Aĭzenštat [@Aizenstat:1958] in 1958 and, some years later, in 1961, Popova [@Popova:1961] established a presentation for the partial transformation monoid $\PT_n$. Over the past decades, several authors determined presentations for many other classes of monoids; see [@Ruskuc:1995] and the survey [@Fernandes:2002] on presentations of monoids related to order-preserving transformations. Finally we must mention the recent work of East and his collaborators [@East:2006; @EasdownEastFitzGerald:2008; @East:2010a; @East:2010b; @East:2011a; @East:2011b]. The paper is organized as follows. We set our notation concerning transformations, partial transformations and wreath products in Section \[notations\]. We establish the relation between $\PT_E$ and wreath products and prove parts (i) and (iii) of Theorem \[main\] in Section \[wreathsec\]. Then in Sections \[sectrank\] we verify Theorem \[main\](ii). Finally in Section \[pressect\] we determine a presentation of $\PT_E$ on a minimal generating set. Notation concerning partial transformations and wreath products {#notations} =============================================================== In this section we establish the notation concerning partial transformations and wreath products that will be used throughout the paper. Let $\Omega$ be a finite set. A [*partial transformation*]{} of $\Omega$ is a map $\alpha:\Gamma\rightarrow\Omega$ where $\Gamma\subseteq\Omega$. The set $\Gamma$ is the [*domain*]{} of $\alpha$ and is denoted by $\dom\alpha$, while the [*image*]{} of $\alpha$ is denoted by $\im\alpha$. If $\Gamma=\Omega$ then $\alpha$ is said to be a [*transformation*]{}. Sometimes we want to emphasize that a partial transformation [*is*]{} or [*is not*]{} a transformation and in such cases we write [*full transformation*]{} and [*strictly partial transformation*]{}, respectively. We allow the case that $\Gamma=\emptyset$ and the corresponding partial transformation is denoted by ${\varnothing}$. In this paper, transformations and partial transformations act on the right unless it is explicitly stated otherwise. That is, if $\alpha$ is a partial transformation on $\Omega$ and $i\in\Omega$, then the image of $i$ under $\alpha$ is denoted by $i\alpha$. The set of all partial transformations of $\Omega$ is a monoid under composition and is denoted by $\PT(\Omega)$. Given an equivalence relation $E$ on $\Omega$ we say that a partial transformation $\alpha$ [*preserves*]{} $E$ if for all $i, j\in\dom \alpha$ such that $(i,j)\in E$ we have that $(i\alpha,j\alpha)\in E$. The equivalence relation $E$ is said to be [*uniform*]{} if its equivalence classes have the same size. The equivalence relations $\Omega\times\Omega$ and $\{(i,i)\mid i\in\Omega\}$ are preserved by $\PT(\Omega)$, and these relations are said to be [*trivial*]{}. Hence an equivalence relation is non-trivial if and only if it has at least two classes and at least one class has size 2. For a natural number $n$, let $\underline n$ denote the set $\{1,\ldots,n\}$. The symbols $\PT_n$, $\T_n$, and $\S_n$ denote the monoid of partial transformations, the monoid of transformations, and the group of permutations, respectively, acting on $\underline n$. We represent a transformation $\alpha\in\T_n$ by the list $[1\alpha,\ldots,n\alpha]$ of images. If $\alpha$ is a partial transformation and $i\not\in\dom\alpha$ then we put ${\varnothing}$ in the place of $i\alpha$. A permutation is usually written as a product of cycles. The identity map in $\T_n$ will be written as $1$. Let us now review the concept of wreath products of monoids; see [@strh Section 1.2.2] for details. Let $m\in\N$, let $T$ be a submonoid of $\T_m$, and let $S$ be any monoid. The *wreath product* of $S$ by the transformation monoid $T$, denoted by $S\wr T$, is the semidirect product $S^m\sdp T$ with respect to the left action of $T$ on the $m$-fold direct power $S^m$ defined as $$\label{action} {\kern1pt}^\tau{\kern-.5pt}(s_1,s_2,\ldots,s_m)=(s_{1\tau},s_{2\tau},\ldots, s_{m\tau}),$$ for all $\tau\in T$ and $s_1,s_2,\ldots,s_m\in S$. Therefore, the wreath product $S\wr T$ is the monoid with underlying set $S^m\times T$ and multiplication defined by $$\begin{gathered} \label{product} (s_1,s_2,\ldots,s_m;\sigma)(t_1,t_2,\ldots,t_m;\tau)= ((s_1,s_2,\ldots,s_m){\kern1pt}^\sigma{\kern-.5pt}(t_1,t_2,\ldots,t_m);\sigma\tau)= \\(s_1t_{1\sigma},s_2t_{2\sigma},\ldots,s_mt_{m\sigma};\sigma\tau),\end{gathered}$$ for all $(s_1,s_2,\ldots,s_m;\sigma),(t_1,t_2,\ldots,t_m;\tau)\in S^m\times T$. An element $\beta$ of a wreath product $S\wr T$ can be written uniquely in the form $({\beta^{(1)}},\ldots, {\beta^{(m)}};{\overline{\beta}})$ where ${\beta^{(1)}},\ldots,{\beta^{(m)}}\in S$ and ${\overline{\beta}}\in T$. In the rest of the paper, for an element $\beta\in S\wr T$, these components of $\beta$ will be denoted by ${\beta^{(1)}},\ldots,{\beta^{(m)}},{\overline{\beta}}$. Wreath products and partial endomorphisms of a uniform partition {#wreathsec} ================================================================ Let $m, n\in\N$ and set $\Omega=\underline n\times\underline m$. Let $E$ denote the equivalence relation on $\Omega$ that is defined by the rule that $((i,j),(k,l))\in E$ if and only if $j=l$. Then $E$ has $m$ equivalence classes each of which has size $n$, and hence $E$ is uniform. Let $\PT_{n\times m}$, $\T_{n\times m}$, and $\S_{n\times m}$ denote the monoid of all partial transformations, the monoid of all transformations, and the group of all permutations on $\Omega$ preserving the equivalence relation $E$. An isomorphism between $\T_{n\times m}$ and the wreath product $\T_n\wr\T_m$ can easily be constructed as follows (see ). Let $\alpha\in \T_{n\times m}$. The transformation $\alpha$ induces a transformation ${\overline{\alpha}}$ on the set of equivalence classes of $E$. Since these equivalence classes are indexed by the elements of $\underline m$, the transformation $\overline\alpha$ can be viewed as an element of $\T_m$. Further, for each $j\in\underline m$, we can define a transformation ${\alpha^{(j)}}$ on $\underline n$ as follows: $i{\alpha^{(j)}}=k$ if $(i,j)\alpha=(k,j\overline\alpha)$. It is proved in that the map $\alpha\mapsto ({\alpha^{(1)}},\ldots,{\alpha^{(m)}};{\overline{\alpha}})$ is an isomorphism between the monoids $\T_{n\times m}$ and $\T_n\wr\T_m$. One would think that the monoids $\PT_{n\times m}$ may be isomorphic to a wreath product constructed from $\PT_n$ and $\T_m$, but unfortunately this is not the case. Nonetheless, the monoid $\PT_{n\times m}$ can be described as a quotient of the wreath product $\PT_n\wr\T_m$ as follows. We define a homomorphism $\varphi$ from $\PT_n\wr\T_m$ to $\PT_{n\times m}$ by defining a “partial action” of $\PT_n\wr\T_m$ on $\Omega=\underline n\times\underline m$. Let $\alpha\in \PT_n\wr\T_m$. We define $\alpha\varphi\in\PT_{n\times m}$ as follows. Let $$\label{domain} \dom\alpha\varphi=\{(i,j)\mid i\in\dom {\alpha^{(j)}}\}$$ and for $(i,j)\in\dom\alpha\varphi$ we set $(i,j)(\alpha\varphi)= (i{\alpha^{(j)}},j\overline\alpha)$. It is routine calculation to check that $\varphi:\PT_n\wr\T_m\rightarrow \PT_{n\times m}$ is a homomorphism. For $j\in\underline m$, let ${\varepsilon^{(j)}}$ denote the homomorphism $\PT_n\rightarrow \PT_n\wr\T_m$ defined by the rule $\alpha\mapsto(1,\ldots,1,\alpha,1,\ldots,1;1)$ where the non-trivial factor of the image appears in the $j$-th position. Further, let ${\overline{\varepsilon}}$ denote the homomorphism $\T_m\rightarrow \PT_n\wr\T_m$ mapping $\alpha\mapsto(1,\ldots,1;\alpha)$. The homomorphisms ${\varepsilon^{(j)}}$ and ${\overline{\varepsilon}}$ are clearly injective. \[l21\] (i) The homomorphism $\varphi$ defined above is surjective. \(ii) For $\alpha,\beta \in\PT_n\wr\T_m$, we have that $\alpha\varphi=\beta\varphi$ if and only if ${\alpha^{(j)}}={\beta^{(j)}}$ for all $j\in\underline m$ and $j{\overline{\alpha}}=j{\overline{\beta}}$ whenever ${\alpha^{(j)}}\neq {\varnothing}$. \(iii) Let $\tau$ denote the element $[2,2,3,\ldots,m]{\overline{\varepsilon}}$ of $\PT_n\wr\T_m$. Then the congruence relation $\ker\varphi$ is generated by the pair $({\varnothing}{\varepsilon^{(1)}},({\varnothing}{\varepsilon^{(1)}})\tau)$. \(i) Let $\alpha\in\PT_{n\times m}$. We define ${\alpha^{(1)}},\ldots,{\alpha^{(m)}}\in\PT_n$ and ${\overline{\alpha}}\in \T_m$ similarly as in the argument that $\T_n\wr\T_m\cong \T_{n\times m}$ before the lemma. The only difference is that, for some $j$, the restriction of $\alpha$ to the equivalence class $\{(i,j)\mid i\in \underline n\}$ may be equal to ${\varnothing}$. In this case we let $j{\overline{\alpha}}=1$. Then it is clear that $({\alpha^{(1)}},\ldots,{\alpha^{(m)}};{\overline{\alpha}})\varphi= \alpha$, and hence $\varphi$ is surjective. \(ii) Let us first assume that $\alpha,\beta\in\PT_n\wr\T_m$ such that $\alpha\varphi=\beta\varphi$. If ${\alpha^{(j)}}\neq{\beta^{(j)}}$, for some $j$, then there is $i\in\underline n$ such that $i{\alpha^{(j)}}\neq i{\beta^{(j)}}$ (including the possibility that either $i\not\in\dom\alpha$ or $i\not\in\dom\beta$). Then $(i,j)(\alpha\varphi)\neq(i,j)(\beta\varphi)$, which is a contradiction. Similarly, if $j\in\underline m$ is such that ${\alpha^{(j)}}\neq {\varnothing}$, but $j{\overline{\alpha}}\neq j{\overline{\beta}}$, then we have, for $i\in\dom{\alpha^{(j)}}$, that $(i,j)(\alpha\varphi)\neq(i,j)(\beta\varphi)$. Therefore we obtain a contradiction again. Thus one direction of assertion (ii) is valid. To show the other direction, let us now suppose that $\alpha, \beta\in\PT_n\wr\T_m$ such that ${\alpha^{(j)}}={\beta^{(j)}}$ for all $j\in\underline m$ and $j{\overline{\alpha}}=j{\overline{\beta}}$ whenever ${\alpha^{(j)}}\neq {\varnothing}$. Equation  implies that $\dom(\alpha\varphi)=\dom(\beta\varphi)$. Let $(i,j)\in\dom(\alpha\varphi)$. Then ${\alpha^{(j)}}\neq {\varnothing}$ and ${\beta^{(j)}}\neq {\varnothing}$ and so $$(i,j)(\alpha\varphi)=(i{\alpha^{(j)}},j{\overline{\alpha}})=(i{\beta^{(j)}},j{\overline{\beta}})=(i,j)(\beta\varphi).$$ Hence $\alpha\varphi=\beta\varphi$ as required. \(iii) Let $Q$ be the smallest congruence relation on the monoid $\PT_n\wr\T_m$ such that $({\varnothing}{\varepsilon^{(1)}},({\varnothing}{\varepsilon^{(1)}})\tau)\in Q$. By the description of $\ker\varphi$ in statement (ii), we have that $({\varnothing}{\varepsilon^{(1)}},({\varnothing}{\varepsilon^{(1)}})\tau)\in \ker\varphi$ and hence $Q\subseteq \ker\varphi$. It remains to show that $\ker\varphi\subseteq Q$. For $i, s\in\underline m$, let $\tau_{i,s}$ denote the transformation $[1,2,\ldots,i-1,s, i+1,i+2,\ldots,m]$. We claim that $$\label{qeq} ({\varnothing}{\varepsilon^{(i)}},({\varnothing}{\varepsilon^{(i)}})(\tau_{i,s}{\overline{\varepsilon}}))\in Q\quad \mbox{for all}\quad i, s\in\underline m.$$ First we verify this claim in the case when $i=1$. Let $s\in\{2,\ldots,m\}$. As $$((2\,s){\overline{\varepsilon}}) ({\varnothing}{\varepsilon^{(1)}})((2\,s){\overline{\varepsilon}})= {\varnothing}{\varepsilon^{(1)}}$$ while $$((2\,s){\overline{\varepsilon}}) ({\varnothing}{\varepsilon^{(1)}})\tau((2\,s){\overline{\varepsilon}})= ({\varnothing}{\varepsilon^{(1)}})(\tau_{1,s}{\overline{\varepsilon}}),$$ we find that $({\varnothing}{\varepsilon^{(1)}},({\varnothing}{\varepsilon^{(1)}})(\tau_{1,s}{\overline{\varepsilon}}))\in Q$. Now conjugating this pair with $(1\,i){\overline{\varepsilon}}$ gives . Suppose that $\alpha, \beta\in \PT_n\wr\T_m$ such that $(\alpha,\beta)\in \ker\varphi$. We are required to show that $(\alpha,\beta)\in Q$. We have, for all $j\in\underline m$, that ${\alpha^{(j)}}={\beta^{(j)}}$ and that $j\overline\alpha=j\overline\beta$ whenever ${\alpha^{(j)}}\neq{\varnothing}$. Let $I=\{j\mid j\in\underline m\mbox{ with } j\overline\alpha\neq j\overline\beta\}$. We show by induction on $|I|$ that $(\alpha,\beta)\in Q$. If $I=\emptyset$, then $\alpha=\beta$, and in this case the claim is clearly valid. Suppose that $I\neq\emptyset$ and that $(\alpha_1,\beta_1)\in \ker\varphi$ implies that $(\alpha_1,\beta_1)\in Q$ whenever $|\{j\mid j\overline\alpha_1\neq j\overline\beta_1\}|{\leqslant}|I|-1$. Let $j_0$ be the minimal element of $I$. After possibly replacing $\alpha$ and $\beta$ with $((1\,j_0){\overline{\varepsilon}})\alpha$ and $((1\,j_0){\overline{\varepsilon}})\beta$, we may assume that $j_0=1$. Thus we have that $1\overline\alpha=r$ and $1\overline\beta=s$ with $r\neq s$ and that ${\alpha^{(1)}}={\beta^{(1)}}={\varnothing}$. Suppose first that $s\in\im{\overline{\alpha}}$; that is, $s=u\overline\alpha$ with some $u\in\underline m$. Then $$({\varnothing}{\varepsilon^{(1)}})\alpha=\alpha$$ while $$({\varnothing}{\varepsilon^{(1)}})(\tau_{1,u}{\overline{\varepsilon}})\alpha= ({\alpha^{(1)}},{\alpha^{(2)}},\ldots,{\alpha^{(m)}};\overline\alpha')$$ where $1\overline\alpha'=s$ and $j\overline\alpha'=j\overline\alpha$ for all $j\in\{2,\ldots,m\}$. Hence $(\alpha, ({\alpha^{(1)}},{\alpha^{(2)}},\ldots,{\alpha^{(m)}};\overline\alpha'))\in Q$. Since $|\{j\mid j\overline\alpha'\neq j\overline\beta\}|{\leqslant}|I|-1$, we obtain from the induction hypothesis that $(\beta,({\alpha^{(1)}},{\alpha^{(2)}},\ldots,{\alpha^{(m)}};\overline\alpha'))\in Q$. Thus $(\alpha,\beta)\in Q$ as claimed. If $s\not\in\im\overline\alpha$, but $r\in\im\overline\beta$ then, interchanging the role of $\alpha$ and $\beta$, the same argument shows that $(\alpha,\beta)\in Q$. Suppose now that $s\not\in\im\overline\alpha$ and $r\not\in\im\overline\beta$. If there is some $j\in\underline m$ such that ${\alpha^{(j)}}\neq {\varnothing}$ then $j\overline\alpha=j\overline\beta$. Set $u=j\overline\alpha$ and let $\overline\alpha'=[u,2\overline\alpha,3\overline\alpha,\ldots,m\overline\alpha]$ and $\overline\beta'= [u,2\overline\beta,3\overline\beta,\ldots,m\overline\beta]$. By the argument above, we have that $(\alpha,({\alpha^{(1)}},\ldots,{\alpha^{(m)}};\overline\alpha'))\in Q$ and $(\beta,({\beta^{(1)}},\ldots,{\beta^{(m)}};\overline\beta'))\in Q$. By the induction hypothesis, $(({\alpha^{(1)}},\ldots,{\alpha^{(m)}};\overline\alpha'),({\beta^{(1)}},\ldots,{\beta^{(m)}};\overline\beta'))\in Q$ which implies that $(\alpha,\beta)\in Q$ in this case. Finally we prove the claim when ${\alpha^{(j)}}={\varnothing}$ for all $j\in \underline m$. In this case we are required to show that $(({\varnothing},\ldots,{\varnothing};\overline\alpha),({\varnothing},\ldots,{\varnothing};\overline\beta))\in Q$ for all $\overline\alpha, \overline\beta\in\T_m$. The elements $\alpha$ with the property that ${\alpha^{(j)}}={\varnothing}$ for all $j\in\underline m$ form a submonoid $T$ in $\PT_n\wr \T_m$ that is isomorphic to $\T_m$ via the isomorphism $\alpha\mapsto{\overline{\alpha}}$. The relation $Q_T=Q\cap (T\times T)$ is a congruence relation on $T$. Equation  implies that $$\left(\prod_{j\in\underline m}{\varnothing}{\varepsilon^{(j)}},\prod_{j\in\underline m}({\varnothing}{\varepsilon^{(j)}})({\tau_{j,1}}{\overline{\varepsilon}})\right)\in Q_T.$$ Now $\prod_{j\in\underline m}{\varnothing}{\varepsilon^{(j)}}=({\varnothing},\ldots,{\varnothing};1)$ while $\prod_{j\in\underline m}({\varnothing}{\varepsilon^{(j)}})({\tau_{j,1}}{\overline{\varepsilon}})= ({\varnothing},\ldots,{\varnothing};[1,\ldots,1])$. Therefore $(({\varnothing},\ldots,{\varnothing};1),({\varnothing},\ldots,{\varnothing};[1,\ldots,1]))\in Q_T$. Now [@GO Theorem 6.3.10] gives that $Q_T$ must be the universal congruence which shows that $(({\varnothing},\ldots,{\varnothing};\overline\alpha),({\varnothing},\ldots,{\varnothing};\overline\beta))\in Q_T$ for all $\overline\alpha, \overline\beta\in\T_m$. Therefore $(({\varnothing},\ldots,{\varnothing};\overline\alpha),({\varnothing},\ldots,{\varnothing};\overline\beta))\in Q$ for all $\overline\alpha, \overline\beta\in\T_m$ as required. Part (iii) of Lemma \[l21\] implies Theorem \[main\](iii). Next we determine the order of $\PT_{n\times m}$ by counting the equivalence classes of the relation $\ker\varphi$ defined in Lemma \[l21\]. Let $\alpha\in\PT_n\wr\T_m$. We say that $\alpha$ is in [*canonical form*]{} if ${\alpha^{(j)}}={\varnothing}$ implies that $j{\overline{\alpha}}=1$ for all $j\in\underline m$. \[ordlemma\] Every equivalence class of $\ker\varphi$ contains precisely one element in canonical form. Consequently, $$|\PT_{n\times m}|=(m(n+1)^n-m+1)^m.$$ Let $\alpha\in\PT_n\wr\T_m$. Then $\alpha= ({\alpha^{(1)}},\ldots,{\alpha^{(m)}};{\overline{\alpha}})$ where ${\alpha^{(1)}},\ldots,{\alpha^{(m)}} \in\PT_n$ and ${\overline{\alpha}}=\T_m$. For $j\in\underline m$, set ${\beta^{(j)}}={\alpha^{(j)}}$ and define ${\overline{\beta}}\in\T_m$ by the rule that $j{\overline{\beta}}=j{\overline{\alpha}}$ if ${\alpha^{(j)}}\neq {\varnothing}$ and $j{\overline{\beta}}=1$ otherwise. Then the element $\beta=({\beta^{(1)}},\ldots,{\beta^{(m)}};{\overline{\beta}})$ is in canonical form and $\alpha\varphi=\beta\varphi$. Hence every congruence class of $\ker\varphi$ contains an element in canonical form. Assume that $\alpha, \beta\in\PT_n\wr\T_m$ are in canonical form such that $\alpha\varphi=\beta\varphi$. We have by Lemma \[l21\] that ${\alpha^{(j)}}={\beta^{(j)}}$ for all $j\in\underline m$ and that $j{\overline{\alpha}}=j{\overline{\beta}}$ for all $j\in\underline m$ such that ${\alpha^{(j)}}\neq {\varnothing}$. On the other hand, if ${\alpha^{(j)}}={\varnothing}$ then, as $\alpha$ and $\beta$ are in canonical form, $j{\overline{\alpha}}=j{\overline{\beta}}=1$. Hence we obtain that $\alpha=\beta$ and so each of the congruence classes contains precisely one element in canonical form. To compute the order of $\PT_{n\times m}$, let us compute the number of elements in canonical form. For $r\in\underline m$, let $A_r$ denote the number of elements $\alpha$ in canonical form such that $|\{j\in\underline m\mid {\alpha^{(j)}}\neq {\varnothing}\}|=r$. Then $$A_r=\binom{m}r\left(|\PT_n|-1\right)^rm^r.$$ Thus $$|\PT_{n\times m}|=\sum_{r=0}^m\binom{m}r\left(|\PT_n|-1\right)^rm^r= \sum_{r=0}^{m}\binom{m}{r}(m(n+1)^n-m)^r= (m(n+1)^n-m+1)^m,$$ as claimed. The second assertion of Lemma \[ordlemma\] implies Theorem \[main\](i). The table below shows the order of the monoid $\PT_{n \times m}$ for small values of $m$ and $n$. [$$\begin{array}{r|rrrrrr} m&|\PT_m|& |\PT_{2\times m}|&|\PT_{3\times m}|&|\PT_{4\times m}|&|\PT_{5\times m}|\\ \hline 1& 2& 9& 64& 625& 7776\\ 2& 9& 289& 16129& 1560001&115856201\\ 3& 64& 15625& 6859000& 6570725617 &3150905752576\\ 4& 625& 1185921& 4097152081& 38875337230081&296120751810639601\\ 5& 7776& 115856201& 3150905752576& 296120751810639601&88798957515761812069376 \end{array}$$]{} The rank {#sectrank} ======== In this section we determine the ranks of $\PT_{n\times m}$ and $\PT_n\wr\T_m$ defined in Section \[wreathsec\]. \[rankmain\] If $m\ge2$ and $n\ge2$, then $\rank\PT_n\wr\T_m=\rank\PT_{n\times m}=5$. Before proving the theorem, we review some known facts about the ranks and generating sets of the monoids $\PT_n$, $\T_n$, $\T_n\wr\T_m$ and fix some notation. Let us consider the following elements of $\T_n\wr\T_m$: $$\begin{aligned} \xi_1&=\left\{\ \begin{array}{ll} ((1\;2){\varepsilon^{(2)}})((2\;3\cdots m){\overline{\varepsilon}}) & \mbox{if $m$ and $n$ are both even}\\ ((1\;2){\varepsilon^{(2)}})((1\;2\cdots m){\overline{\varepsilon}}) & \mbox{otherwise;} \end{array}\right.\\ \xi_2&=((1\;2\cdots n){\varepsilon^{(1)}})((1\;2){\overline{\varepsilon}});\\ \tau&= [2,2,3,\ldots,n]{\varepsilon^{(1)}};\\ \overline\tau&=[2,2,3,\ldots,m]{\overline{\varepsilon}}.\end{aligned}$$ Then the following lemma is valid. \[ranklemma\] 1. If $\gamma\in\PT_n$ then $\PT_n=\left<\T_n,\gamma\right>$ if and only if $\gamma\in\PT_n\setminus\T_n$ and $\rank\gamma=n-1$. 2. We have that $\rank\S_n\wr\S_m=2$ and $\rank\T_n\wr\T_m=4$. Further, $\S_n\wr\S_m=\left<\xi_1,\xi_2\right>$ and $\T_n\wr\T_m=\left<\xi_1,\xi_2,\tau,\overline\tau\right>$. 3. Let $P$ be a monoid of partial transformations of a finite set and let $G$ be the group of permutations in $P$. If $X$ is a generating set of $P$ then $X\cap G$ is a generating set of $G$. Statement (i) follows from [@GO Theorem 3.1.5]. Combining  gives statement (ii). (Note that the generators that correspond to $\tau$ and $\overline\tau$ appear in  in a slightly different form.) Finally statement (iii) can be easily proved using the argument in . Now we can prove Theorem \[rankmain\]. First we show that the monoids $\PT_{n\times m}$ and $\PT_n\wr\T_m$ are generated by 5 elements. Since, by Lemma \[l21\], $\PT_{n\times m}$ is a homomorphic image of $\PT_n\wr\T_m$, it suffices to show that $\PT_n\wr\T_m$ is generated by 5 elements. Let $\sigma_1=[{\varnothing},2,3,\ldots,n]$; then, by Lemma \[ranklemma\](i), $\PT_n=\left<\T_n,\sigma_1\right>$. Set $\sigma=\sigma_1{\varepsilon^{(1)}}=(\sigma_1,1,\ldots,1;1)$. We claim that $\PT_n\wr\T_m=\left<\T_n\wr\T_m,\sigma\right>$. Once this assertion is proved, the claim that $\PT_n\wr\T_m$ is generated by 5 elements will follow, as $\rank \T_n\wr\T_m=4$ (Lemma \[ranklemma\](ii)). Let us now prove that $\PT_n\wr\T_m=\left<\T_n\wr\T_m,\sigma\right>$. Set $W=\left<\T_n\wr\T_m,\sigma\right>$. First we note that $\T_m{\overline{\varepsilon}}{\leqslant}W$. The fact that $\PT_n=\left<\T_n,\sigma_1\right>$ implies that $\PT_n{\varepsilon^{(1)}}=\left<\T_n{\varepsilon^{(1)}},\sigma\right>$. As $\T_n{\varepsilon^{(1)}}{\leqslant}W$, we obtain that $\PT_n{\varepsilon^{(1)}}{\leqslant}W$. For $j{\geqslant}1$, it is a consequence of equation  that $((1\,j){\overline{\varepsilon}})(\PT_n{\varepsilon^{(1)}})((1\,j) {\overline{\varepsilon}})=\PT_n{\varepsilon^{(j)}}$. As $\PT_n{\varepsilon^{(1)}}{\leqslant}W$ and $\S_m{\overline{\varepsilon}}{\leqslant}W$, we find that $\PT_n{\varepsilon^{(j)}}{\leqslant}W$ for all $j\in\underline m$. Since the monoids $\T_m{\overline{\varepsilon}}$ and $\PT_n{\varepsilon^{(j)}}$ with $j\in\underline m$ are contained in $W$ and they generate $\PT_n\wr\T_m$, we obtain that $\PT_n\wr\T_m{\leqslant}W$. As clearly, $W{\leqslant}\PT_n\wr\T_m$ we conclude that $\PT_n\wr\T_m=W$, as claimed. Next we show that the monoids $\PT_{n\times m}$ and $\PT_n\wr\T_m$ cannot be generated by fewer than 5 elements. We note, as above, that $\PT_{n\times m}$ is a homomorphic image of $\PT_n\wr\T_m$, and we need only prove this for $\PT_{n\times m}$. Let $G$ denote the unit group of $\PT_{n\times m}$. Let $X$ be a subset of $\PT_{n\times m}$ such that $\PT_{n\times m}=\left<G,X\right>$. The group $G$, being isomorphic to $\S_n\wr\S_m$, is generated by 2 elements (Lemma \[ranklemma\](ii)), and hence, by Lemma \[ranklemma\](iii), it suffices to show that $|X|{\geqslant}3$. Since the rank of a product is not greater than the minimum of the ranks of its factors and elements of rank $nm$ form a subgroup, we obtain that $X$ must contain at least one element of rank $nm-1$. We claim that $X$ must in fact contain at least two elements of rank $nm-1$, namely one strictly partial transformation, and one full transformation. Assume by contradiction that $X$ only contains one element, $\alpha$ say, with rank $nm-1$. Then an element $\beta\in\PT_{n\times m}$ with rank $nm-1$ can be written as a product in $X$ and $\S_n\wr\S_m$ and such a product cannot contain a transformation with rank less than $nm-1$. Hence $\beta$ must be written as a product in $\S_n\wr\S_m$ and $\alpha$. Now, if $\alpha$ were a strictly partial transformation, then no full transformation $\beta$ with rank $nm-1$ could be written in this form, which is a contradiction. Similarly, if $\alpha$ were a full transformation, then no strictly partial transformation with rank $nm-1$ could be written as a product in $\S_n\wr\S_m$ and $\alpha$. Thus $X$ must contain at least two elements with rank $nm-1$, namely one full and one strictly partial, as claimed. Finally we prove that $X$ must contain a third element. Assume by contradiction that $X=\left<\alpha_1,\alpha_2\right>$. We have, by the previous paragraph, that $\rank\alpha_1=\rank\alpha_2=nm-1$ and we may assume without loss of generality that $\alpha_1$ is full and $\alpha_2$ is strictly partial. Let $\beta_1\in\T_m$ be an element of rank $m-1$ and set $\beta=\beta_1{\overline{\varepsilon}}\varphi$ where $\varphi:\PT_n\wr\T_m\rightarrow \PT_{n\times m}$ is the homomorphism defined before Lemma \[l21\]. By Lemma \[l21\], the elements $\alpha_1$, $\alpha_2$, and $\beta$ have unique preimages $\alpha_1'$, $\alpha_2'$, and $\beta'$ in $\PT_n\wr\T_m$, respectively, under the homomorphism $\varphi$. Since $\beta$ must be written as a word in $G$, $\alpha_1$, and $\alpha_2$, we have that $\beta'$ must be written as a word in $\S_n\wr\S_m$, $\alpha_1'$ and $\alpha_2'$. However, the map $\alpha\mapsto\overline\alpha$, defined before Lemma \[l21\], is a homomorphism $\PT_n\wr\T_m\rightarrow \T_m$ by equation . As $\overline{\alpha_1'}, \overline{\alpha_2'}\in\S_m$, we obtain that $\overline{\beta'}\in\S_m$. Since $\overline\beta'=\beta_1$ and, by assumption, $\rank\beta_1=m-1$, this is a contradiction. Thus $X$ must contain a third element as claimed. The last lemma implies Theorem \[main\](ii). Lemma \[ranklemma\](ii) and the argument of the proof of Theorem \[rankmain\] yield explicit generating sets for $\PT_n\wr\T_m$ and $\PT_{n\times m}$. Recall that the homomorphism $\varphi:\PT_n\wr\T_m\rightarrow \PT_{n\times m}$ was defined before Lemma \[l21\]. \[5gen\] Let $\xi_1$, $\xi_2$, $\tau$, $\overline\tau$ be the elements of $\T_n\wr\T_m$ defined before Lemma $\ref{ranklemma}$ and let $\sigma$ be the element defined in the proof of Theorem $\ref{rankmain}$. Then $\{\xi_1,\xi_2,\tau,\overline\tau,\sigma\}$ and $\{\xi_1\varphi,\xi_2\varphi,\tau\varphi,\overline\tau\varphi,\sigma\varphi\}$ are minimal generating sets for $\PT_n\wr\T_m$ and $\PT_{m\times n}$ respectively. Presentations for $\PT_n\wr\T_m$ and $\PT_{n\times m}$ {#pressect} ====================================================== Let $S$ be a monoid with generating set $X$ and let $X^*$ denote the free monoid on $X$. A word $w$ in $X$ can be considered as an element of $S$ and also as an element of $X^*$, but this will cause no confusion. Let $R$ denote the equivalence relation on $X^*$ defined by the rule that $(w_1,w_2)\in R$ if and only if $w_1=w_2$ holds in $S$. Then $R$ is in fact a congruence relation. Suppose, further, that ${\mathcal R}\subseteq X^*\times X^*$ such that the smallest congruence relation on $X^*$ containing ${\mathcal R}$ is $R$. In this case we say that ${\mathcal R}$ is a [*set of defining relations*]{} for $S$ with respect to the generating set $X$ and write that $S=\left<X\mid{\mathcal R}\right>$. The pair $\left<X\mid{\mathcal R}\right>$ is called a [*presentation*]{} for $S$. A relation $(w_1,w_2)\in{\mathcal R}$ is usually written as $w_1=w_2$. Let $n, m{\geqslant}2$. To discuss presentations of $\PT_n\wr\T_m$ and $\PT_{n\times m}$ let us fix generators of these monoids. As in Section \[wreathsec\], for $i\in\underline m$, ${\varepsilon^{(i)}}$ denotes the $i$-th coordinate embedding $\PT_n\rightarrow \PT_n\wr\T_m$ and ${\overline{\varepsilon}}$ denotes the embedding $\T_m\rightarrow \PT_n\wr\T_m$ into the $(m+1)$-th component. Let us consider the following elements of $\PT_n\wr\T_m$: $$\begin{aligned} \pi&=(1\,2){\varepsilon^{(1)}};\\ \varrho&=(1\,2\ldots n){\varepsilon^{(1)}};\\ \tau &=[2,2,3,4,\ldots,n]{\varepsilon^{(1)}};\\ \sigma &=[{\varnothing},2,3,\ldots,n]{\varepsilon^{(1)}};\\ \overline\pi&=(1\,2){\overline{\varepsilon}};\\ \overline\varrho&=(1\,2\ldots m){\overline{\varepsilon}};\\ \overline\tau&=[2,2,3,4,\ldots,m]{\overline{\varepsilon}}.\\\end{aligned}$$ We have that $\left<\pi,\varrho,\tau,\sigma\right>\cong \PT_n$ under the map ${\varepsilon^{(1)}}$ and that $\left<\overline\pi,\overline\varrho,\overline\tau\right>\cong\T_m$ under the embedding ${\overline{\varepsilon}}$. Hence we may identify $\PT_n$ with $\left<\pi,\varrho,\tau,\sigma\right>$ and $\T_m$ with $\left<\overline\pi,\overline\varrho,\overline\tau\right>$. Let ${\mathcal R}_P$ and ${\mathcal R}_T$ be sets of defining relations for $\PT_n$ and $\T_m$ with respect to the generating sets $\{\pi,\varrho,\tau,\sigma\}$ and $\{\overline\pi,\overline\varrho,\overline\tau\}$, respectively. Explicit expressions for ${\mathcal R}_T$ and ${\mathcal R}_P$ can be found in [@Fernandes:2002 Section 2]. Let $S$ and $Q$ be two monoids defined by the presentations $\langle X_S\mid {\mathcal R}_S\rangle$ and $\langle X_Q\mid {\mathcal R}_Q\rangle$, respectively. Let $S\sdp Q$ be a semidirect product of $S$ by $Q$ with respect to a left action of $Q$ on $S$. For each $a\in X_S$ and $b\in X_Q$, denote by ${\kern1pt}^b{\kern-.5pt}a$ a (fixed) word of $X_S^*$ that represents the image of $a$ under $b$ in $S$. Then Lavers proved in [@Lavers:1998 Corollary 2] (see also [@Araujo:2001]) that $$\label{semidir} S\rtimes Q=\langle X_S\cup X_Q\mid {\mathcal R}_S\cup{\mathcal R}_Q\cup\{ba=({\kern1pt}^b{\kern-.5pt}a)b\mid a\in X_S,b\in X_Q\} \rangle.$$ Using this result in the case when the left action is trivial, we obtain that $$\label{dirpr} S\times Q=\langle X_S\cup X_Q\mid {\mathcal R}_S\cup {\mathcal R}_Q\cup\{ ba=ab\mid a\in X_S,b\in X_Q\} \rangle.$$ Next define the following sets of relations in $\left\{\pi,\varrho,\tau,\sigma,\overline \pi,\overline \varrho,\overline \tau\right\}$: $$\begin{aligned} {\mathcal R}_1 =&\{ \overline \varrho^{m-j+1} u \overline \varrho^{m+j-k} v \overline \varrho^{k-1} = \overline \varrho^{m-k+1} v \overline \varrho^{m+k-j} u \overline \varrho^{j-1}\mid j, k\in\underline m,\ j<k,\ u, v\in\{\pi,\varrho,\tau,\sigma\}\};\\ {\mathcal R}_2=&\{\overline \pi \overline \varrho^{m-1} u \overline \varrho = u \overline \pi, \overline \pi \overline \varrho^{m-j+1} u \overline \varrho^{j-1} = \overline \varrho^{m-j+1} u \overline \varrho^{j-1} \overline \pi\mid 3{\leqslant}j{\leqslant}m,\ u\in\{\pi,\varrho,\tau,\sigma\}\};\\ {\mathcal R}_3=&\{\overline \tau u = \overline \tau, \overline \tau \overline \varrho^{m-1} u \overline \varrho = u\overline\varrho^{m-1}u\overline\varrho\overline \tau,\ \overline \tau \overline \varrho^{m-j+1} u \overline \varrho^{j-1} = \overline \varrho^{m-j+1} u \overline \varrho^{j-1} \overline \tau \mid \\ &3{\leqslant}j{\leqslant}m,\, u\in\{\pi,\varrho,\tau,\sigma\}\}.\end{aligned}$$ The next result states presentations for $\PT_n\wr\T_m$ and for $\PT_{n\times m}$ with respect to the generating sets $\left\{\pi,\varrho,\tau,\sigma,\overline \pi, \overline \varrho,\overline \tau\right\}$ and $\left\{\pi\varphi,\varrho\varphi,\tau\varphi,\sigma\varphi, \overline \pi\varphi,\overline \varrho\varphi, \overline \tau\varphi\right\}$, respectively, where $\varphi:\PT_n\wr\T_m\rightarrow \PT_{n\times m}$ is the epimorphism defined before Lemma \[l21\]. To simplify notation, we will omit “$\varphi$” from the description of the elements of $\PT_{n\times m}$, and hence we will consider an element of $\PT_n\wr\T_m$ as an element of $\PT_{n\times m}$. \[presth\] (i) $\PT_n\wr\T_m=\left<\pi,\varrho,\tau,\sigma,\overline \pi,\overline \varrho, \overline \tau\mid {\mathcal R}_P\cup{\mathcal R}_T\cup{\mathcal R}_1\cup{\mathcal R}_2\cup{\mathcal R}_3\right>;$ \(ii) $\PT_{n\times m}=\left<\pi,\varrho,\tau,\sigma,\overline \pi,\overline \varrho, \overline \tau\mid {\mathcal R}_P\cup{\mathcal R}_T\cup{\mathcal R}_1\cup{\mathcal R}_2\cup{\mathcal R}_3\cup\{(\varrho\sigma)^n=(\varrho\sigma)^n\overline\tau\}\right>$. \(i) By the definition of the wreath product, $$\label{wrdec} \PT_n\wr\T_m=\left((\PT_n){\varepsilon^{(1)}}\times\cdots\times (\PT_n){\varepsilon^{(m)}}\right)\rtimes(\T_m{\overline{\varepsilon}}).$$ For $i\in\{1,\ldots,m\}$ set $\pi_i$, $\varrho_i$, $\tau_i$ and $\sigma_i$ as $(1\,2){\varepsilon^{(i)}}$, $(1\,2\ldots n){\varepsilon^{(i)}}$, $[2,2,3,4,\ldots,n]{\varepsilon^{(i)}}$, and $[{\varnothing},2,3,\ldots,n]{\varepsilon^{(i)}}$, respectively. Set ${\mathcal G}_i=\{\pi_i, \varrho_i, \tau_i,\sigma_i\}$ and ${\overline{{\mathcal G}}}= \{\overline\pi, \overline\varrho, \overline\tau\}$. Then, by , the set ${\mathcal G}_1\cup\cdots\cup{\mathcal G}_m\cup{\overline{{\mathcal G}}}$ is a generating set of $\PT_n\wr\T_m$ and combining , and the definition of the left action of $\T_m$ on $(\PT_n)^m$ given in Section \[notations\], we obtain that $$\label{pres1} \PT_n\wr\T_m=\left<{\mathcal G}_1\cup\cdots\cup{\mathcal G}_m\cup{\overline{{\mathcal G}}}\mid {\mathcal R}_{1P}\cup\cdots\cup{\mathcal R}_{mP}\cup{\mathcal R}_T\cup{\mathcal R}\right>$$ where, for $i\in\{1,\ldots,m\}$ the set ${\mathcal R}_{iP}$ is the set of relations that we obtain from ${\mathcal R}_P$ by substituting $\pi_i$, $\varrho_i$, $\tau_i$ and $\sigma_i$ in the places of $\pi$, $\varrho$, $\tau$ and $\sigma$, respectively. Moreover, ${\mathcal R}$ is the set consisting of the following relations: $$\label{eqc} uv=vu\quad u\in\ {\mathcal G}_i,\ v\in {\mathcal G}_j\mbox{ with }i<j;$$ $$\begin{aligned} \label{eqp1}\overline\pi\pi_1&=\pi_2\overline\pi,\ &\overline\pi\varrho_1&=\varrho_2\overline\pi,\ &\overline\pi\tau_1&=\tau_2\overline\pi,\ &\overline\pi\sigma_1&=\sigma_2\overline\pi;&\\ \label{eqp2}\overline\pi\pi_2&=\pi_1\overline\pi,\ &\overline\pi\varrho_2&=\varrho_1\overline\pi,\ &\overline\pi\tau_2&=\tau_1\overline\pi,\ &\overline\pi\sigma_2&=\sigma_1\overline\pi;&\\ \label{eqp3}\overline\pi\pi_i&=\pi_i\overline\pi,\ &\overline\pi\varrho_i&=\varrho_i\overline\pi,\ &\overline\pi\tau_i&=\tau_i\overline\pi,\ &\overline\pi\sigma_i&=\sigma_i\overline\pi& 3{\leqslant}i{\leqslant}m;\\ \label{eqr1}\overline\varrho\pi_i&=\pi_{i-1}\overline\varrho,\ &\overline\varrho\varrho_i&=\varrho_{i-1}\overline\varrho,\ &\overline\varrho\tau_i&=\tau_{i-1}\overline\varrho,\ &\overline\varrho\sigma_i&=\sigma_{i-1}\overline\varrho& 2{\leqslant}i{\leqslant}m;\\ \label{eqr2}\overline\varrho\pi_1&=\pi_{m}\overline\varrho,\ &\overline\varrho\varrho_1&=\varrho_{m}\overline\varrho,\ &\overline\varrho\tau_1&=\tau_{m}\overline\varrho,\ &\overline\varrho\sigma_1&=\sigma_{m}\overline\varrho;&\\ \label{eqt1}\overline\tau\pi_1&=\overline\tau,\ &\overline\tau\varrho_1&=\overline\tau,\ &\overline\tau\tau_1&=\overline\tau,\ &\overline\tau\sigma_1&=\overline\tau;&\\ \label{eqt2}\overline\tau\pi_2&=\pi_1\pi_1\overline\tau,\ &\overline\tau\varrho_2&=\varrho_1\varrho_2\overline\tau,\ &\overline\tau\tau_2&=\tau_1\tau_2\overline\tau,\ &\overline\tau\sigma_1&=\sigma_1\sigma_2\overline\tau;&\\ \label{eqt3}\overline\tau\pi_i&=\pi_i\overline\tau,\ &\overline\tau\varrho_i&=\varrho_i\overline\tau,\ &\overline\tau\tau_i&=\tau_i\overline\tau,\ &\overline\tau\sigma_i&=\sigma_i\overline\tau&3{\leqslant}i{\leqslant}m.\end{aligned}$$ Note that $\pi_1=\pi$, $\varrho_1=\varrho$, $\tau_1=\tau$, and $\sigma_1=\sigma$. Further, $i{\geqslant}2$, we have that $\pi_i={\overline{\varrho}}^{m-i+1}\pi{\overline{\varrho}}^{i-1}$ and we have analogous expressions for $\varrho_i$, $\tau_i$, and $\sigma_i$. Hence for $i{\geqslant}2$, we can remove the elements of ${\mathcal G}_i$ from the generating set given in  and we can also remove the relation sets ${\mathcal R}_{iP}$. Further we replace the occurrences in ${\mathcal R}$ of the generators belonging to ${\mathcal G}_i$ by the corresponding expressions in $\pi$, $\varrho$, $\tau$ and $\sigma$. This way the relations in – will transform into the trivial relation. As $\pi^2=1$ follows from the relations in ${\mathcal R}_P$, the relations in  can be removed, as they follow from the relations in . This way, the relations in  lead to the relations in ${\mathcal R}_1$, the relations in  and  give the relations in ${\mathcal R}_2$, and finally the relations in – will result in the relation set ${\mathcal R}_3$. This shows that the presentation in part (i) is valid. \(ii) As $(\varrho\sigma)^n={\varnothing}$, the second statement follows from part (i) and Lemma \[l21\](iii). Theorem \[presth\] gives a presentation of $\PT_n\wr\T_m$ and $\PT_{n\times m}$ in terms of 7 generators, even though we proved in Theorem \[rankmain\] that these monoids are generated by 5 generators. In order to pass to the 5-element generating set, we replace the 4-generator set $\{\pi,\varrho,\overline\pi,\overline\varrho\}$ of $\S_n\wr\S_m$ used in Theorem \[presth\] with the generating set $\{\xi_1,\xi_2\}$ given at the beginning of Section \[sectrank\]. In order two write down a presentation that is satisfied by the generating set $\{\xi_1,\xi_2,\tau,\sigma,\overline\tau\}$, we need to express the generators $\pi,\varrho,\overline\pi,\overline\varrho$ in terms of the generators $\xi_1$ and $\xi_2$. We do this in the next lemma, whose proof, though rather technical, is routine, and so we omit the details. We note that the following lemma also implies  that $\S_n\wr\S_m$ is generated by two elements. Let $m, n{\geqslant}2$. If $m$ and $n$ are both even then $$\begin{aligned} \label{eq1}\pi&=(\xi_1^{m-1}\xi_2^2)^{(m-2)(n-1)^2} (\xi_1\xi_2^2)^{(m-1)(n^2-n-1)} (\xi_1\xi_2)^m; \\ \label{eq2}\varrho&=\left( (\xi_1^{m-1}\xi_2^2)^{(m-2)(n-1)^2} (\xi_1\xi_2^2)^{(m-1)(n^2-n-1)} \right)^{n-1}; \\ \label{eq3}\overline\pi&=\xi_2^{2n-1} \left( (\xi_1^{m-1}\xi_2^2)^{(m-2)(n-1)^2} (\xi_1\xi_2^2)^{(m-1)(n^2-n-1)} \right)^{n-1};\\ \label{eq4}\overline\varrho&=\xi_2^{2n-1} \left( (\xi_1^{m-1}\xi_2^2)^{(m-2)(n-1)^2} (\xi_1\xi_2^2)^{(m-1)(n^2-n-1)} \right)^n (\xi_1\xi_2)^m \xi_2^{2n-1} \xi_1 \xi_2. \end{aligned}$$ If either $m$ or $n$ is odd then $$\begin{aligned} \label{eq5}\pi&=(\xi_1\xi_2^{2n-1})^{(m-1)n} (\xi_1^{m+1}\xi_2^{2n-1})^{(m-1)(n-1)} \xi_1^m;\\ \label{eq6}\varrho&=(\xi_1\xi_2^{2n-1})^{n-1} (\xi_1^{m+1}\xi_2^{2n-1})^{(m-2)(n-1)};\\ \label{eq7}\overline \pi&=\xi_1^{2m-1} (\xi_1\xi_2^{2n-1})^n (\xi_1^{m+1}\xi_2^{2n-1})^{(m-2)(n-1)};\\ \label{eq8}\overline \varrho&=\xi_1^{2m-1} (\xi_1\xi_2^{2n-1})^{(m-1)n} (\xi_1^{m+1}\xi_2^{2n-1})^{(m-1)(n-1)} \xi_1^{m+2}.\end{aligned}$$ Suppose that $w_1$, $w_2$, $w_3$ and $w_4$ are the words on the right hand sides of the equations –, when both $m$ and $n$ are even, or of the equations – otherwise, respectively. Let $\overline {\mathcal R}_P$, $\overline {\mathcal R}_T$, $\overline {\mathcal R}_1$, $\overline {\mathcal R}_2$ and $\overline {\mathcal R}_3$ the sets of relations that are constructed by substituting in ${\mathcal R}_P$, ${\mathcal R}_T$, ${\mathcal R}_1$, ${\mathcal R}_2$ and ${\mathcal R}_3$ the words $w_1$, $w_2$, $w_3$, $w_4$ in the places of $\pi$, $\varrho$, $\overline\pi$, $\overline\varrho$, respectively. Further, let $\overline r$ be the relation obtained by substituting $w_2$ in the relation $(\varrho\sigma)^n=(\varrho\sigma)^n\overline\tau$ into the place of $\varrho$. Then we obtain the following main result of this section. \[relcor\] (i) $\PT_n\wr\T_m=\left<\xi_1,\xi_2,\tau,\sigma,\overline \tau\mid \overline {\mathcal R}_P\cup\overline{\mathcal R}_T\cup\overline {\mathcal R}_1\cup\overline{\mathcal R}_2\cup\overline {\mathcal R}_3\right>$; \(ii) $\PT_{n\times m}=\left<\xi_1,\xi_2,\tau,\sigma,\overline \tau\mid \overline {\mathcal R}_P\cup\overline{\mathcal R}_T\cup\overline {\mathcal R}_1\cup\overline{\mathcal R}_2\cup\overline {\mathcal R}_3\cup\{\overline r\}\right>$. Acknowledgement {#acknowledgement .unnumbered} =============== Cicalò gratefully acknowledges the support of FCT and PIDDAC, within the project PTDC/MAT/69514/2006 of CAUL and the support of the Regione Sardegna, within the project Master & Back, PR-MAB-A2009-837. She also thanks the Department of Mathematics of the University of Trento for its hospitality from 2011. Fernandes gratefully acknowledges support of FCT and PIDDAC, within the projects ISFL-1-143 and PTDC/MAT/69514/2006 of CAUL. Schneider was supported by the FCT project PTDC/MAT/101993/2008. [Eas11b]{} A. Ja. A[ĭ]{}zen[š]{}tat. Defining relations of finite symmetric semigroups. , 45 (87):261–280, 1958. Isabel M. Araújo. . PhD thesis, University of St Andrews, 2001. Jo[ã]{}o Ara[ú]{}jo and Csaba Schneider. The rank of the endomorphism monoid of a uniform partition. , 78(3):498–510, 2009. James East. A presentation of the singular part of the symmetric inverse monoid. , 34(5):1671–1689, 2006. James East. A presentation for the singular part of the full transformation semigroup. , 81(2):357–379, 2010. James East. Presentations for singular subsemigroups of the partial transformation semigroup. , 20(1):1–25, 2010. James East. Generators and relations for partition monoids and algebras. , 339:1–26, 2011. James East. On the singular part of the partition monoid. , 21(1-2):147–178, 2011. David Easdown, James East, and D. G. FitzGerald. A presentation of the dual symmetric inverse monoid. , 18(2):357–374, 2008. V[í]{}tor H. Fernandes. Presentations for some monoids of partial transformations on a finite chain: a survey. In [*Semigroups, algorithms, automata and languages ([C]{}oimbra, 2001)*]{}, pages 363–378. World Sci. Publ., River Edge, NJ, 2002. Vítor H. Fernandes and Teresa M. Quinteiro. On the ranks of certain monoids of transformations that preserve a uniform partition. submitted. V[í]{}tor H. Fernandes and Teresa M. Quinteiro. On the monoids of transformations that preserve the order and a uniform partition. , 39(8):2798–2815, 2011. Vítor H. Fernandes and Teresa M. Quinteiro. The cardinal of various monoids of transformations that preserve a uniform partition. , 35(4):885–896, 2012. Olexandr Ganyushkin and Volodymyr Mazorchuk. , volume 9 of [ *Algebra and Applications*]{}. Springer-Verlag London Ltd., London, 2009. An introduction. Pei Huisheng and Zou Dingyu. Green’s equivalences on semigroups of transformations preserving order and an equivalence relation. , 71(2):241–251, 2005. Pei Huisheng. On the rank of the semigroup [$T\sb E(X)$]{}. , 70(1):107–117, 2005. T. G. Lavers. Presentations of general products of monoids. , 204(2):733–741, 1998. Eliakim Hastings Moore. Concerning the abstract groups of order $k!$ and $\mbox{\textonehalf} k!$ holohedrically isomorphic with the symmetric and the alternating substitution-groups on $k$ letters. , 28(1):357–367, 1896. L. M. Popova. Defining relations of a semigroup of partial endomorphisms of a finite linearly ordered set. , 238:78–88, 1962. John Rhodes and Benjamin Steinberg. . Springer Monographs in Mathematics. Springer, New York, 2009. Nikola Ru[š]{}kuc. . PhD thesis, University of St Andrews, 1995. Lei Sun, Huisheng Pei, and Zhengxing Cheng. Regularity and [G]{}reen’s relations for semigroups of transformations preserving orientation and an equivalence. , 74(3):473–486, 2007.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We address the problem of video representation learning without human-annotated labels. While previous efforts address the problem by designing novel self-supervised tasks using video data, the learned features are merely on a frame-by-frame basis, which are not applicable to many video analytic tasks where spatio-temporal features are prevailing. In this paper we propose a novel self-supervised approach to learn spatio-temporal features for video representation. Inspired by the success of two-stream approaches in video classification, we propose to learn visual features by regressing both motion and appearance statistics along spatial and temporal dimensions, given only the input video data. Specifically, we extract statistical concepts (fast-motion region and the corresponding dominant direction, spatio-temporal color diversity, dominant color, ) from simple patterns in both spatial and temporal domains. Unlike prior puzzles that are even hard for humans to solve, the proposed approach is consistent with human inherent visual habits and therefore easy to answer. We conduct extensive experiments with C3D to validate the effectiveness of our proposed approach. The experiments show that our approach can significantly improve the performance of C3D when applied to video classification tasks. Code is available at <https://github.com/laura-wang/video_repres_mas>.' author: - | Jiangliu Wang$^1 \thanks{Work done during an internship at Tencent AI Lab.}$     Jianbo Jiao$^{2\dagger}$     Linchao Bao$^3 \thanks{Corresponding authors.}$     Shengfeng He$^4$     Yunhui Liu$^1$     Wei Liu$^{3*}$\ $^1$The Chinese University of Hong Kong $^2$University of Oxford\ $^3$Tencent AI Lab$^4$South China University of Technology title: | Self-supervised Spatio-temporal Representation Learning for Videos\ by Predicting Motion and Appearance Statistics --- Introduction ============ Learning powerful spatio-temporal representations is the most fundamental deep learning problem for many video understanding tasks such as action recognition [@carreira2017quo; @jiang2015human; @liu2016video], action proposal and localization [@chao2018rethinking; @shou2017cdc; @shou2016temporal], video captioning [@wang2018reconstruction; @wang2018bidirectional],   Great progresses have been made by training expressive networks with massive human-annotated video data [@tran2015learning; @r2plus1d_cvpr18]. However, annotating video data is very laborious and expensive, which makes the learning from unlabeled video data important and interesting. ![The main idea of the proposed approach. Given a video sequence, we design a novel task to predict several numerical labels derived from motion and appearance statistics for spatio-temporal representation learning, in a self-supervised manner. Each video frame is first divided into several spatial regions using different partitioning patterns like the grid shown above. Then the derived statistical labels, such as *the region with the largest motion and its direction* (the red patch), *the most diverged region in appearance and its dominant color* (the yellow patch), and *the most stable region in appearance and its dominant color* (the blue patch), are employed as supervision during the learning.[]{data-label="fig:teas"}](image/teas){width="\columnwidth"} Recently, several approaches [@misra2016shuffle; @fernando2017self; @lee2017unsupervised; @gan2018geometry] have emerged to learn transferable representations for video recognition tasks with unlabeled video data. In these approaches, a CNN is first pre-trained on unlabeled video data using novel self-supervised tasks, where supervision signals can be easily derived from input data without human labors, such as solving puzzles with perturbed video frame orders [@misra2016shuffle; @fernando2017self; @lee2017unsupervised] or predicting flow fields or disparity maps obtained with other computational approaches [@gan2018geometry]. Then the learned representations can be directly applied to other video tasks as features, or be employed as initialization during succeeding supervised learning. Unfortunately, although these work demonstrated the effectiveness of self-supervised representation learning with unlabeled videos, their approaches are only applicable to a CNN that accepts one or two frames as inputs, which is not a recommended way for tackling video tasks. In most video understanding tasks, spatio-temporal features that can capture information of both appearances and motions are proved to be vital in many recent studies [@cao2013mining; @simonyan2014two; @tran2015learning; @carreira2017quo; @r2plus1d_cvpr18]. In order to extract spatio-temporal features, a network architecture that can accept multiple frames as inputs and perform operations along both spatial and temporal dimensions is needed. For example, the popular C3D network [@tran2015learning], which accepts 16 frames as inputs and employs 3D convolutions along both spatial and temporal dimensions to extract features, is becoming more and more popular for many video tasks [@shou2017cdc; @shou2016temporal; @krishna2017dense; @li2018jointly; @wang2018bidirectional]. Vondrick [@vondrick2016generating] proposed to address the representation learning by C3D-based networks, while motion and appearance are not explicitly incorporated thus the performance is not satisfactory when transferring the learned features to other video tasks. In this paper, we propose a novel self-supervised learning approach to learn spatio-temporal video representations by predicting motion and appearance statistics in unlabeled videos. The idea is inspired by Giese and Poggio’s work on human visual system [@giese2003cognitive], in which the representation of motion is found to be based on a set of learned patterns. These patterns are encoded as sequences of ‘snapshots’ of body shapes by neurons in the *form pathway*, and by sequences of complex optic flow patterns in the *motion pathway*. In our work, the two pathways are the appearance branch and motion branch respectively. Besides, the abstract statistical concepts are also inspired by the biological hierarchical perception mechanism. The main idea of our approach is shown in Figure \[fig:teas\]. We design several spatial partitioning patterns to encode each spatial location and its motion and appearance statistics over multiple frames, and use the encoded vectors as supervision signals to train the spatio-temporal representation network. The novel objectives are simple to learn and informative for the motion and appearance distributions in video, , the spatial locations of the most dominant motions and their directions, the most consistent and the most diverse colors over a certain temporal cube,   We conduct extensive experiments with C3D network to validate the effectiveness of the proposed approach. We show that, compared with training from scratch, pre-training C3D without labels using our proposed approach gives a large boost to the performance of the action recognition task (, $45.4\%$ *v.s.* $61.2\%$ on UCF101). By transferring the learned representations to other video tasks on smaller datasets, we demonstrate significant performance gains on various tasks like dynamic scene recognition, action similarity labeling, Related work ============ Self-supervised representation learning is proposed to leverage the huge amounts of unlabeled data to learn useful representations for various problems, for example, image classification, object detection, video recognition,   It has been proved that lots of deep learning methods can benefit from pre-trained models on large labeled datasets, , ImageNet [@deng2009imagenet] for image tasks and Kinetics [@kay2017kinetics] or Sports-1M [@karpathy2014large] for video tasks. The basic motivation behind self-supervised representation learning is to replace the expensive labeled data with “free” unlabeled data. A common way to achieve self-supervised learning is to derive easy-to-obtain supervision signals without human annotations, to encourage the learning of useful features for regular tasks. Various novel tasks are proposed to learn image representations from unlabeled image data, , re-ordering perturbed image patches [@doersch2015unsupervised; @noroozi2016unsupervised], colorizing grayscale images [@zhang2016colorful], inpainting missing regions [@pathak2016context], counting virtual primitives [@noroozi2017representation], classifying image rotations [@gidaris2018unsupervised], predicting image labels obtained using a clustering algorithm [@caron2018deep], . There are also studies that try to learn image representations from unlabeled video data. Wang and Gupta [@wang2015unsupervised] proposed to derive supervision labels from unlabeled videos using traditional tracking algorithms. Pathak [@pathak2017learning] instead obtained labels from videos using conventional motion segmentation algorithms. Recent studies leveraging video data try to learn transferable representations for video tasks. Misra [@misra2016shuffle] designed a binary classification task and asked the CNN to predict whether the video input is in right order or not. Fernando [@fernando2017self] and Lee [@lee2017unsupervised] also designed tasks based on video frame orders. Gan [@gan2018geometry] proposed a geometry-guided network that force the CNN to predict flow fields or disparity maps between two input frames. Although these work demonstrated the effectiveness of self-supervised representation learning with unlabeled videos and showed impressive performances when transferring the learned features to video recognition tasks, their approaches are only applicable to a CNN that accepts one or two frames as inputs and cannot be applied to network architectures that are suitable for spatio-temporal representations. The most related work to ours are Vondrick [@vondrick2016generating] and Kim [@kim2018self]. Vondrick [@vondrick2016generating] proposed a GAN model for videos with a spatio-temporal 3D convolutional architecture, which can be employed as a self-supervised approach for video representation learning. Kim [@kim2018self] proposed to learn spatio-temporal representations with unlabeled video data, by solving space-time cubic puzzles, which is a straightforward extension of the 2D puzzles [@noroozi2016unsupervised]. Our Approach ============ We design a novel task for self-supervised video representation learning by predicting the motion and appearance statistics in a video sequence. The task is bio-inspired and consistent with human visual habits [@giese2003cognitive] to capture high-level concepts of videos. In this section, we first illustrate the statistical concepts and motivations to design the task (Sec. \[sec.concept\]). Next, we formally define the proposed statistical labels (Sec. \[sec.motionstat\] and \[sec.rgbstat\]). Finally, we present the whole learning framework when applying the self-supervised task to the C3D [@tran2015learning] network (Sec. \[sec.learnc3d\]). Statistical Concepts {#sec.concept} -------------------- Given a video clip, humans usually first notice the moving proportion of the visual field [@giese2003cognitive]. By observing the foreground motion and the background appearance, we can easily tell the motion class based on prior knowledge. Inspired by human visual system, we break the process of understanding videos into several questions and encourage a CNN to answer them accordingly: (1) Where is the largest motion in the video? (2) What is the dominant direction of the largest motion? (3) Where is the largest color diversity and what is its dominant color? (4) Where is the smallest color diversity, , the potential background of a scene and what is its dominant color? The approach to quantify these questions into annotation-free training labels will be described in details in the following sections. Here, we introduce the statistical concepts for motion and appearance. Figure \[fig:concepts\] illustrates an example of a three-frame video clip with two moving objects (blue circle and yellow triangle). A typical video clip normally consists of much more frames. We here instead use the three-frame clip for better understanding of the key ideas. To accurately represent the location and quantify “where”, each frame is divided into 4-by-4 blocks and each block is assigned to a number in an ascending order. The blue circle moves from block four to block seven, and the yellow triangle moves from block 12 to block 11. Comparing the moving distance, we can easily tell that the motion of the blue circle is larger than the motion of the yellow triangle. And the largest motion lies in block seven since it contains moving-in motion between frame one and two, and moving-out motion between frame two and three. As for the question “*what is the dominant direction of the largest motion?*”, it can be easily observed from Figure \[fig:concepts\] that the blue circle is moving towards lower-left. To quantify the directions, the full angle is divided into eight angle pieces, with each piece covering a motion direction range. And similar to location quantification, each angle piece is assigned to a number in an ascending order counterclockwise. The corresponding angle piece number of “lower-left” is five. For the appearance statistics, the largest spatio-temporal color diversity area is also block seven, as it changes from the background color to the circle color. The dominant color is the same as the moving circle color, , blue. As for the smallest color diversity location, most of the blocks stay the same and the background color is white. Keeping the above concepts and motivation in mind, we next present the proposed novel self-supervised approach. We assume that by training a spatio-temporal CNN to predict the motion and appearance statistics mentioned above, better spatio-temporal representations can be learned, by which the video understanding tasks could be benefited consequentely. Specifically, we design a novel regression task to predict a group of numbers related to motion and appearance statistics, such that by correctly predicting them, the following queries could be roughly derived: the largest motion location and the dominant motion direction in the video, the most consistent colors over the frames and their spatial locations, and the most diverse colors over the frames and their spatial locations. ![A simple illustration of statistical concepts in a three-frame video clip. See explanations in Sec. \[sec.concept\] for more details. []{data-label="fig:concepts"}](image/statistical_concepts){width="1\linewidth"} Motion Statistics {#sec.motionstat} ----------------- \[motion\_s\] We use optical flow computed by classic coarse-to-fine algorithms [@brox2004high] to derive the motion statistical labels to be predicted in our task. Optical flow is a motion representation feature that is commonly used in many video recognition methods. For example, the classic two-stream network [@simonyan2014two] and the recent I3D network [@carreira2017quo], both of which use stack of optical flow as their inputs for action recognition tasks. However, optical flow based methods are sensitive to camera motion, since they represent the absolute motion [@carreira2017quo; @wang2011action]. To suppress the influence of camera motion, we instead seek a more robust feature, motion boundary [@dalal2006human], to capture the video motion information. ![image](image/motion_boundary_new){width="90.00000%"} \[fig:motion\_boudary\] #### Motion Boundary. Denote optical flow horizontal component and vertical component as $u$ and $v$, respectively. Motion boundaries are calculated by computing x- and y- derivatives of $u$ and $v$, , $u_x=\frac{\partial u}{\partial x}$, $u_y=\frac{\partial u}{\partial y}$, $v_x=\frac{\partial v}{\partial x}$, $v_y=\frac{\partial v}{\partial y}$. As motion boundaries capture changes in the flow field, constant or smoothly varied motion, such as motion caused by camera view change, will be cancelled out. Only motion boundaries information is kept, as shown in Figure \[fig:motion\_boudary\]. Specifically, for an $N$-frame video clip, $(N-1)*2$ motion boundaries are computed. Diverse video motion information can be encoded into two summarized motion boundaries by summing up all these $(N-1)$ sparse motion boundaries of each component as follows: $$%\begin{array}{l} M_u=( \sum\limits_{i=1}^{N-1}u_x^i, \sum\limits_{i=1}^{N-1}u_y^i), ~ M_v=( \sum\limits_{i=1}^{N-1}v_x^i, \sum\limits_{i=1}^{N-1}v_y^i), %\end{array} \label{sum_up}$$ where $M_u$ denotes the motion boundaries on horizontal optical flow $u$, and $M_v$ denotes the motion boundaries on vertical optical flow $v$. Figure \[fig:motion\_boudary\] shows the visualization of the two sum-up motion boundaries images. #### Spatial-aware Motion Statistical Labels. In this section, we describe how to design the spatial-aware motion statistical labels to be predicted by our self-supervised task: 1) where is the largest motion; 2) what is the dominant orientation of the largest motion, based on motion boundaries. Given a video clip, we first divide it into several blocks using simple patterns. Although the pattern design is an interesting problem to be investigated, here, we introduce three simple yet effective patterns as shown in Figure \[fig:pattern\]. For each video block, we assign a number to it for representing its location. Then we compute $M_u$ and $M_v$ as described above. The motion magnitude and orientation of each pixel can be obtained by casting motion boundaries $M_u$ and $M_v$ from the Cartesian coordinates to the Polar coordinates. As for the largest motion statistics, we compute the average magnitude of each block and use the number of the block with the largest average magnitude as the largest motion location. Note that the largest block number computed from $M_u$ and $M_v$ can be different. Therefore, we use two labels to represent the largest motion locations of $M_u$ and $M_v$ separately. While for the dominant orientation statistics, an orientation histogram is computed based on the largest motion block, similar to the computation motion boundary histogram (MBH) [@dalal2006human]. Note that we do not have the normalization step since we are not computing a descriptor. Instead, we divide into 8 bins, with each bin containing angle range and again assign each bin to a number to represent its orientation. For each pixel in the largest motion block, we first use its orientation angle to determine which angle bin it belongs to and then add the corresponding magnitude number into the angle bin. The dominant orientation is the number of the angle bin with the largest magnitude sum. ![Three different partitioning patterns (from left to right: 1 to 3) used to divide video frames into different types of spatial regions. Pattern 1 divides each frame into 4$\times$4 blocks. Pattern 2 divides each frame into 4 different non-overlapped areas with the same gap between each block. Pattern 3 divides each frame by the two center lines and the two diagonal lines. The indexing strategies of the labels are shown in the bottom row. []{data-label="fig:pattern"}](image/Patterns){width="\columnwidth"} #### Global Motion Statistical Labels. We also propose a set of global motion statistical labels to provide complementary information to the local motion statistics described above. Instead of focusing on the local patch of video clips, a CNN is asked to predict the largest motion frame. That is given an $N$-frame video clip, the CNN is encouraged to understand the video evolution from a global perspective and find out between which two frames, contains the largest motion. The largest motion is quantified by $M_u$ and $M_v$ separately and two labels are used to represent the global motion statistics. ![image](image/network_long.pdf){width="100.00000%"} Appearance Statistics {#sec.rgbstat} --------------------- #### Spatio-temporal Color Diversity Labels. Given an $N$-frame video clip, same as motion statistics, we divide it into several video blocks by patterns described above. For an $N$-frame video block, we first compute the 3D distribution $V_i$ in 3D color space of each frame $i$. We then use the Intersection over Union (IoU) along temporal axis to quantify the spatio-temporal color diversity as follows: $$IoU_{score}=\frac{V_1\cap V_2 \cap ... \cap V_i ... \cap V_N}{V_1\cup V_2 \cup ... \cup V_i ... \cup V_N}.$$ The largest color diversity location is the block with the smallest $IoU_{score}$, while the smallest color diversity location is the block with the largest $IoU_{score}$. In practice, we calculate the $IoU_{score}$ on R,G,B channels separately and compute the final $IoU_{score}$ by averaging them. #### Dominant Color Labels. After we compute the largest and smallest color diversity locations, the corresponding dominant color is represented by another two labels. In the 3-D RGB color space, we evenly divide it into 8 bins. For the two representative video blocks, we assign each pixel a corresponding bin number by its RGB value, and the bin with the largest number of pixels is the dominant color. #### Global Appearance Statistical Labels. We also design a global appearance statistics to provide supplementary information. Particularly, we use the dominant color of the whole video as the global statistics. The computation method is the same as described above. Learning with Spatio-temporal CNNs {#sec.learnc3d} ---------------------------------- We adopt the popular C3D network [@tran2015learning] as the backbone for video spatio-temporal representation learning. Instead of using 2D convolution kernel $k \times k$, C3D proposed to use 3D convolution kernel $k \times k \times k$ to learn spatial and temporal information together. To have a fair comparison with other self-supervised learning methods, we use the smaller version of C3D as described in [@tran2015learning]. It contains 5 convolutional layers, 5 max-pooling layers, 2 fully-connected layers and a soft-max loss layer in the end to predict the action class, which is similar to CaffeNet [@jia2014caffe]. We followed the same video pre-processing procedure as C3D. Input video samples are first split into non-overlapped 16-frame video clips. And for each input video clip, it is first reshaped into 128 $\times$ 171 and then randomly cropped into 112 $\times$ 112 for spatial jittering. Thus, the input size of C3D is 16 $\times$ 112 $\times$ 112 $\times$ 3. Temporal jittering is also adopted by randomly flipping the whole video clip horizontally. We model our self-supervised task as a regression problem. The whole framework of our proposed method is shown in Figure \[fig:network\]. When pre-training the C3D network with the self-supervised labels introduced in the previous section, after the final convolutional layer, we use two branches to regress motion statistical labels and appearance statistical labels separately. For each branch, two fully connected layers are used similarly to the original C3D model design. And we replace the final soft-max loss layer with a fully connected layer, with 14 outputs for the motion branch and 13 outputs for the appearance branch. Mean squared error is used to compute the differences between the target statistics labels and the predicted labels. Experiments =========== In this section, we evaluate the effectiveness of our proposed approach. We first conduct several ablation studies on the local and global, motion and appearance statistics design. Specifically, we use motion statistics as our auxiliary task and appearance statistics acts the similar way. The activation based attention map of different video samples is visualized to validate our proposed methodology. Second, we compare our method with other self-supervised learning auxiliary tasks on action recognition problem based on two popular dataset UCF101 [@soomro2012ucf101] and HMDB51 [@kuehne2011hmdb]. Our method achieves the state-of-the-art result. Finally, we conduct two more experiments on action similarity [@kliper2012action] and dynamic scene recognition [@derpanis2012dynamic] to validate the transferability of our self-supervised spatio-temporal features. Datasets and Evaluations ------------------------ In our experiment, we incorporate five datasets: the UCF101 [@soomro2012ucf101], the Kinetics [@kay2017kinetics], the HMDB51 [@kuehne2011hmdb], the ASLAN [@kliper2012action], and the YUPENN [@derpanis2012dynamic]. Unless specifically state, we use UCF101 dataset for our model pre-training. UCF101 dataset [@soomro2012ucf101] consists of 13,320 video samples, which fall into 101 action classes. Actions in it are all naturally performed as they are collected from YouTube. Videos in it are quite challenging due to the large variation in human pose and appearance, object scale, light condition, camera view and It contains three train/test splits and in our experiment, we use the first train split to pre-train C3D. Kinetics-400 dataset is a very large human action dataset [@kay2017kinetics] proposed recently. It includes 400 human action classes, with 400 or more video clips for each class. Each sample is collected from YouTube and is trimmed into a 10-seconds video clip. This dataset is very challenging as it contains considerable camera motion/shake, illumination variations, shadows, . We use the training split for pre-training, which contains around 240k videos. HMDB51 dataset [@kuehne2011hmdb] is a smaller dataset which contains 6766 videos and 51 action classes. It also consists of three train/test splits. In our experiment, to have fair comparison with others, we use HMDB51 train split 1 to finetune the pre-trained C3D network and test the action recognition accuracy on HMDB51 test split 1. When pre-training on UCF101 train split 1 video data, we set the batch size to 30 and use the SGD optimizer with learning rate 0.001. We divide the leaning rate every 5 epochs by 10. The training process is stopped at 20 epochs. When pre-training on the Kinetics-400 train split, the batch size is 30 and we use the SGD optimizer with learning rate 0.0005. The learning rate is divided by 10 for every 7 epochs and the model is also trained for 20 epochs. When finetuning the C3D, we retain the conv layers weights from the pre-trained network and initialize three fully-connected layers. The entire network is finetuned with SGD on 0.001 learning rate. The learning schedule is the same as the pre-training procedure. When testing, average accuracy for action classification is computed on all videos to obtain the video-level accuracy. Initialization Accuracy (%) ------------------ -------------- Random 45.4 Motion pattern 1 53.8 Motion pattern 2 53.2 Moiton pattern 3 54.2 : Comparison the performance of different patterns of motion statistics for action recognition on UCF101. \[pattern\] Ablation Analysis ----------------- In this section, we analyze the performance of our local and global statistics, motion and appearance statistics on extensive experiments. Particularly, we first pre-train the C3D using different statistics design on UCF101 train split 1. For local and global statistics ablation studies, we finetune the pre-train model on UCF101 train split 1 data with human annotated labels. For the high-level appearance and motion statistics studies, we also finetune the C3D with HMDB51 train split 1 to get more understanding of the design. #### Pattern. The objective of this section is to investigate the performance of different pattern design. Specifically, we use the motion statistics and appearance statistics follow the same trend. As shown in Table \[pattern\], all the three patterns outperform the random initialization, , train from scratch setting, by around 8%, which strongly proves that our motion statistics is a very useful task. The performance of the three patterns are quite similar, indicating that we have balanced pattern design. #### Local *v.s.* Global. In this section, we compare the performance of local statistics, *where is the largest motion video block?*, global statistics, *where is the largest motion frame?* and their combination. As can be seen in Table \[global\], only global statistics serves as a useful auxiliary task for action recognition problem, with a improvement of 3%. And when all the three motion patterns are combined together, we can further get around 1.5% improvement, compared with single pattern. Finally, all motion statistics labels can achieve 57.8% accuracy, which is a significant improvement compared with train from scratch. #### Motion, RGB, and Joint Statistics. We finally compare all motion statistics, all RGB statistics, and their combination on UCF101 and HMDB51 dataset as shown in Table \[table:joint\]. From the table, we can find that both the appearance and motion statistics serve as a useful self-supervised signals for UCF101 and HMDB51 dataset. The motion statistics is more powerful as the temporal information is more important for video understanding. It is also interesting to note that although UCF101 only improves 1% when combined motion and appearance, the HMDB51 dataset benefits a lot from the combination, with a 3% improvement. Initialization Accuracy (%) ----------------------------- -------------- Random 45.4 Motion global 48.3 Motion pattern all 55.4 Motion pattern all + global 57.8 : Comparison of local and global motion statistics for action recognition on the UCF101 dataset. \[global\] Domain UCF101 acc.(%) HMDB51 acc. (%) -------------- ---------------- ----------------- From scratch 45.4 19.7 Appearance 48.6 20.3 Motion 57.8 29.95 Joint 58.8 32.6 : Comparison of different supervision signals on the UCF101 and the HMDB51 datasets. \[table:joint\] Action Recognition ------------------ In this section, we compare our method with other self-supervised learning methods on the action recognition problem. Particularly, we compare the results with RGB video input and directly quote the number from [@gan2018geometry]. As shown in Table \[comparison\], our method can achieve significantly improvement compared with the state-of-the-art both on UCF101 and HMDB51. Compared with methods that are pre-trained on UCF101 dataset, we improve 9.3% accuracy on HMDB51 than [@gan2018geometry] and 2.5% accuracy on UCF101 than [@lee2017unsupervised]. Compared with the method proposed recently [@kim2018self] that are pre-trained on Kinetics dataset using 3D CNN models, we can also achieve 0.6% improvement on UCF101 and 5.1% improvement on HMDB51. And please note that [@kim2018self] used various regularization techniques during pre-training, such as channel replication, rotation with classification and spatio-temporal jittering while we do not use these techniques. The results strongly support that our proposed predicting motion and appearance statistics task can really drive the CNN to learn powerful spatio-temporal features. And our method can generate multi-frame spatio-temporal features transferable to many other video tasks. [max width=]{} Method UCF101 acc.(%) HMDB51 acc.(%) -------------------------------------- ---------------- ---------------- DrLim [@hadsell2006dimensionality] 38.4 13.4 TempCoh [@mobahi2009deep] 45.4 15.9 Object Patch [@wang2015unsupervised] 42.7 15.6 Seq Ver.[@misra2016shuffle] 50.9 19.8 VGAN [@vondrick2016generating] 52.1 - OPN [@lee2017unsupervised] 22.1 Geometry [@gan2018geometry] 55.1 **Ours (UCF101)** **58.8** **32.6** ST-puzzle (Kinetics) [@kim2018self] 60.6 28.3 **Ours (Kinetics)** **61.2** **33.4** : Comparison with the state-of-the-art self-supervised video representation learning methods on UCF101 and HMDB51. \[comparison\] #### Visualization. To further validate that our proposed method really helps the C3D to learn video related features, we visualize the attention map [@zagoruyko2016paying] on several video frames as shown in Figure \[fig:attention\]. It is interesting to note that for similar actions: *Apply eye makeup* and *Apply lipstick*, C3D is just sensitive to the location that is exactly the largest motion location as quantified by the motion boundaries as shown in the right. For different scale motion, for example, the *balance beam* action, the pre-trained C3D is also able to focus on the discriminative location. ![Attention visualization. From left to right: A frame from a video clip, activation based attention map of conv5 layer on the frame by using [@zagoruyko2016paying], motion boundaries $M_u$ of the whole video clip, and motion boundaries $M_v$ of the whole video clip.[]{data-label="fig:attention"}](image/spatial_attention){width="\columnwidth"} Action Similarity Labeling -------------------------- We validate our learned spatio-temporal features on ASLAN dataset [@kliper2012action]. This dataset contains 3,631 video samples of 432 classes. The task is to predict whether the given two videos are of the same class or not. We use C3D as a feature extractor, followed by a *linear* SVM to do the classification. Each video sample is split into several 16 frames video clips with 8 frames overlapped and then go through a feed-forward pass on C3D to extract features from the last conv layer. The video-level spatio-temporal feature is obtained by averaging the clip feature, followed by l2-normalization. When testing on the ASLAN dataset, we follow the same 10-fold cross validation with leave-one-out evaluation protocol in each fold. Given a pair of videos, we first extract C3D feature from each video and then compute the 12 different distances described in [@kliper2012action]. The 12 (dis-)similarity are finally concatenated together to obtain a video-pair descriptor which is then fed into a linear SVM classifier. Since the scales of each distance are different, we normalize the distances separately into zero-mean and unit-variance as described in [@tran2015learning]. As no previous self-supervised learning methods have done experiment on this dataset, to validate that our self-supervised task can drive C3D to learn powerful spatio-temporal features, we design 4 scenarios to extract features from ASLAN dataset: (1) Use the random initialization C3D as feature extractor. (2) Use the C3D pre-trained on UCF101 with labels as feature extractor. (3) Use the C3D pre-trained on UCF101 with our self-supervised task as feature extractor. (4) Use the C3D finetuned on UCF101 on our self-supervised model as feature extractor. Table \[ASLAN\] shows the performance of different feature extractors. The random initialization model can achieve 51.4% accuracy as the problem is a binary classification problem. What surprises us is that although our self-supervised pre-trained C3D has never seen the ASLAN dataset before, it can still do well in this problem and outperforms the C3D trained with human-annotated labels by 1.1%. Such results strongly support that our proposed self-supervised task is able to learn powerful and transferable spatio-temporal features. This can be explained by the internal characteristics of the action similarity labeling problem. Different from the previous action recognition problem, the goal of ASLAN dataset is to predict video similarity instead of predicting the actual label. To achieve good performance, C3D must understand the video context, which is just what we try to drive the C3D to do with our self-supervised method. When finetuned our self-supervised pre-trained model with labels on UCF101, we can further get around 3% improvement. It outperforms the handcrafted features STIP [@kliper2012action], which is the combination of three popular features: HOG, HOF, and HNF (a composition of HOG and HOF). Features Accuracy (%) ------------------------------------ -------------- HOF [@kliper2012action] 56.68 HOG [@kliper2012action] 59.78 STIP [@kliper2012action] 60.9 C3D, random initialization 51.7 C3D, train from scratch with label 58.3 **C3D, self-supervised training** **59.4** C3D, finetune on self-supervised 62.3 : Comparison with different handcrafted features and our proposed four scenarios performance on the ASLAN dataset. \[ASLAN\] Dynamic Scene Recognition ------------------------- The performance on UCF101, HMDB51 and ASLAN dataset shows that our proposed self-supervised learning task can drive the C3D to learn powerful spatio-temporal features for action recognition problem. One may wonder that can action-related features be generalized to other problems? We investigate this question by transferring the learned features to the dynamic scene recognition problem based on the YUPENN dataset [@derpanis2012dynamic], which contains 420 video samples of 14 dynamic scenes. For each video in the dataset, first split it into 16 frames clips with 8 frames overlapped. The spatio-temporal features are then extracted based on our self-supervised C3D pre-trained model from the last conv layer. The video-label representations are obtained by averaging each video-clip features, followed with l2 normalization. A *linear* SVM is finally used to classify each video scene. We follow the same leave-one-out evaluation protocol as described in [@derpanis2012dynamic]. We compared our methods with both hand-crafted features and other self-supervised learning tasks as shown in Table \[YUPENN\]. Our self-supervised C3D outperforms both the traditional features and self-supervised learning methods. It shows that although our self-supervised C3D is trained on a action dataset, the learned weights has impressive transferability to other video-related tasks. [max width=]{} Method [@feichtenhofer2013spacetime] [@derpanis2012dynamic] [@wang2015unsupervised] [@misra2016shuffle] [@gan2018geometry] **Ours** -------------- ------------------------------- ------------------------ ------------------------- --------------------- -------------------- ---------- Accuracy (%) 86.0 80.7 70.47 76.67 86.9 **90.2** : Comparison with hand-crafted features and other self-supervised representation learning methods for dynamic scene recognition problem on the YUPENN dataset. \[YUPENN\] Conclusions =========== In this paper, we presented a novel approach for self-supervised spatio-temporal video representation learning by predicting a set of statistical labels derived from motion and appearance statistics. Our approach is bio-inspired and consistent with human visual systems. We demonstrated that by pre-training on unlabeled videos with our method, the performance of C3D network is improved significantly over random initialization on the action recognition problem. Compared with other self-supervised representation learning approaches, our method achieves state-of-the-art performances on UCF101 and HMDB51 datasets. This strongly supports that our method can drive C3D network to capture more crucial spatio-temporal information. We also showed that our pre-trained C3D network can be used as a powerful feature extractor for other tasks, such as action similarity labeling and dynamic scene recognition, where we also achieve state-of-the-art performances on public datasets. #### Acknowledgements: This work is supported in part by the Natural Science Foundation of China under Grant U1613218 and 61702194, in part by the Hong Kong ITC under Grant ITS/448/16FP, and in part by the VC Fund 4930745 of the CUHK T Stone Robotics Institute. Jianbo Jiao is supported by the EPSRC Programme Grant Seebibyte EP/M013774/1.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Motivated by the recent results on impurity effects in MgB$_2$, we present a theoretical model for a two-band superconductor in which the character of quasiparticle motion is ballistic in one band and diffusive in the other. We apply our model to calculate the electronic structure in the vicinity of an isolated vortex. We assume that superconductivity in the diffusive ($\pi$) band is induced by that in the clean ($\sigma$) band, as suggested by experimental evidence for MgB$_2$. We focus our attention to the spatial variations of the order parameter, the current density, and the vortex core spectrum in the two bands. Our results indicate that the coupling to the $\pi$ band can lead to the appearance of additional bound states near the gap edge in the $\sigma $ band that are absent in the single-band case.' author: - 'K. Tanaka$^{1,2}$, D. F. Agterberg$^{1,3}$, J. Kopu$^{4,5}$, and M. Eschrig$^{5}$' date: 'December 6, 2005' title: 'Ballistic and diffusive: theory of vortices in the two-band superconductor MgB$_2$' --- It is now well established that $\rm MgB_2$ is a two-gap superconductor, and its essential superconducting properties are well described by an isotropic $s$-wave two-band model [@nagamatsu01; @kwok03; @shulga01; @liu01; @bouguet01; @schmidt02; @iavarone02]. The two ‘bands’ in $\rm MgB_2$ are the ‘strong’ $\sigma$ band that arises from the boron $\sigma $-orbitals (with the energy gap $\Delta_\sigma\approx 7.2 {\rm meV}$), and the ‘weak’ $\pi$ band that derives from the boron $\pi $ orbitals ($\Delta_\pi\approx 2.3 {\rm meV}$). Despite the fact that each ‘band’ consists in reality of a pair of bands, the variation of the gap within each corresponding pair of Fermi-surface sheets can be neglected [@choi02; @mazin04]. The two energy gaps are observed to vanish at a common transition temperature $T_c$: no second transition has been observed, and there is evidence of induced superconductivity in the $\pi$ band [@schmidt02; @eskildsen02; @geerk05]. There have been considerable efforts to understand impurity effects in $\rm MgB_2$ [@kwok03]. Besides the potential applications of $\rm MgB_2$ as magnetic devices [@gurevich03], these studies have aimed at testing a prediction of the two-band model: the reduction of $T_c$ and the gap ratio by non-magnetic impurities. However, Mazin [*et al.*]{} [@mazin02; @erwin03] have shown that this does not apply to $\rm MgB_2$, i.e. to Mg-site impurities or defects, which are more favourable energetically than those at B sites. While the $\pi$ band is strongly affected by such impurities, the $\sigma$ band is more robust, and there is little mixing of the two bands because of negligible interband scattering [@mazin02; @erwin03]. This is consistent with accumulating experimental evidence suggesting that the $\sigma$ and $\pi$ bands are essentially in the ballistic and diffusive limit, respectively [@clean_sigma] (see also references in Refs. [@mazin02; @erwin03]). Nevertheless, so far in theoretical studies both of the two bands have been assumed to be either in the clean [@nakai02] or in the dirty limit [@koshelev03]. In this Letter, we examine theoretically the effects of induced superconductivity and impurities on the vortex-core structure in a two-band superconductor. Our work is motivated by the experimental investigation of the vortex state in $\rm MgB_2$ using scanning tunnelling spectroscopy [@eskildsen02]. We present a novel formulation of coupled quasiclassical Eilenberger and Usadel equations to describe a multiband superconductor with one ballistic and one diffusive band with negligible interband scattering: the two bands are assumed to be coupled only by the pairing interaction. We apply our model to calculate numerically the local density of states (LDOS) and supercurrent density around an isolated vortex. We examine in detail the intriguing spatial variation of these quantities and the order parameter in the two bands. A particularly interesting result emerging from our studies is the possibility of additional bound states near the gap edge in the $\sigma$ band. Our model is based on the equilibrium quasiclassical theory of superconductivity, where the physical information is contained in the Green function, or propagator, $\hat{g}(\epsilon, {\bf p}_{F\alpha}, {\bf R})$. Here $\epsilon$ is the quasiparticle energy measured from the chemical potential, ${\bf p}_{F\alpha}$ the quasiparticle momentum on the Fermi surface of band $\alpha \in \{\sigma, \pi\}$, and ${\bf R}$ is the spatial coordinate (the hat refers to the 2$\times$2 matrix structure of the propagator in the particle-hole space). In the clean $\sigma$ band, $\hat g_\sigma(\epsilon, {\bf p}_{F\sigma}, {\bf R})$ satisfies the Eilenberger equation [@eilenberger] $$\left[ \epsilon \hat \tau_3 - \hat \Delta_\sigma ,\; \hat g_\sigma \right] +i {\bf v}_{F\sigma} \cdot {\nabla } \hat g_\sigma = \hat 0, \label{eil}$$ where ${\bf v}_{F\sigma}$ is the Fermi velocity and $\hat \Delta_\sigma$ the (spatially varying) order parameter in the $\sigma$ band. Throughout this work, we ignore the external magnetic field (this is justified because MgB$_2$ is a strong type-II superconductor). The coherence length in the $\sigma $ band is defined as $\xi_{\sigma }=v_{F\sigma }/ 2\pi T_c$. In the presence of strong impurity scattering, the Green function has no momentum dependence and the Eilenberger equation reduces to the Usadel equation [@usadel70]. We assume this to be the appropriate description for the dirty $\pi$ band, and take the propagator $\hat g_\pi(\epsilon,{\bf R})$ to satisfy $$\left[ \epsilon \hat \tau_3 - \hat \Delta_\pi ,\; \hat g_\pi \right] + \frac{D}{\pi} {\nabla } \cdot (\hat g_\pi {\nabla } \hat g_\pi ) = \hat 0.$$ The diffusion constant $D$ defines the $\pi$-band coherence length as $\xi_{\pi }=\sqrt{D/2\pi T_c}$. Additionally, both propagators are normalized according to $\hat g_\sigma^2=\hat g_\pi^2=-\pi^2 \hat 1$. We assume that the quasiparticles in different bands are coupled only through the pairing interaction, neglecting interband scattering by the impurities. The gap equations for the multiband system are given as $$\Delta_\alpha ({\bf R}) = \sum_{\beta} V_{\alpha\beta} N_{F\beta} {\cal F}_\beta ({\bf R}), \label{gapeq}$$ where $\alpha,\beta \in \{\sigma,\pi\}$, $\hat\Delta_\alpha=\hat\tau_1~\rm{Re}~\Delta_\alpha-\hat\tau_2~\rm{Im}~\Delta_\alpha$, the coupling matrix $V_{\alpha\beta}$ determines the strength of the pairing interaction, $N_{F\beta}$ is the Fermi-surface density of states on band $\beta$, and $$\begin{aligned} {\cal F}_{\sigma }({\bf R})&\equiv& \int_{-\epsilon_c}^{\epsilon_c} {d\epsilon\over 2 \pi i}\, \langle f_\sigma (\epsilon , {\bf p}_{F\sigma},{\bf R}) \rangle_{{\bf p}_{F\sigma}} \, {\rm tanh}\left({\epsilon\over 2 T}\right), \label{F1} \nonumber \\ {\cal F}_{\pi }({\bf R})&\equiv& \int_{-\epsilon_c}^{\epsilon_c} {d\epsilon\over 2 \pi i}\, f_\pi (\epsilon, {\bf R} ) \, {\rm tanh}\left({\epsilon\over 2 T}\right). \label{F2}\end{aligned}$$ Here $f_\alpha$ is the upper off-diagonal (1,2) element of the matrix propagator $\hat g_\alpha$, $\langle\cdots\rangle_{{\bf p}_{F\sigma}}$ denotes averaging over the $\sigma$ band Fermi surface, and $\epsilon_c$ is a cut-off energy. We solve the system of equations (\[eil\])–(\[F2\]) numerically. The normalization condition is taken into account with the Riccati parameterisation for the Green functions [@schopohl; @matthias1]. After self-consistency has been achieved for the order parameter, the (for the $\sigma$ band angle-resolved) LDOS in each band can be calculated as $$\begin{aligned} N_{\sigma }(\epsilon,{\bf p}_{F\sigma}, {\bf R})/N_{F\sigma}&=&- ~{\rm Im}~g_\sigma (\epsilon , {\bf p}_{F\sigma},{\bf R}) /\pi , \nonumber \\ N_{\pi }(\epsilon,{\bf R})/N_{F\pi}&=&- ~{\rm Im}~g_\pi (\epsilon,{\bf R})/\pi, \label{LDOS}\end{aligned}$$ where $g_\alpha$ is the upper diagonal (1,1) element of $\hat g_\alpha$. The current density around the vortex has contributions from both the $\pi$ band and the $\sigma $ band. The corresponding expressions are ($e=-|e|$ is the electron charge) $$\begin{aligned} \frac{{\bf j}_{\sigma }({\bf R})}{2eN_{F\sigma }}&=& \int_{-\infty}^{\infty} {d\epsilon\over 2\pi }\, \langle {\bf v}_{F\sigma } {\rm Im}~g_\sigma \rangle_{{\bf p}_{F\sigma}} \tanh\left( \frac{\epsilon}{2T}\right), \; \nonumber \\ \frac{{\bf j}_{\pi}({\bf R})}{2eN_{F\pi}}&=& \frac{D}{\pi} \int_{-\infty}^{\infty} {d\epsilon\over 2\pi }\, {\rm Im}~[f_\pi^{\ast } {\nabla } f_\pi] \tanh\left( \frac{\epsilon}{2T}\right). \; \label{CD}\end{aligned}$$ Throughout this work, we focus on the effects of purely induced superconductivity in the $\pi$ band. Diagonalization of the coupling matrix in Eq. (\[gapeq\]) decouples the gap equations. The larger of the two eigenvalues of the matrix $V_{\alpha \beta }N_{F\beta }$, denoted by $\lambda^{(0)}$, determines the critical temperature $T_c$ and can be eliminated together with $\epsilon_c$. The smaller eigenvalue $\lambda^{(1)}$ is parameterized by the cut-off independent combination $\Lambda= \lambda^{(0)}\lambda^{(1)}/ (\lambda^{(0)}-\lambda^{(1)})$. The pairing interactions have been calculated with ab initio methods, employing an electron-phonon coupling model [@liu01; @choi02]; from these studies, taking Coulomb repulsion into account [@mazin04], we estimate $\Lambda <0.3 $ for MgB$_2$. We present results for $\Lambda =-0.1$ (implying a weak repulsion in the subdominant $\lambda^{(1)}$ channel). The ratio of the magnitude of the bulk gaps $\rho =|\Delta_\pi^{\rm bulk}|/|\Delta_\sigma^{\rm bulk}|$ near $T_c$ parameterizes the strength of the induced superconductivity in the $\pi $ band; experimentally $\rho \approx 0.3$ in MgB$_2$. For simplicity, we have set $N_{F\sigma}=N_{F\pi}$ in our calculations; the densities of states of the two bands are indeed of comparable size [@choi02; @mazin04]. A cylindrical Fermi surface was used for the two-dimensional $\sigma$ band. For the ratio of the coherence lengths in the two bands we present results for $\xi_\pi/\xi_\sigma$=1, 3, and 5. An estimate from experiments [@eskildsen02; @serventi04] gives for MgB$_2$ a value between 1 and 3. ![ Magnitude of the order parameter, $|\Delta_{\sigma,\pi} (x)|$, in the $\sigma$- and $\pi $ band, as a function of coordinate $x$ on a path through the vortex center, at $T=0.1 T_c$: (a) for different ratios $\xi_\pi/\xi_\sigma $ and fixed strength of the induced $\pi$-band gap, parameterized by the mixing ratio (see text), $\rho =0.3$; (b) for different $\rho$ and fixed $\xi_\pi/\xi_\sigma=5$. A weak effective repulsion due to Coulomb interaction was assumed in the subdominant $\lambda^{(1)}$ pairing channel, parameterized by $\Lambda=-0.1$ (see text). []{data-label="fig1"}](Delta_of_r_xi.eps "fig:"){width="0.48\columnwidth"} ![ Magnitude of the order parameter, $|\Delta_{\sigma,\pi} (x)|$, in the $\sigma$- and $\pi $ band, as a function of coordinate $x$ on a path through the vortex center, at $T=0.1 T_c$: (a) for different ratios $\xi_\pi/\xi_\sigma $ and fixed strength of the induced $\pi$-band gap, parameterized by the mixing ratio (see text), $\rho =0.3$; (b) for different $\rho$ and fixed $\xi_\pi/\xi_\sigma=5$. A weak effective repulsion due to Coulomb interaction was assumed in the subdominant $\lambda^{(1)}$ pairing channel, parameterized by $\Lambda=-0.1$ (see text). []{data-label="fig1"}](Delta_of_r_a.eps "fig:"){width="0.48\columnwidth"} In Fig. \[fig1\] we present the order-parameter magnitudes for each band as a function of coordinate $x$ along a path through the vortex center. In Fig. \[fig1\] (a) we show the order parameter variation at $T=0.1 T_c$ for several coherence-length ratios $\xi_\pi/\xi_\sigma $ and gap ratio $\rho=0.3$. Surprisingly, we find that an increase of $\xi_\pi$ results in an increase of the recovery length of the order parameter (the characteristic length over which the order parameter changes from zero to the bulk value) in the $\sigma$ band, while the recovery length in the $\pi$ band is barely affected. Thus, the recovery lengths in the two bands can differ considerably, even though the superconductivity in the $\pi $ band is induced by the $\sigma $ band. In Fig. \[fig1\] (b), the gap ratio $\rho $ is varied for fixed $\xi_\pi/\xi_\sigma=5$. One can see that, apart from the well-known increase of the bulk $\Delta_\sigma/T_c$ ratio with increasing bulk $\Delta_\pi$, the recovery length in the $\sigma $ band increases considerably with increasing $\rho $. The $\pi$-band recovery length is also enhanced together with that of the $\sigma$ band if $\rho >0.3$, as clearly seen for $\rho=0.5$. For $\rho=0.1$, $\Delta_\pi$ near the vortex core is enhanced from its bulk value. ![ (a) Local density of states (LDOS) for the $\pi $ band, as a function of energy $\epsilon $ for different distances $r$ from the vortex center. The spectra are shifted in vertical direction for convenience. (b) LDOS at the chemical potential, $\epsilon=0$, as a function of $r$ for different $\rho$ and fixed $T=0.5T_c$, $\xi_\pi/\xi_\sigma =3$. []{data-label="fig3"}](Npi_xi3_r3_t5.eps "fig:"){width="0.48\columnwidth"} ![ (a) Local density of states (LDOS) for the $\pi $ band, as a function of energy $\epsilon $ for different distances $r$ from the vortex center. The spectra are shifted in vertical direction for convenience. (b) LDOS at the chemical potential, $\epsilon=0$, as a function of $r$ for different $\rho$ and fixed $T=0.5T_c$, $\xi_\pi/\xi_\sigma =3$. []{data-label="fig3"}](NFp_of_r_xi3.eps "fig:"){width="0.48\columnwidth"} To understand these effects we have performed calculations for different values of $\Lambda $ that characterizes the Coulomb repulsion in the $\pi$ band. We have found that for $\Lambda >0$ the ratio $|\Delta_\pi|/|\Delta_\sigma |$ is reduced in the core region with respect to its bulk value, however is enhanced for $\Lambda <0$. This leads to a renormalization of the recovery lengths of the order parameters in both bands. In Fig. \[fig3\] we show the spectral properties of the $\pi$ band. The LDOS, shown in Fig. \[fig3\] (a), is flat at the vortex center ($r=0$), in agreement with the experiment of Ref. [@eskildsen02]. Outside the vortex core the BCS density of states is recovered. The decay of the zero-bias LDOS as a function of radial coordinate is shown in Fig. \[fig3\] (b), and compared with that of the $\sigma$ band \[obtained from the data in Fig. \[fig2\] (a)\]. The decay length of the zero-bias DOS is clearly different for the two bands. For the $\pi$ band it is given by $\xi_\pi \sqrt{\Delta_\sigma /\Delta_\pi }$, and thus dominated by the parameter $\rho$. In the $\sigma $ band the length scale of the decay is $\xi_\sigma$, and thus shorter than that in the $\pi$ band. The existence of two apparent length scales in the LDOS was also reported in the case of two clean bands [@nakai02] and two dirty bands [@koshelev03]. ![ (a) LDOS for the $\sigma $ band as a function of energy $\epsilon $ for different distances $r$ from the vortex center. (b) The development of an extra bound state near the gap edge in the $\sigma $ band for different parameter combinations. The bound state develops for sufficiently large mixing ratio $\rho $ and coherence length ratio $\xi_\pi/\xi_\sigma $. (c) Angle-resolved LDOS for quasiparticles with momentum along the $x$ axis at positions along the $y$ axis in the $\sigma $ band, showing the bound states as a function of $y$. The dispersion of the bound states as a function of $y$ is also shown in (d) for $\rho=0.5$, $\xi_\pi/\xi_\sigma =3$ and $T=0.5T_c$. Spectra in (a)-(c) are shifted in vertical direction for convenience. []{data-label="fig2"}](Nsigma_xi3_r3_t5_a.eps "fig:"){width="0.5\columnwidth"} ![ (a) LDOS for the $\sigma $ band as a function of energy $\epsilon $ for different distances $r$ from the vortex center. (b) The development of an extra bound state near the gap edge in the $\sigma $ band for different parameter combinations. The bound state develops for sufficiently large mixing ratio $\rho $ and coherence length ratio $\xi_\pi/\xi_\sigma $. (c) Angle-resolved LDOS for quasiparticles with momentum along the $x$ axis at positions along the $y$ axis in the $\sigma $ band, showing the bound states as a function of $y$. The dispersion of the bound states as a function of $y$ is also shown in (d) for $\rho=0.5$, $\xi_\pi/\xi_\sigma =3$ and $T=0.5T_c$. Spectra in (a)-(c) are shifted in vertical direction for convenience. []{data-label="fig2"}](Nsigma_r_0_a.eps "fig:"){width="0.45\columnwidth"} \ ![ (a) LDOS for the $\sigma $ band as a function of energy $\epsilon $ for different distances $r$ from the vortex center. (b) The development of an extra bound state near the gap edge in the $\sigma $ band for different parameter combinations. The bound state develops for sufficiently large mixing ratio $\rho $ and coherence length ratio $\xi_\pi/\xi_\sigma $. (c) Angle-resolved LDOS for quasiparticles with momentum along the $x$ axis at positions along the $y$ axis in the $\sigma $ band, showing the bound states as a function of $y$. The dispersion of the bound states as a function of $y$ is also shown in (d) for $\rho=0.5$, $\xi_\pi/\xi_\sigma =3$ and $T=0.5T_c$. Spectra in (a)-(c) are shifted in vertical direction for convenience. []{data-label="fig2"}](ArDos_x3r05t05.eps "fig:"){width="0.5\columnwidth"} ![ (a) LDOS for the $\sigma $ band as a function of energy $\epsilon $ for different distances $r$ from the vortex center. (b) The development of an extra bound state near the gap edge in the $\sigma $ band for different parameter combinations. The bound state develops for sufficiently large mixing ratio $\rho $ and coherence length ratio $\xi_\pi/\xi_\sigma $. (c) Angle-resolved LDOS for quasiparticles with momentum along the $x$ axis at positions along the $y$ axis in the $\sigma $ band, showing the bound states as a function of $y$. The dispersion of the bound states as a function of $y$ is also shown in (d) for $\rho=0.5$, $\xi_\pi/\xi_\sigma =3$ and $T=0.5T_c$. Spectra in (a)-(c) are shifted in vertical direction for convenience. []{data-label="fig2"}](Bound_x3_r05t05_rot.eps "fig:"){width="0.48\columnwidth"} In Fig. \[fig2\] we show the vortex-core spectra in the $\sigma $ band. The LDOS, $ N_{\sigma }(\epsilon,{\bf R})= \langle N_{\sigma }(\epsilon,{\bf p}_{F\sigma}, {\bf R}) \rangle_{{\bf p}_{F\sigma}} $, as a function of energy for different distances from the vortex center is plotted in Fig. \[fig2\] (a), showing the well-known Caroli-de Gennes-Matricon bound-state bands at low energies. The new feature of our model is the additional bound states in the vortex-core region near the gap edges, that are clearly visible in Fig. \[fig2\] (a). This is in strong contrast with the case of a single-band superconductor, where the spectrum near the vortex center is suppressed at the gap edges, showing neither a coherence peak nor additional bound states. We note that the self-consistency of the order-parameter profile is essential for discussing the presence of the bound states at the gap edges. We illustrate the development of these additional bound states in terms of the spectrum at the vortex center. We find that for a given $\xi_\pi/\xi_\sigma $, the bound state exists for $\rho$ larger than a certain critical value, as seen in the left panel of Fig. \[fig2\] (b). For a given $\rho$, on the other hand, the bound state develops if $\xi_\pi/\xi_\sigma $ exceeds a critical value, which is between 1 and 3 for $\rho=0.3$, as can be inferred from the right panel of Fig. \[fig2\] (b). The bound states are, e.g., clearly resolved for $\rho=0.3$ and $\xi_\pi/\xi_\sigma =3$, values appropriate for MgB$_2$. We have found a similar bound state spectrum also for $\Lambda >0 $. The bound states move with the gap edge as a function of temperature. ![ Contribution to the current density as a function of distance from the vortex center, shown separately for the $\pi$ band and the $\sigma $ band for fixed mixing ratio $\rho=0.3 $ and different temperatures $T$; (a) for $\xi_\pi/\xi_\sigma =1$, (b) $\xi_\pi/\xi_\sigma =3$, and (c) $\xi_\pi/\xi_\sigma =5$. In (d) the $\sigma $-band contribution to the total current density is shown as a function of radial coordinate, for $T=0.1 T_c$, $\rho=0.3 $, and different $\xi_\pi/\xi_\sigma $. []{data-label="fig4"}](curr_x1_r0.3.eps "fig:"){width="0.48\columnwidth"} ![ Contribution to the current density as a function of distance from the vortex center, shown separately for the $\pi$ band and the $\sigma $ band for fixed mixing ratio $\rho=0.3 $ and different temperatures $T$; (a) for $\xi_\pi/\xi_\sigma =1$, (b) $\xi_\pi/\xi_\sigma =3$, and (c) $\xi_\pi/\xi_\sigma =5$. In (d) the $\sigma $-band contribution to the total current density is shown as a function of radial coordinate, for $T=0.1 T_c$, $\rho=0.3 $, and different $\xi_\pi/\xi_\sigma $. []{data-label="fig4"}](curr_x3_r0.3.eps "fig:"){width="0.48\columnwidth"} \ ![ Contribution to the current density as a function of distance from the vortex center, shown separately for the $\pi$ band and the $\sigma $ band for fixed mixing ratio $\rho=0.3 $ and different temperatures $T$; (a) for $\xi_\pi/\xi_\sigma =1$, (b) $\xi_\pi/\xi_\sigma =3$, and (c) $\xi_\pi/\xi_\sigma =5$. In (d) the $\sigma $-band contribution to the total current density is shown as a function of radial coordinate, for $T=0.1 T_c$, $\rho=0.3 $, and different $\xi_\pi/\xi_\sigma $. []{data-label="fig4"}](curr_x5_r0.3.eps "fig:"){width="0.48\columnwidth"} ![ Contribution to the current density as a function of distance from the vortex center, shown separately for the $\pi$ band and the $\sigma $ band for fixed mixing ratio $\rho=0.3 $ and different temperatures $T$; (a) for $\xi_\pi/\xi_\sigma =1$, (b) $\xi_\pi/\xi_\sigma =3$, and (c) $\xi_\pi/\xi_\sigma =5$. In (d) the $\sigma $-band contribution to the total current density is shown as a function of radial coordinate, for $T=0.1 T_c$, $\rho=0.3 $, and different $\xi_\pi/\xi_\sigma $. []{data-label="fig4"}](Curr_x_ratio.eps "fig:"){width="0.48\columnwidth"} The bound-state spectrum is most clearly discussed in terms of the angle-resolved spectra, shown in Fig. \[fig2\] (c). Here the spectrum of quasiparticles moving in the $x$ direction is shown as a function of position along the $y$ axis. The position of the bound states as a function of $y$ is shown in Fig. \[fig2\] (d). The main bound state branch crosses the chemical potential in the vortex center. The additional bound states are seen near the gap edge, and show a weak dispersion. In fact, a close inspection reveals that there are two additional branches; however, only one of them is present at the vortex center. Finally, in Fig. \[fig4\] we show contributions from the two bands to the supercurrent density around the vortex separately. We show the results for $\rho=0.3$, and (a) $\xi_\pi/\xi_\sigma =1$, (b) $\xi_\pi/\xi_\sigma =3$, and (c) $\xi_\pi/\xi_\sigma =5$, at $T/T_c=0.1$, 0.3 and 0.5. The current density from the $\sigma $ band is enhanced for low temperatures near the vortex center, and the maximum approaches the center for $T\to 0$. This is due to the well-known Kramer-Pesch effect for the clean $\sigma $ band. The current density due to the induced superconductivity in the $\pi$ band is also enhanced by decreasing temperature, but the maximum does not approach the vortex center as in the $\sigma $ band. With increasing $r$, the contribution of the $\sigma $ band is reduced, and the contribution of the $\pi $ band becomes considerable when $\xi_\pi/\xi_\sigma \gtrsim 1$. At the same time, ${\bf j}_\sigma$ shows temperature dependence only in the core area, whereas ${\bf j}_\pi $ is temperature dependent also far outside the core. To discuss this effect, we note that at a large distance $r$ from the vortex center (but small compared to the London penetration depth), the current-density magnitudes are approximately given by $j_{\sigma }({r}) \sim eN_{F\sigma} v_{F\sigma }^2/2r $ and $j_{\pi}({r}) \sim eN_{F\pi} \pi D |\Delta_{\pi }|/r$. The temperature dependence of $j_\pi$ is thus dominated by that of $|\Delta_\pi |$. Another important observation is that, for the parameters appropriate for MgB$_2$, the contribution of $j_\pi$ is considerable outside the vortex core. In fact, already for a moderate ratio $\xi_\pi/\xi_\sigma =3$ the $\sigma $-band contribution is restricted to the region very close to the vortex center and is negligible outside the core for $\rho \ge 0.3$. To understand this effect, shown in Fig. \[fig4\] (d), we note that the current-density ratio far away from the vortex center approaches $$\begin{aligned} \lim_{r\gg \xi_{\pi,\sigma} }\frac{j_{\pi}({r})}{ j_{\sigma}({r})} =\frac{2D \pi |\Delta_{\pi }|}{v_{F\sigma }^2} = \frac{|\Delta_{\pi }|}{T_c} \left(\frac{\xi_\pi }{\xi_\sigma }\right)^2 \label{ratio} .\end{aligned}$$ Clearly, the $\pi $-band contribution to the current density dominates when $(\xi_\pi /\xi_\sigma )^2 > (|\Delta_\pi |/T_c)^{-1}$, which is a case relevant for MgB$_2$. In conclusion, we have formulated a model for coupled ballistic and diffusive bands in terms of coupled Eilenberger and Usadel equations. We have studied the effects of induced superconductivity in the ‘weak’ diffusive band on the order parameter, the current density and the spectral properties of the ‘strong’ ballistic band. We have found that (a) the recovery lengths of the order parameters in the two bands are renormalized by Coulomb interactions; (b) the vortex core spectrum in the $\sigma $ band shows additional bound states at the gap edges; and (c) the current density is dominated in the vortex core by the $\sigma $-band contribution, and outside the vortex core the $\pi $ band contribution is substantial, or even dominating, for parameters appropriate for MgB$_2$. Our predictions concerning the vortex-core spectrum of the $\sigma $ band can be tested in future tunnelling experiments. We acknowledge discussions with B. Jankó, M. Iavarone, M.W. Kwok, and H. Schmidt, and support by the NSERC of Canada, the U.S. DOE, Basic Energy Sciences (W-7405-ENG-36), the NSF (DMR-0381665), and the Deutsche Forschungsgemeinschaft within the CFN. [999]{} J. Nagamatsu [*et al.*]{}, Nature (London) [**410**]{}, 63 (2001). S.V. Shulga [*et al.*]{}, cond-mat/0103154 (2001). A. Y. Liu, I. I. Mazin, J. Kortus, Phys. Rev. Lett. [**87**]{}, 087005 (2001); A. A. Golubov [*et al.*]{}, J. Phys.: Condens. Matter [**14**]{}, 1353 (2002). F. Bouquet, [*et al.*]{}, Europhys. Lett. [**56**]{}, 856 (2001). H. Schmidt [*et al.*]{}, Phys. Rev. Lett. [**88**]{}, 127002 (2002). M. Iavarone [*et al.*]{}, Phys. Rev. Lett. [**89**]{}, 187002 (2002). Special issue on MgB$_2$, Physica (Amsterdam) [**385C**]{}, issues 1-2, (2003); T. Dahm in [*Frontiers in Superconducting Materials*]{}, ed. A.V. Narlikar, Springer (2005). H. J. Choi [*et al.*]{}, Nature (London) [**418**]{}, 758 (2002). I. I. Mazin [*et al.*]{}, Phys. Rev. B [**69**]{}, 056501 (2004). M. R. Eskildsen [*et al.*]{}, Phys. Rev. Lett. [**89**]{}, 187003 (2002); Phys. Rev. B [**68**]{}, 100508(R) (2003). J. Geerk [*et al.*]{}, Phys. Rev. Lett. [**94**]{}, 227005 (2005). A. Gurevich [*et al.*]{}, Supercond. Sci. Technol. [**17**]{}, 278 (2004). I. I. Mazin [*et al.*]{}, Phys. Rev. Lett. [**89**]{}, 107002 (2002). S. C. Erwin, I. I. Mazin, Phys. Rev. B [**68**]{}, 132505 (2003). M. Putti [*et al.*]{}, Phys. Rev. B [**67**]{}, 064505 (2003); [*ibid.*]{} [**70**]{}, 052509 (2004); [*ibid.*]{} [**71**]{}, 144505 (2005); J. W. Quilty, [*et al.*]{}, Phys. Rev. Lett. [**90**]{}, 207006 (2003); M. Ortolani [*et al.*]{}, Phys. Rev. B [**71**]{}, 172508 (2005); A. Carrington [*et al.*]{}, Phys. Rev. B [**72**]{}, 060507(R) (2005). N. Nakai, M. Ichioka, K. Machida, J. Phys. Soc. Japan [**71**]{}, 23 (2002). A. E. Koshelev, A. A. Golubov, Phys. Rev. Lett. [**90**]{}, 177002 (2003). G. Eilenberger, Z. Phys. [**214**]{}. 195 (1968). K. Usadel, Phys. Rev. Lett. [**25**]{}, 507 (1970). N. Schopohl and K. Maki, Phys. Rev. B [**52**]{}, 490 (1995). M. Eschrig, Phys. Rev. B [**61**]{}, 9061 (2000); M. Eschrig [*et al.*]{}, Adv. in Sol. State Phys. [**44**]{}, 533, Springer (2004); M. Eschrig [*et al.*]{} in [*Vortices in Unconventional Superconductors and Superfluids*]{}, eds. R.P. Huebener, N. Schopohl, and G.E. Volovik, Springer (2002). S. Serventi [*et al.*]{}, Phys. Rev. Lett. [**93**]{}, 217003 (2004).
{ "pile_set_name": "ArXiv" }
--- abstract: 'We discuss the notions of spin-orbital polarization and ordering in paramagnetic materials, and address their consequences in transition metal oxides. Extending the combined density functional and dynamical mean field theory scheme to the case of materials with large spin-orbit interactions, we investigate the electronic excitations of the paramagnetic phases of Sr$_2$IrO$_4$ and Sr$_2$RhO$_4$. We show that the interplay of spin-orbit interactions, structural distortions and Coulomb interactions suppresses spin-orbital fluctuations. As a result, the room temperature phase of Sr$_2$IrO$_4$ is a paramagnetic spin-orbitally ordered Mott insulator. In Sr$_2$RhO$_4$, the effective spin-orbital degeneracy is reduced, but the material remains metallic, due to both, smaller spin-orbit and smaller Coulomb interactions. We find excellent agreement of our [*ab-initio*]{} calculations for Sr$_2$RhO$_4$ with angle-resolved photoemission, and make predictions for spectra of the paramagnetic phase of Sr$_2$IrO$_4$.' author: - Cyril Martins - Markus Aichhorn - Loïg Vaugier - Silke Biermann title: 'Reduced effective spin-orbital degeneracy and spin-orbital ordering in paramagnetic transition metal oxides: Sr$_2$IrO$_4$ vs. Sr$_2$RhO$_4$' --- Probably the most important consequence of relativistic quantum mechanics in solids is the coupling of spin and orbital degrees of freedom. The concept of spin-orbit coupling (SOC) has been known for more than half a century, and its importance in magnetic materials has been recognized early on. Nevertheless, SOC in non-magnetic materials has only recently become a hot topic in condensed matter physics, in the context of topological insulators [@KanePRL95-2005], within the search for systems for quantum computing applications. Still, the interplay of spin-orbit (SO) interactions with electronic Coulomb interactions in paramagnetic materials remains a largely unexplored field. This is at least partially due to the seemingly mutually exclusive regimes of their acting: SO interactions are strongest in heavy elements, that is in transition metal compounds with $5d$ (and, to a lesser extent $4d$) electrons. Coulomb interactions, on the other hand, are believed to be most efficient in $3d$ compounds, due to the more localized $3d$ orbitals. It is only recently that this separation has been questioned, e.g. in [@MoonPRL101-2008; @KimPRL101-2008; @KimScience323-2009; @PesinNP-2010]. In this Letter, we present a particularly striking example for the interplay of SOC and Coulomb interactions: in the room-temperature phase of the $5d$ transition metal oxide (TMO)  the combined effect of SOC and distortions from the ideal K$_2$NiF$_4$-structure is strong enough to suppress the effective degeneracy of the coupled spin-orbital degrees of freedom. Even moderate Coulomb interactions then induce a Mott localized state. The resulting state – despite of being paramagnetic and paraorbital, i.e. in the absence of any magnetic or orbital order – displays a “spin-orbital order” in the sense that the only hole in the t$_{2g}$ manifold occupies a state of well-defined t$_{2g}$-projected total angular momentum $J^2$. This notion generalizes the concept of orbital order, well-known in TMOs, to the case where neither spin nor orbitals are good quantum numbers. We present – entirely from first-principle calculations – a scenario for the paramagnetic insulating state of , in comparison to the isostructural and isoelectronic $4d$ compound . For the latter, we find that the smaller SOC and Coulomb interactions induce a partially spin-orbitally polarized metallic state. We calculate photoemission spectra for both materials and find excellent agreement with available experiments.  is a $5d$- with a tetragonal crystal structure whose symmetry is lowered from the K$_2$NiF$_4$-type, well-known in  or La$_2$CuO$_4$, by 11$^\circ$ rotation of its IrO$_6$ octahedra around the **c**-axis as in its $4d$-counterpart  [@HuangJSSC112-1994; @ItohJSSC118-1995]. Although each Ir site accomodates $5$ electrons,  exhibits an insulating behavior at all temperatures with an optical gap of about $0.26$ eV at room temperature [@MoonPRB80-2009]. Below $240$K, canted-antiferromagnetic (AF) order sets in, with an effective paramagnetic moment of $0.5$$\mu_B$/Ir and a saturation moment of $0.14$$\mu_B$/Ir [@CaoPRB57-1998]. This phase has triggered much experimental and theoretical work recently [@KimPRL101-2008; @KimScience323-2009; @JinPRB80-2009; @WatanabePRL105-2010], highlighting in particular the importance of the SOC $\zeta_{SO}$. SOC was also shown to be relevant in , which is a paramagnetic metal down to $36$mK [@MoonPRB74-2006]: density functional calculations within the local density approximation (LDA), augmented by Coulomb interactions within the  method, reproduce the Fermi surface only if SO interactions are taken into account [@HaverkortPRL101-2008; @LiuPRL101-2008]. ![\[fig2\] LDA band structures of  without distortions and without SOC (a), with distortions without SOC (b), without distortions with SOC (c) and with both, distortions and SOC (d).](fig1.eps){width="0.78\linewidth"} Here, we focus on the [*paramagnetic*]{} insulating phase of  above $240$ K, which has not been addressed by theory before. We analyze the electronic properties by a combination of the LDA with dynamical mean field theory (DMFT). Our method is a generalization of the LDA+DMFT scheme as implemented in [@AichhornPRB80-2009] based on the Wien2k package [@ReferenceWien2k], extended to include SO interactions [@Martins-tbp]. The Coulomb interactions are calculated from the constrained random phase approximation (cRPA) [@FerdiPRB70-2004], taking matrix elements in the same set of localized orbitals as used in the DMFT calculations [@Vaugier-tbp]. For  we estimate the Slater integrals as $F^0$=$2.2$eV and $J$=$0.3$eV, giving an intraorbital Hubbard U in the half-filled 12 orbitals of $U$=$2.25$eV [@footnote1]. ![\[fig1\] LDA band structure of  (left) and  (right), projected on the 12 (top), 32 $|m_j|$=$3/2$ (middle), and 32 $|m_j|$=$1/2$ (bottom) spin-orbitals.](fig2.eps){width="0.9\linewidth"} Within an LDA band picture, Fig. \[fig2\] (d), a metallic solution – at variance with experiments – is obtained for . We construct Wannier functions for the t$_{2g}$ orbitals from the LDA bands within the energy window $[-3.0,0.5]$eV. The large SOC of about $\zeta_{\textrm{SO}}$$\approx$$0.4$eV splits these states into a quartet of states, commonly labeled as 32, and a higher lying doublet 12. Each state is twice degenerate in $\pm m_j$. The four 32 states are almost filled; we find $n_{3/2,|1/2|}$=$1.98$ and $n_{3/2,|3/2|}$=$1.84$ for the 32 orbitals. The 12 states thus slightly exceed half-filling with $n_{1/2}$=$1.16$. In the left panel of Fig. \[fig1\], we plot the $j_{\textrm{eff}}$ character of the Wannier orbitals. The four bands that cross the Fermi level are not purely formed by the 12 orbitals, but there is a slight overlap of the 12 and 32 $|m_j|$=$3/2$ characters, particularly obvious at the $\Gamma$ point. Similar conclusions were drawn for the AF phase in Ref. [@WatanabePRL105-2010]. We now redefine the 12 and 32 $|m_j|$=$1/2$ atomic states, yielding a diagonal density matrix of the local problem. Thus, our definition takes into account the structural distortions and a tetragonal crystal field in each IrO$_6$ octahedra. Our coefficients for the 12 state are similar to those obtained for the AF phase in Ref. [@JinPRB80-2009]. We now turn to our LDA+DMFT results. In Fig. \[fig3\], we display the corresponding spectral functions. Within LDA+DMFT, supplemented by the   interactions, an insulating solution with a Mott gap of the size of the optical gap measured at room temperature (about $0.26$ eV [@MoonPRB80-2009]) is obtained. Despite the fact that the  measurements of Ref. [@KimPRL101-2008] were performed in the AF phase, a comparison of the total spectral function to the experimental energy distribution curves shows qualitative and even quantitative agreement along the direction $\Gamma$-$X$ and $\Gamma$-$M$ (between -$0.5$eV and -$1.5$eV). Along the $M$-$X$ direction a band around -$1$eV is identified. From the orbitally-resolved spectral functions (Fig. \[fig3\] b and c), one can locate the lower 12 Hubbard band at about -$0.5$eV, in agreement with the  data. ![\[fig3\] Momentum-resolved spectral functions $A(\mathbf{k},\omega)$ of the paramagnetic phase of  from  at $T$=$300$K. ](fig3.eps){width="0.8\linewidth"} The spin-orbital polarization is enhanced when taking into account Coulomb correlations: the Wannier orbital 12 is now exactly half-filled, and the upper Hubbard band is of 12 type only. We are thus dealing with a state exhibiting a “spin-orbital order” in the sense of a well-defined $j_{\textrm{eff}}$ quantum number. In contrast to the AF phase, where it is also the 12 state that carries the hole, we do [*not*]{} have an ordering of the corresponding $m_{j_\textrm{eff}}$=$\pm 1/2$ quantum number. Also, the orbital occupations of the original t$_{2g}$ orbitals to the 12 spin-orbital are roughly equal. We are thus facing the remarkable situation of “spin-orbital order”, with [*neither orbital nor magnetic order*]{}. As it is well-known from studies of multi-orbital Hubbard models the critical interaction for the formation of the insulating state is lowest for the one-band case, and increases with degeneracy (for a review see [@GeorgesJdeP-2004]). In , the Hubbard interactions calculated from are large enough to induce a Mott insulating state in a half-filled one-orbital but not in a quarter-filled two-orbital or a 1/6-filled three-orbital system. We can thus conclude that the reduced effective spin-orbital degeneracy is the reason for the insulating nature of . An analogous suppression of [*orbital*]{} degeneracy, albeit purely induced by cristal field splittings, has been studied in the model context in [@ManiniPRB66-2002; @PoteryaevPRB78-2008], and has been found to make LaTiO$_3$ and YTiO$_3$ Mott-insulating [@PavariniPRL92-2004]. In  the suppression of [*spin-orbital*]{} fluctuations is a consequence of the combined effect of SOC and the tetragonal distortions. To demonstrate this, we do a numerical experiment in which we switch on or off the two effects separately and monitor the critical value of $U$ for the Mott metal-insulator transition. In Fig. \[fig2\] (a) both the SOC and the structural distortions are omitted (we use the lattice parameters of Ref. [@RandallJACS79-1957]). The band structure of  is then very similar to the one of , except for the increased filling compared to the $d^4$ Ru-compound. The t$_{2g}$ orbitals almost equally accomodate the $5$ electrons ($n_{d_{xy}}$=$1.612$, $n_{d_{xz},d_{yz}}$=$1.696$). The insulating Mott state is reached for $F^0$=$3.6$eV and $J$=$0.3$eV with enhanced orbital polarization. The  band gets filled with increasing correlations, approaching a completely filled  band and two $3/4$-filled d$_{xz/yz}$ bands (which is the atomic ground state). In Fig. \[fig2\] (b), the SOC is neglected but the structural distortions are included. This lowers the symmetry of  and results in a four-times larger unit cell and four-folded bands. Similarly as in  [@HaverkortPRL101-2008], a -2y2 hybridization gap opens between the  and the t$_{2g}$ bands (between $0.4$ and $1.3$eV), and the  band becomes almost filled. The orbital polarization is thus enhanced compared to the undistorted case, and the Mott transition occurs for smaller values of $F^0$. Indeed,  gives the insulating state for $F^0$=$3.0$eV with $J$=$0.3$eV. In Fig. \[fig2\] (c), we show the band structure of “undistorted”  but with SOC. Three $j_{\textrm{eff}}$ bands can be identified: the 12 one with an occupation of $n_{1/2}$=$1.20$ lies above the two 32 ones, with fillings $n_{3/2,|3/2|}$=$1.92$ and $n_{3/2,|1/2|}$=$1.96$. We find this three band system to be insulating at $F^0$=$3.0$eV with $J$=$0.3$eV. Although the problem is effectively reduced to an almost half-filled one-band model, the large bandwidth of the 12 band prevents a smaller value for the critical $F^0$. Finally, panel Fig. \[fig2\]-(d) depicts the band structure of the true compound. In this case, the Mott transition occurs between $F^0$=$2.1$ and $2.2$eV for $J=0.3$eV, the values that have been estimated by our cRPA calculations. The combination of both the structural distortions and the SOC is thus necessary for  to be insulating. ![\[fig4\] Momentum-resolved spectral function $A(\mathbf{k},\omega)$ of  (a). Bottom pannels display the orbitally-resolved spectral densities for 32 $|m_j|$=$1/2$ (b), 32 $|m_j|$=$3/2$ (c) and 12 (d) at $T$=$300$K. ](fig4.eps){width="0.8\linewidth"} In particular, we would expect  to be metallic, if either the distortions or the SOC were smaller. Such a situation may be realized if the material is strained, or grown on a substrate [@footnote2]. It also occurs in , the isostructural and isoelectronic $4d$ counterpart of . As expected for $4d$-orbitals, the bandwidth is smaller. As before, we construct Wannier $j_{\textrm{eff}}$ orbitals from the bands within the energy window $[-2.67,0.37]$eV; the Wannier $j_{\textrm{eff}}$ character along the bands of  is shown on the right hand side of Fig. \[fig1\]. Again, there is a mixture of the 12 and 32 $|m_j|$=$3/2$ orbital character along the four bands crossing the Fermi level. However, since the SOC is weaker in this compound ($\zeta_{SO}$=$0.161$eV) than in its $5d$-counterpart, the effective splitting between the bands 12 and 32 $|m_j|$=$3/2$ is smaller and these two states have similar filling ($n_{1/2}$=$1.42$, $n_{3/2,|3/2|}$=$1.64$), whereas the 32 $|m_j|$=$1/2$ state is filled ($n_{3/2,|1/2|}$=$1.96$). The system is thus close to a $3/4$-filled two-band model, which favours a metallic state. Moreover, the Hubbard interactions are smaller than in : a cRPA calculation as above gives $F^0$=$1.6$eV and $J$=$0.3$eV for . Indeed, the weaker hybridization of the Rh-$4d$ states with the O-$2p$ locates the latter about $1$eV higher in energy than in . As a result, the Coulomb interactions are screened more efficiently in  compared to . By computing the interaction within cRPA for an artificial  system in which we shift the O-$p$-states down by $1$eV we verify that $F^0$ is indeed increased. Our  calculations, Fig. \[fig4\], are in good agreement with energy distribution curves obtained by  at $10$K [@BaumbergerPRL96-2006; @PerryNJP8-2006]. We oberve a large electron-like pocket of radius $0.65-0.69$Å$^{-1}$ and a smaller hole pocket of radius $0.19-0.18$Å$^{-1}$ in the undistorted Brillouin zone. Further structures are found between -$0.1$ and -$0.2$eV along $\Gamma$-$X$ and around -$0.1$eV along $\Gamma$-$M$. The orbitally-resolved spectral functions allow to attribute both of these structures to the 32 $|m_j|$=$1/2$ orbital, whereas the hole-like $\alpha$-pocket around $\Gamma$ is of 32 $|m_j|$=$3/2$ type. The two other pockets, $\beta_M$ and $\beta_X$, are (mostly) of type 12. We find that  is close to the Mott transition (which would occur for interaction strengths of $F^0$=$1.8$-$2.0$ eV while keeping $J$ fixed). The occupancies of the 12 and 32 $|m_j|$=$3/2$ spin-orbitals are $0.63$ and $0.89$, resp. The quasi-particle weights for 12 and 32 $|m_j|$=$3/2$ have been estimated as $Z_{{1}/{2}}$=$0.5$ and $Z_{{3}/{2},|{3}/{2}|}$=$0.6$. In conclusion, we have shown that only the cooperative effect of SOC, lattice distortions and Coulomb correlations drives the 5$d$ compound  insulating, due to a complete spin-orbital polarization resulting in an effective half-filled one-orbital (more precisely, two-spin-orbital) system. The isostructural and isoelectronic 4$d$ compound   has smaller SOC and Coulomb interactions, leading to a less dramatic reduction of spin-orbital fluctuations and, as a result, a two-orbital (or four spin-orbital) metal. The structural and electronic similarities between and La$_2$CuO$_4$ – both are Mott insulators with one hole in an effective one-orbital system – together with the pronounced difference in their magnetic structure, suggest that the transport properties of [*doped*]{}   may give valuable information about the role of magnetic fluctuations in high-T$_c$ superconductivity. Oxygen-deficient  is not superconducting [@KornetaPRB82-2010]. Our electronic structure calculations show, however, that doping by impurities with extremely weak SOC should act as giant perturbations to the spin-orbital structure, in analogy to introducing a [*magnetic*]{} impurity into a system with weak SOC. It would be most interesting to introduce carriers into the system while conserving as much as possible the very specific electronic structure, e.g. by doping with heavy atoms such as Os or Re. This work was supported by the French Agence Nationale de la Recherche under project CorrelMat and GENCI/IDRIS under project 20111393. We acknowledge useful discussions with R. Arita, J.-S. Bernier, M. Ferrero, M. Imada, J. Kunes, J. Mravlje, O. Parcollet, and H. Takagi. [29]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ** (, ). , , , . , ****, (). , , , . , , , ****, (). , ****, (). , ****, (). , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, ().
{ "pile_set_name": "ArXiv" }
--- abstract: 'Probabilistic tractography based on diffusion weighted MRI has become a powerful approach for quantifying structural brain connectivities. In several works the similarity of probabilistic tractography and path integrals was already pointed out. This work investigates this connection more closely. For the so called Wiener process, a Gaussian random walker, the equivalence is worked out. We identify the source of the asymmetry of usual random walkers approaches and show that there is a proper symmetrization, which leads to a new symmetric connectivity measure. To compute this measure we will use the Fokker-Planck equation, which is an equivalent representation of a Wiener process in terms of a partial differential equation. In experiments we show that the proposed approach leads a symmetric and robust connectivity measure.' author: - | Marco Reisert\ Dept. of Radiology, Medical Physics, University Hospital Freiburg, Germany title: 'Probabilistic tractography, Path Integrals and the Fokker Planck equation' --- Introduction ============ Diffusion MRI has become a very important tool for understanding the living brain tissue ([@Jones2010_diff_book]). It can reveal both macro- and microstructural features of the neuronal network of the human brain. Tractography tries to characterize the structural connectome to understand the details of the interregional relationships of the human brain. Tractography algorithms may be divided in deterministic, streamline-based methods ([@Mori99; @Basser2000]), probabilistic approaches ([@Hagmann2003; @Parker2003]) and global approaches ([@Mangin2002; @reisert2010]). In probabilistic tractography one basically draws samples from a distribution over paths and computes some statistics over these paths, for example, recording their endpoints. In some work the similarity to the notion of path integrals appearing in quantum mechanics [@feynman2012quantum] and statistical physics [@kleinert2009path] was already pointed out [@tuch2000path; @bjornemo2002regularized; @friman2006bayesian; @Behrens2007]. Rigorous mathematical investigations show that the basic stochastic process behind path integrals is a so called Wiener process, a continuous Gaussian random walker [@vanKampen1992stochastic]. In this work we will recap the foundations of the theory of Wiener processes and path integrals. Based on this we build a path integral that leads to a symmetric brain connectivity measure. We will see that the source of asymmetry of conventional walker principles is due to particle conservation. By dropping particle conservation and starting from the path integral perspective we will find a symmetric brain connectivity measure. Besides the walker perspective and path integrals, there is a third, equivalent approach, which describes expectation values of Wiener walkers by partial differential equations (PDE). This equivalence will be used to give an algorithm for the computation of the connectivity measure. It will be based on solving a large linear system, describing a mixed diffusion/convection process. We also introduce a novel discretization scheme to avoid the heavy directional dependency, which usually appears for discretized convection operators. PDEs in the context of diffusion MRI have been previously used for regularization [@reisert2011], connectivity estimation [@tournier2003diffusion; @hageman2009diffusion; @zhang2013logical] and fiber density estimation [@miccai12]. Our proposed PDE contains a convection operator, but in the joint position/orientation space, in [@batchelor2002fibre] also convection was used for tractography, however just in position space. In [@zhang2013logical] probabilistic tractography was put into a logical framework and and an algorithm was derived which is based on solving a large linear system. This approach also works just in position space such that crossings cannot handled adequately as long as only local neighborhoods are considered. The system matrix describes a anisotropic diffusion process. Method ====== Apart from a few examples most methods to estimate brain connectivity are based on the walker principle. Fiber tracts are initiated from certain seed points and are iteratively built by following locally defined directions. While the deterministic tracking approaches are more used for illustrative purposes, the probabilistic ones are more related to quantitative connectivity analysis. The mathematical principle behind both approaches is an integration of streamlines along the underlying data. In probabilistic tracking the process can be seen as a Markov process. The iteration process is simple: if we call ${\mathbf{s}}(t)$ the current state of the tracker at step $t$, then the next state ${\mathbf{s}}(t+1)$ is drawn from some transition probability density $W({\mathbf{s}}(t+1) | {\mathbf{s}}(t))$, which may depend in some way on the DW-measurement. This process is basically a random walk in the state space of the tracker. If $W$ is Gaussian, it is possible to formulate the limit for very small time steps, which results in a Wiener process, which we will now concentrate on. Wiener Process -------------- In physics and mathematics there exists a huge collection of concepts for the analysis and characterization of Wiener random walkers. Basically, there are three perspectives which are mostly equivalent. The first one, which is closest to the above described Markov process is the concept of stochastic calculus and stochastic differential equations (SDE) [@vanKampen1992stochastic]. In physics, SDEs are usually written as Langevin equations, which do not have the strong mathematical footing of a SDE, but are simpler to read and more similar to ordinary calculus. To give an example (actually this example already covers everything we need later) we want to consider a simple diffusion process with some additional drift. Let the state of the ’tracker’ be ${\mathbf{s}} \in {\mathbb R}^3$ and ${\mathbf{v}}$ a vector field causing the drift. Then, the corresponding Langevin equation is $${\dot{{\mathbf{s}}}} (t) = {\mathbf{v}}({\mathbf{s}}(t) ) + \eta(t) \label{eq:langevin}$$ where $\eta$ stands for mean free white noise with variance $\sigma^2$ per unit time. It is uncorrelated in time, meaning $\langle\eta(t)\eta(t')\rangle = \sigma^2 \delta(t-t')$. The dot in ${\dot{{\mathbf{s}}}}$ means time differentiation. An approximative numerical integration gives rise to the following propagation scheme: $${\mathbf{s}}(t+\Delta t) = {\mathbf{s}}(t) + {\mathbf{v}}({\mathbf{s}}(t)) \Delta t + {\mathbf{u}}(t) \sqrt{\Delta t}. \label{eq:langdisc}$$ The first term ${\mathbf{v}}({\mathbf{s}}(t)) \Delta t$ looks like ordinary Euler integration. The second, stochastic term ${\mathbf{u}}(t)$ is drawn independently for each time step from $\mathcal{N}(0,\sigma^2)$. The factor $\sqrt{\Delta t}$ expresses the fact that the process $\eta(t)$ is not differentiable: variances add up and not standard deviations. Note the difference between the continuous Wiener process $\eta(t)$ and the discrete process ${\mathbf{u}}(t)$ which is just defined at discrete time points $k\Delta t$. Both are related by $\int_{t-\Delta t}^t \eta(t) dt = {\mathbf{u}}(t) \sqrt{\Delta t}$. The transition probability $W$ as described in the beginning for the discrete process described in is a Gaussian $W(\cdot|{\mathbf{s}}(t)) = \mathcal{N}({\mathbf{s}}(t) + \Delta t\ {\mathbf{v}}({\mathbf{s}}(t)),\Delta t \sigma^2)$. The second perspective describes expectation values of an ensemble $\mathcal{S}$ of random walkers described by . Suppose we have generated such an ensemble $\mathcal S$ of walkers all starting at ${\mathbf{s}}(0) = {\mathbf{x}}_0$ and we want to know the distribution of the states at some time $T$, i.e. $$p_T({\mathbf{x}}|{\mathbf{x}}_0) = \sum_{\{{\mathbf{s}}\} \in \mathcal S} \delta({\mathbf{s}}(T) - {\mathbf{x}}) \label{eq:pdens}$$ Note, that the sum ranges over complete random walks or paths. To make this formally more transparent we always write curly brackets whenever we refer to a path (a chain of states of a walker) as a whole and not just a particular state at a specific time. In fact, the distribution described in is a solution of a partial differential equation, which is called the Fokker-Planck equation (FPE) and is the master equation of the proposed continuous stochastic process. For this example, the FPE takes the form $${\dot{p_t}} = -\nabla \cdot ({\mathbf{v}} p_t) + \frac{\sigma^2}{2} {\Delta}p_t = \mathcal{H} p_t \label{eq:fpe}$$ where $\nabla$ is the usual gradient operator in ${\mathbb R}^3$ and $\Delta$ the Laplacian. If we integrate this equation with respect to the initial condition $p_0({\mathbf{x}}|{\mathbf{x}}_0) = \delta({\mathbf{x}} - {\mathbf{x}}_0)$ we just resemble the distribution given in . The function $p_T({\mathbf{x}}|{\mathbf{x}}_0)$ is also known as the propagator or Green’s function of the corresponding stochastic process. It can formally be written as $p_t = {\mathit{e}^{(}}\mathcal{H}t)$. Note that $p_T({\mathbf{x}}|{\mathbf{x}}_0)$ does not necessarily need to be normalized like a probability, e.g. walkers can die at the boundaries, which manifests in the boundary conditions of the corresponding FPE. The proposed algorithm will be based on discretized solutions of the steady state solutions of a symmetrized FPE. The third perspective is the path integral concept. From the viewpoint of a brain connectivity, the theory of path integrals is probably the most appealing one, however, they are probably the mathematical least understood concept and do not give a constructive way for designing an algorithm. However, it is essential for the understanding of what we are actually doing. The idea is to compute $p_T({\mathbf{x}}|{\mathbf{x}}_0)$ by summing over *all* path starting at ${\mathbf{x}}_0$ and ending at ${\mathbf{x}}$ weighted by its probability. For a rough derivation and motivation recall the discrete time representation. To compute the total probability that a particular path is drawn we take the product along the path $$P({\mathbf{s}}(N\Delta t),\hdots,{\mathbf{s}}(\Delta t)|{\mathbf{s}}(0) = {\mathbf{x}}_0) = \prod_{i=1}^N W({\mathbf{s}}(i \Delta t)|{\mathbf{s}}((i-1) \Delta t),$$ but note that $W$ is a probability *density*. So, it is not totally correct to just multiply them, the problem will be discussed below. By taking the logarithm of the above product it can be converted into a sum and in the continuous time limit $N \rightarrow \infty$ with $N \Delta t =T$ we have the integral form $$-\log(P(\{{\mathbf{s}}\}|{\mathbf{s}}(0) = {\mathbf{x}}_0)) = -\int_0^T \log(W({\mathbf{s}}(t),{\dot{{\mathbf{s}}}}(t)) dt$$ The functional $L({\mathbf{s}}(t),{\dot{{\mathbf{s}}}}(t)) = -\log(W({\mathbf{s}}(t),{\dot{{\mathbf{s}}}}(t)))$ is sometimes called the Onsager-Machlup functional ([@haken1976generalized; @durr1978onsager; @ito1981characterization; @andersson1999finite]) or, in the style of mechanics, the Lagrangian. It describes the cost of a path. Let us further follow the naive way and use and the corresponding $W$ to compute $L$. By disregarding the normalization constant (which actually diverges in the limit), we can show that $$L_\text{sym}({\mathbf{s}},{\dot{{\mathbf{s}}}}) = \frac{1}{2\sigma^2}({\dot{{\mathbf{s}}}} - {\mathbf{v}}({\mathbf{s}}))^2 \label{eq:sym}$$ which seems to be plausible but is wrong. The correct answer is $$L({\mathbf{s}},{\dot{{\mathbf{s}}}}) = \frac{1}{2\sigma^2}({\dot{{\mathbf{s}}}} - {\mathbf{v}}({\mathbf{s}}))^2 + \frac{1}{2} (\nabla \cdot {\mathbf{v}})({\mathbf{s}}) \label{eq:nosym}$$ The error is coming from the faulty assumption that we can simply multiply the conditional transition densities $W$ to get the total probability, which leads to the diverging normalization constant and the missing term $\nabla\cdot {\mathbf{v}}$. Actually we have to consider some small volumes instead of infinitesimal points. We have to consider the portion of walkers starting in a volume around ${\mathbf{s}}((i-1)\Delta t)$ and arriving in a volume around ${\mathbf{s}}(i\Delta t)$. This leads to an additional factor $(1+\frac{\Delta t\ \nabla \cdot {\mathbf{v}}}{2})$ which is the infinitesimal volume change during time $\Delta t$ caused by the drift ${\mathbf{v}}$. In more rigorous derivations of this result the factor is justified by the Jacobian of a change of variables ([@chaichian2001path; @kleinert2009path]), or by the considering the most probable tube around the path ([@durr1978onsager]). Let us consider the final equation stating that the probability for a walker starting at ${\mathbf{s}}(0) = {\mathbf{x}}_0$ and arriving at ${\mathbf{s}}(T) = {\mathbf{x}}$ is the sum over all paths connecting this two points weighted by their probabilities $$p_T({\mathbf{x}}|{\mathbf{x}}_0) = \sum_{\{{\mathbf{s}}\} \in \mathcal{P}_T^{({\mathbf{x}}|{\mathbf{x}}_0)}} {\mathit{e}^{\left}}(-\int_0^T L({\mathbf{s}},{\dot{ {\mathbf{s}}}}) dt\right) \label{eq:pi}$$ where $\mathcal{P}_T^{({\mathbf{x}}|{\mathbf{x}}_0)}$ denotes the set of all paths of length $T$ starting in ${\mathbf{x}}_0$ and terminating at ${\mathbf{x}}$. In fact, the *sum* over all path is an integration in a functional space, for more details see e.g. [@feynman2012quantum; @kleinert2009path; @vanKampen1992stochastic]. For convenience we omit to write here any normalization constant and assume that the measure is appropriately normalized. Reversibility and Symmetry -------------------------- In fact, the additional term $(\nabla \cdot {\mathbf{v}})/2$ in is quite important for us, because it is responsible for the fact that the described random walk is irreversible ([@garrido1979stochastic]). In the context of brain connectivity this is directly related to the symmetry of the connectivity measure, so we will detail this out now. If we call $L'$ the Lagrangian with negative velocities ${\mathbf{v}} \mapsto -{\mathbf{v}}$ and ${\mathbf{s}}'(t) = {\mathbf{s}}(T-t)$ the path traversed in opposite direction. Then, it is easy to see that $\int L'({\mathbf{s}}',{\dot{{\mathbf{s}}}}') dt \neq \int L({\mathbf{s}},\dot{{\mathbf{s}}}) dt$ with the $L$ obtained in . This also leads to $p_T({\mathbf{x}}|{\mathbf{x}}_0) \neq p'_T({\mathbf{x}}_0|{\mathbf{x}})$. This asymmetry is basically caused by drift together with particle conservation. Considering the corresponding FPE in makes it more clear. While the diffusion term ${\Delta}$ is symmetric (selfadjoint, see Appendix \[app1\] for definition), the convection/drift is not antisymmetric, i.e. the operator $\mathcal{H}'$ corresponding to the flipped Lagrangian $L'$ is not equal to adjoint operator $\mathcal{H}^+$ which has to be case to fulfill $p_T({\mathbf{x}}|{\mathbf{x}}_0) = p'_T({\mathbf{x}}_0|{\mathbf{x}})$. However, for $L_\text{sym}$ everything is different. In fact, we have $$\int_0^T L_\text{sym}'({\mathbf{s}}',{\dot{{\mathbf{s}}}}') dt= \int_0^T L_\text{sym}({\mathbf{s}},\dot{{\mathbf{s}}}) dt$$ which should also lead to a symmetric propagator $p_T({\mathbf{x}}|{\mathbf{x}}_0) = p'_T({\mathbf{x}}_0|{\mathbf{x}})$. But how to compute for $L_\text{sym}$ in practice, what is the corresponding FPE? The term $\nabla \cdot {\mathbf{v}}$ acts like an additional potential energy. In fact, the Feynman-Kac (FK) formula [@kac1949distributions] tells us that such a potential can be directly integrated into the FPE: suppose we have a given Lagrangian $L$ together with its path integral and a corresponding FPE with operator $\mathcal{H}$. For a potential field $V({\mathbf{s}})$ we define a new $L'({\mathbf{s}} ,{\dot{{\mathbf{s}}}}) = L({\mathbf{s}} ,{\dot{{\mathbf{s}}}})-V({\mathbf{s}})$, then the FK formula tells us that the corresponding FPE operator is just $\mathcal H' = \mathcal H + V$. Applying this our problem with $V = (\nabla \cdot {\mathbf{v}})/2$ gives $$\begin{aligned} \mathcal{H}_\text{sym}p &=& \mathcal{H}p + \frac{1}{2}(\nabla \cdot {\mathbf{v}})p \\ &=& \frac{\sigma^2}{2} {\Delta}p + \frac{1}{2}(\nabla \cdot {\mathbf{v}})p - \nabla \cdot ({\mathbf{v}} p) \\ &=& \frac{\sigma^2}{2} {\Delta}p - \frac{1}{2}({\mathbf{v}} \cdot \nabla p + \nabla \cdot ({\mathbf{v}} p) ) \end{aligned}$$ From the last line we can also see the symmetry property of $\mathcal{H}_\text{sym}$. The convection part $\nabla \cdot ({\mathbf{v}} p) + {\mathbf{v}} \cdot \nabla p$ is antisymmetric in the sense of the operator adjoint and hence the operator $\mathcal{H}'_\text{sym}$, which corresponds to flipped velocities, is just the adjoint of $\mathcal{H}_\text{sym}$. One can also see that the particle conservation is lost: we cannot write $\mathcal{H}_\text{sym}$ as the a divergence of something. Note the similarity of $\mathcal{H}_\text{sym}$ with the quantum mechanical Hamiltonian of a particle in a vector potential, which is up to some constants $(i \nabla - {\mathbf{v}})^2 = -{\Delta}- i(\nabla {\mathbf{v}} - {\mathbf{v}} \nabla) + {\mathbf{v}}^2$. In quantum mechanics the Hamiltonian is indeed self-adjoint due to the additional complex unit in front of the anti-symmetric part. Steady States and Path Trails ----------------------------- Up to now we have considered only paths with some specific length $T$. However, we are interested in *all* path without any length restriction. That is, we sum up over all paths connecting ${\mathbf{x}}_0$ and ${\mathbf{x}}$ of arbitrary length to get a density $p({\mathbf{x}} | {\mathbf{x}}_0)$ which is independent of $T$, i.e. $$\begin{aligned} p({\mathbf{x}}| {\mathbf{x}}_0) &=& \int_0^\infty dT\ p_T({\mathbf{x}}| {\mathbf{x}}_0) \\ &=& \sum_{\{{\mathbf{s}}\} \in \mathcal{P}^{({\mathbf{x}}|{\mathbf{x}}_0)}} e^{-\int_0^T L({\mathbf{s}},{\dot{ {\mathbf{s}}}}) dt} \\\end{aligned}$$ where $\mathcal{P}^{({\mathbf{x}}|{\mathbf{x}}_0)}$ is the set of paths of any length connecting ${\mathbf{x}}_0$ with ${\mathbf{x}}$. With the assumption $\lim_{T\rightarrow \infty } p_T({\mathbf{x}}|{\mathbf{x}}_0) = 0$, i.e. all walkers eventually die, the function $p({\mathbf{x}} | {\mathbf{x}}_0)$ is the steady state solution of the corresponding FPE, i.e. $p({\mathbf{x}}|{\mathbf{x}}_0)$ is the solution of the equation $$-\mathcal{H} p({\mathbf{x}} | {\mathbf{x}}_0) = \delta ({\mathbf{x}}-{\mathbf{x}}_0).$$ This is the basic type of equation we will solve to estimate brain connectivities. In the following we call $p({\mathbf{x}}|{\mathbf{x}}_0)$ the connectivity amplitude. From ordinary probabilistic tractography we know the so called length bias. With increasing euclidean distance, connectivity values decrease dramatically, usually exponentially. To account for this bias a linear reweighing was proposed in [@Behrens2007], that is, $$p_\text{lin} ({\mathbf{x}}| {\mathbf{x}}_0) = \int_0^\infty dT\ T\ p_T({\mathbf{x}}| {\mathbf{x}}_0). \label{eq:linrew}$$ We will see later how $p_\text{lin}$ can be computed in practice. Due to the exponential decrease one could also tend to use exponential corrections, i.e. use $$p_\kappa({\mathbf{x}}| {\mathbf{x}}_0) = \int_0^\infty dT\ e^{\kappa T} p_T({\mathbf{x}}| {\mathbf{x}}_0) \label{eq:exprew}$$ with positive $\kappa\in{\mathbb R}$. In fact, this results in a spectral shift of the operator $\mathcal H \mapsto \mathcal H + \kappa$. Of course, the choice of $\kappa$ is difficult. Large $\kappa$ will cause bad condition numbers for $\mathcal{H}$ and, in the extreme case, lead to a diverging $p_\kappa({\mathbf{x}}| {\mathbf{x}}_0)$. Until now, we were interested in a measure how strong two points ${\mathbf{x}}_1$ and ${\mathbf{x}}_0$ are connected, but one might also be interested in *how* the actual connection looks like. Instead of just summing over all paths connecting two points, one can additionally image the path itself by counting how often the path visited a specific position ${\mathbf{x}}$. We call this image the trail image of a path ${\mathbf{s}}$, and it is defined by $$\tau_{{\mathbf{s}}}({\mathbf{x}}) = \int_0^T \delta({\mathbf{x}} - {\mathbf{s}}(t)) dt$$ Now, we can write the mean trail image of a path connecting two point ${\mathbf{x}}_1$ and ${\mathbf{x}}_2$ by the expectation value $$\bar{\tau}_{({\mathbf{x}}_1| {\mathbf{x}}_0)} ({\mathbf{x}}) = \sum_{\{{\mathbf{s}}\} \in \mathcal{P}^{({\mathbf{x}}|{\mathbf{x}}_0)}}\tau_{{\mathbf{s}}}({\mathbf{x}}) e^{-\int_0^T L({\mathbf{s}},{\dot{ {\mathbf{s}}}}) dt}.$$ In order to compute $\bar{\tau}$ we have to use the so called Einstein-Smoluchowski-Kolmogorov-Chapman (ESKC) relation, or simply semigroup property, which tells that the path-integral, or the corresponding propagator can always be split like $$p_T({\mathbf{x}}_1|{\mathbf{x}}_0) = \int d{\mathbf{x}}'\ p_{T-t'}({\mathbf{x}}_1|{\mathbf{x}}') p_{t'}({\mathbf{x}}'|{\mathbf{x}}_0)$$ for any given intermediate point $0<t'<T$. This means we can slice the propagator at any given length $t'$ and consider all possible intermediate positions ${\mathbf{s}}(t') = {\mathbf{x}}'$ the path has at $t'$ and integrate over them by accounting for the probability of the path to get from ${\mathbf{x}}_0$ to ${\mathbf{x}}'$ and to reach the target ${\mathbf{x}}_1$ starting from ${\mathbf{x}}'$. To compute $\bar{\tau}$ we first consider the point density of the path at some specific length $t'$. That is, we count how often a path traverses the point ${\mathbf{x}}$ at length $t'$ by $$\begin{aligned} \rho_{T,t'}({\mathbf{x}}) &=& \int d{\mathbf{x}}'\ \delta({\mathbf{x}} -{\mathbf{x}}') p_{T-t'}({\mathbf{x}}_1|{\mathbf{x}}') p_{t'}({\mathbf{x}}'|{\mathbf{x}}_0) \\ &=& p_{T-t'}({\mathbf{x}}_1|{\mathbf{x}}) p_{t'}({\mathbf{x}}|{\mathbf{x}}_0)\end{aligned}$$ If we integrate now $\rho_{t',T}$ over all intermediate times $0<t'<T$ and, again, integrate this over all possible $0<T<\infty$ we get the mean trail image $$\bar{\tau}_{({\mathbf{x}}_1| {\mathbf{x}}_2)} ({\mathbf{x}}) = \int_0^\infty dT \int_0^T dt' \rho_{T,t'}({\mathbf{x}})$$ Plugging this altogether gives the simple final result $$\bar{\tau}_{({\mathbf{x}}_1| {\mathbf{x}}_0)} ({\mathbf{x}}) = p({\mathbf{x}}_1 | {\mathbf{x}}) p({\mathbf{x}} | {\mathbf{x}}_0)$$ and, as we already know how to compute $p$ we are ready. Further note that $$\begin{aligned} \int d{\mathbf{x}}\ \bar{\tau}_{({\mathbf{x}}_1| {\mathbf{x}}_0)} ({\mathbf{x}}) &=& \int_0^\infty dT\ T\ p({\mathbf{x}}_1 | {\mathbf{x}}_0) \\ &=& p_\text{lin}({\mathbf{x}}_1 | {\mathbf{x}}_0),\end{aligned}$$ that is, if we integrate the mean trail image, we just get the linearly reweighed version of and have also a way to compute in practice. Application to DW-MRI data -------------------------- All concepts were introduced by considering a 3D diffusion process with drift. But, how to apply this kind of stochastic process for brain connectivity analysis? The data is basically tensor-like, meaning it is point symmetric, there is no velocity field available. Even when we have orientation distributions, there is still no convection-like force. So, what to do? First of all, we have to realize that usual probabilistic tracking algorithms are not Markovian with respect to the position ${\mathbf{r}}\in{\mathbb R}^3$. The propagation direction usually depends not just on the position, but also on the previous step made. So, the state space of the tracker should be the joint space of position and orientation ${\mathbf{s}} = ({\mathbf{r}},{\mathbf{n}})$, where ${\mathbf{r}}\in {\mathbb R}^3$ and ${\mathbf{n}}\in S_2$. The data we get from the diffusion measurement is basically an fiber orientation distribution $f({\mathbf{r}},{\mathbf{n}})$ defined on this joint position/orientation space. In fact, in this joint space there are several ways to define meaningful connectivity measures. We propose to use the following Lagrangian $$L({\mathbf{r}},{\dot{{\mathbf{r}}}},{\mathbf{n}},{\dot{{\mathbf{n}}}}) = \frac{1}{2\sigma_r^2} ({\dot{{\mathbf{r}}}} - {\mathbf{n}} f({\mathbf{r}},{\mathbf{n}}))^2 + \frac{1}{2\sigma_n^2} {\dot{{\mathbf{n}}}}^2 \label{eq:theLag}$$ as path-costs. Paths with ${\dot{{\mathbf{s}}}} \sim {\mathbf{n}} f$ and small changes in ${\mathbf{n}}$ are assigned with small costs, and hence, with high probability. From the walker perspective, there is a convective force which drives a walker with internal state ${\mathbf{n}}$ into direction ${\mathbf{n}}$, while the speed is proportional to the data term $f$. Additionally there is diffusion on the orientation variable, enabling the walker to change its directions, or conversely, penalizes too strong bending. Imagine the system as a network of pipes oriented in all possible directions carrying a fluid which flows with speed $f({\mathbf{r}},{\mathbf{n}})$. The flow is not perfectly convective, neighboring parallel pipes can exchange fluid (this is spatial diffusion) and also crossing pipes exchange some fluid (this is the angular diffusion). In fact, we can also let spatial diffusion $\sigma_r \rightarrow 0$ such that there is only pure convection. However, as we will see later, the discretization of the FPE will force us to have finite values for $\sigma_r$. The FPE operator is not totally straight-forward, because there is an additional term coming from the curvature of the sphere ([@ito1981characterization; @andersson1999finite]). It looks like $$\mathcal{H} = \frac{1}{2} ( \sigma_r^2 {\Delta}_{{\mathbf{r}}} + \sigma_n^2 {\Delta}_{{\mathbf{n}}}) - \frac{1}{2} ({\mathbf{v}} \cdot \nabla_{{\mathbf{r}}} + \nabla_{{\mathbf{r}}} \cdot {\mathbf{v}}) + \frac{\sigma_n^2 R}{12}$$ where the velocity is defined by ${\mathbf{v}}({\mathbf{r}},{\mathbf{n}}) = {\mathbf{n}} f({\mathbf{r}},{\mathbf{n}})$. The additional constant $\frac{R}{12}$ is coming from the curvedness of the sphere which is $R=2/\rho^2$ for a 2-sphere of radius $\rho$. However, there is actually no metric connection between the position space ${\mathbb R}^3$ of the orientation space $S_2$. The ’radius’ of the sphere $S_2$ has no actual meaning. So, the value of $R/12$ is rather arbitrary. In fact, $R/12$ acts like an additional exponential weighting, like the $\kappa$ we already introduced above and could be absorbed by that. So, we exclude $R/12$ in the following, or, imagine $\rho$ to be pretty large. Let us now look at the symmetry properties of $\mathcal{H}$. Let $\mathcal{Z}$ the operator that does a point reflection on some function $a({\mathbf{r}},{\mathbf{n}})$ in the sense $$(\mathcal{Z} a)({\mathbf{r}},{\mathbf{n}}) := a({\mathbf{r}},-{\mathbf{n}})$$ Then, we can easily show that $$\mathcal{H Z} = \mathcal{Z H^+}$$ because the data $f$ is symmetric $\mathcal{Z} f = f$. In other words, the operator $\mathcal{H Z}$ is selfadjoint. Formally we can write the connectivity amplitudes as $$p({\mathbf{r}},{\mathbf{n}}|{\mathbf{r}}_0,{\mathbf{n}}_0) = -\langle e_{({\mathbf{r}},{\mathbf{n}})}| \mathcal{H}^{-1} e_{({\mathbf{r}}_0,{\mathbf{n}}_0)} \rangle$$ where $e_{({\mathbf{r}}',{\mathbf{n}}')}({\mathbf{r}},{\mathbf{n}}) = \delta({\mathbf{r}}' -{\mathbf{r}}) \delta({\mathbf{n}}' - {\mathbf{n}})$. For the connectivity amplitude the symmetry means $$p({\mathbf{r}},{\mathbf{n}}|{\mathbf{r}}_0,{\mathbf{n}}_0) = p({\mathbf{r}}_0,-{\mathbf{n}}_0|{\mathbf{r}},-{\mathbf{n}}),$$ because $\mathcal{H}^{-1} = \mathcal{Z}(\mathcal{H^+})^{-1}\mathcal{Z}$. If we now think of some function $a({\mathbf{r}},{\mathbf{n}})$ describing a seed point density, and $b({\mathbf{r}},{\mathbf{n}})$ a terminal point density. For both we can think of indicator functions describing some cortical region of interests. Then, we can define the symmetric connectivity measure $c(a,b)$ to be $$c(a,b) = -\langle a | \mathcal{Z H}^{-1} | b \rangle$$ which can be written in terms of the connectivity amplitudes $p({\mathbf{r}},{\mathbf{n}}|{\mathbf{r}}_0, {\mathbf{n}}_0)$ as $$c(a,b) = \int \int a({\mathbf{r}},-{\mathbf{n}}) p({\mathbf{r}},{\mathbf{n}}|{\mathbf{r}}_0, {\mathbf{n}}_0) b({\mathbf{r}}_0,{\mathbf{n}}_0)$$ So, $c(a,b)$ is the path integral over all path connecting region $a$ with region $b$ according to the Lagrangian $L$. Accordingly, we can also define $c^\kappa$ and $c^\text{lin}$ corresponding to $p_\text{lin}$ in equation and $p_\kappa$ in equation , respectively. The path trail image is $$\begin{aligned} \bar{\tau}_{(a|b)}({\mathbf{r}},{\mathbf{n}}) &=& \langle a | \mathcal{H}^{-1} e_{{\mathbf{r}},{\mathbf{n}}} \rangle \langle e_{{\mathbf{r}},{\mathbf{n}}} | \mathcal{H}^{-1} b \rangle \\ &=& \langle e_{{\mathbf{r}},{\mathbf{n}}} | \mathcal{ZH}^{-1}\mathcal{Z} a \rangle \langle e_{{\mathbf{r}},{\mathbf{n}}} | \mathcal{H}^{-1} b \rangle \\ &=& \langle e_{{\mathbf{r}},-{\mathbf{n}}} | \mathcal{H}^{-1}\mathcal{Z} a \rangle \langle e_{{\mathbf{r}},{\mathbf{n}}} | \mathcal{H}^{-1} b \rangle\end{aligned}$$ Note that $\mathcal{Z} \bar{\tau}_{(a|b)} = \bar{\tau}_{(\mathcal{Z}b|\mathcal{Z}a)}$. Usually the seed functions $a$ and $b$ do not depend on ${\mathbf{n}}$ because we do not have any preference for the starting orientation, so $\mathcal{Z}a = a$ and $\mathcal{Z}b = b$. Several relationships are then simplified, e.g. $\mathcal{Z} \bar{\tau}_{(a|b)} = \bar{\tau}_{(b|a)}$ and $c(a,b) = \langle a | \mathcal{H}^{-1} | b \rangle$. For the special case if $a_{{\mathbf{r}}_0}({\mathbf{r}},{\mathbf{n}}) = \delta({\mathbf{r}}-{\mathbf{r}}_0)$ we can also define the connectivity measure between two spatial locations ${\mathbf{r}}$ and ${\mathbf{r}}_0$ as $$c({\mathbf{r}}_0,{\mathbf{r}}_1) = c(a_{{\mathbf{r}}_0},a_{{\mathbf{r}}_1})\label{eq:conampspat}$$ In all experiments we will report the normalized connectivity amplitudes $$c_n(a,b) = \frac{c(a,b)}{\sqrt{c(a,a) c(b,b)}}$$ which is quite natural and reminds one of Pearson’s correlation coefficient. Angular Constraints, the Data Term and Boundary Conditions ---------------------------------------------------------- The Lagrangian penalizes bending, which is a good prior for straight fibers, but underestimates curved fibers. In [@Behrens2007] this problem does not appear, because the walker is always back projected onto the nearest fiber direction in a voxel. Such a behavior cannot be realized by a Wiener process. However, we can mimic such a behavior by an angular constraint such that the walkers do not deviate too strong from the main fiber directions. That is, we only simulate paths where ${\mathbf{n}}$ is not too far away from the main fiber directions, which is similar to the maximum angle thresholds known from ordinary streamline algorithms, but not so rigorous (it is still possible to find paths where the spatial tangent ${\dot{{\mathbf{r}}}}$ to the path is not perfectly ${\mathbf{n}}$). Now, we are able to set $\sigma_n$ pretty large, i.e. low penalty on bending, but the walker will be along the main fiber directions. In fact, this idea will also help to keep the running time and memory consumption of our algorithm in a reasonable range, because we only have to simulate the domain in the neighborhood of the fiber directions and not on the complete sphere. Note that the angular constraint may also be seen as an additional cost in the Lagrangian, i.e. the forbidden path are assigned with infinite costs. There are lots of ways to determine the speed function $f$, which represents the DW-data. For example, we could directly use the fiber orientation distributions (FOD), e.g. estimated by spherical deconvolution [@Tournier2007]. Or, one could estimate the main fiber directions by a local maxima detection of an FOD, and use them as anchor directions to construct the speed function. We opt for latter, primarily because the angular constraints can be much beter controlled. Let ${\mathbf{d}}_i$ be the local maximas of the FOD. Then, we construct the speed function by $$f({\mathbf{n}}) = \sum_{i=1}^N ({\mathbf{n}} \cdot {\mathbf{d}}_i)^{2n} \label{eq:speedfun}$$ with some fixed $n>0$. The simulation domain is now defined by thresholding this speed function, i.e. we only consider regions with $f>\epsilon$ for some $\epsilon>0$. To solve the FPE we have to set some boundary conditions. It is natural to let the walkers die once they touch the boundary of the domain, which is equivalent to a zero Dirichlet conditions at the boundary. So the complete problem we want to solve is $$-\mathcal{H} p = a \ \text{ where }\ p(\partial \Omega) = 0 \label{eq:problem}$$ where $\partial \Omega$ denotes the boundary of the simulation domain, $p$ the unknown steady state solution and $a$ the seed region where the walkers are emitted. Of course, other boundary condition, like Neumann conditions, might be applied, for example at the transitions to the ventricles. However, we restrict here all considerations to the simple Dirichlet conditions. Implementation, Discretization and Solver ----------------------------------------- In order to solve the equation $-\mathcal{H} a = b$ we have to discritize the operator $\mathcal{H}$. The convective part of $\mathcal{H}$ makes this a pretty hard problem, usual Finite Element Methods are only applicable for diffusion dominated problems. So, we decided to use a finite difference upwind-downwind scheme. However, such schemes are known to introduce errors (known as false diffusion) when the convection direction is not aligned with the underlying discretization grid. In particular for steady state solutions these problems become severe and the solutions show heavy orientation dependency. Fortunately our problem is special, the convection direction is always fixed for a volume $p({\mathbf{r}},{\mathbf{n}})$ with fixed ${\mathbf{n}}$, namely ${\mathbf{n}}$. So, we propose to steer the grid for each direction along the current convection direction. Practically, we discretize the sphere into $N$ direction homogeneously distributed over the full sphere (in the experiments we used $N=128$ directions). So, we have $N$ volumes, where the discretization grid of the volume associated with the $i$th direction ${\mathbf{n}}_i$ is steered such that one of the coordinate axis is along direction ${\mathbf{n}}_i$. In this way we are able to produce quite clean convection, however the angular diffusion gets disturbed. A voxel at some discrete coordinate in volume $i$ does not have a unique partner in volume $j$ at exactly the same discrete position. We need this partner in order to implement the spherical Laplace-Beltrami operator ${\Delta}_{{\mathbf{n}}}$, which only acts on the spherical coordinate. The simplest way to account for that is to use trilinear interpolation, which, unfortunately causes another unwanted spatial diffusion effect. However, this false diffusion is not so disturbing: it is independent of the orientation and the error is proportional to the angular diffusion and not to the speed function. Second, we can reduce the error by a finer discretization of the spatial grid. We quantified this additional spatial diffusion in an experiment and found that a angular displacement of $\pi = 180^\circ$ (for a $N=128$ points on the sphere) causes an additional spatial displacement of about $5$ spatial grid units, i.e. the cost of a $180^\circ$ turn on the spot compares to a 5 voxel jump. So, for some given theoretical $\sigma_n$ and $\sigma_r$, the actual $\sigma_r^\text{act}$ turns out to be approximately $\sigma^\text{act}_r = \sqrt{\sigma_r^2 + (\sigma_n 5/\pi)^2}$, if $\sigma_r$ is given in grid units. As this is already pretty much we decided to set the theoretical $\sigma_r$ to zero in all experiments. In general, for arbitrary number of points on the sphere, we found the approximation $\sigma^\text{act}_r = \sqrt{\sigma_r^2 + (\sigma_n \sqrt N/7)^2}$, i.e. with a finer sphere discretization the effects get stronger. But again, note that the ’physical’ amount of this false diffusion depends also on the spatial grid size. All the gritty details about the discretization scheme are given in Appendix \[app2\]. The whole matrix is scattered with MATLAB’s sparse matrix capabilities. To solve the discretized equation we found the GMRES algorithm the most stable and efficient [@saad1986gmres], we just used the MATLAB’s `gmres` command. As number of restarts we used $5$. The tolerance value is set to $10^{-6}$ in all experiments. The running time depends on the size of the problem. If we use the setting as described below in the experiments we get usually a system with about $2-3\cdot 10^6$ variables, which is scattered in $1-2$ minutes on a common Desktop PC (Intel I7, 16GB). Solving the equation also takes about one minute. Experiments =========== We want to begin with a small demonstration of the proposed discretization scheme. Let ${\mathbf{d}}$ some arbitrary direction and the speed function is constructed according to with $n=25$ and $\epsilon = 0.02$. In Figure \[fig1\] we choose ${\mathbf{d}} = (cos(\varphi),sin(\varphi),0)$ with $\varphi=0.3 \pi$ (a,b,e) and $\varphi=\pi$ (c,d,f). Further we choose $\sigma_r=0$ and $\sigma_n=\pi/12$. Figure \[fig1\] (a-d) shows the solution to with $a({\mathbf{r}}, {\mathbf{n}}) = \delta({\mathbf{r}}-{\mathbf{r}}_0)$ where ${\mathbf{r}}_0$ is located in the lower left corner of the simulation domain. The spatial intensity maps in Figure \[fig1\] show the solution $p$ integrated over the angular variable $\int_{S_2} p({\mathbf{r}},{\mathbf{n}}) d{\mathbf{n}}$, i.e. we actually show the connectivity amplitudes $c({\mathbf{r}},{\mathbf{r}}_0)$ as defined in . The matrix size is of size $30\times30\times7$. In Figure \[fig1\]a,b,c) we used a sphere with N=128 directions in (d) with N=256 directions. Further, we demonstrate the effect of spatial oversampling, i.e. a upsampling by a factor of $2$ means that the actual matrix size is in the range of $60\times60\times14$ (for details see Appendix \[app2\]). By comparing (a) and (b) we can obviously see how the the upsampling reduces false diffusion. Similarly comparing (b), (c) and (d), we see that the amount of false diffusion does not depend on the direction, but the directions are not perfect due to the discretization of the sphere. For comparison we also simulated an ordinary up/downwind scheme in (e),(f) with $\sigma_n = \sigma_r = 0$. The amount of false diffusion depends dramatically on the direction. Numerical Phantom ----------------- For further demonstration we constructed a numerical phantom consisting of crossing and bending configurations (see Figure \[fig2\]d)). Six seed locations ${\mathbf{r}}_1,\hdots,{\mathbf{r}}_6$ were selected such that they are pairwise connected as (1-2), (3-4) and (5-6). We did not simulate the MR-signal, but created the underlying directions directly. The directions ${\mathbf{d}}_i$ are created continuously (not aligned with the discretized sphere directions), and a pseudo FOD is generated using $\sum_i {\mathit{e}^{(}}\alpha (({\mathbf{d}}_i {\mathbf{n}})^2 - 1)$, which is shown in Figure \[fig2\]d). For robustness analysis we distorted the directions by Gaussian noise of standard deviation $\sigma_{\text{nz}}$. Figure \[fig2\]a-c) gives some first results of our approach. We show the connectivity amplitudes $c({\mathbf{r}},{\mathbf{r}}_i)$ as images, where ${\mathbf{r}}_i$ is one of the fixed seed point. Figure \[fig2\]a) shows $c$ without noise and b) with noise ($\sigma_{\text{nz}}=0.1$). Figure \[fig2\]c) shows the same experiment but with exponential length bias correction of $\kappa = 0.01$. Examples for the path trails images are shown in Figure \[fig2\]e) that correspond to the experiment in Figure \[fig2\]b). Trail images for all pairs of seed combinations are shown, the diagonal shows again the connectivity amplitudes but in logarithmic scaling. To get a quantitative picture we repeated the above experiment for different noise levels $\sigma_{\text{nz}} = 0.1,0.2$ and different discretization scheme, namely spherical discretization with $N=128$ and $N=256$ directions and with and w/o spatial upsampling (with a factor of $2$). Additionally, the orientation of the discrete sphere was randomly changed for each run. As reference approach we followed [@Behrens2007] (for details see Appendix \[app3\]), with $\Delta s = 1, \sigma=0.2$, without revisits and with length bias correction. Figure \[fig3\] shows barplots of the normalized connectivity amplitudes $c_n({\mathbf{r}}_i,{\mathbf{r}}_j)$ for all pairs of seeds averaged over 50 runs together with the standard deviation. In-Vivo Human Brain ------------------- To investigate our approach on real DTI data we considered 28 scans of healthy volunteers at a b-value of $1 \text{ms}/\mu m^2$ with 61 diffusion directions and an isotropic resolution of $2mm^3$. A white matter probability map was generated with SPM (Version 8, *http://www.fil.ion.ucl.ac.uk/spm/*, [@SPMbook2007]) on a T1-weighted scan, which was co-registered to the $b_0$-scan of the diffusion sequence. For each subject the scans were repeated two times (in two different sessions) to allow to investigate the robustness of the approach. To compute the FOD a constrained spherical deconvolution was used together with a $L_1$-regularizer [@Michailovich2011]. The fiber response function (FRF) was chosen as $FRF({\mathbf{n}}) = {\mathit{e}^{(}}-b D_1 (n^\text{fib}_i n_i)^2)$ with $D_1 = 1 \mu m^2/ms$ for all subjects and reconstructed on a $N=128$ sphere. The local maximas of the FOD are determined by fitting a quadratic form to the neighborhood of the discrete local maxima (the same approach as used in [@reisert2011]). Spurious local maximas were filtered out by neglecting all maximas smaller than $20\%$ of the global one. The speed function is constructed in the same way as for the numerical phantom. The spherical discretization is kept at $N=128$ and no spatial up sampling is applied. The probabilistic white matter segmentation of SPM was thresholded at $0.5$ and used as white matter mask. To compare and analyze the group, the AAL atlas [@tzourio2002automated] was registered to the subjects’ native space, normalization was done with SPM 8. The first 90 ROIs (distributed over the neocortex) of the AAL atlas were used to compute connectivity matrices. Figure \[fig4\]a-e) show examples of the mean connectivity matrices (CM) over the whole group. The first 45 ROIs belong to the right hemisphere, the last 45 the right hemisphere. The CM is shown in logarithmic scaling. To get a comparable contrast we used the following formula $c_\text{log}(a,b) = \log(t+c_n(a,b))$ where $t$ is the $20\%$ quantile of $c_n$ over all regions. The CMs obtained by our approach are compared to the reference approach [@Behrens2007] (Figure \[fig4\]d-e)). To investigate the robustness we computed the intraclass correlation coefficient (ICC)[^1] for each connectivity value (c-value) in the CM. In Figure \[fig4\]h) we show the ICC over all c-values in the CM sorted by ICC magnitude for different settings: we varied $\sigma_n$ and show one result with exponential weighting ($\kappa=0.01$). For all results we also show the corresponding linear weighted result $c_n^\text{lin}$. Figure \[fig4\]i) compares our approach with [@Behrens2007] with different parameters: $\sigma = 0.2,0.5, \Delta s = 1$, with (w/o) revisits and length bias correction (lbc). In Figure \[fig4\]f,g) we show for our approach and the reference the ICCs for the individual c-values as a matrix. Table \[tab1\] shows mean and median ICCs for our approach and the reference with more different settings. It is common to put thresholds to get a binary decision of connectivity. To investigate the robustness of such a thresholding operation an agreement measure is calculated as follows: let $t$ be some threshold, then $a = |c_1<t \wedge c_2 <t|$ denotes the number of regions that are not connected in scan $1$ and in scan $2$, where the number is counted over all possible ROI pairs and subjects. The agreement value $a$ is normalized by the total number of non-connected regions $s = |c_1 <t| + |c_2<t|$, and hence, $a/s$ is a number between $0$ and $1$. If we now vary over all thresholds $t$ and plot $a/s$ as a function of $s$ we get plots displayed in Figure \[fig4\]j). Finally, Figure \[fig5\] shows examples of path trail images for two ROI pairs of the AAL atlas. The two ROIs are highlighted in green and blue, respectively. The white matter mask is depicted in dark gray. The path trail $ \bar{\tau}_{(a|b)}({\mathbf{r}},{\mathbf{n}})$ is obviously a function in the joint position/orientation space, which is shown in glyph representation, but note that the glyphs are not pointsymmtric as one used to observe. The traversal direction is in this example from the blue to the green ROI. **Method** **Mean** **Median** ------------------------------------------------ ---------- ------------ Our Approach $\sigma_n = 0.2$ 0.47 0.51 $\sigma_n = 0.5$ 0.64 0.69 $\sigma_n = 1.25$ 0.73 0.78 $\sigma_n = 0.5, \kappa = 0.01$ 0.67 0.72 $\sigma_n = 0.2$, lin. weighted 0.50 0.54 $\sigma_n = 0.5$, lin. weighted 0.67 0.71 $\sigma_n = 1.25$, lin. weighted 0.73 0.78 $\sigma_n = 0.5, \kappa = 0.01$, lin. weighted 0.69 0.73 Reference [@Behrens2007] $\sigma=0.2, \Delta s = 1$ 0.65 0.68 $\sigma=0.5, \Delta s = 1$ 0.55 0.63 $\sigma=0.2, \Delta s = 1$, w/o revisits 0.61 0.64 $\sigma=0.2, \Delta s = 1$, no lbc 0.64 0.67 $\sigma=0.2, \Delta s = 0.25$ 0.55 0.58 $\sigma=0.2, \Delta s = 0.5$ 0.60 0.63 : Quantitative Comparison by ICC[]{data-label="tab1"} Discussion ========== We proposed to use a Wiener process for the estimation of structural brain connectivities. The path integral perspective gave us the insight why the usual random walker principle leads to asymmetric connectivities. To symmetrize the solution an additional potential was introduced, which symmetrized the path integral and the corresponding Fokker Planck equation. However, particle conservation and, hence, the usual walker perspective is lost. To solve the FPE we introduced a novel discretization scheme which avoids the directional dependency of the false diffusion introduced by the discretization of the convection operator. Initial experiments (in Figure \[fig1\]) show that the new discretization approach is valid. In Figure \[fig2\] the visual results on the numerical phantom show the approach works also nicely under noisy conditions and that the exponential length bias correction is reasonable. Quantitative experiments show that it leads to a robust connectivity measure. Compared to [@Behrens2007] the new approach shows smaller connectivity values for bending configurations, which is due to the bending penalty which is inherent to the new approach (see Figure \[fig3\]). One can also see that the robustness of our approach does not really depend too much on the discretization granularity. Compared to [@Behrens2007] it is in most cases higher (lower standard deviation over the trials). For stronger noise ($\sigma_\text{nz}=0.2$) our approach gives relatively higher c-values to false connections, but the c-values itself are more stable. Another difference to [@Behrens2007] is the way the curvature is controlled. If we want for our approach a higher connectivity for bending fibers (ROIs (5,6)), then we also get higher c-values for (1,3) and (2,4). For [@Behrens2007] it is different, the step directions are always back projected onto the nearest most likely fiber direction, which results in higher c-values for the bending (5,6). To investigate the new approach in-vivo, we considered a group of volunteers which were all scanned twice. We did a ROI-based approach based on the AAL-atlas. Overall, the CMs obtained from our approach and the reference are similar. As one might expect, with our approach strong bending fibers (like the connection between ’Occipital Sup L’ to ’Occipital Sup R’) are assigned smaller connectivity values. Analysis of the intraclass correlations of c-values shows that the new approach is quite robust. Usually higher angular diffusion $\sigma_n$ gives more robust c-values. Robustness is further increased by a length bias correction, either linear or exponentially. Compared to [@Behrens2007] our approach shows good performance. For [@Behrens2007] increasing angular diffusion does not help so much. One can see that in regions with low connectivity [@Behrens2007] has problems (see Figure \[fig4\]f,g)), which goes inline with our finding in the numerical phantom. We found [@Behrens2007] to work best for $\Delta s = 1$, $\sigma = 0.2$, revisits are allowed and with length bias correction (see Table \[tab1\]). In this setting the median ICC is $0.68$ which is acceptable, $50\%$ of the c-values have a ICC higher than $0.68$. The highest value we can achieve with our approach is a median of $0.78$. The agreement scores obtained from thresholded c-values show a similar but more pronounced picture (see Figure \[fig4\]j)). From Figure \[fig5\] it can also be seen that the estimated path trail images are meaningful and follow the expected pattern. Appendix {#appendix .unnumbered} ======== Operator Formalism {#app1} ------------------ The space of square integrable function on ${\mathbb R}^n$ is a Hilbert space $H$ with the inner-product $$\langle q|p \rangle = \int_{{\mathbb R}^n} {{q}^*}({\mathbf{x}}) p({\mathbf{x}}) d{\mathbf{x}}$$ Sometimes we adopt Dirac’s bra-ket notation, i.e. by $|p\rangle$ we denote a vector in $H$, by $\langle p|$ a linear 1-form. The adjoint of a linear operator $\mathcal{A}$ is defined to be the operator $\mathcal{A}^+$ such that $\langle q|\mathcal{A} p \rangle = \langle \mathcal{A^+} q|p \rangle$ for all functions $q,p \in H$. Discretization {#app2} -------------- Let $m({\mathbf{r}})$ some predefined white matter mask. For each direction ${\mathbf{n}}_i$ on the sphere, we create a voxel grid ${\mathbf{r}}^i_{(x,y,z)}$ such that the $x$-axis is aligned with direction ${\mathbf{n}}_i$. Only valid triples $(x,y,z) \in {\mathbb N}^3$ are considered where $m({\mathbf{r}}^i_{(x,y,z)})$ is on. The voxel spacing $h = |{\mathbf{r}}^i_{(x+1,y,z)}-{\mathbf{r}}^i_{(x,y,z)}|$ can be chosen arbitrarily. In the experiments we used $h=2mm$ which is resolution of the DTI scan, and $h=1mm$, that is an oversampling factor of 2. The sphere has a natural neighborhood system $\mathcal{N}(i)$ obtained by Voronoi tessellation. Let $p_{i,x,y,z}$ be some function on the introduced discrete domain. To discretize the spherical Laplace-Beltrami ${\Delta}_n$ operator we define the discrete operator $\tilde {\Delta}_n$ as $$(\tilde {\Delta}_n p)_{i,x,y,z} = \sum_{k\in \mathcal{N}(i)} \left( p_{i,x,y,z} -\sum_{(a,b,c)\in \atop \mathcal{T}({\mathbf{r}}^k_{(x,y,z)})} w_{k,a,b,c}\ p_{k,a,b,c} \right)$$ where $\mathcal{T}({\mathbf{r}}^k_{(x,y,z)})$ is neighborhood system needed for trilinear interpolation at position ${\mathbf{r}}^k_{(x,y,z)}$ in the $k$-th volume, and $w_{k,a,b,c}$ are the corresponding weights. However, the matrix $\tilde {\Delta}_n$ is not symmetric in the sense of the matrix transpose due to the one-sided interpolation. So the final discretized spherical Laplacian is $\frac{\sigma_n^2}{2A}(\tilde {\Delta}_n + \tilde {\Delta}_n^{T})$. The factor $A$ is empirically determined to get standard deviations of approximately $\sigma_n$. To implement the convection part, the speed function $f({\mathbf{r}},{\mathbf{n}})$ given in the native DTI frame is also transferred by trilinear interpolation onto the new coordinate frame $f_{i,x,y,z} = f({\mathbf{r}}^i_{(x,y,z)},{\mathbf{n}}_i)$. The simulation domain is determined by thresholding $f_{i,x,y,z}$. For the convection operator we have to distinguish between ’real’ boundaries and boundaries coming from the thresholding. The discretized convection generator $\tilde \nabla$ is defined as $$-(\tilde \nabla p)_{i,x,y,z} = \left\{ \begin{array}{ll} \frac{p_{i,x-1,y,z} - p_{i,x,y,z}}{h} & \text{otherwise}\\ p_{i,x-1,y,z}/h & \text{if }b_{i,x+1,y,z}=1 \\ -p_{i,x,y,z}/h & \text{if }b_{i,x-1,y,z}=-1 \\ -p_{i,x,y,z}/h & \text{if }b_{i,x-1,y,z}=1 \\ \end{array} \right.$$ where $b_{i,x,y,z}$ is an boundary indicator, where $b=1$ means the boundary is coming from thresholding and $b=-1$ from a ’real’ boundary. Finally, the convection part of the equation is $-(F \tilde \nabla + \tilde \nabla F)$, where $F$ is the matrix with $f_{i,x,y,z}$ on the diagonal. Reference Method {#app3} ---------------- For comparison we followed [@Behrens2007]. Suppose for some given fiber directions, which are always extracted in the same way as for our approach. The state of a walker is its current position ${\mathbf{r}}$ and its last step direction ${\mathbf{n}}$. The propagation law for a single walker is: The current voxel contains a set of fiber directions ${\mathbf{d}}_i$. Determine the fiber direction ${\mathbf{d}}$ which is closest to the current ${\mathbf{n}}$. If the angle between those is above a threshold of $80^\circ$ tracking is stopped. Otherwise, disturb this fiber direction by Gaussian noise ${\mathbf{d}}_n = {\mathbf{d}} + \eta_\sigma$ with deviation $\sigma$. Renormalize ${\mathbf{d}}_n$ to unity and track along this direction with step width $\Delta s$, i.e. ${\mathbf{r}}^\text{new} = {\mathbf{r}} + \Delta s {\mathbf{d}}_n$ and set ${\mathbf{n}}^\text{new} = {\mathbf{d}}_n$. If the new position is out of the tracking mask, stop tracking. If desired, one can also stop tracking whenever a voxel is visited more than once. During tracking we count how often a walker has visited a voxel resulting in a probabilistic map (PM). Alternatively, to account for the length bias, one can also weight with visitation count by the mean length of all walkers visited the voxel. To compute the connectivity value between ROI A and ROI B, $N$ walkers are ejected in each voxel in ROI A. The connectivity value is then the sum over the PM in ROI B. To get a symmetric value the connectivity from A to B is averaged with the connectivity from B to A. To set $N$ we computed the intra-scan ICC, i.e. for each subject *and* scan in the group we started the algorithm twice and computed the ICC between the two algorithm runs. We set $N$, such that the median ICC over all ROI pairs is above $0.95$, which resulted in $N=5000$ per voxel. [10]{} Lars Andersson and Bruce K Driver. Finite dimensional approximations to wiener measure and path integral formulas on manifolds. , 165(2):430–498, 1999. P. J. Basser, S. Pajevic, C. Pierpaoli, J. Duda, and A. Aldroubi. In vivo fiber tractography using [DT-MRI]{} data. , 44(4):625–32, 2000. 20481737 0740-3194 Journal Article. PG Batchelor, DLG Hill, D Atkinson, F Calamante, and A Connelly. Fibre-tracking by solving the diffusion-convection equation. In [*Proc. ISMRM*]{}, volume 10, 2002. T. E. J. Behrens, H. J. Berg, S. Jbabdi, M. F. S. Rushworth, and M. W. Woolrich. Probabilistic diffusion tractography with multiple fibre orientations: What can we gain? , 34(1):144–155, 2007. English Article 1053-8119. Mats Bj[ö]{}rnemo, Anders Brun, Ron Kikinis, and Carl-Fredrik Westin. Regularized stochastic white matter tractography using diffusion tensor mri. In [*Medical Image Computing and Computer-Assisted Intervention—MICCAI 2002*]{}, pages 435–442. Springer, 2002. M Chaichian and A Demichev. Path integrals in physics, vol. 1: Stochastic processes and quantum mechanics. , 2001. Detlef D[ü]{}rr and Alexander Bach. The onsager-machlup function as lagrangian for the most probable path of a diffusion process. , 60(2):153–170, 1978. Richard P Feynman and Albert R Hibbs. . Courier Dover Publications, 2012. Ola Friman, Gunnar Farneback, and C-F Westin. A bayesian approach for stochastic white matter tractography. , 25(8):965–978, 2006. Karl J. Friston, John T. Ashburner, Stefan J. Kiebel, Thomas E. Nichols, and William D. Penny, editors. . Elsevier, 2007. L Garrido, D Luri[é]{}, and M San Miguel. Stochastic quantization and detailed balance in fokker-planck dynamics. , 21(3):313–335, 1979. Nathan S Hageman, Arthur W Toga, Katherine L Narr, and David W Shattuck. A diffusion tensor imaging tractography algorithm based on navier–stokes fluid mechanics. , 28(3):348–360, 2009. P. Hagmann, J. P. Thiran, L. Jonasson, P. Vandergheynst, S. Clarke, P. Maeder, and R. Meuli. Dti mapping of human brain connectivity: statistical fibre tracking and virtual dissection. , 19(3):545–554, 2003. English 1053-8119. H Haken. Generalized onsager-machlup function and classes of path integral solutions of the fokker-planck equation and the master equation. , 24(3):321–326, 1976. Hidemi Ito. A characterization of the detailed balance from a viewpoint of the onsager-machlup theory. , 66(2):454–460, 1981. Derek K Jones, editor. . Oxford University Press, 2010. Mark Kac. On distributions of certain wiener functionals. , 65:1–13, 1949. Hagen Kleinert. . World Scientific, 2009. J. Mangin. A framework based on spin glass models for the inference of anatomical connectivity from diffusion-weighted mr data - a technical review. , 15(7-8):481–492, 2002. O. Michailovich. Spatially regularized compressed sensing for high angular resolution diffusion imaging. , 30:1100–1115, 2011. S. Mori, B. J. Crain, V. P. Chacko, and P. C. van Zijl. Three-dimensional tracking of axonal projections in the brain by magnetic resonance imaging. , 45(2):265–9, 1999. 99142739 0364-5134 Journal Article. G. J. M. Parker, H. A. Haroon, and C. A. M. Wheeler-Kingshott. A framework for a streamline-based probabilistic index of connectivity (pico) using a structural interpretation of mri diffusion measurements. , 18(2):242–254, 2003. English 1053-1807. M. Reisert and V. Kiselev. Fiber continuity: An anisotropic prior for odf estimation. , 30(6):1274–1283, June 2011. M. Reisert, I. Mader, C. Anastasopoulos, M. Weigel, S. Schnell, and V. Kiselev. Global fiber reconstruction becomes practical. , 54(2):955–62, 2011. M. Reisert, H. Skibbe, and V. Kiselev. Fiber density estimation by tensor divergence. , pages 297–304, 2012. Youcef Saad and Martin H Schultz. Gmres: A generalized minimal residual algorithm for solving nonsymmetric linear systems. , 7(3):856–869, 1986. J Tournier, Fernando Calamante, David G Gadian, Alan Connelly, et al. Diffusion-weighted magnetic resonance imaging fibre tracking using a front evolution algorithm. , 20(1):276–288, 2003. J.D. Tournier, F. Calamante, D.G. Gadian, and A. Connelly. Robust determination of the fibre orientation distribution in diffusion mri: Non-negativity constrained super-resolved spherical deconvolution. , 35(4):1459–1472, 2007. David S Tuch, John W Belliveau, and Van J Wedeen. A path integral approach to white matter tractography. In [*Proceedings of the 8th Annual Meeting of ISMRM, Denver*]{}, page 791, 2000. Nathalie Tzourio-Mazoyer, Brigitte Landeau, Dimitri Papathanassiou, Fabrice Crivello, Olivier Etard, Nicolas Delcroix, Bernard Mazoyer, and Marc Joliot. Automated anatomical labeling of activations in spm using a macroscopic anatomical parcellation of the mni mri single-subject brain. , 15(1):273–289, 2002. Nicolaas Godfried Van Kampen. , volume 1. Elsevier, 1992. Myron Zhang, Ken E Sakaie, and Stephen E Jones. Logical foundations and fast implementation of probabilistic tractography. , 2013. [^1]: If $c_1$ and $c_2$ are c-values in the CM for scan 1 and scan 2, respectively, then the ICC is $ICC(c) = \langle(c_1-\bar c)(c_2-\bar c)\rangle / \langle(c-\bar c)^2\rangle)$, where $\langle\rangle$ denotes the expectation over the whole group and $\bar c = \langle (c_1 + c_2)/2 \rangle$. The ICC is 1 if all the variance of the c-value can be explained by differences amongst the subjects.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We derive formulas for the matrix elements of the lattice Green function for the discrete Poisson equation on an infinite square lattice. The partial difference equation for the matrix elements is solved by reducing it to a series of first order difference equations, which can then be solved sequentially. These formulas are useful in solving two dimensional Poisson equation problems using the finite difference approximation.' author: - Stefan Hollos - Richard Hollos bibliography: - '/doc/articles/lgfint/gf.bib' title: The Lattice Green Function for the Poisson Equation on an Infinite Square Lattice --- introduction ============ In this paper we will derive formulas for the matrix elements of the lattice Green function for the discrete Poisson equation on an infinite square lattice. The discrete Poisson equation arises when the finite difference approximation is applied to the continuous Poisson equation. The results derived here will therefore be of considerable interest for solving any problem that can be modeled by a two dimensional Poisson equation [@cserti00]. One particularly large area of application is in the solution of two dimensional electrostatics and magnetostatics problems. Calculating capacitance, inductance, and charge distributions on conductors with boundary conditions at infinity are problems of great practical as well as some theoretical interest [@exstrom2005]. We will begin by introducing some notation and a formulation of the problem. We let $L$ denote the lattice Laplacian operator and then write the discrete Poisson equation as:$$L\vert\phi\rangle=\vert f\rangle\label{eq:1}$$ In the lattice basis we have basis vectors $\vert n\rangle$ associated with the lattice point $\vec{r}_{n}=n_{1}\vec{a}_{1}+n_{2}\vec{a}_{2}$, where $\vec{a}_{i}=a\hat{x}_{i}$, $\hat{x}_{i}\cdot\hat{x}_{j}=\delta(i,j)$, and $n_{i}$ is an integer. In this basis eq. \[eq:1\] becomes$$\sum_{n}L_{ln}\phi(\vec{r}_{n})=f(\vec{r}_{l})\label{eq:2}$$ The Laplacian matrix elements are given by$$L_{ln}=-4\delta(\vec{r_{l},}\vec{r}_{n})+\delta(\vec{r}_{l}+\vec{a}_{1},\vec{r}_{n})+\delta(\vec{r}_{l}-\vec{a}_{1},\vec{r}_{n})+\delta(\vec{r}_{l}+\vec{a}_{2},\vec{r}_{n})+\delta(\vec{r}_{l}-\vec{a}_{2},\vec{r}_{n})\label{eq:3}$$ The Green function is defined by the equation $LG=-I$ which in the lattice basis is$$\sum_{n}L_{ln}G_{nm}=\sum_{n}L(\vec{r}_{l}-\vec{r}_{n})G(\vec{r}_{n}-\vec{r}_{m})=-\delta(l,m)\label{eq:4}$$ Using eq. \[eq:3\] gives the following recurrence for the matrix elements of $G$.$$4G(\vec{r}_{l}-\vec{r}_{m})-\sum_{i=1}^{2}\left[G(\vec{r}_{l}+\vec{a}_{i}-\vec{r}_{m})+G(\vec{r}_{l}-\vec{a}_{i}-\vec{r}_{m})\right]=\delta(l,m)\label{eq:5}$$ With the notation $\vec{r}_{l}-\vec{r}_{m}=(l_{1}-m_{1})\vec{a}_{1}+(l_{2}-m_{2})\vec{a}_{2}=p_{1}\vec{a}_{1}+p_{2}\vec{a}_{2}$, and $G(\vec{r}_{l}-\vec{r}_{m})=G(p_{1},p_{2})$, eq. \[eq:5\] becomes$$4G(p_{1},p_{2})-G(p_{1}+1,p_{2})-G(p_{1}-1,p_{2})-G(p_{1},p_{2}+1)-G(p_{1},p_{2}-1)=\delta(p_{1},0)\delta(p_{2},0)\label{eq:6}$$ For an infinite lattice, an eigenbasis expansion gives the following formula for the matrix elements of $G$ [@hollos05_2]$$G(p_{1},p_{2})=\frac{1}{2\pi^{2}}\int_{0}^{\pi}\int_{0}^{\pi}\frac{\cos x_{1}p_{1}\cos x_{2}p_{2}}{2-\cos x_{1}-\cos x_{2}}dx_{1}dx_{2}\label{eq:7}$$ The problem with this equation is that the integral is divergent for all values of $p_{1}$and $p_{2}$. We get around this problem by using the origin referenced Green function $g(p_{1},p_{2})=G(0,0)-G(p_{1},p_{2})$ which is then given by the following integral$$g(p_{1},p_{2})=\frac{1}{2\pi^{2}}\int_{0}^{\pi}\int_{0}^{\pi}\frac{1-\cos x_{1}p_{1}\cos x_{2}p_{2}}{2-\cos x_{1}-\cos x_{2}}dx_{1}dx_{2}\label{eq:8}$$ This integral is finite for all finite values of $p_{1}$ and $p_{2}$. $g(p_{1},p_{2})$ will also provide a solution to eq. \[eq:1\] provided that the sum of the source terms, $f(\vec{r}_{n})$ over all the lattice sites is equal to zero. To see this, note that the solution to eq. \[eq:1\] in the lattice basis is given by$$\phi(\vec{r}_{l})=-\sum_{n}G_{ln}f(\vec{r}_{n})\label{eq:9}$$ Now if we have$$\sum_{n}f(\vec{r}_{n})=0\label{eq:10}$$ Then clearly$$\phi(\vec{r}_{l})=\sum_{n}g_{ln}f(\vec{r}_{n})\label{eq:11}$$ Since the coefficients of the recurrence in eq. \[eq:6\] add to zero we see that $g(p_{1},p_{2})$ must obey the equation$$4g(p_{1},p_{2})-g(p_{1}+1,p_{2})-g(p_{1}-1,p_{2})-g(p_{1},p_{2}+1)-g(p_{1},p_{2}-1)=-\delta(p_{1},0)\delta(p_{2},0)\label{eq:12}$$ We will now show how to solve this equation for $g(p_{1},p_{2})$. general solution of the difference equation =========================================== The key to solving eq. \[eq:12\] is the use of the symmetry $g(p_{1},p_{2})=g(p_{2},p_{1})$. This will allow eq. \[eq:12\] to be reduced to a series of first order difference equations for the matrix elements along the subdiagonals. The value of the elements along the main diagonal will then provide a unique solution to the problem. These values can be directly calculated via eq. \[eq:8\] [@hollos05_1]. We begin by letting $p_{1}=p_{2}=n$ in eq. \[eq:12\] and using the fact that $g(p_{1},p_{2})=g(p_{2},p_{1})$, we get$$g(n+1,n)+g(n,n-1)=2g(n,n)+\frac{1}{2}\delta(n,0)\label{eq:13}$$ This equation is really just a first order difference equation for the first subdiagonal matrix elements. To make this observation explicit we introduce the following notation: $r_{0}(n)=g(n,n)$, and $r_{1}(n)=g(n,n-1)$, $r_{1}(n+1)=g(n+1,n)$. Eq. \[eq:13\] then becomes$$r_{1}(n+1)+r_{1}(n)=2r_{0}(n)+\frac{1}{2}\delta(n,0)\label{eq:14}$$ This equation for the first subdiagonal matrix elements, $r_{1}(n)$, is easily solved once the diagonal matrix elements, $r_{0}(n)$, are known. The initial condition for the equation is $r_{1}(1)=r_{1}(0)$. This comes from the definition, $r_{1}(n)=g(n,n-1)$, from which we get $r_{1}(1)=g(1,0)$ and $r_{1}(0)=g(0,-1)=g(1,0)$. The value of the diagonal matrix elements was previously determined to be [@hollos05_1]$$g(n,n)=\frac{1}{\pi}\sum_{k=1}^{n}\frac{1}{2k-1}\label{eq:15}$$ The solution to eq. \[eq:14\] can of course be found as the sum of the solutions to the two equations:$$r_{1}(n+1)+r_{1}(n)=2r_{0}(n)\label{eq:16A}$$ $$r_{1}(n+1)+r_{1}(n)=\frac{1}{2}\delta(n,0)\label{eq:16B}$$ The solution that comes from eq. \[eq:16A\] is the complementary solution of eq. \[eq:12\]. The solution corresponding to eq. \[eq:16B\] will have $r_{0}(n)=0$ for all $n$. For the next subdiagonal we let $p_{1}=n$ and $p_{2}=n-1$ in eq. \[eq:12\] and we let $r_{2}(n)=g(n,n-2)$ to get:$$r_{2}(n+1)+r_{2}(n)=4r_{1}(n)-r_{0}(n)-r_{0}(n-1)\label{eq:17}$$ and in general for the $k$th subdiagonal elements we will have $$r_{k}(n+1)+r_{k}(n)=4r_{k-1}(n)-r_{k-2}(n)-r_{k-2}(n-1)\label{eq:18}$$ where $r_{k}(n)=g(n,n-k)$. This equation can easily be solved to give $r_{k}(n)$ in terms of $r_{k-1}(n)$ and $r_{k-2}(n)$. $$r_{k}(n)=(-1)^{n-k+1}(r_{k-2}(k-2)+r_{k-2}(k-1))-r_{k-2}(n-1)-4\sum_{j=1}^{n-1}(-1)^{j}r_{k-1}(n-j)\label{eq:19}$$ We apply this equation iteratively with $r_{1}(n)$ given by either eq. \[eq:16A\] or eq. \[eq:16B\] to find a closed form solution for $r_{k}(n)$. First we look at the case where $r_{1}(n)$ is given by eq. \[eq:16B\]. This leads to the following solution:$$\begin{aligned} r_{0}(n) & = & 0\label{eq:20}\\ r_{1}(n) & = & (-1)^{n+1}\frac{1}{4}\nonumber \\ r_{2}(n) & = & (-1)^{n}(n-1)\nonumber \\ & \vdots\nonumber \\ r_{k}(n) & = & (-1)^{n+k}\frac{1}{4}\sum_{j=0}^{\lfloor\frac{k-1}{2}\rfloor}4^{k-2j-1}\begin{array}{c} \left(\begin{array}{c} k-j-1\\ j\end{array}\right)\left(\begin{array}{c} n-j-1\\ k-2j-1\end{array}\right)\end{array}\nonumber \end{aligned}$$ At this point we introduce a new function $s(n,m)$, defined as$$s(n,m)=\sum_{j=0}^{\lfloor\frac{n}{2}\rfloor}4^{n-2j}\begin{array}{c} \left(\begin{array}{c} n-j\\ j\end{array}\right)\left(\begin{array}{c} n-j+m/2\\ j+m/2\end{array}\right)\end{array}\label{eq:21}$$ In terms of this new function, eq. \[eq:20\] becomes$$r_{k}(n)=(-1)^{n+k}\frac{1}{4}s(k-1,2(n-k))\label{eq:22}$$ Now the complementary solution, that starts from eq. \[eq:16A\], can also be expressed in terms of this new function as:$$r_{k}(n)=r_{0}(n-l)+(-1)^{k}\frac{1}{\pi}\left[\sum_{j=1}^{n-l}(-1)^{n-j}\frac{s(k-1,2(n-j-k)+1)}{2j-1}+\sum_{j=1}^{l}(-1)^{n+j}\frac{s(k-2j,2(n+j-k)-1)}{2j-1}\right]\label{eq:23}$$ where $l=\lfloor\frac{k}{2}\rfloor$. We can verify that these two solutions for $r_{k}(n)$ satisfy eq. \[eq:18\] by direct substitution and the use of the following identities for the function $s(n,m)$.$$s(n,m)=s(n-1,m)+s(n-1,m+1)+s(n,m-1)\label{eq:24}$$ $$4s(n,m)=s(n-1,m)+s(n+1,m)-s(n-1,m+2)-s(n+1,m-2)\label{eq:25}$$ Eq. \[eq:24\] can be derived from the definition of $s(n,m)$ given in eq. \[eq:21\]. Eq. \[eq:25\] is gotten by applying eq. \[eq:24\] four times. Adding the two solutions for $r_{k}(n)$ and using $r_{k}(n)=g(n,m)$, with $m=n-k$, gives us the following equation for the general matrix element of the Green function:$$\begin{aligned} g(n,m) & = & h(l)+\frac{(-1)^{m}}{4}s(n-m-1,2m)\nonumber \\ & + & \frac{(-1)^{m}}{\pi}\sum_{j=1}^{l}(-1)^{j}\frac{s(n-m-1,2(m-j)+1)}{2j-1}\nonumber \\ & + & \frac{(-1)^{m}}{\pi}\sum_{j=1}^{\lfloor\frac{n-m}{2}\rfloor}(-1)^{j}\frac{s(n-m-2j,2(m+j)-1)}{2j-1}\label{eq:26}\end{aligned}$$ where $l=\lfloor\frac{n+m+1}{2}\rfloor$ and $h(l)=\frac{1}{\pi}\sum_{j=1}^{l}\frac{1}{2j-1}$. conclusion ========== The symmetry of the square lattice means that we have $g(n,m)=g(m,n)$. This allowed the difference equation for $g(n,m)$ to be reduced to a series of first order equations for the subdiagonal matrix elements. Solving these equations gave us an expression for the general Green function matrix element $g(n,m)$. The value of the diagonal matrix elements $g(n,n)$ determine a unique solution. This same procedure can also be used to find the lattice Green function for the discrete Helmholtz equation as long as the diagonal matrix elements can be calculated directly. An initial investigation indicates that it may also be possible to apply this same type of procedure to cubic and higher dimensional lattices. For a cubic lattice, the values along the main diagonal $g(n,n,n)$ will give a unique solution to the problem, and so on for higher dimensional lattices. Work on these problems is currently being pursued. The authors acknowledge the generous support of Exstrom Laboratories and its president Istvan Hollos.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Understanding the internal representations of deep neural networks (DNNs) is crucal to explain their behavior. The interpretation of individual units, which are neurons in MLPs or convolution kernels in convolutional networks, has been paid much attention given their fundamental role. However, recent research [@morcos2018importance] presented a counterintuitive phenomenon, which suggests that an individual unit with high class selectivity, called interpretable units, has poor contributions to generalization of DNNs. In this work, we provide a new perspective to understand this counterintuitive phenomenon, which makes sense when we introduce *Representative Substitution* (RS). Instead of individually selective units with classes, the RS refers to the independence of a unit’s representations in the same layer without any annotation. Our experiments demonstrate that interpretable units have high RS which are not critical to network’s generalization. The RS provides new insights into the interpretation of DNNs and suggests that we need to focus on the independence and relationship of the representations.' author: - | Li Chen$^\dagger$, Hailun Ding$^\ddagger$, Qi Li$^\uparrow$, Zhuo Li$^\uparrow$, Jian Peng$^\dagger$, Haifeng Li$^{\dagger*}$\ $^\dagger$School of Geosciences and Info-Physics , $^\ddagger$School of Software\ $^\uparrow$School of Information Science and Engineering\ Central South University\ Changsha Hunan 410083 P.R. China\ {vchenli, hlding, dsjliqi, pengj2017, lihaifeng}@csu.edu.cn,buebirzhuo@gmail\ $^*$Corresponding author bibliography: - 'references.bib' title: | Understanding the Importance of Single Directions\ via Representative Substitution --- Introduction ============ Deep neural networks (DNNs) have made remarkable progress in various domains, such as image classification, word recognition, and language translation. Despite the substantial success of DNNs, their working mechanism, such as the capability to memorize the entire datasets with low generalization error, is still an open issue. Inspired by neuroscience, recent works [@morcos2018importance] have demonstrated *high interpretability units* (HIUs), which are important response to a class of objects and were believed to play a crucial role in the performance of DNNs in the past, actually have an inconsiderable contribution to accuracy and even damage a network’s generalization performance. Instead, *weak interpretability units* (WIUs) may be important to improve the accuracy and generalization of DNNs. Different experiments, such as ablation and perturbation, show that a single direction has minimal importance to generalization. Therefore, previous studies of understanding the working mechanism of a network by analyzing HIUs may be misleading [@radford2017learning; @nguyen2016synthesizing; @zeiler2014visualizing]. The relationship among HIUs, WIUs, accuracy, and generalization remains unclear. Regardless of whether a unit is sensitive to class, latent patterns are found with the corresponding unit when the unit activation is large. Therefore, we take the activation maximization (AM) [@visualization_techreport] of units as their representation in the feature space. This representation propagates the image initialized with random noise to compute the activation of certain unit and then passes the activation backward to update the best direction of the input image. This method prevents the impacts of class and limited data samples in the activation space. We propagate the generated image forward the network and record all activation in the same layer. Although the generated image is optimized for the target unit, many units’ activation is larger than that of a certain unit. In this optimization direction, other units also have the same representations. We define *Representative Substitution* (RS), the ratio of units with activation greater than that of a target unit to the full layer, to quantify the independence of a unit’s representations in the same layer for understanding the behavior of HIUs and WIUs and the influence of their interactions on the accuracy and generalization of DNNs. We also improve AM to obtain independent representations, which reduces the influences of other units in the image generation process. Regardless of what generates an image, the HIUs’ RS is relatively high. When a unit is ablated, a large number of other units can complement the loss representations in the same layer; therefore, HIUs are inconsiderably important. We find why the relationship between class selectivity and unit importance is negatively correlated or irrelevant. The first few layers of the network are for feature extraction, whereas the remaining layers are for feature integration. HIUs and WIUs’ importance differs in different layers. The main contributions are summarized as follows: \(1) We forge a concept, called *Representative Substitution*, to indicate the independence of a unit’s representations in the same layer. The importance of HIUs to the network is explained through the perspective of feature space. \(2) The importance of HIUs is not a universal phenomenon; instead, it heavily depends on the layers they are located given different convolutional layers have different functions, feature extraction, and feature integration. \(3) Our method is built on the representations of convolution kernels without any information on classes or labels, which give us a potential opportunity to analyze the intrinsic features of convolution kernels. Our research suggests that independent representations of convolution kernels may be a promising way to explain the behavior of DNNs. Methods ======= We propose an *independent activation maximization* (IAM) which represents a unit in the feature space, and *Representative Substitution* (RS) which represents the independence of the unit’s representations in the same layer, to understand the characteristics of HIUs and WIUs and their impacts on the network importance. We introduce basic concepts before the IAM and RS are discussed. **Class Selectivity.** A metric is inspired by system neuroscience to quantify the class selectivity of units [@morcos2018importance]. The selectivity index is calculated as follows: $$\label{sel} selectivity(i) = \frac{\bar{x}^i_{max}-\bar{x}^i_{-max}}{\bar{x}^i_{max}+\bar{x}^i_{-max}},$$ where $\bar{x}^i_{max}$ is the maximum mean activation of the $i$-th unit under class and $\bar{x}^i_{-max}$ is the mean activation of the $i$-th unit across all other classes. The $i$-th unit mean activation of all samples under each class is calculated, and the vector of dimension equals to the number of classes. The class selectivity of the $i$-th unit can be obtained by Eq.\[sel\]. The metric varies from 0 to 1; 0 denotes that a unit activates identically for all classes, whereas 1 denotes that a unit only activates for a single class. In other words, a unit with high class selectivity value is an HIU and a unit with low class selectivity value is a WIU. **Activation Maximization.** As shown in [@visualization_techreport], the purpose of the algorithm is to visualize the representations of a network’s unit after the end of training. The *activation maximization* (AM) of the $i$-th unit is $$AM(i) = \arg\max {x^i}, %{h(x^i, \Theta^\ast)}$$ where $x^i$ is the activation of the $i$-th unit. In detail, we put a random noise image forward the network to compute the activation of the $i$-th unit and then propagate activation backward through the network to compute the gradient of the input image. After a few iterations, we can acquire the generated image; thus, the activation of the $i$-th unit becomes large. The generated image is considered as visual representations of the unit in the feature space. **Independent Activation Maximization.** In the process of image generation, the AM algorithm only follows consideration of the maximum activation value of the target unit. However, other units may also have high activation to the input image, which makes the visual representations entangled with other units. Thus, we modify the objective function to generate the following IAM formula, $$\label{iam} IAM(i) = \arg\max {(x^i - \bar x^{-i})}, %{(h(x^i, \Theta^\ast) - \bar h(x^{-i}, \Theta^\ast)) }$$ where $\bar x^{-i}$ is the mean activation of all units without the $i$-th unit in the same layer. The images are obtained by Eq.\[iam\], which can maximize the activation value of the $i$-th unit while the activation of other units is suppressed. Therefore, images become the independent visual representations of the unit. **Representative Substitution.** We generate an image for a certain unit by the IAM algorithm to obtain the unit’s representations and then pass the generated image to the network for quantifying the impacts of other units on the representation of the target unit in the same layer. The activation of all units in the same layer of a certain unit is recorded. The RS is defined by the number of units greater than the activation of the certain unit divided by the total number of units in that layer. The formula is $$RS(i) = \frac{|\{x^{'}:x^{'} > x^{i} \} |}{|\{x \} |} \quad s.t. \quad x^{i} = f(IAM(i)), %\footnote{what's f }$$ where $|\{x \} |$ represents the number of activation or units and $f(image)$ represents networks’ feedforward computation that the image can be obtained by IAM. The RS indicates the degree to which visual representations of a unit is activated by other units in the same layer. This metric varies from 0 to 1; 0 indicates that the visual representations of the $i$-th unit is independent and unique, whereas 1 indicating the all units have the same representations with the $i$-th unit. Experiments =========== **Experimental settings** In the experiments, we use various DNNs, including an MLP, a shallow convolutional network, VGG16 [@Simonyan2014Very] under CIFAR-10 [@Krizhevsky2009Learning], and ImageNet [@Deng2009ImageNet] datasets. Each layer of MLP contains 128, 512, 2048 and 2048 neurons. The shallow convolutional network is trained on CIFAR-10 for 100 epochs. Its convolutional layer sizes are 64, 64, 128, and 128. All kernels are 3$\times$3, with strides of 1. CIFAR-10 contains 10 classes with 60,000 images, and ImageNet has 1.2 million images from 1,000 classes. **Relationship with class selectivity** We train a shallow convolutional network on Cifar-10 and compute the class selectivity of all units. We then obtain the RS results under the IAM and the AM algorithms for comparison, and select two layers of the network, as shown in Fig.\[fig:vgg\_1\_3\]. The experimental results show that the RS is high when the class selectivity of the convolution kernels in the two layers is high. The WIUs with low class selectivity also have low RS. A low RS indicates independent and irreplaceable representations in the same layer. Therefore, the WIUs are important to network’s generalization. When the RS is obtained by the AM algorithm, the lower left corner of the figure is dense, whereas the upper right corner is sparse. This finding shows that the representations of WIUs in this layer are difficult to replace, whereas HIUs are easy to replace. We use the improved AM algorithm, the IAM algorithm, to compute the units’ RS. Although the overall RS declines slightly, the basic trend is consistent with that of AM, and the same conclusions are reached. HIUs are unimportant because other alternative representations exist. WIUs represent several independent features, and its ablation largely impacts the network generalization. ![**The HIU’s representations are simple compared with the representations of other units.** The HIU has high class selectivity and RS.[]{data-label="fig:vis"}](vis.jpg) We visualize an HIU and two units with higher activation than the HIU by the IAM algorithm, as shown in Fig.\[fig:vis\]. The representations of other units cover the HIU’s representations, but they are also given their own independent features. Thus, HIUs with simple representations are unsuitable for network generalization. In the latter experiment, we utilize the IAM algorithm to acquire the RS. **Relationship with different layers** Unfortunately, a unit of a good network must be a single class on the fully connected layer before the classifier. Although a HIU is given, its representations may not be substituted in the same layer. Therefore, we compute the RS and class selectivity of the fully connected layer of the shallow convolutional network, as shown in Fig.\[fig:vggful\]. The results indicate that the class selectivity of units in the fully connected layer before the classifier is large, and the RS is mostly zero. \[fig:vgg4\_c\_t\] HIUs are consequently inappropriate to indicate the quality of network generalization. The importance of HIUs is related to which layer it locates. We conduct a Spearman’s correlation analysis of the RS and class selectivity of each convolutional layer of the shallow convolutional network, as shown in Fig.\[fig:tend\]. The results demonstrate that the RS and class selectivity are positively correlated, and the correlation gradually decreases as the number of layers increases. So in [@morcos2018importance], the importance of unit ablation under different convolutional layers on the network is thus difficult to determine because the way of ablation is random. From a feature space prospective, as the number of layers increases, the independence of features on the unit gradually becomes unrelated to the interpretability of the unit. @morcos2018importance clarified that class selectivity has a strong relationship with importance in the shallow layer, but it gradually decays as layer deepens. The first few layers of the convolutional network have feature extraction capabilities; hence, the HIUs in these layers are unimportant because their RS are high and other convolution kernels can replace their representations. Therefore, too many HIUs will reduce the generalization performance of the network. As the number of layers increases, feature extraction is no longer part of the main influencing factor of network performance. The network needs to integrate features and provide semantic information. This finding is consistent with theoretical observations [@raghu2016expressive]. We conduct the same experiments on VGG16 trained on Cifar-10, as shown in Fig.\[fig:vggcifar\]. Similarly, we conclude that WIUs have low RS in shallow layers. However, the units’ RS presents a small value and no longer has diversity in deep layers. This phenomenon starts from a certain layer, and its division of the network layer is evident. The first few layers of the network are mainly used for feature extraction, and the main functions of the latter layers are feature integration for classification. We conduct experiments on an MLP trained on Cifar-10, as shown in Fig.\[fig:mlp\]. The RS is mostly the same and does not change considerably in all layers. Depending on previous conclusions, the function of the MLP is feature integration. The MLP is prone to overfitting and poor generalization. This finding explains that MLP performs poorly because it has no layer to extract features. We use the pre-trained VGG16 model and ImageNet to perform the same experiments, as shown in Fig.\[fig:vgg3-11\]. The results imply that a large number of WIUs assemble at the bottom left of the figure, which provide many independent representations of the network. The RS and class selectivity in all layers are shown in Fig.\[fig:vgg-r\]. The RS and class selectivity show no correlation. This finding differs from our observation. Further work will need to consider the impacts of RS and class selectivity on the network. Conclusions =========== This study clarifies the insignificant role of HIUs in DNNs from a feature space perspective, which inspires us to pay considerable attention to the independence and relationship of representations for explaining the working mechanism of DNNs. Convolutional layers actually have distinct functions, although they have the same structure. We need to consider the impact of the network structure when we study the interpretability of the network. However, the correlation between the RS and HIUs in a large dataset is not evident at the end of the experiment. Using a single metric to understand the internal representations of DNNs is thus unreasonable. Considering that negative values of images are meaningless, we analyze outputs by positive values. Nevertheless, negative activation in neuroscience can indicate an inhibition, which may be considered in the future.
{ "pile_set_name": "ArXiv" }
--- address: | Institute for Nuclear Research of the Russian Academy of Sciences,\ 60th October Anniversary prospect 7a, Moscow 117312, Russia author: - 'F. L. Bezrukov, D. S. Gorbunov' title: 'T-ODD CORRELATIONS IN $\pi\to e\nu_e\gamma$ DECAY' --- Introduction ============ T-violation beyond the Standard Model is usually searched for in decays forbidden by time reversal symmetry. Another way to probe T-violation is measurement of T-odd observables in allowed decays of mesons. Widely considered T-odd observables are transverse muon polarizations ($P_T$) in $K\to\pi\mu\nu$ and $K\to\mu\nu\gamma$ decays. There is no tree level SM contribution to $P_T$ in these decays, so they are of a special interest for search for new physics. Unfortunately, $P_T$ is not exactly zero in these decays even in T-invariant theory — electromagnetic loop corrections contribute to $P_T$ and should be considered as a background. There are no experimental evidence for nonzero $P_T$ in these processes at present time, but the sensitivity of the experiments has not yet reached the level of SM loop contributions. In this talk we discuss the decay $\pi\to e\nu\gamma$. Within the Standard Model, T-violation in this process does not appear at tree level, but interactions, contributing to it, emerge in various extensions of SM. We shall demonstrate that $\pi_{e2\gamma}$ decay is an attractive probe of new physics beyond the Standard Model. Depending on the model, $\pi_{e2\gamma}$ decay may be even more attractive than usually considered $K_{l2\gamma}$ decays. The similar decay $\pi_{\mu2\gamma}$ is analized in the article [@Bezrukov:2002zc], where one can also find the detailed bibliography. Though $\pi_{e2\gamma}$ decay has very small branching ratio (it is of order $10^{-7}$), we find that the distribution of the transverse electron polarization over the Dalitz plot significantly overlaps with the distribution of differential branching ratio, as opposed to the situation with $K_{\mu2\gamma}$ decay. Moreover, the contribution of FSI (final-state interactions related to SM one-loop diagrams) to the observable asymmetry, being at the level of $10^{-3}$, becomes even smaller in that region of the Dalitz plot, where the contribution from new effective T-violating interaction is maximal. Thus, $\pi_{e2\gamma}$ decay is potentially quite interesting probe of T-violation. To demonstrate that pion decays may be relevant processes where the signal of new physics may be searched for, we present a simple model of heavy pseudoscalar particle exchange leading in the low energy limit to the T-violating four fermion interaction. We find the constraints on the parameters of this model coming from various other experiments and describe regions of the parameter space which result in large T-violating effects in $\pi_{l2\gamma}$ decays. Depending on the parameters of the model, an experiment measuring transverse lepton polarization with pion statistics of $10^5\div10^{10}$ pions for $\pi_{\mu2\gamma}$ decay and $10^8\div10^{13}$ pions for $\pi_{e2\gamma}$ decay is needed to detect the T-violating effects (taking into account statistical uncertainty only and assuming ideal experimental efficiencies). T-violating effect in $\pi^+_{e2\gamma}$ decay {#sec:2} ============================================== ----- ----- ----- (a) (b) (c) ----- ----- ----- Let us consider the simplest effective four-fermion interaction $$\label{4-fermion} {\cal L}_{\mathrm{eff}} = G^e_{{{\scriptscriptstyle \rm P}}}\bar{d}\gamma_5u\cdot\bar{\nu}_e(1+\gamma_5)e +\mathrm{h.c}.$$ that may be responsible for the T-violating effects in pion physics beyond the Standard Model. Indeed, the imaginary part of the constant $G^e_{{{\scriptscriptstyle \rm P}}}$ contributes to transverse lepton polarization. For the transverse electron polarization asymmetry (polarization in the direction $\vec{e}_T = {\vec{q}\times \vec{p}_l}/|\vec{q}\times \vec{p}_l|$) $$\begin{gathered} P_T(x,y)={d\Gamma(\vec{e}_T)-d\Gamma(-\vec{e}_T) \over d\Gamma(\vec{e}_T)+d\Gamma(-\vec{e}_T)} =[\sigma_V(x,y)-\sigma_A(x,y)]\cdot{\mathop\mathrm{Im}\nolimits}[\Delta^l_P]\\ \Delta^e_P = \frac{\sqrt{2}G^e_P}{G_F\cos\theta_c}\cdot \frac{B_0}{m_l}\;,\qquad B_0=-{2\over (f^0_\pi)^2}{\langle}0|\bar{q}q|0{\rangle}= {m_\pi^2\over m_u+m_d}\approx2\mathrm{GeV}\;. \end{gathered}$$ The contour-plot of $[\sigma_V-\sigma_A]$ as a function of $x$ and $y$ is presented in the Figure (a). As one can see, in a large region of kinematic variables, $[\sigma_V-\sigma_A]$ is about $0.5$. This means that transverse electron polarization $P_T$ for this process, is of the same order as ${\mathop\mathrm{Im}\nolimits}[\Delta_P^e]\simeq5\times10^3\cdot{\mathop\mathrm{Im}\nolimits}[G_P^e/G_F]$. It is worth noting that the region of the Dalitz plot where large T-violating effect might be observed, significantly overlaps with the region where the partial decay width $\Gamma(\pi_{e2\gamma})$ is saturated (cf. Figures (a) and (b)). This is in contrast to the situation with T-violation in $K_{\mu2\gamma}$ decay, where the analogous overlap is small, so the differential branching ratio in the relevant region is smaller than on average. Theoretical background to the observation of $P_T$ caused by interaction of the type appeares from the contribution from final-state interactions (FSI) — one-loop diagrams with virtual photons. The value of this contribution in the Standard Model is presented in Figure (b). For the $\pi_{e2\gamma}$ decay, the ($x,y$)-distributions of FSI contribution and the contribution from the four-fermionic interaction (\[4-fermion\]) differ in shape. Specifically, part of the region with maximal $P_T$ from four-fermion interaction corresponds to the region of small $P_T$ from FSI. This implies that if measured, $P_T$ distribution could probe T-violating interaction (\[4-fermion\]) with an accuracy higher than ${\mathop\mathrm{Im}\nolimits}[\Delta^e_P]\sim 10^{-3}$ (${\mathop\mathrm{Im}\nolimits}[G_P^e/G_F]\sim2\times10^{-7}$). Again, this is not the case for $K_{\mu2\gamma}$ decay. Constraint from $\pi\to l\nu$ decays {#sec:2.2} ------------------------------------ The interaction term (\[4-fermion\]) not only gives rise to T-violation in $\pi\to e\nu_e\gamma$ decay but also contributes to the rate of $\pi\to l\nu_l$ decays. Since the ratio of leptonic decays of the pion has been accurately measured, the coupling constants $G_P^e$ is strongly constrained. This constraint can be evaded by introducing a similar coupling to muon and muon neutrino with the corresponding constant $G_P^\mu = \frac{m_\mu}{m_e}G_P^e$. Note that to the leading order in $\Delta_P$, only the real parts of the coupling constants $G_P^\mu$ and $G_P^e$ are constrained, while constraints on imaginary parts are weaker. Thus, for general $G_P^\mu$ and $G_P^e$ (if the mentioned hierarchy does not hold) one obtains $|{\mathop\mathrm{Re}\nolimits}[\Delta_P]|\lesssim10^{-3}$ and $|{\mathop\mathrm{Im}\nolimits}[\Delta_P]|\lesssim0.03$. Hence in this case experiments aimed at searching for T-violation in $\pi_{e2\gamma}$ decays should have sufficiently large statistics: the total number of charged pions should be $N_\pi\gtrsim10^{11}$. In models with ${\mathop\mathrm{Re}\nolimits}[G_P]\sim{\mathop\mathrm{Im}\nolimits}[G_P]$ and without $\mu-e$ hierarchy, the bound from $\pi\to l\nu$ decays implies $|{\mathop\mathrm{Im}\nolimits}[\Delta_P]|\lesssim10^{-3}$, which significantly constrains possible contribution of the new interaction (\[4-fermion\]) to T-odd correlation in $\pi_{e2\gamma}$ decay. Namely, the contribution to the $\pi_{e2\gamma}$ decay should be of the same order or weaker than one from the Standard Model FSI. Nevertheless, as we discussed, in the case of $\pi_{e2\gamma}$ the difference in ($x,y$) distributions of FSI and four-fermion contributions may allow one to discriminate between the two if they are of the same order of magnitude, and even if the contribution of four-fermion interaction (\[4-fermion\]) is somewhat weaker. To test four-fermion interaction (\[4-fermion\]) at the level allowed by $\pi\to e\nu_e$, i.e., at the level of $10^{-3}$, one has to collect not less than $10^{13}$ charged pions, assuming statistical uncertainty only. Overall, in the case of the hierarchy muon and electron couplings, decays $\pi\to l\nu$ do not constrain new T-violating interactions which can be searched for in relatively low statistics experiments, $N_\pi\gtrsim10^8$ for $\pi_{e2\gamma}$ and $N_\pi\gtrsim10^5$ for $\pi_{\mu2\gamma}$. In the worst case of no hierarchy and ${\mathop\mathrm{Re}\nolimits}G_P\sim{\mathop\mathrm{Im}\nolimits}G_P$, new T-violating interactions have little chance to be observed, and need high statistics experiments. Acknowledgments {#acknowledgments .unnumbered} =============== The authors are indebted to V. N. Bolotov, Yu. G. Kudenko and V. A. Rubakov for stimulating discussions. The work is supported in part by CPG and SSLSS grant 00-15-96626, by RFBR grant 02-02-17398 and by the program SCOPES of the Swiss National Science Foundation, project No. 7SUPJ062239. The work of F.B. is supported in part by CRDF grant RP1-2364-MO-02. The work of D.G. is supported in part by the RFBR grant 01-02-16710 and INTAS YSF 2001/2-142. References {#references .unnumbered} ========== [1]{} F. Bezrukov and D. Gorbunov. hep-ph/0205158.
{ "pile_set_name": "ArXiv" }
--- author: - 'J. Olofsson' - 'J.-C. Augereau' - 'E. F. van Dishoeck' - 'B. Merín' - 'F. Lahuis' - 'J. Kessler-Silacci' - 'C. P. Dullemond' - 'I. Oliveira' - 'G. A. Blake' - 'A. C. A. Boogert' - 'J. M. Brown' - 'N. J. Evans II' - 'V. Geers' - 'C. Knez' - 'J.-L. Monin' - 'K. Pontoppidan' bibliography: - 'biblio.bib' date: 'Received ; accepted ' title: | C2D Spitzer-IRS spectra of disks around T Tauri stars\ IV. Crystalline silicates --- [Dust grains in the planet forming regions around young stars are expected to be heavily processed due to coagulation, fragmentation and crystallization. This paper focuses on the crystalline silicate dust grains in protoplanetary disks, for a statistically significant number of TTauri stars (96).]{} [As part of the Cores to Disks (c2d) Legacy Program, we obtained more than a hundred Spitzer/IRS spectra of TTauri stars, over a spectral range of 5–35$\mu$m where many silicate amorphous and crystalline solid-state features are present. At these wavelengths, observations probe the upper layers of accretion disks up to distances of a dozen AU from the central object.]{} [More than 3/4 of our objects show at least one crystalline silicate emission feature that can be essentially attributed to Mg-rich silicates. Fe-rich crystalline silicates are largely absent in the c2d IRS spectra. The strength and detection frequency of the crystalline features seen at $\lambda > 20\,\mu$m correlate with each other, while they are largely uncorrelated with the observational properties of the amorphous silicate $10\,\mu$m feature. This supports the idea that the IRS spectra essentially probe two independent disk regions: a warm zone ($\leq$1AU) emitting at $\lambda \sim 10\,\mu$m and a much colder region emitting at $\lambda > 20\,\mu$m ($\leq$10AU). We identify a [*crystallinity paradox*]{}, as the long-wavelength ($\lambda > 20\,\mu$m) crystalline silicate features are 3.5 times more frequently detected ($\sim$55% vs. $\sim$15%) than the crystalline features arising from much warmer disk regions ($\lambda\sim10\,\mu$m). This suggests that the disk has an inhomogeneous dust composition within $\sim$10AU. The analysis of the shape and strength of both the amorphous $10\,\mu$m feature and the crystalline feature around 23$\mu$m provides evidence for the prevalence of $\mu$m-sized (amorphous and crystalline) grains in upper layers of disks.]{} [ The abundant crystalline silicates found far from their presumed formation regions suggests efficient outward radial transport mechanisms in the disks around TTauri disks. The presence of $\mu$m-sized grains in disk atmospheres, despite the short time-scales for settling to the midplane, suggests efficient (turbulent) vertical diffusion, likely accompanied by grain-grain fragmentation to balance the efficient growth expected. In this scenario, the depletion of submicron-sized grains in the upper layers of the disks points toward removal mechanisms such as stellar winds or radiation pressure.]{} Introduction ============ The silicate dust grains that are originally incorporated into stellar nebulae and eventually constitute planet forming disks are thought to be of interstellar medium (ISM) size and composition, namely sub-micron in diameter and extremely amorphous ($>$99%) in structure [e.g. @Gail1998]. Solar System comets, on the other hand, show high crystallinity fractions. The silicate grains in Oort comet Hale-Bopp are for instance composed of 40 to 60% of Mg-rich crystalline grains [@Wooden1999; @Wooden2007a; @Crovisier1997], while Jupiter Family comets have slightly lower crystalline fractions [of about 25–35%, e.g. comet 9P/Tempel 1, @Harker2007]. Although the actual crystalline fraction may depend somewhat on the methodology used to calculate the dust optical properties (e.g. [@Min2008]), the clear detection of crystalline silicates features in comet spectra indicates their mass fraction to be high compared to the ISM. The amorphous silicate grains that were incorporated into the Solar nebula must have been processed in the very early stages of planet formation before being incorporated into comets. Because of the high temperatures required to thermally anneal amorphous silicates and modify their lattice structure, it is generally accepted that the crystallization of amorphous silicates occured close to the young Sun. The crystalline silicates were then transported outwards due to processes such as turbulence or winds, leading to possible radial variations that may explain some of the mineralogic differences observed between small Solar System bodies today. Other processes could also explain the presence of crystals in the outer regions, such as shock waves triggered by gravitational instabilities in the solar nebula as described in @Desch2002 and @Harker2002. Locally, such shocks could sufficiently heat up surrounding materials to anneal and crystallize amorphous grains. In all these scenarii, silicate crystalline grains can be considered as tracers of the history of the protoplanetary, circumsolar disk. The detection of silicates in the disks around young stars was notoriously difficult until the [*Infrared Space Observatory*]{} (ISO) and, later, the [*Spitzer Space Telescope*]{} were launched. Thanks to ISO, mid- to far-IR spectroscopy of disks around young intermediate mass stars (Herbig Ae/Be stars, hereafter, HAeBe) could be performed. @Acke2004 showed that 52% of HAeBe stars exhibit emission in the $10\,\mu$m (amorphous) silicate feature that arises from the upper layers of the disks, while 23% of these stars show features at $11.3\,\mu$m, which is associated with forsterite, a Mg-rich crystalline silicate. Using the mid-IR capabilities of MIDI on the VLT Interferometer, @2004 spatially resolved the 10$\mu$m emission zone for three HAeBe stars. They found a gradient of crystallinity, with a larger fraction of enstatite and forsterite crystalline grains located close to the star (1–2AU) than further away in the disk (2–20AU). This was among the first observations where the disk dust content was proven to be radially inhomogeneous. Although TTauri stars were too faint for observations with the ISO/SWS and LWS, @Natta2000 employed the lower spectral resolution ISOPHOT instrument to detect the broad amorphous silicate emission feature at around $10\,\mu$m for 9 stars in the Chameleon I cloud; but the low resolution, the low signal-to-noise ratio and the limited spectral range did not allow firm detection of any crystalline features in the disks around young solar mass stars. Pioneering ground-based observations of TTauri stars (hereafter TTs) nevertheless showed evidence for forsterite emission features in the terrestrial spectral window around $10\,\mu$m. @Honda2003, for example, clearly identified the presence of crystalline silicate emission features (forsterite at $10.1$, $10.5$ and $11.2\,\mu$m, and possibly enstatite at $10.9\,\mu$m) in the N-band spectrum ($R\sim 250$) of , and argued that the crystalline grains represent about 50% (30% enstatite, 20% forsterite) of the total mass of grains emitting at $10\,\mu$m. Because of the difficulty of disentangling between grain growth, the emission from PAH (Polycyclic Aromatic Hydrocarbon) or from forsterite as the main cause of the $11.3\,\mu$m “feature” in some disks (see e.g. [@Sitko2000] and the discussion in [@Przygodda2003]), an unambiguous assessment of the presence of crystalline silicates in large samples of TTauri disks has awaited the launch of Spitzer. The high sensitivity and large spectral range ($5$–$35\,\mu$m) of the IRS spectroscopic instrument onboard Spitzer permits routine mid-infrared spectroscopy of TTauri disks in nearby molecular clouds. Surveys of young solar analogs in different star forming regions such as those by @Furlan2006 and @Watson2009 for the Taurus cloud, and of amorphous silicate features in other clouds by @Kessler-Silacci2006, were performed. Detailed mineralogic fits analyzing the amorphous and crystalline silicate fractions in TTauri disks were performed by @Bouwman2008 [7 TTs] and @Sargent2009a [65 TTs], as well as for disks around much lower mass objects like the borderline brown dwarf (hereafter, BD) [@Mer'in2007], and the BD in the Taurus cloud [@Bouy2008], both of which show abundant crystalline silicate grains despite their low (sub-)stellar temperatures. According to @Bouwman2008, the crystalline content of the inner regions of TTauri disks appear to be mostly dominated by enstatite silicates while the crystalline content of the outer region instead seems dominated by forsterite grains. Such data point toward an inhomogeneous dust content within the disks which could result from a radial dependence of crystallisation processes, or a difference in the initial conditions under which these crystals formed. The evolution of crystalline silicates has also been studied by @Watson2009 [84 TTs], who find that within the same environmental conditions, the silicate crystalline mass fraction varies greatly, from none to nearly 100%. They find that 90% of the objects in their sample present crystalline features in the 10$\mu$m region (either the 9.2$\mu$m enstatite feature or the 11.3$\mu$m forsterite feature). Regarding other crystalline features at longer wavelengths, they found that 50% of the objects surveyed had detectable crystalline features at $\sim$33$\mu$m. Also, no strong correlations between crystalline indices and stellar parameters have been found, suggesting that some mechanisms (e.g. X-ray emission, giant planets formation or migration) operate to erase the correlations that are expected from the standard models of crystallisation processes. Here we present a comprehensive study of crystalline silicates in more than a hundred disks around young stars based on 5–35$\,\mu$m IRS spectroscopic observations obtained as part of the c2d Spitzer Legacy Program “From Molecular Clouds to Planets” [@Evans2003]. This work is the continuation of a series of c2d/IRS papers, studying grains, PAHs and gas in inner disk regions around young stellar objects (class II). Paper I by @Kessler-Silacci2006 focuses on amorphous silicate features as a proxy for grain growth in a subsample of the c2d IRS Class II objects. They find that the grain size is a function of spectral type. Paper II by @Geers2006 shows evidence for low fractions of PAHs around TTs, indicating low gas phase PAHs abundance in disks as compared to the ISM. Paper III by @Lahuis2007 reports the first detections of \[Ne II\] and \[Fe I\] gas-phase lines in TTauri disks. In this paper, we present in Sec.\[secobs\] IRS observations of the entire c2d/IRS sample of class II objects, and we provide in Sec.\[sec:overview\] an overview of the detection statistics of solid-state emission features, especially those from crystalline silicates. In Sec.\[sec:crystals\] we study the properties of crystalline silicate grains, including their size, and search for correlations between amorphous and crystalline silicates and between the SED shape and grain size. We further investigate the difference in crystallinity between the inner and outer regions of disks. In Sec.\[sec:discussion\] we discuss the implications of our results for the dynamics of disks at planet-forming radii, and we summarize our results in Sec.\[sec:summary\]. In a companion paper (Olofsson et al. 2009b, in prep.), we analyse in more detail the dust mineralogy based on compositional fitting of the complete spectra. Spitzer/IRS observations {#secobs} ======================== We present in this study the infrared spectra of disks around 108 young stellar objects obtained as part of the c2d Spitzer Legacy program. The spectra were obtained using the IRS instrument onboard the Spitzer Space Telescope. It was used to expand the early spectroscopic studies of HAeBe stars with ISO, and of a few TTs observed from the ground in the $\lambda \sim$ 10$\mu$m atmospheric window. Thanks to the high sensitivity of Spitzer, a large sample of disks around low-mass and solar-mass stars could be efficiently observed. The IRS spectrograph is composed of four different modules that enable a wavelength coverage from 5 to 37$\mu$m. Two modules, Short-Low and Long-Low (SL and LL hereafter), provide a spectral resolution between 60-127 over ranges of 5.2-14.5$\mu$m and 14.0-38.0$\mu$m for SL and LL, respectively. The remaining two modules, Short-High and Long-High (SH and LH hereafter), have a spectral resolution of about 600. They cover more limited spectral ranges of 9.9-19.6$\mu$m and 18.7-37.2$\mu$m for SH and LH, respectively. Stellar sample -------------- In this paper, we focus on Class II objects – young stars no longer embedded in their protostellar envelope and surrounded by a circumstellar disk, where the dust may already be significantly processed. Our selection contains objects classified as Class II according to the literature, but also 10 objects with no known spectral type. In this sample, 30 objects are sources newly discovered with the instrument IRAC onboard Spitzer as part of the mapping phase of the c2d project. All such objects are named with the prefix SSTc2d, and are classified as Class II objects according to @Evans2009 except for one source with no known classification (SSTc2dJ182909.8+03446). The object is characterized as ClassI, but shows amorphous 10$\mu$m emission without additional absorption features. Two other SSTc2d objects have no classification. However, for the 10 objects with no classification in the literature, they were retained in the full list as they clearly show amorphous silicate features in emission. The final sample contains 108 stars including 60 TTs, 9 HAeBe and 1 BD. The 38 remaining objects have no spectral classification in the literature, or are the new SSTc2d stars discovered with Spitzer. This sample is distributed toward several clouds: Perseus (16 objects), Taurus (9), Chamaeleon (23), Ophiuchus (25), Lupus (16), Serpens (15), and includes 4 isolated stars: 3 HAeBes (, and ) and one TT star (). All six clouds are young star forming regions (1–5Myr), located at distances within 140 and 260pc from the Sun, with star densities (number of young stellar objects per square parsec) lying between 3 and 13. Some clouds are therefore quite extended, like Taurus or Lupus. The main cloud properties are given in @Evans2009. The complete list of IRS targets analysed in this paper, their classifications and spectral types, as well as details about the observations, are given in Table\[obs\]. Out of the 108 objects, four display spectra unusual for ClassII objects. These objects are , , and . The first three sources all display spectra reminiscent of those observed toward Oxygen-rich Asymptotic Giant Branch stars (see [@Sloan2003] for example spectra) while the fourth object displays a spectrum similar to a C-rich AGB star. Both SSTc2d objects located in the Serpens cloud are objects newly discovered using Spitzer/IRAC surveys. DLCha is a known variable star in Chamaeleon and HD132947 is an isolated Herbig star, classified in the literature as a Pre-Main Sequence star (see e.g. [@Valenti2003]). We do not question here the classification of these four stars but as they display unusual spectra (see Fig.\[sp:agb\]) we decided to remove them from the rest of the study. Data reduction -------------- The spectra presented in this paper were extracted from S13 pre-reduced (BCD) data using the c2d legacy team pipeline [@Lahuis2006] which makes available two different extraction methods: full aperture extraction and optimal Point Spread Function (PSF) extraction. The full aperture extraction method is applied to both low and high spectral resolution modules. For the low resolution modules, the c2d pipeline implements an extraction with a fixed-width aperture over the whole order. The source position in the slit is determined and the aperture is then centered on the source. The width is such that 99% of the flux of a point-source falls within the window. For the high resolution modules, the full slit width is used for the extraction. One disadvantage of this method is that the reduced spectra present a number of spikes (mainly due to bad pixels response, or “hot” pixels on the array, see below) that cannot be easily removed by hands-off pipelines. For the PSF extraction, the observed signal is assumed to be that of a point source or slightly resolved source plus an uniform zero level. This zero level, in most cases, represents the local extended emission close to the source, and the typical PSF profile is calculated from high signal-to-noise ratio calibrator data. As this zero-level may also contain residuals from, for example, raw pipeline dark current, we remove this contribution. PSF fitting is less sensitive to bad data samples and unidentified bad pixels than the aperture extraction method. The [*Spitzer Science Center*]{} provides masks with known bad pixels, but they are not all identified. Our pipeline detects such remaining pixels. The LH array is particularly affected by this issue, but other modules require attention. With the PSF extraction method, bad pixels are simply eliminated, whereas their values are simply interpolated for full aperture extraction method. The PSF extraction method, in addition, provides an estimate of the data zero level or sky contribution to the observed spectrum. An important disadvantage of the PSF fitting method, though, is that for some modules (mainly for the $2^{\mathrm{nd}}$ orders), the PSF is subpixel in size and the extraction method can become unstable. Thus, for some sources, the $2^{\mathrm{nd}}$ order of the SL module and the $2^{\mathrm{nd}}$ order of the LL module are unavailable, leading to incomplete spectra. In this paper, we adopt the PSF extraction method in order to obtain spectra with as few spikes as possible. Where the PSF extraction method is not available or turns out to be unstable, we use the full aperture extraction method to build spectra. We then opt for a median filtering scheme to remove the remaining spikes: for the SL and SH modules we use a median smoothing over three channels while we use a median smoothing over five channels for the LL and LH modules. Some problems were encountered with the data reduction for two objects in our sample: EC69 and ISO-Cha237. For the latter source there are two possible explanations: either the LL1 module spectrum poorly matches the LL2 module spectrum, or the LL2 module pipeline reduction failed and its slope is off. The second explanation seems the most likely, given the shape of the red side of the 10$\mu$m feature. For EC69, flux values for some orders were null. We decided to keep these two objects in our sample, first because ISO-Cha237 shows a 10$\mu$m feature that can be analyzed. For EC69, the spectrum appears featureless, and as it is classified as a TTauri star, we choose not to bias our results by removed “problematic” spectra. Module merging and offset correction\[offset\_corr\] ---------------------------------------------------- The c2d pipeline corrects for telescope pointing errors that may lead to important flux offsets between the different modules. Some small offsets can nevertheless remain between modules even after pointing correction, and an additionnal post-processing offset correction is thus applied. This correction depends on the modules available for every spectrum. Table\[mod\_cas\] gives a summary of all the possible configurations we encountered. A tick-mark ($\surd$) means that data provided by this module are used to build the final spectrum. Priority was given to SL and LL, first because the spectral resolution is sufficient for detecting crystalline features, and second, because the high resolution modules provide noisier spectra than those from the low resolution modules. For each star, a final spectrum is then obtained by merging all the chosen modules and correcting manually for small remaining offsets. The spectra themselves are presented in Fig.\[spectre\], Fig.\[fig:C23\] and Figs.\[sp:perseus\]–\[sp:others\]. Case SL1 SL2 SH LL1 LL2 LH Number of stars ------ --------- --------- --------- --------- --------- --------- ----------------- 0 - - $\surd$ - - $\surd$ 9 1 - - $\surd$ $\surd$ - $\surd$ 2 2 $\surd$ $\surd$ - $\surd$ $\surd$ - 51 3 $\surd$ $\surd$ $\surd$ - - $\surd$ 18 4 $\surd$ $\surd$ $\surd$ $\surd$ - $\surd$ 22 5 $\surd$ $\surd$ $\surd$ $\surd$ $\surd$ - 6 : \[mod\_cas\]Overview of the modules used in all the different cases encountered to obtain the final spectra (see Sec.\[offset\_corr\] for details). The exact modules used for each individual star are given in Table\[obs\]. Note: SL1: $7.4$–$14.5\,\mu$m , SL2: $5.2$ – $7.7\,\mu$m , LL1: $19.5$ – $38.0\,\mu$m, LL2: $14.0$ – $21.3\,\mu$m Estimating the uncertainties {#sec:uncertainties} ---------------------------- The uncertainties propagated through the c2d pipeline are dominated by the flux calibration error estimates. As our study is oriented toward the detection of features in emission from silicates, we are only concerned with relative uncertainties which are evaluated by dividing the original spectrum by a smoothed version of itself obtained using a Savitzky-Golay filter [@Savitzky1964]. This provides an estimate of the signal-to-noise ratio (SNR) on the spectrum, and is used to assess the presence of features as follows: The Savitzky-Golay filtering process calculates a local 3$^{\rm{rd}}$-order polynomial for every point of the spectrum, based on the four left and right neighbors. The main advantage of this filter is that it maintains the shape and contrast of the features, while “pointy” features tend to be rounded out with classical average smoothing, thereby increasing the uncertainties at the top of every feature and downgrading the actual quality of the data. The Savitzky-Golay filter has only been used to calculate the uncertainties displayed as grey envelopes in Fig.\[spectre\], Fig.\[fig:C23\] and Figs.\[sp:perseus\]–\[sp:others\], not affecting the spectra themself presented on these figures and used for the scientific analysis. [ccccc]{} Crystal & Peak & Measured & Measured $\Delta \lambda$\ & $\lambda$ \[$\mu$m\] & peak $\lambda$ \[$\mu$m\] & Eq. Width \[$\mu$m\]\ --------------- Enstatite (MgSiO$_{3}$) --------------- : \[feat\]Summary of the different crystalline silicate features examined in this study. & ---------------- 9.2$^{(1,2)}$ 21.6$^{(1)}$ 23.0$^{(1,2)}$ 24.5$^{(1,2)}$ 28.2$^{(1,2)}$ ---------------- : \[feat\]Summary of the different crystalline silicate features examined in this study. & ---------------- 9.24$\pm$0.08 21.47$\pm$0.16 C23 C23 C28 ---------------- : \[feat\]Summary of the different crystalline silicate features examined in this study. & --------------- 0.49$\pm$0.11 0.45$\pm$0.25 C23 C23 C28 --------------- : \[feat\]Summary of the different crystalline silicate features examined in this study. \ --------------------- Forsterite (Mg$_{2}$SiO$_{4}$) --------------------- : \[feat\]Summary of the different crystalline silicate features examined in this study. & -------------- 11.3$^{(2)}$ 16.2$^{(2)}$ 23.8$^{(2)}$ 27.6$^{(2)}$ 33.6$^{(2)}$ -------------- : \[feat\]Summary of the different crystalline silicate features examined in this study. & ---------------- 11.26$\pm$0.12 16.08$\pm$0.16 C23 C28 33.1$\pm$0.63 ---------------- : \[feat\]Summary of the different crystalline silicate features examined in this study. & --------------- 0.60$\pm$0.33 0.38$\pm$0.16 C23 C28 1.19$\pm$0.47 --------------- : \[feat\]Summary of the different crystalline silicate features examined in this study. \ ----------------------- Diopside (CaMgSi$_{2}$O$_{6}$) ----------------------- : \[feat\]Summary of the different crystalline silicate features examined in this study. & 25.0$^{(2)}$ & 25.14$\pm$0.18 & 0.58$\pm$0.22\ References: $^{(1)}$@Koike2000, $^{(2)}$@Molster2002b Overview of observed solid-state features\[sec:overview\] ========================================================= Observed crystalline silicates ------------------------------ The 104 spectra analyzed in this paper are displayed in Figures \[sp:perseus\] to \[sp:others\], for wavelengths between 5 and 35$\mu$m and in units of Jy ($F_{\nu}$). They are sorted by cloud and increasing Right Ascension. The spectra of the four isolated stars in our sample are displayed in Fig.\[sp:others\]. Broad emission features at 10 and 20$\mu$m from amorphous silicates can be identified in many spectra, as well as narrower emission features. The latter are more easily identified beyond 20$\mu$m, and are due to crystalline silicates. Figure\[spectre\] illustrates the diversity of the spectra seen in our sample. (bottom spectrum) shows many crystalline features, with a weak 10$\mu$m band. In contrast, (top spectrum) is almost “pristine”, with strong amorphous bands and no evidence of crystalline features[^1]. The third spectrum, for , has both amorphous and crystalline features. The crystalline features in disks can, in most cases, be attributed to Mg-rich silicate minerals (with a possible contribution from Ca-rich silicates); in agreement with some model predictions (e.g. [@Gail2004] and ref. therein), and as observed in solar system comets (e.g. [@Wooden2007a]). to ### Mg-rich crystalline silicates: Forsterite & Enstatite {#sec:forsens} Table\[feat\] presents the different crystalline silicate features we searched for in every spectrum – in addition to the amorphous silicate feature at 10$\mu$m. In Table\[feat\], “Peak $\lambda$” is taken from the literature, “Measured $\lambda$” and “$\Delta \lambda$” values are the peak positions and equivalent widths measured by our observations. “C23” and “C28” refer to silicate features blended in complexes. Uncertainties are standard deviations. Here we consider features arising from Mg-rich silicates, specifically enstatite (inosilicate belonging to the pyroxene group, with chemical formula MgSiO$_{3}$) and forsterite, the nesosilicate Mg-rich end-member of the olivine group (Mg$_{2}$SiO$_{4}$). Blowups of representative spectral features are shown in Fig.\[fig:C23\]. Using IRS spectra, we cannot disentangle differences between crystalline structures, for instance, clino-enstatite or ortho-enstatite. Still, according to @Koike2000, enstatite features are most likely attributed to ortho-enstatite. The example spectrum for ISO-ChaII54, displayed in Fig.\[spectre\] as well as the blowups in the right panel of Fig.\[fig:C23\], shows that the 23.0, 24.5$\mu$m enstatite and the 23.8$\mu$m forsterite features can be blended into one single complex. Similarly, the 27.6$\mu$m forsterite feature can also be blended with the 28.2$\mu$m enstatite feature. In the following, we will treat these features as two complexes, at reference positions of 23 and 28$\mu$m, independent of the actual crystals responsible for the observed emission. We shall note these two complexes, C23 and C28, which are by far the most frequently detected crystalline features in disks (Sec.\[sec:measure\]). Fig.\[fig:C23\] displays eight example spectra of the C23 and C28 complexes. Clearly, the shape and peak positions may vary from one object to another. We will show in Sec.\[sec:gsize\] that some of these differences may relate to the mean size of the emitting grains, although the composition (more or less enstatite compared to forsterite), grain shape, and strengh of the continuum might also contribute to the diversity of the C23 and C28 features observed. Beside the features listed in Table\[feat\], other enstatite bands at 10.6, 11.7 and 19.6$\mu$m, as well as forsterite features at 10.1 and 19.7$\mu$m were also searched for with little success in a subsample of 47 high SNR spectra. The detection rate of these crystalline features is very low and not significant. For the two enstatite and forsterite features at $\sim$19.6 and 19.7$\mu$m, the main difficulty lies in the added noise where SH and LH modules merge, which prevents any firm detection of these two features. ### Fe-rich components: Troilite & Fayalite {#sec:ferich} The formation of Mg-rich crystalline silicates by direct condensation from a gas of solar composition (as opposed to formation by thermal annealing) is expected to additionally yield iron-bearing condensates, in particular FeS (or troilite, see [@Wooden2007a]). This condensate is of great interest and, while a detailed study of the ISO spectra of two HAeBe stars (including , Fig.\[sp:chama\_2\]) carried out by [@Keller2002] indicates FeS in these systems, it remains difficult to detect, in general, because the main feature (between 17-26$\mu$m), centered at 23.5$\mu$m, is broad, and falls in regions where the IRS spectra are often noisy (due to module merging around 20$\mu$m). Further, the disk spectra are often dominated by the presence of the 18$\mu$m amorphous silicate feature as well as by frequent C23 emission features. Therefore, studying the presence or absence of FeS emission in the c2d sample is beyond the scope of this paper. Another Fe-rich component that can potentially be detected in IRS spectra is fayalite, a crystal belonging to the olivine class (Fe$_2$SiO$_4$). Theoretical opacities from @Fabian2001 show strong emission features around 30$\mu$m. For small grains, three narrow features are expected at $\sim$27, 29.3 and 31.6$\mu$m. We find there are 4 objects where fayalite may tentatively be present, all four in the Chameleon cloud (Fig.\[sp:chama\_1\]): TWCha, VZCha, WXCha, and Sz62. Interestingly, the Sz62 spectrum also shows other features that can be attributed to fayalite, at roughly 11$\mu$m and 19$\mu$m. All these objects notably display a feature centered at around 30.5$\mu$m, which is at shorter wavelength compared to the feature from pure fayalite which lies at 31.6$\mu$m (e.g. [@Fabian2001] or [@Koike2003]). This may reflect a small departure from pure Fe-rich crystalline olivine (e.g. $\sim$10–15% Mg fraction) as illustrated by the three top spectra in Fig.1 from @Koike2003 where it is seen that the feature shifts to smaller wavelengths with increasing Mg-content. We note that there are other spectra that show some similar behaviours in the 30$\mu$m region, but they are of lower signal-to-noise ratio and contain possible data reduction artifacts (see e.g. RXJ0432.8+1735 spectrum, Fig.\[sp:taurus\]). These factors make it difficult to conclude whether fayalite is present or not. Because of the small number of positive detections and because of the ambiguity between possible noise and real features, we will not investigate the presence of this component further. We do note, however, that the compositional fits to seven Spitzer/IRS spectra of TTs by @Bouwman2008 show no evidence of Fe-rich materials. However, the previous conclusion means that we do not observe pure fayalite in our spectra. Considering the following formula for crystalline olivine (Mg$_x$Fe$_{1-x}$)$_2$SiO$_4$ (with $0 \leq x \leq 1$), features in the 20-30$\mu$m spectral range become weaker (especially around 23$\mu$m) for Mg fractions below 0.6. This means that we are actually probing crystalline olivine with at least more than 60% of Mg compared to Fe. Still, as crystalline pyroxene is also contributing to C23 and C28 complexes, further interpretations remain difficult. ### Other silicates: Diopside & Silica {#sec:diotrosil} Diopside, a calcium magnesium silicate, and a member of the pyroxene group (CaMg(SiO$_{3}$)$_{2}$), can be searched for at $\sim$$25\,\mu$m. Because of the presence of the C23 complex, however, the diopside feature can sometimes be blended with the complex. Even worse, the shoulder of the complex can be mis-interpreted as a 25$\mu$m feature. We therefore obtain only a lower limit on the frequency at which this feature is present. The low detection rate of an additional (but weaker) $20.6\,\mu$m diopside feature does not strengthen the confidence level of the detection of the diopside in TTauri star disks. Finally, silica (SiO$_{2}$) has also been identified in our spectra, with a feature arising at 12.5$\mu$m. Attribution of this single feature to amorphous or crystalline silica is not straightforward. @Sargent2009a showed that amorphous silica tends to produce a feature around 12.3–12.4$\mu$m, while various crystalline silica polymorphs produce a feature centered at or slightly to longer wavelengths of 12.5$\mu$m. Due to the presence of the broad amorphous 10$\mu$m feature, disentangling the different contributions is beyond the scope of our statistical study of the most prominent crystalline silicates. Example silica features are shown on Fig.\[spectre\] and the left panel of Fig.\[fig:C23\]. Silicate feature statistics for TTauri stars -------------------------------------------- ### The fraction of disks showing silicate(s) feature(s) {#sec:measure} We developed a routine that measures the characteristics of both crystalline and amorphous (10$\mu$m) features. For each feature, it assumes a local continuum which is built using a two point Lagrange polynomial joining the feet of the feature, computed as follows: $$\begin{aligned} F_{\nu,\mathrm{c}} = F_{\nu_1} \times \left( \frac{\lambda - \lambda_2}{\lambda_1 - \lambda_2} \right) + F_{\nu_2} \times \left( \frac{\lambda - \lambda_1}{\lambda_2 - \lambda_1}\right), \end{aligned}$$ where $F_{\nu,\mathrm{c}}$ is the final continuum, $\lambda_1$ and $\lambda_2$ the blue and red feet of the feature, respectively, and $F_{\nu_1}$ and $F_{\nu_2}$, the observed fluxes at $\lambda_1$ and $\lambda_2$, respectively. This routine returns the band flux, its peak position, the mean wavelength and the peak flux. It also returns the 1-$\sigma$ uncertainty on the band flux. The results given by this procedure are presented in Table\[IDfeat\], in the form of the SNR on the measured feature fluxes using the above procedure. Results for only 101 objects out of total 104 are shown since 3 of them (2.8% of total sample) do not display any crystalline or amorphous silicate features. Those objects are: , and . Thus, the overwhelming majority ($\sim$97%) of the disks do show at least one feature that can be attributed to silicates. Since $10\,\mu$m-sized or larger grains are essentially featureless in the IRS spectral range (see, for example, the theoretical opacities in Fig.\[blowopac\]), our results qualitatively indicate that inner disk atmospheres are populated by grains a few micrometers or less in size, independent of the age of our stars, and of cloud membership. Some differences between clouds are nevertheless addressed in Sec.\[sec:prevalence\], and a discussion concerning the size of the amorphous and crystalline silicate grains can be found in Sec.\[sec:gsize\]. In the following, we remove the 8 HAeBe stars (plus HD132947 that we already removed) in our survey from the statistical analysis so that the sample is mainly composed of objects classified as TTs. The statistics in the paper are therefore made on the remaining 96 objects. ### The prevalence of crystalline features {#sec:prevalence} Feature-by-feature statistics are displayed in Figure\[fig:tt\_feat\], which shows detection fractions of crystalline features ranging between $\sim$10% and $\sim$60%. A visual inspection of the 28–29$\mu$m and 33–35$\mu$m features for a subset (40 TTs and 7 HAeBes) of the target sample analyzed in this paper led @Kessler-Silacci2006 to conclude that $\sim$50% of the spectra show crystalline silicate features. Considering now the full 96 star sample and only secure detections with SNR $> 20$ for the C28 and 33.6$\mu$m complexes, we obtain a fraction of $\sim$65% (63 out of 96) for objects displaying at least one of the two features. Including all the other crystalline features (still with SNR $> 20$), this fraction rises to $\sim$78% (75 out of 96 stars). We have checked that these results are not corrupted with possible PAH emission features, and that the 11.3$\mu$m can be associated with forsterite. Indeed, @Geers2006 identified only three objects in our sample that have potential PAHs emission at 11.3$\mu$m: , and . All the other objects cited in @Geers2006 are HAeBe stars and, therefore, not considered in our sample. With additional data, @Geers2007 could confirm the presence of PAHs for , and , and so these three objects have then been removed from our statistics for the analysis of the 11.3$\mu$m forsterite feature. One other object, is identified by @Geers2007 to have a 6.2$\mu$m PAH feature but with no 11.3$\mu$m counterpart, thereby avoiding any possible contamination in this particular case. Note that we have been as conservative as possible in identifying features, especially for the crystalline features around 10$\mu$m. Some spectra display a shoulder at around 11.3$\mu$m that could possibly be interpreted as a forsterite feature (see e.g. spectrum, Fig.\[fig:C23\], label [ *f*]{}). We choose not to attribute such kind of shoulder to forsterite. Therefore, we conclude from our analysis that about 3/4 of the protoplanetary disks in our sample of young solar-analogs show at least one crystalline feature in their 5–35$\mu$m spectrum, meaning that dust crystallization is not a marginal process but is instead widespread at this stage of disk evolution. ![\[fig:tt\_feat\]Detection statistics for crystalline silicate features (yellow and red bars) for the 10.0$\mu$m amorphous silicate feature (blue) and for silica at 12.5$\mu$m (green). Darker color bars are positive detections (SNR $> 20 $), and lighter color bars stand for tentative detections (SNR $\leq 20 $). The uncertainties are those due to Poisson statistics.](fig4.ps){width="1.1\columnwidth"} It is interesting to compare these results to those of @Watson2009 who analyzed 84 TTs in the Taurus-Auriga association, also using Spitzer/IRS. Developing indices measuring the departure from an ISM-like pristine 10$\mu$m feature, they found that 90% of the Taurus-Auriga disks show either the 9.2$\mu$m enstatite feature or the 11.3$\mu$m forsterite feature, while we obtain about 20% for both features (Fig.\[fig:tt\_feat\]). In @Watson2009, any departure from the 10$\mu$m pristine reference spectrum is considered as being evidence for crystalline features, while we argue that this frequent departure that we also observe has rather more to do with a grain size effect (see Sec.\[sec:gsize\] and Sec.\[sec:depl\] for further discussions). There is therefore a difference in the interpretation of the measurements that leads to inconsistent conclusions. @Watson2009 used the @Dalessio2006 model to investigate the impact of dust sedimentation on the observed dispersion in equivalent width for the 10$\mu$m amorphous feature. They could reproduce a broad range of observed equivalent widths by depleting the dust fraction (that consists of grains $\leq 0.25\,\mu$m in their model) in the upper layers of the disks, thereby mimicking dust settling. In a more recent paper, @Dullemond2008 find that sedimentation has, nevertheless, a very limited effect on the shape of the 10$\mu$m feature as long as a continuous grain size distribution is considered (as opposed to fixed grain sizes). For longer wavelength crystalline features, comparison with the work of @Watson2009 is possible for the 33.6$\mu$m forsterite feature, but not for the features around 23 and 28$\mu$m as they did not study these two complexes. They find that 50% of the Taurus-Auriga objects show the 33.6$\mu$m crystalline feature, compared to 28% for our sample, but this detection statistic is consistent with the statistics for the C23 (55%) and C28 (54%) crystalline complexes. Figure\[fig:clouds\] shows feature-by-feature statistics sorted by clouds. The largest occurence of the C23 complex is found in the Ophiuchus cloud with about 80% of the objects showing this feature. Disks in Serpens have a prevalence of crystalline features significantly smaller than in the Perseus, Chamaeleon, Ophiuchus and Lupus clouds which show comparable fractions of amorphous and crystalline features. The Taurus cloud shows a noticeably low frequency of the 10$\mu$m amorphous feature compared to the other clouds and does not appear to be balanced by a remarkably high fraction of crystalline grains when compared to the other regions. Our sample contains only eight objects in the Taurus cloud, however, which introduces a strong bias. @Watson2009 found that the amorphous 10$\mu$m feature is present for a large majority of the objects in their survey of 84 TTs. Therefore, discussing the differences is not statistically significant for our Taurus sample. One can notice that, except for Taurus, the five remaining clouds show very similar behavior. This points toward a rather homogeneous distribution of amorphous and crystalline silicate dust grains in disks throughout clouds. Detection statistics in the Serpens cloud are overall smaller, but the trends are the same as that in the other clouds. Our current result on the Serpens cloud will be revisited in a forthcoming paper by Oliveira et al. (in prep) who analyze a much larger sample of more than a hundred Serpens objects observed with Spitzer/IRS. In the following, we will treat all the objects as a single sample. The first reason for this is that cloud-to-cloud differences suffer from low statistics and merging the entire sample will increase the significance of our inferences (96 objects in total). The second reason is that the trends of the different clouds seem to be similar (except for our Taurus cloud smaple which has a low number of stars). Merging the sample will strenghten our conclusions thanks to the large number of ClassII objects. ![\[fig:clouds\]Same as Fig.\[fig:tt\_feat\] but now sorted by clouds. Uncertainties are again Poisson noise, derived from the number of objects per cloud: 15 for Perseus, 8 for Taurus, 20 for Chamaeleon, 16 for Lupus, 24 for Ophiuchus and 12 for Serpens.](fig5.ps){width="1.1\columnwidth"} Properties of crystalline silicates in the disks around young stars {#sec:crystals} =================================================================== One may expect that the relatively large spectral range of the IRS instrument (5–35$\,\mu$m) implies that the detected short- and long-wavelength emission features could arise from physically distinct regions in the disk. @Kessler-Silacci2006 and @Pinte2008, among others, have shown that for a typical TTs, the $10\,\mu$m emission comes from a region about 10 times smaller than the $20\,\mu$m emitting zone. We examine in this section the correlations between the detected features. In particular, we inspect the relation between the properties of the $10\,\mu$m silicate amorphous feature (its energy, the size of its carriers, its apparition frequency) and those of the crystalline silicate features at $\lambda > 20\,\mu$m. Growth of warm amorphous silicates {#sec:gsize} ---------------------------------- The amorphous 10$\mu$m feature is present at a rate of about 70% in our sample. @Bouwman2001, @2003 and @Kessler-Silacci2006 have shown that detailed studies of this feature reveal much information on dust characteristics. Especially noteworthy is that the observed shape versus strength relation for the 10$\mu$m feature can be reproduced by varying the characteristic grain size over the range of $0.1$–$3\,\mu$m. Here we analyse the 10$\mu$m feature by adopting the same continuum normalization and the same computational method for the shape (flux ratio $S_{11.3}/S_{9.8}$) and strength ($S_{\mathrm{Peak}}^{10 \mu \mathrm{m}}$) of the 10$\mu$m feature as in @Kessler-Silacci2006. The different values $S_{11.3}$, $S_{9.8}$ and $S_{\mathrm{Peak}}$ are obtained by normalizing the observed flux $F_{\nu}$ as follows: $ S_{\mathrm{\nu}} = 1+ (F_{\mathrm{\nu}} - F_{\mathrm{\nu,c}})/\langle F_{\mathrm{\nu,c}} \rangle$, where $F_{\mathrm{\nu,c}}$ is the estimated linear continuum between the red and blue feet of the feature and $\langle F_{\mathrm{\nu,c}} \rangle$ its mean value between the two feet (all fluxes expressed in Jy). We choose a wavelength range of $\pm 0.1$$\mu$m around 9.8$\mu$m and 11.3$\mu$m to calculate $S_{11.3}$ and $S_{9.8}$. Figure\[kurt\_sp\] (left panel) shows the known correlation between the flux ratio $S_{11.3}/S_{9.8}$ ($10\,\mu$m feature shape) and $S_{\mathrm{Peak}}$ ($10\,\mu$m feature strength). To quantify this correlation we use a Kendall $\tau$ test. The Kendall $\tau$ rank correlation coefficient measures the degree of correspondance between two datasets. If the agreement between the two sets of values is perfect then $\tau = 1$, if the disagreement is perfect then $\tau = -1$. The Kendall $\tau$ procedure also returns a probability $P$, computed using the complementary error function of $| \tau |$ (see [@Press1992] for further informations on the Kendall $\tau$ and $P$ calculations). The $P$ probability is the two-sided significance; the smaller the $P$ value, the more significant the correlation. For the shape versus strength of the amorphous 10$\mu$m feature, we find an anti-correlation ($\tau = -0.42$) with a significance probability $P = 3.58 \times 10^{-7}$. This correlation can be interpreted as larger grains producing a flatter 10$\mu$m amorphous feature, while smaller grains produce narrower features (e.g. [@2003]). A thorough analysis by @Kessler-Silacci2006 and @Kessler-Silacci2007 of the shape-strength trend showed that more massive stars (HAeBe) tend to have sharper $10\,\mu$m features than do TTs and BDs, which cluster in a region more consistent with micron-diameter grains as the typical emitting grain size. The comparison of our observations to synthetic $10\,\mu$m amorphous features calculated for different grain sizes (see Sec.\[sec:contrast\] for details on the computation method of grain opacities) confirms this trend as the bulk of our sample clusters in a region consistent with grains that have grown to sizes larger than about $2\,\mu$m (left panel of Fig.\[kurt\_sp\]). to Growth of cold crystalline silicates {#sec:gsize2} ------------------------------------ Although a similar analysis of the $20\,\mu$m amorphous emission features is rendered more difficult because of the broadness of the feature, the narrower longer-wavelength crystalline silicate features, especially the C23 complex, show a grain size-dependent behaviour that can be tested with our disk sample (see the theoretical opacities in Fig.\[blowopac\]). Similar trends could not be exploited for the C28 complex as its shape is not much sensitive to grain size (over the 0.1-3$\mu$m range). For the C23 complex, we build a feature shape index, the $S_{24}/S_{23}$ flux ratio, computed using the same approach as for the amorphous 10$\mu$m feature. The $S_{23}$ normalized flux is obtained by averaging the normalized (local) continuum-subtracted flux between $23.5$ and $23.8\,\mu$m, and between $24.1$ and $24.4\,\mu$m for $S_{24}$. Figure\[kurt\_sp\] (right panel) shows the $S_{24}/S_{23}$ ratio as a function of the strength of the C23 complex (C23 $S_{\mathrm{Peak}}$). The Kendall test gives $\tau = -0.54$ with a significance probability below $10^{-38}$, indicative of a clear anticorrelation between these two quantities. To better evaluate the relationship between grain size and the C23 shape, we calculate theoretical $S_{24}/S_{23}$ ratios for synthetic crystalline silicates opacity curves obtained using the Distribution of Hollow Spheres (DHS) method [@Min2005]. We consider grains between $0.1$ and $2.7\,\mu$m in radius, and two different compositions: 100% forsterite (Mg$_{2}$SiO$_{4}$, using optical constants from [@Servoin1973]) and a mixture composed of 50% forsterite plus 50% enstatite (MgSiO$_{3}$, optical constants taken from [@Jaeger1998]). Grains larger than $\sim3.0\,\mu$m, for either enstatite or forsterite, do not give striking results, as the feature disappears into the continuum. Also, a pure enstatite composition turns out to be incompatible with the observed shape for the C23 complex as it shows two unblended features at 23.8 and 24.5$\mu$m. This indicates that we are not observing pure enstatite grains in our sample, a result which seems to first order consistent with @Bouwman2008 who found that the very inner ($\sim$ 1AU) warm dust population is dominated by enstatite while the dust population at large radii is dominated by forsterite. The overall shape-strength trend for the C23 crystalline complex is better reproduced for a forsterite/enstatite mixture, with characteristic grain sizes larger than 1.0$\mu$m in order to to match the bulk of the measured values. This suggests that not only have the warm amorphous silicate grains grown in our sample of TTs (left panel of Fig.\[kurt\_sp\]), but also that the colder crystalline silicate grains have coagulated into $\mu$m-sized particles and remain suspended in the upper layers of the disk atmospheres. The offset in the right panel of Fig.\[kurt\_sp\] between the data points distribution and the curves obtained using theoretical opacities may come from various effects. First, we consider pure crystalline dust grains, which may be a too restrictive assumption to reproduce the observations. It allows us to see the impact of grain size, however, and to show that it matches reasonably well the observed trend. Recently, @Min2008 studied the impact of inclusions of small crystals into larger amorphous grains, but this is too much of a refinement for our purposes. Second, we only consider pure Mg-rich crystals (no iron). According to @Chihara2002, the 24.5$\mu$m feature of crystalline pyroxenes shifts towards slightly larger wavelengths with increasing Fe-content. Also, according to @Koike2003, the peak position of the 23.8$\mu$m crystalline olivine feature does not shift unless the Fe-content is larger than 20%. As the $S_{24}$ index is computed integrating between 24.1 and 24.4$\mu$m, raising the Fe-content will produce a smaller $S_{24}$ and therefore a smaller $S_{24}/S_{23}$, while the strength of the complex will slightly decrease as the 24.5$\mu$m feature shifts to larger wavelengths. To conclude, even if there is a small shift when comparing the observations to theoretical points, the trend is well reproduced by varying the grain size. Interestingly, we find no correlation between the grain size proxy for the amorphous 10$\mu$m feature and the grain size proxy for the C23 complex. Considering the sets of values $(S_{11.3}/S_{9.8})/S_{\mathrm{Peak}}^{10}$ as a function of $(S_{24}/S_{23})/S_{\mathrm{Peak}}^{23}$, we indeed obtain a Kendall’s $\tau$ of 0.07, with a significance probability $P$ of 0.44. This means that even if grain growth took place in both components, the growth of the amorphous and crystalline grains, respectively, does not seem to be correlated. Relationship between the 10$\mu$m and C23 features with disk properties ----------------------------------------------------------------------- ![\[fig:slope\] [*Left panel:*]{} Correlations between the slope of spectra ($F_{30}/F_{13}$) as a function of the strength ($S_{Peak}$) of the amorphous 10$\mu$m feature. The two outliers (blue open circles) are cold disks: and . [*Right panel*]{}: Spectral slope ($F_{30}/F_{13}$) versus strength ($S_{\mathrm{Peak}}$) for the C23$\mu$m crystalline complex.](fig8.ps){width="1.\columnwidth"} Young flared disks are expected to show rising spectra in the mid-IR, as opposed to more evolved, settled systems having self-shadowed disks with much flatter mid-IR spectra [e.g. @Dominik2003]. Here we measure the slope of the spectra by the ratio of the $F_{30}$ and $F_{13}$ indexes, where $F_{13}$ is the mean flux value (in Jy) in the range 13$\mu$m$\pm 0.5\,\mu$m, and between 29$\mu$m and 31$\,\mu$m for $F_{30}$. The left panel of Figure \[fig:slope\] shows the correlation between the strength of the amorphous 10$\mu$m feature and the slope of the spectra ($F_{30}/F_{13}$ ratio). Taking out the two outliers (blue open circles), which are known cold disks [ and , @Brown2007], we find a correlation coefficient $\tau = 0.29$ with a significance probability $P$ smaller than $5.2 \times 10^{-4}$ suggestive of a correlation. A similar trend is found by @Bouwman2008 within their FEPS sample of seven stars, and by @Watson2009 with the Taurus-Auriga sample of 84 TTs. This can be interpreted as smaller grains (large $S_{\mathrm{Peak}}^{10 \mu \mathrm{m}}$ values) being probed in flared disks (high values of $F_{30}/F_{13}$), while flattened disk (low $F_{30}/F_{13}$ values) atmospheres are characterized by larger grains. If dust sedimentation is the main process responsible for disk flattening, one would expect the upper layers of the disks to be depleted in large grains, hence the mid-infrared spectroscopic signature to be that of very small grains (peaked 10$\,\mu$m features). The observed trend (Fig.\[fig:slope\]) very clearly refutes such a possible trend in the relation between the SED shape and that of the 10$\,\mu$m feature. The effect of dust sedimentation on the shape of the 10$\,\mu$m amorphous silicate feature was recently discussed by @Dullemond2008 based on models. They showed that the sedimentation of the larger grains toward the disk midplane is generally not enough to transform a flat 10$\,\mu$m feature into a peaked feature. Our observations tend to support these results as sedimentation, if assumed to be revealed by disk flattening, is generally associated with flat 10$\,\mu$m features. This trend is in fact valid for the majority of the TTs in our sample as the objects cluster in the region with flat SEDs and flat silicate features in Figure \[fig:slope\]. For the C23 crystalline complex (right panel of Fig.\[fig:slope\]), we find a rather weak anticorrelation, with $\tau = -0.21$ and a significance probability $P <$0.01. The small crystalline grains with $S_{\mathrm{Peak}}$ larger than about 1.2 are only found in flat or slowly rising disks ($F_{30}/F_{13}$ smaller than about 2), while the larger grains can be found in all kinds of disks. Interestingly, the very flared disks which have large crystalline grains emitting at 23$\,\mu$m also show small amorphous grains in their warm component (that is, the region probed at 10$\,\mu$m). Seven out of the eight objects (, , , , , , ) whith $F_{30}/F_{13} > 3$ on the right panel of Fig.\[fig:slope\] indeed have their counterpart in the left panel with $S_{\rm Peak}^{10}$ larger than about 2 (the last object being ). to Relationship between the 10$\mu$m feature and crystallinity\[sec\_cryst\_10\] ----------------------------------------------------------------------------- We find in Sec.\[sec:gsize2\] that the growth of the warm amorphous silicates (probed by the $10\,\mu$m feature) is statistically disconnected from the growth of the cold crystalline silicates (probed by the C23 complex). We further investigate the relationship between the amorphous and crystalline silicates features by searching for a correlation between the energy contained in the amorphous $10\,\mu$m feature and the energy contained in the C23, C28 and the $33.6\,\mu$m features. To get rid of possible distance or brightness effects we normalized the line fluxes as follows: $\mathrm{Line\,\,flux\,\,[unitless]} = \mathrm{Line\,\,flux} [\rm{W.m}^{-2}] / (\lambda \langle F_{\mathrm{\lambda, c}} \rangle)$, where $\langle F_{\mathrm{\lambda, c}} \rangle$ is the mean value of the local feature continuum and $\lambda$ is the measured central wavelength of the considered feature. The unitless line fluxes are displayed in the left panel of Figure\[LF10\], which shows that when both the C23 and the 10$\mu$m feature are present with SNR $> 20$, their emission energies are not correlated ($\tau = 0.06$, significance probability $P = 0.55$), although the plot would suggest that objects with low 10$\mu$m line flux usually also have a low C23 line flux. The middle panel of Figure\[LF10\] shows a similar result when both the C28 and the 10$\mu$m feature are present with SNR $> 20$: $\tau = 0.023$ (no correlation between these features) with a significance probability $P = 0.83$, reflecting the dispersion. Finally, the right panel of Fig.\[LF10\] shows a similar trend for the 33.6$\mu$m and the amorphous 10$\mu$m feature: $\tau = 0.22 $ and a significance probability $P = 0.15$. Therefore, the flux radiated by the 10$\,\mu$m amorphous feature is largely unrelated to the energy contained in the crystalline silicates features that appear at wavelengths longer than 20$\,\mu$m. Similarly, the shape of the amorphous 10$\mu$m feature (that is, the $S_{11.3}/S_{9.8}$ ratio, which is a proxy for grain size, Sec.\[sec:gsize\]) seems uncorrelated with the emergence of the 23$\mu$m complex, the 28$\mu$m complex, or the 33.6$\mu$m forsterite feature. The right panel of Fig.\[LF10\], for example, shows the number of positive detections of the C23, C28 and 33.6$\mu$m features per $S_{11.3}/S_{9.8}$ bin, where the errors bars correspond to Poisson noise (square root of the number of detections). The fraction of disks with C23 and C28 features is found to be independent of the mean grain size probed by the 10$\mu$m amorphous feature, further suggesting that the 10$\mu$m feature and the C23, C28 features essentially probe disconnected populations. For the 33.6$\mu$m feature, the fraction of objects per bin seems less independent because of the smaller numbers of detections (23 objects with both the 10 and 33.6$\mu$m features). Overall, these results suggest that the crystalline features appearing at wavelengths longer than 20$\mu$m and the 10$\mu$m amorphous feature arise from unrelated populations. On the other hand, Figure\[LF\] shows the correlation between the normalized line fluxes of the C23 and C28 complexes where both are present with SNR $> 20$, and also their correlation with the 33.6$\mu$m feature. The Kendall $\tau$ tests and associated significance probabilities confirm the expected result that the crystalline features at wavelengths longer than 20$\mu$m have emission fluxes that correlate to each other. Feature correlation coefficients -------------------------------- 9.2$\mu$m 10.0$\mu$m 11.3$\mu$m 12.5$\mu$m 16.2$\mu$m 21.6$\mu$m C23 25.0$\mu$m C28 33.6$\mu$m ------------ ------------ ------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ 9.2$\mu$m [**15**]{} 0.008 0.238 0.012 0.003 0.622 0.000 0.887 0.000 0.195 10.0$\mu$m 0.187 (14) [**69**]{} 0.000 0.293 0.049 0.088 0.468 0.203 0.858 0.441 11.3$\mu$m -0.084 (1) -0.447 (3) [**12**]{} 0.556 0.557 0.056 0.001 0.363 0.000 0.182 12.5$\mu$m 0.179 (4) 0.075 (10) 0.042 (2) [**12**]{} 0.087 0.308 0.113 0.003 0.004 0.182 16.2$\mu$m 0.211 (6) -0.140 (12) 0.042 (3) 0.121 (4) [**19**]{} 0.000 0.000 0.732 0.115 0.068 21.6$\mu$m 0.035 (2) 0.121 (9) -0.135 (1) 0.072 (2) 0.253 (5) [**10**]{} 0.053 0.025 0.025 0.963 C23 0.354 (15) 0.051 (43) 0.244 (11) 0.112 (9) 0.299 (17) 0.137 (8) [**56**]{} 0.000 0.000 0.000 25.0$\mu$m -0.010 (2) 0.090 (11) -0.064 (1) 0.214 (4) 0.024 (3) 0.159 (3) 0.325 (13) [**13**]{} 0.178 0.318 C28 0.260 (13) 0.013 (40) 0.332 (12) 0.202 (10) 0.112 (13) 0.158 (8) 0.619 (46) 0.095 (9) [**53**]{} 0.001 33.6$\mu$m 0.092 (6) 0.055 (22) 0.095 (5) 0.095 (5) 0.129 (8) -0.003 (3) 0.288 (23) 0.071 (5) 0.233 (21) [**28**]{} The one-to-one correlation probabilities between the detection frequencies of all the features examined in this study with SNR $> 20$ are displayed in Table\[correl\_f\]. The diagonal shows the total number of positive detections for any given feature (these numbers being close to percentages of detections as we have 96 objects). The lower part of the Table displays the Kendall $\tau$ values as well as the number of spectra with common detections of the considered features between parenthesis. The Kendall significance $P$ probabilities are displayed in the upper part. At this point, it is interesting to understand the meaning of the inferred Kendall $\tau$ values in this Table. Some correlations can possibly be found according to the Kendall $\tau$ values but only because both features are absent from spectra. Consider as an example the 11.3$\mu$m forsterite feature and the C23 complex, for which we obtain $\tau = 0.244$ and $P = 0.001$, for 12 detections of the 11.3$\mu$m feature and 56 detections of the C23 complex, indicating a weak correlation. Out of the 96 stars, 11 spectra show both features, one spectrum shows the 11.3$\mu$m feature only, and 45 objects show the C23 complex only. The other 39 ($96-45-1-11$) objects show neither the 11.3$\mu$m feature nor the C23 complex, and these objects account for concordant pairs. This explains why we find a weak correlation (non-zero $\tau$ value), as the number of concordant pairs (50) is of the order of the number of discordant pairs (46) for this example. In case we would eliminate from the statistics objects that do not show any of these two features, we would obtain an anti-correlation with $\tau = -0.25$ and $P = 0.004$. Therefore, in order to have comparable numbers for all the features, irrespective of the number of detections, we decide to consider the entire sample for the calculation of the correlation coefficients, even when spectra do not show any of the considered features. The detection frequencies of the 10$\mu$m amorphous feature and the C23 complex are found to be essentially uncorrelated ($\tau = 0.051$ with $P = 0.47$). Similar results are obtained when considering the 10$\mu$m feature and the C28 complex ($\tau = 0.013$ with $P = 0.86$), and also the 10$\mu$m and 33.6$\mu$m features ($\tau = 0.055$ with $P = 0.44$). For the C23, C28 complexes and the 33.6$\mu$m feature, we have, respectively, 43, 40 and 22 common detections with the amorphous 10$\mu$m feature (which is observed in 69 spectra). The C23 and C28 complexes, as well as the 33.6$\mu$m forsterite feature, do therefore appear in spectra independently from the presence of the 10$\mu$m feature. On the other hand, the C23 complex appears to be coincident with the C28 complex in our spectra ($\tau = 0.619$ with $P < 10^{-38}$), which concurs with our results based on line fluxes (Sec.\[sec\_cryst\_10\]). The 33.6$\mu$m feature also appears to be coincident, to a slightly lesser extent with the C23 complex ($\tau = 0.288$ with $P < 10^{-38}$), and the C28 features ($\tau = 0.233$ with $P = 3 \times 10^{-3}$). The number of common detections also confirm these trends. Table \[correl\_f\] further shows that the forsterite 11.3$\mu$m feature is anti-correlated with the amorphous 10$\mu$m feature ($\tau = -0.44$, only 3 objects show both the 10$\mu$m and the 11.3$\mu$m features). The anti-correlation between the forsterite 11.3$\mu$m feature and the amorphous 10$\mu$m feature could be explained by a contrast effect. The amorphous feature is indeed rather stronger than the 11.3$\mu$m feature which could therefore potentially be present but hidden in the amorphous feature (see discussion in Sec.\[sec:contrast\] for a more detailed analysis of this effect). Interestingly, the crystalline features in the 10$\mu$m range (9.2 and 11.3$\mu$m) seem to be correlated with both the C23 and C28 complexes: if crystalline features are present in the 10$\mu$m spectral range then the presence of the two complexes is highly probable. Still the reverse consideration is not true, if one of the two complexes (C23 or C28) is present, it is not guaranteed that we will find the 9.2 or 11.3$\mu$m features. This reflects that the correlation is dominated by objects where none of the considered features is present. Furthermore, as we find no strong correlation for the line fluxes regarding these features, this would mean that there is a general increase of crystallinity for several objects. Also, the two features at 9.2 and 11.3$\mu$m do not seem to be correlated with the 33.6$\mu$m forsterite feature. The declining quality of the data (lower SNR) at the very end of the IRS spectra may explain this trend in part, as some (lower contrast) 33.6$\mu$m features may not have been detected. Another possible explanation would be that the 33.6$\mu$m forsterite feature may probe an even colder disk region than the population probed by C23 and C28 complexes. To summarize, our results show that the C23, C28 and 33.6$\mu$m crystalline features probably arise from the same population of grains. Also, the detections of the 9.2 and 11.3$\mu$m crystalline features are correlated with the detections of C23 and C28 (but the presence of one of the longer wavelengths complexes does not involve the detection of 9.2 or 11.3$\mu$m features), which can be interpreted as a general increase of crystallinity in the inner and outer regions of disks. However, this crystalline population seems to be disjoint from the population of amorphous silicate grains that produce the 10$\mu$m feature. Degree of crystallinity of the warm versus cold disk regions {#sec:contrast} ------------------------------------------------------------ ![image](fig13.ps){width="105.00000%"} ![image](fig14.ps){width=".525\textwidth"} ![image](fig15.ps){width=".525\textwidth"} Both Figures \[fig:tt\_feat\] and \[fig:clouds\] show larger occurrences for the C23 and C28 emission features than for smaller wavelength crystalline silicate features, especially the 9.2 (enstatite) and 11.3$\mu$m (forsterite) features. As silicates are expected to thermally anneal close to the star, this result may reflect a counterintuitive positive gradient of the crystalline silicate fraction toward colder temperatures. The presence of the broad, and generally strong, amorphous $10\,\mu$m feature may lower the contrast of (or even hide) the shortest wavelength crystalline features, however, preventing direct conclusions. We examine in this section this apparent [*crystallinity paradox*]{} by comparing our Spitzer/IRS observations to synthetic spectra. To better gauge the actual differences between the degrees of crystallinity of the warm ($\lambda\sim10\,\mu$m) versus cold ($\lambda >20\,\mu$m) disk regions, we generate synthetic spectra representative of the observations, and from which we extract line fluxes as was performed for the IRS spectra in Sec.\[sec:measure\]. We consider amorphous species that include silicates of olivine stoichiometry (glassy MgFeSiO$_{4}$, optical constants from [@Dorschner1995]), and silicates of pyroxene stoichiometry (glassy MgFeSi$_{2}$O$_{6}$, optical constants from [@Dorschner1995]). For the crystalline species, we use enstatite (MgSiO$_{3}$) optical constants from @Jaeger1998 and forsterite (Mg$_{2}$SiO$_{4}$) optical constants from @Servoin1973. Theoretical mass opacities $\kappa_\lambda$ are computed using Mie theory (valid for hard spheres) for the amorphous species and DHS theory (Distribution of Hollow Spheres, [@Min2005]) for the crystalline silicates. Twelve example opacity curves for the four compositions and three typical (in terms of spectroscopic signature) grain sizes ($0.1\,\mu$m, $1.5\,\mu$m and $6.0\,\mu$m) are displayed in Fig.\[blowopac\]. The synthetic spectra of Fig.\[fig:contrast\] have been generated by adding to a representative continuum ($C_{\nu}$), the mass opacities of amorphous and crystalline silicates, multiplied by relative masses and by a blackbody at reference temperatures (350K for the warm region, and 100K for cold region). The amorphous content is a 50:50 mixture of olivine and pyroxene and the crystalline content is a 50:50 mixture of enstatite plus forsterite. The adopted $10\,\mu$m representative continuum is given by the median values (mean slopes and mean offsets: $C_{\nu} = \bar{a} \times \lambda + \bar{b}$) of all the local continua linearly ($C_{\nu}^{i} = a ^{i} \times \lambda + b ^{i}$) estimated in Sec.\[sec:measure\] for the observed amorphous 10$\mu$m features: $C_{\nu} = 0.01 \times \lambda + 0.097$, with $C_{\nu}$ in Jy and $\lambda$ in $\mu$m. Similarly, a representative C23/C28 continuum has been derived for cold synthetic spectra: $C_{\nu} =-0.001 \times \lambda + 0.43$ ($C_{\nu}$ in Jy and $\lambda$ in $\mu$m). To obtain as representative synthetic spectra as possible, we compute the median observed continuum over feature ratio (i.e., the continuum flux divided by the height of the feature) for the 10$\mu$m feature and the C23 complex. For the amorphous 10$\mu$m feature, we obtain a ratio of 2 and we therefore normalize the opacities, before adding them to the continuum ($C_{\nu}$), so that the synthetic spectra show similar ratios. For the C23, we measure a median continuum over feature ratio of 20$\pm$10, and we normalize the opacities in order to get a ratio between 10 and 30 in the synthetic spectra. This exercise could not be reproduced for the large 6.0$\mu$m grains because the shapes of their C23/C28 features are not representative of the observed features (especially for the forsterite grains, see right panel of Fig.\[blowopac\]), suggesting such large grains are not the main carriers of these features. As a final step, the synthetic spectra are degraded by adding random noise, with a maximum amplitude of $1.8 \times 10^{-3}$ Jy for the warm spectra and $7.2 \times 10^{-3}$ for the cold spectra, representative of the quadratic sum of the measured uncertainties on the Spitzer data over in the 10$\mu$m region and in the C23/C28 spectral range. The synthetic observations displayed in Fig.\[fig:contrast\] are then processed as in Sec.\[sec:measure\] to derive the band fluxes of the crystalline features. to The theoretical SNRs for the $9.2$, $11.3\,\mu$m, C23 and C28 crystalline silicate features, assuming crystalline mass fractions between 0% and 50% and characteristic grain sizes, are displayed in Figure\[fig:snr\_vs\_crys\]. It is seen that the SNR reached in the simulations compare well with the measured SNR values, especially for micron-sized grains (Table\[IDfeat\]). The diagonal-striped areas correspond to crystalline mass fractions where the continuum over C23/C28 feature ratios could hardly match the median observed values, or which totally fail to reproduce them. This means that we are not able to discuss the detectability of features for crystalline mass fractions smaller than 5 and 15% for grains sizes of 0.1 and 1.5$\mu$m, respectively. In the particular case of the 6.0$\mu$m grains, as explained above, we cannot compute SNR for C23 and C28 complexes as opacities do not match typical observations. We see in Fig.\[fig:snr\_vs\_crys\] that the SNR of the crystalline features drops with increasing grain size (especially the $9.2\,\mu$m, $11.3\,\mu$m features) and with decreasing crystalline mass fraction, as expected. A crystalline fraction of $\sim 5$% or larger seems to be enough such that all four of the studied crystalline features are observable if the grains are sub-micron in size. For the $1.5\,\mu$m-sized grains, crystalline mass fractions of about 15% are required, (1) to obtain spectra consistent with observations, and (2) to get the all the features studied detected with SNR$>$20. For higher crystalline mass fractions, they all present very similar behaviors. For $6.0\,\mu$m-sized grains, it becomes more difficult to trace the detectability of the C23 and C28 features as opacities become much less consistent with typical observations. For 9.2$\mu$m and 11.3$\mu$m, crystalline fractions larger than 20% and 15%, respectively, are required in order to detect these features with SNR$>$20. Since the observed features in our sample are mainly produced by $\mu$m-sized grains (Sec.\[sec:gsize\]), which in particular produce a shoulder at the location of the $11.3\,\mu$m crystalline feature, we conclude that the physical processes that preferentially place $\mu$m-sized grains in the disk atmospheres do not contribute much to the measured differences between the number of detections of the $11.3\,\mu$m feature compared to the number of detections of the C23 complex (Figs.\[fig:tt\_feat\] and \[fig:clouds\]). To summarize, we find an apparent [*crystallinity paradox*]{}, namely a noticeable difference in detection statistics for the $11.3\,\mu$m and C23 features, with more than 3 times more detections for C23 compared to the number of detections of the 11.3$\mu$m forsterite feature. This seems counterintuitive as shorter wavelengths are expected to probe warmer disk regions where grains can more efficiently crystallize. We investigated the possibility that crystalline features could be hidden by amorphous features. The effect seems insufficient to explain the observations, but we cannot yet firmly predict the real impact of this effect until the Spitzer observations are confronted by a more detailed compositional analysis. This is deferred to a future paper (Olofsson et al. 2009b, in prep.). Discussion {#sec:discussion} ========== On the lack of Fe-rich silicates -------------------------------- Because of the large number of objects analysed in this paper, our results raise challenging questions on the dynamics and (chemical) evolution of planet forming disks regions that are in principle not statistically biased. For instance, the lack of Fe-rich silicates in our 96 star sample[^2] reinforces the result by @Bouwman2008 obtained on their 7-star sample. Silicates crystals studied in this analysis may contain iron, as seen in Sec\[sec:ferich\] but magnesium is the most abundant element. Several scenarios can explain this apparent lack of iron. First, @Davoisne2006 found that a reduction reaction during the thermal annealing of ferro-magnesian amorphous silicates at temperatures below 1000K can produce pure forsterite crystals plus spheroidal metallic particles. The iron is therefore locked inside a metallic nanophase that cannot be observed. Secondly, according to @Nuth2006, simple thermal annealing in the warm inner regions of the disk cannot produce Fe-rich silicate crystals. As the annealing timescale for iron silicates is larger than for magnesium silicates ([@Hallenbeck2000]), the lifetime of such Fe-rich silicates against the evaporation timescale is expected to be too short for these grains to crystallize and then be transported outward in the disk. Still, thermal annealing induced by shocks may be able to produce both Fe-rich and Mg-rich crystals at the same time. On the need for turbulent diffusion and grain-grain fragmentation\[sec:frag\] ----------------------------------------------------------------------------- ![\[fig:gs\]Correlation between shape and strength for the amorphous 10$\mu$m feature compared to models assuming a power law grain size distributions with indexes between -4 and -3. For each colored curve, from right to left, open circles correspond to mimimum grain sizes $a_{\mathrm{min}}$ equal to 0.1, 0.5, 0.75, 1.0, 1.5, 2.0 and 4.0$\mu$m. For blue points only the last four mininum grain sizes are represented, for red points the 0.1$\mu$m point is not shown, as their $S_{\mathrm{Peak}}^{10\,\mu m}$ is larger than $3.5$.](fig17.ps){width="1.\columnwidth"} The new grain size proxy introduced here for cold crystalline silicates (via the C23 complex) indicates that the upper layers of the disks are preferentially populated by $\mu$m-sized grains. This result also holds for the warm disk region using the classical size proxy for the carriers of the 10$\mu$m amorphous feature (Fig.\[kurt\_sp\]). Adopting a purely sequential evolutionary approach for the solid constituents in disks, where interstellar-like grains coagulate to form micron-sized dust aggregates, then pebbles and planetesimals, a straighforward conclusion from these results would be that grain agglomeration occured in the first few AU and that we are witnessing an intermediate, regular step of the growth of solid particles along their way to form planets. Nevertheless, this picture is most certainly an oversimplified view of the processes at work in planet forming regions around young stars as it does not fit with basic dynamical considerations. Indeed, the short settling time-scale of $\mu$m-sized grains in a laminar disk, and their expected fast growth which will accelerate sedimentation [@Dullemond2005; @Laibe2008], both argue in favor of some processes that sustain the disk atmosphere with grains that have mid-infared spectroscopic signatures (grains smaller than about 10$\,\mu$m) over a few Myr. Therefore, our results support the replenishment of the disk upper regions by the vertical transport of dust particles [e.g. turbulent diffusion, @Fromang2009], likely combined with grain-grain fragmentation to balance the expected efficient growth of solid particulates [that accelerates sedimentation, @Dullemond2005; @Laibe2008]. In this picture, the grains observed at mid-IR wavelengths would then result from the destruction of much larger solid particles [an assumption consistent with the presence of millimeter-sized grains in most TTauri disks, e.g. review by @Natta2007], rather than from the direct growth from interstellar-like grains. On the depletion of submicron-sized grains {#sec:depl} ------------------------------------------ ![\[fig:S11\_cr\]Correlation between shape and strength for the amorphous 10$\mu$m feature compared to models assuming a variation in crystallinity fraction, for three different grain sizes. For each grain size, increasing crystallinity goes from bottom left to top right, from 0% to 50% crystallinity (0, 5, 10, 15, 20, 25, 30, 40, 50%)](fig18.ps){width="1.\columnwidth"} A consequence of the above conclusion is that one would have to explain the general apparent depletion of submicron-sized grains in the disk zones probed by mid-IR spectroscopy. The fact that we see a flat, boxy 10$\mu$m feature profile for most TTs, indeed indicates that there cannot be many submicron-sized grains, because their emission would overwhelm the few large grains. Grain-grain collisions are nevertheless expected to produce fragments with a range of sizes that are likely to extend to the submicron domain (i.e. the size of the monomers that form $\mu$m-sized aggregates). This effect might be enhanced by the large velocity fluctuations (a fraction of the local sound speed) found at high altitudes above the disk midplane in disk simulations with MHD turbulence [@Fromang2009]. Qualitatively, submicron-sized grains, lifted in the disk atmospheres by turbulent diffusion or produced locally, are therefore expected to be abundant, which contradicts the IRS observations. ### Impact of the size distribution The above conclusion is based on our single, characteristic grain size approach (Secs.\[sec:gsize\] and \[sec:gsize2\]) and one may question whether this assumption affects our interpretation of the observations. To answer that question, we calculate mean cross-sections as follows: $$\begin{aligned} \sigma_{\mathrm{abs}} = \int_{a\mathrm{_{min}}}^{a\mathrm{_{max}}} \pi a^2 Q_{\mathrm{abs}}(a)\,\mathrm{d}n(a) / \int_{a\mathrm{_{min}}}^{a\mathrm{_{max}}}\mathrm{d}n(a),\end{aligned}$$ where $a$ is the grain radius, $Q_{\mathrm{abs}}$ is the dimensionless Mie emission/absorption coefficient for a 50:50 mixture of amorphous pyroxene and olivine, and $\mathrm{d}n(a)\propto a^p \mathrm{d}a$ is the differential grain size distribution. We arbitrarily choose $a\mathrm{_{max}}\,=\,100\,\mu$m (large enough such that it does not affect the calculations for MRN-like size distributions), and we vary the minimum grain size $a\mathrm{_{min}}$ between 0.1 and 4.0$\mu$m. Using a classical MRN ($p=-3.5$) grain size distribution, we find that the diversity of $S_{11.3}/S_{9.8}$ and $S_{\mathrm{Peak}}^{10\mu m}$ values is well reproduced by adjusting the minimum grain size $a_{\mathrm{min}}$, and that the bulk of the observations is reproduced for $a_{\mathrm{min}}$ larger than 2$\mu$m (Fig.\[fig:gs\]), a result consistent with our single grain size analysis. Nevertheless, another way to avoid having peaked 10$\mu$m amorphous features typical of pristine grains is to use a flatter slope for the size distribution. For example, $p$ values larger than $-3$ reduce the contribution of submicron-sized grains to the mean absorption cross-section (see Fig.\[fig:gs\] with $p=-3$), creating an amorphous $10\,\mu$m feature that is large and boxy despite the presence of some submicron-sized grains. Therefore, the majority of the points is located at the left side of Fig.\[fig:gs\], which either probes $\mu$m-sized grains as the small end of the size distribution, or much flatter grain size distributions compared to the MRN, meaning that the amount of small grains is at least severely diminished in the regions probed by the Spitzer/IRS. ### Impact of the crystallinity Several studies (e.g. [@Bouwman2001], [@Honda2006] or [@2005]) mention the possible impact of the crystallinity on the shape versus strength correlation for the 10$\mu$m feature. According to these studies, the presence of crystalline grains can mimic the flattening of the feature. @Min2008 recently investigated in detail this effect by considering inhomogeneous aggregates. Their Fig.2 shows that $S_{\mathrm{Peak}}^{\mathrm{10\,\mu m}}$ remains approximately constant for crystalline fractions up to 20% (the maximum fraction considered in their study) while $S_{11.3}/S_{9.8}$ rises, suggesting a shift along the $y$–axis of Fig.\[fig:gs\] due to crystallinity rather than along the observed trend. In order to further investigate the impact of crystallinity on this correlation, we use the synthetic spectra described in Sec.\[sec:contrast\]. The only difference is that we do not normalize opacities anymore to obtain a constant ratio between feature strength and continuum. Normalizing opacities would lead to a constant $S_{\mathrm{Peak}}^{\mathrm{10\,\mu m}}$, which is not what we want to characterize here. We therefore normalize the opacities, for each grain size, so that only the strongest feature reaches a ratio of 2 between the continuum and the feature strength. Other features are normalized with the same factor and obviously present a higher ratio between continuum and feature strength. This normalization is done in order to reproduce spectra similar to observations, and still conserve the diversity of feature strengths. Subsequently, values for $S_{9.8}$, $S_{11.3}$ and $S_{\mathrm{Peak}}^{\mathrm{10\,\mu m}}$ are derived in the same way as in Sec\[sec:gsize\]. Fig.\[fig:S11\_cr\] shows the impact of crystallinity on the shape versus strength for the 10$\mu$m feature. We consider three grain sizes (0.1, 1.5 and 6.0$\mu$m) and the same composition as in Sec.\[sec:contrast\]: 50:50 mixture of olivine and pyroxene for the amorphous content and 50:50 of enstatite and forsterite for the crystalline population. The crystallinity fraction varies from 0% to 50% (0, 5, 10, 15, 20, 25, 30, 40 and 50%), and increasing crystallinity goes from bottom left to top right on the plot. We note that for 6.0$\mu$m-sized grains the effect is negligible, while it is more dramatic for 0.1$\mu$m-sized grains. The spread induced by an increase of crystallinity is almost orthogonal to the direction of the correlation. This result was also found by @Kessler-Silacci2006 (Fig.10e). The direction of this deviation can simply be understood when considering opacities for amorphous and crystalline grains. For the same mass of dust, crystalline grain emission will be stronger compared to amorphous grains. Therefore adding some crystalline grains to a purely amorphous dust content will give a 10$\mu$m feature stronger [*and*]{} broader (because of crystalline features at, e.g., 11.3$\mu$m in the case of forsterite). This will thus produce a higher value of $S_{11.3}/S_{9.8}$ and at the same time a stronger $S_{\mathrm{Peak}}^{\mathrm{10\,\mu m}}$. Crystallinity therefore does impact the correlation and may be responsible for a dispersion in the direction orthogonal to the correlation, but according to the opacities, it cannot solely explain the observed trend. ### Hypothesis for the depletion of submicron-sized grains Several solutions could be envisaged to explain the apparent depletion of submicron-sized grains at high altitudes above the disk midplane. The fragmentation time-scale of grains in disk atmospheres may be longer than the grain growth timescale, thereby quickly eliminating the smallest and most freshly produced submicronic grains. The high velocity fluctuations observed in MHD simulations at high altitudes, do not, in principle, work in favor of this scenario. An alternative possibility may be that the submicron-sized grains are evacuated from the upper layers of the disks by stellar winds or radiation pressure. The impact of radiation pressure on grains in young disks has been explored by @Takeuchi2003 and present several advantages. First, it acts on short timescales as long as the grain size reaches a threshold, usually referred to as the blowout size limit, below which the radiation pressure force overcomes the gravitational and gas drag forces. Second, the blowout size limit falls in the right regime of grain sizes (close to a micrometer for silicates) for solar-type stars in the presence of some gas. It is interesting to apply the latter scenario to the case of more luminous stars. According to @Kessler-Silacci2007, HAeBe stars tend to show a widest variety of 10$\,\mu$m silicate features, including the peaked features attributed to submicron-sized grains. Because of their higher temperature and luminosity, HAeBe stars are expected to have upper layers that are even more ionized than those of TTauri stars, thereby enhancing the turbulence if it is supported by disk magnetic fields, as well as velocity fluctuations. This would therefore encourage the systematic production of submicron dust fragments in disk atmospheres. On the other hand, silicate grains of a few micrometers and smaller, are in principle more prone to be eliminated by radiation pressure. Nevertheless, the latter reasoning assumes an optically thin medium at the wavelengths where grains absorb most, namely in the UV and visible. The optically thin region at mid-IR wavelengths is likely to be essentially opaque in the UV/visible, and this opacity will affect more strongly the most luminous stars, thereby possibly reducing the blowout size limit. Clearly, the depletion of submicron-sized grains in the disk atmospheres of most TTauri stars should be further investigated by models to examine the likelihood of the scenarios discussed above (competition between coagulation and fragmentation, stellar winds, radiation pressure, ...) On radial mixing ---------------- The coldest disk component probed by our IRS observations shows a rather high crystallinity fraction. According to Fig.\[fig:tt\_feat\], the 23 and 28$\,\mu$m complexes are detected in more than 50% of the spectra. In a scenario where silicate crystals form by thermal annealing, the production of crystalline silicates can only occur on reasonably short time-scales near the star (T $>$ 900K), namely in regions much hotter than those emitting at $\lambda > 20\,\mu$m. The observed cold crystals would thus have been radially transported there from the inner disk zone, which according to two-dimensional models can occur around the midplane of accretion disks [@Ciesla2009 and references therein]. @Keller2004, for example, find that $\sim 5$ to 30% of the total mass that is transported inward at high disk altitudes is transported outward near the disk midplane. For a solar mass star with an accretion rate of $\dot{M} = 10^{-6}M_{\odot}$/yr, @Ciesla2009 estimates the crystalline fraction at 10AU to be about $\sim 40$% by the end of his simulations. As the accretion rate becomes smaller, the crystalline fraction drops accordingly: $\sim 10$% and $\sim 1$% for $\dot{M} = 10^{-7}M_{\odot}$/yr and $10^{-8}M_{\odot}$/yr, respectively. However, @Watson2009 find only a weak correlation between the accretion rate and the crystalline features arising at around 10$\,\mu$m, and no correlation with the 33$\,\mu$m feature, leading the authors to suggest that some mechanisms may be erasing such trends. Alternatively, other crystallisation processes at colder temperatures (e.g. via nebular shock, [@Desch2002]) could also contribute to the presence of crystals at significant distances from the central object [e.g. @Kimura2008]. It is noteworthy that the warmer disk zone seems significantly less crystalline than do the colder regions, although detailed compositional analysis of the IRS spectra is required before firmly concluding this must be so (Olofsson et al. 2009b, in prep.) Interestingly, compositional fits to some individual IRS spectra by @Bouy2008 and @Sargent2009a do show the need for larger amounts of silicate crystals in the outer regions compared to inner regions (e.g. and ). If radial mixing leads to the large crystalline fraction in the cold disk component, it leaves the disks with an inhomogeneous chemical composition for what concerns the solid phase. Some re-amorphisation processing of the crystalline grains, by X-ray emission or cosmics rays for example, could also contribute to the apparent lack of crystals in the inner regions of disks. This result of a cold component being more crystalline compared to the inner warm regions seems, at first glance, to be in contradiction with the results by @2004 who observed three HAeBe stars with the mid-IR instrument MIDI for the VLTI interferometer. Using different baselines, they found that the inner regions (1–2AU) of the disks are more crystalline compared to the outer regions (2–20AU). Beside the fact that their sample is limited to 3 intermediate mass objects while we analyse a statistically significant TTs sample, direct comparison of their observations with ours is not straightforward. They do have spatial information, but only for the 10$\mu$m spectral region, while we cover a larger spectral range. We, on the other hand, are observing the entire disk with Spitzer/IRS, with no direct spatial information. An ideal and challenging way to confront such kinds of interferometric data with IRS would be to repeat similar observations for TTs and even better, at wavelengths larger than 20$\mu$m. This will ultimately require space-born interferometers because of the poor atmosphere transmission in the mid-IR. With such observations, however, we would then be able to constrain the crystalline fraction in the inner and outer regions of the disk more directly. On the similarities with Solar System objects --------------------------------------------- Studies of Solar Sytem asteroids and comets show strong similarities with our results. First of all, laboratory studies of asteroidal and cometary solids have show that they contain very little ISM-like materials. The 81P/Wild 2 samples from the Stardust mission @McKeegan2006, for example, show that this comet contains high-temperature silicates and oxide minerals. These authors conclude that such materials could not have been formed via annealing of presolar amorphous silicates in the Kuiper belt, and also that they could not have been formed from a single isotopic reservoir. Overall, this suggests that comet 81P/Wild 2 sampled different regions of the inner solar protoplanetary disk during its formation. One possible transportation mechanism proposed is the combination of winds associated with bipolar outflows. This hypothesis is also supported by laboratory measurements from @Toppani2006 on condensation under high temperature, low-pressure conditions. Mg-rich silicates crystals can rapidly ($\sim 1$h at $4 \times 10^{-3}$ bar) be produced from the gas phase through condensation, confirming the fact that crystalline grains can even be produced in bipolar outflows from evolved stars. In still further work on the 81P/Wild 2 Stardust mission, @Zolensky2006, found large abundances of crystalline materials. Olivine is present in most of the studied particles, with grain sizes ranging between submicron-sized to over 10$\mu$m, and it is concluded that some materials in the comet have seen temperatures possibly higher than 2000K. Similar to @McKeegan2006, large-scale radial transportation mechanisms inside the disk are required to match the derived mineralogy of the olivine samples. Similar results were found studying other solar system bodies, for exemple, comet Hale-Bopp. Spectroscopic studies of this object ([@Wooden1999], [@Wooden2000]) revealed the presence of crystalline olivine, crystalline ortho-pyroxene and crystalline clino-pyroxene, in significant quantities. All these results echo the [*crystallinity paradox*]{} we find in our Spitzer observations, in the sense that the disk dust content is far from being homogeneous, and even far from the simple picture with crystalline grains close to the central object and amorphous content in the outer regions. to Summary and conclusion {#sec:summary} ====================== We have conducted in this paper a comprehensive statistical study of crystalline silicates in proto-planetary (Class II) disks which was made possible thanks to the high sensitivity of the spectroscopic instrument IRS (5–35$\,\mu$m) onboard the Spitzer Space Telescope. This unprecedented work is among the first statistical studies of crystalline silicates emission features over a large sample of young solar-analog stars (96 objects). We find the following results: 1. Crystallinity is not a marginal phenomenon in Class II disks around young solar analogs, as more than 3/4 of the objects show at least one crystalline feature in the IRS spectra. Crystalline features emitting at wavelengths larger than 20$\mu$m are in addition widely present: both the C23 and C28 complexes, produced by enstatite plus forsterite features, are present at a frequency larger than 50%. 2. The crystalline silicates revealed by the IRS spectra are essentially Mg-rich silicates (forsterite, enstatite), with a small contribution of CaMg-rich silicates (diopside). We find no evidence of Fe-rich silicates (marginal evidence for fayalite) nor troilite (FeS) in our spectra. 3. We find that the amorphous 10$\mu$m feature is not correlated with the C23, C28 and 33.6$\mu$m crystalline features, neither on the basis of the frequencies at which they are detected nor on the fluxes they emit. Crystalline features emitted at wavelengths larger than 20$\mu$m are, on the other hand, correlated to each other (C23, C28 and 33.6$\mu$m), based either on detection frequencies and band fluxes. 4. From previous results, we conclude that within the spectral range of the Spitzer/IRS instrument, we are essentially probing two independent dust populations (see the illustration in Fig.\[fig:disksum\]): a warm component, which is likely located close to the central object ($\leq 1$AU), emitting features in the 10$\mu$m region, and a second component, located at larger radii ($\leq 10$AU) or deeper inside the disk; the latter cold component being responsible for the emission of both C23 and C28 complexes plus the 33.6$\mu$m forsterite feature. 5. For a large majority of objects, the upper layers of the disks are essentially populated by micron-sized grains, suggesting that submicron grains are largely depleted in disks atmospheres. For the amorphous 10$\mu$m feature, we find the known correlation between the shape and the strength of the feature used as a grain size proxy and showing that the grains are $\mu$m-sized or larger. We adapted this relationship to the C23 complex and find similar results for the crystalline grains producing this complex. We also noted that the grain size proxies for the 10$\,\mu$m and C23 features are uncorrelated. 6. We find a correlation between the shape of the SED ($F_{30}/F_{13}$) and the grain size proxy for the 10$\mu$m feature. We see rather small, warm amorphous silicates for flared disks and grains becoming larger as disks are flattening. This trend is almost reversed for the cold, C23 crystalline grains as small crystals are only found for flattened SEDs. 7. We identify an apparent [*crystallinity paradox*]{} since the cold crystalline features ($\lambda>20\,\mu$m) are much more frequently detected than the warm ones ($\lambda\sim10\,\mu$m). We show that crystalline features can be hidden by amorphous features, especially near $\lambda \simeq 10\,\mu$m, but this contrast effect is not sufficient to explain the difference regarding the apparition frequencies of the C23 complex and the 11.3$\mu$m forsterite feature, suggesting that the warm grain population is, on average, intrinsically more amorphous than the cold grain population. We argue that our Spitzer spectroscopic observations of disks around young stars reveal ongoing dynamical processes in planet forming regions of disks that future theoretical disk evolution and planet formation models will have to explain: - some mechanisms must supply the upper layers of the disks with micron-sized grains that should otherwise grow and settle rapidly. Vertical turbulent diffusion, accompanied by grain-grain fragmentation, would be natural candidates, - some mechanisms must act efficiently to remove the submicron-size grains from the regions which are optically thin at IRS wavelengths. Stellar winds and radiation pressure have been suggested as possible processes, - and, some mechanisms are responsible for rendering the colder disks regions much more crystalline than the warmer zones ([ *crystallinity paradox*]{}). If this result has to do with dust radial mixing, it teaches us that the mixing does not result in a disk with a homogenous grain composition, and that it must transport crystals efficiently into the “comet-forming regions”. The authors warmly thank Bram Acke for his help on the IDL routine to extract the crystalline features characteristics, Jérome Aléon and Matthieu Gounelle for sharing with us their precious knowledge of small solid Solar System bodies, and Joel Green for the rich discussions we had. We also thank the anonymous referee for the very useful comments that helped improving this study. It is also a pleasure to acknowledge fruitful discussions with the Grenoble FOST team members, as well as the participants to the ANR (Agence Nationale de la Recherche of France) project ANR-07-BLAN-0221. This research is partly supported by the [*Programme National de Physique Stellaire*]{} (PNPS). [^1]: Extensive, narrow features not discussed in this paper, but seen between $10$ and $20\,\mu$m in Fig.\[spectre\] are attributed to water emission lines [@Salyk2008] [^2]: except 5 marginal detections discussed in Sec.\[sec:ferich\]
{ "pile_set_name": "ArXiv" }
--- abstract: | Data collection designs for social network studies frequently involve asking both parties to a potential relationship to report on the presence of absence of that relationship, resulting in two measurements per potential tie. When inferring the underlying network, is it better to estimate the tie as present only when both parties report it as present or do so when either reports it? Employing several data sets in which network structure can be well-determined from large numbers of informant reports, we examine the performance of these two simple rules. Our analysis shows better results for mutual assent across all data sets examined. A theoretical analysis of estimator performance shows that the best rule depends on both underlying error rates and the sparsity of the underlying network, with sparsity driving the superiority of mutual assent in typical social network settings. *Keywords:* network inference, measurement, locally aggregated structures, informant accuracy, error author: - 'Francis Lee[^1] and Carter T. Butts[^2]' bibliography: - 'bibhome/las\_compare.bib' date: '1/27/18' title: ' Mutual Assent or Unilateral Nomination? A Performance Comparison of Intersection and Union Rules for Integrating Self-reports of Social Relationships [^3] ' --- Network inference is the problem of inferring an unknown graph from a set of error and/or missingness-prone observations. This problem is of fundamental importance in the study of social networks, where relationships between individuals, organizations, or other entities must typically be inferred from self or proxy reports, archival materials, or other imperfect source of information. Arguably, the most basic and familiar example of the network inference problem arises when attempting to integrate self-reports from subjects, each of whom is asked to identify all others with whom he or she has a particular relationship (or, in the case of a directed relationship, all others to/from whom he or she respectively sends and/or receives ties). Such data has been widely collected [see, e.g. @drabek.et.al:bk:1981; @reitz:white:1989; @bernard.et.al:ara:1984; @killworth.bernard:sn:1979; @pattison.et.al:jmp:2000], and poses a basic challenge for the analyst: given two reports on the state of a given relationship, what is to be done when the subjects disagree? @krackhardt:sn:1987 famously formalized two basic strategies for the analysis of such data (leading to respective estimators of the underlying network): regard an edge as present if either party reports it (the *union rule*); or regard an edge as present if and only if both parties report it (the *intersection rule*). While one or another rule has in some cases been argued to be preferred on substantive grounds, there has been little systematic investigation of how the rules perform on empirical data, and in particular on the relative performance of these rules in inferring network structure under realistic conditions. This paper seeks to address this gap, employing interpersonal networks whose complete structures can be well-estimated through hierarchical Bayesian models [@butts:sn:2003] to assess the accuracy of these simpler (but more widely applicable) rules. Our findings demonstrate that the intersection rule (“mutual assent”) generally outperforms the union rule (“unilateral nomination”) for the networks studied here - a surprising result, given that our informants are not prone to making one particular type of error over another. We resolve this discrepancy by showing that the sparsity of the network is key to the performance of the two rules, with the intersection rule dominating the union rule for networks in which the opportunities for false positives greatly outweigh the opportunities for false negatives. Background ========== Social network analysis is intrinsically and trivially dependent on the ability to accurately measure the structure of social relationships. Despite the rise of network measurement via online social networks, mobile devices, and other sources of observational data, collection of self-reports via sociometric surveys continues to be a popular method for network measurement in a wide range of settings. At least since the seminal studies of Bernard, Killworth, and Sailer– who provocatively (if hyperbolically) concluded that “there is no evidence that people know who their network connections are” [@bernard.et.al:ara:1984]–error from informant observations has been known to be a major challenge in network data collection and subsequent inference. Though subsequent studies [e.g. @freeman:1987; @romney:1982] have tempered the extremity of this conclusion, it is clear that error rates are substantial enough to warrant concern for social network research. With self-report (i.e informants reporting on their own ties) remaining a popular method of network data collection, there is an ongoing need for simple methods that can maximize the accuracy of networks inferred from this type of information. Although highly accurate estimates of network structure can be obtained when many measures of each potential tie are available [e.g., from cognitive social structure data - see @butts:sn:2003], simple self-report designs allow only two observations per edge variable (one for each party involved in the potential edge). The question, then, is how best to integrate these reports to infer the underlying network. Although many techniques are possible, we here focus on simple, easily used methods of aggregation that (1) estimate the state of an edge variable as being consistent with informants’ reports where they agree, and (2) resolve disagreements via a simple uniform rule. Such strategies lead to estimators[^4] referred to by @krackhardt:sn:1987 as *locally aggregated structures* (LAS), the union and intersection rules (U-LAS, I-LAS) being the special cases mentioned above. LAS estimators can be employed for both both undirected relations (when both parties report on the presence/absence of a single undirected edge) and directed relations (when both parties report on their incoming/outgoing ties); indeed, many networks collected in the former manner are erroneously treated as directed, where a LAS or other estimator of an underlying directed relation should be employed. The present work is thus applicable to any situation in which we obtain edge observations associated with both potential endpoints. Formal Framework {#sec_formal} ---------------- To formalize the above, our network inference problem may be summarized as follows. Let $G=(V,E)$ represent an unknown network of interest, with fixed and known vertex set $V$ and unknown edge set $E$. Without loss of generality, we will represent $G$ via its adjacency matrix, $\Theta$; where $G$ is undirected, $\Theta$ is constrained to be symmetric. Here, we assume the vertex set to be fixed and the edge set unknown. Informant reports are represented via an informant by sender by receiver adjacency array, $Y$, such that $Y_{ijk}=1$ if $i$ reports that the edge from $j$ to $k$ is present (with 0 otherwise). In our setting, we assume that informants report only on their own ties, and hence only $Y_{iij}$ and $Y_{jij}$ (and their reciprocating edge variables) are employed. The locally aggregated structure (LAS) introduced by @krackhardt:sn:1987 has been a popular network inference tool for aggregating an ego’s and alter’s judgments from such data. While this has traditionally been explored through cognitive social structures (which collect an informant’s perception of the entire social structure), the only responses *needed* from an informant are the ties they report sending out and the ties they perceive others send to them. In terms of the above, the union and intersection LAS estimators are defined as follows: $$\begin{gathered} {\ensuremath{{\hat{\Theta}^U}}}_{ij} = 1-(1-Y_{iij})(1-Y_{jij}) \label{e_ulas} \\ {\ensuremath{{\hat{\Theta}^I}}}_{ij} = Y_{iij} Y_{jij} \label{e_ilas}\end{gathered}$$ As noted above, [${\hat{\Theta}^U}$]{}estimates an edge as being present when either party reports it, while [${\hat{\Theta}^I}$]{}does so only when both parties agree. Both rules are simple and easily understood, but may lead to very different estimates of network structure. If one must employ either [${\hat{\Theta}^U}$]{}or [${\hat{\Theta}^I}$]{}, which should one use? To determine this, we consider the accuracy of each estimator under realistic conditions. LAS Accuracy: Some Basic Theory {#sec_theory} =============================== It is not immediately clear which LAS method would provide a more accurate estimate of the unknown graph. If informants uniformly make more false positive errors (reporting that an edge exists when it does not) relative to their false negative rate (reporting that an edge does not exist when it does) in their edge observations, then it would seem the LAS intersection would obtain an estimated graph close to the unknown graph. Conversely, if informants make a greater number of false negative errors relative to the false positive rate, then the LAS union would be expected to provide a closer estimate to the unknown graph. This intuition follows from the response of the respective rules to informant error rates on a per-edge basis. Define the false positive and false negative error rates for an arbitrary informant $i$ by $$\begin{gathered} {\ensuremath{{e^{{}+{}}}}}_i=\Pr(Y_{iij}=1|\Theta_{ij}=0)=\Pr(Y_{iji}=1|\Theta_{ji}=0)\\ {\ensuremath{{e^{{}-{}}}}}_i=\Pr(Y_{iij}=0|\Theta_{ij}=1)=\Pr(Y_{iji}=0|\Theta_{ji}=1),\end{gathered}$$ with ${\ensuremath{{e}}}=({\ensuremath{{e^{{}+{}}}}},{\ensuremath{{e^{{}-{}}}}})$ being the full set of error rates. Assuming that errors occur independently, it then immediately follows that the per-edge error rates for [${\hat{\Theta}^I}$]{}and [${\hat{\Theta}^U}$]{}are given by $$\begin{gathered} \Pr({\ensuremath{{\hat{\Theta}^I}}}_{ij}\neq\Theta_{ij}|{\ensuremath{{e}}}) = \begin{cases} {\ensuremath{{e^{{}+{}}}}}_i{\ensuremath{{e^{{}+{}}}}}_j & \Theta_{ij}=0\\ 1-(1-{\ensuremath{{e^{{}-{}}}}}_i)(1-{\ensuremath{{e^{{}-{}}}}}_j)& \Theta_{ij}=1\end{cases} \\ \Pr({\ensuremath{{\hat{\Theta}^U}}}_{ij}\neq\Theta_{ij}|{\ensuremath{{e}}}) = \begin{cases} 1-(1-{\ensuremath{{e^{{}+{}}}}}_i)(1-{\ensuremath{{e^{{}+{}}}}}_j) & \Theta_{ij}=0\\ {\ensuremath{{e^{{}-{}}}}}_i{\ensuremath{{e^{{}-{}}}}}_j& \Theta_{ij}=1\end{cases}. \end{gathered}$$ Where error rates are approximately equal across informants, [${\hat{\Theta}^I}$]{} false positive rates scale with the square of the individual false positive rates, with the same holding mutatis mutandis for [${\hat{\Theta}^U}$]{}and false negative rates; while this can result in substantial suppression for these types of errors, errors of the opposite type (false negatives for [${\hat{\Theta}^I}$]{}, false positives for [${\hat{\Theta}^U}$]{}) are correspondingly magnified. This is illustrated graphically in Figure \[f\_base\_las\_err\], which shows the “worst case” probabilities of a correct inference as a function of informant accuracy (with both informants assumed to have the same error rates). In the typical setting for which ${\ensuremath{{e^{{}-{}}}}}>{\ensuremath{{e^{{}+{}}}}}$—i.e., omission of true ties, due e.g. to forgetting, is more common than fabrication or confabulation of nonexistent ties—-this analysis suggests that ${\ensuremath{{\hat{\Theta}^U}}}$ should be more accurate than ${\ensuremath{{\hat{\Theta}^I}}}$ (perhaps by a large margin). ![Single edge-variable error rates for union and intersection rules, as a function of informant error rates. The intersection rule is robust to false positive errors but vulnerable to false negative errors, with the reverse holding for the union rule. \[f\_base\_las\_err\]](las_union_err.png "fig:"){width="2.5in"}![Single edge-variable error rates for union and intersection rules, as a function of informant error rates. The intersection rule is robust to false positive errors but vulnerable to false negative errors, with the reverse holding for the union rule. \[f\_base\_las\_err\]](las_intersect_err.png "fig:"){width="2.5in"} There is, however, another aspect to this problem. Consider the expected total (Hamming) errors for [${\hat{\Theta}^I}$]{}and [${\hat{\Theta}^U}$]{}given the true graph state: $$\begin{gathered} \mathbf{E} \sum_{(i,j)\in \mathcal{D}} \left|{\ensuremath{{\hat{\Theta}^I}}}-\Theta\right| = \sum_{(i,j)\in \mathcal{D}} \left[\Theta_{ij} \left({\ensuremath{{e^{{}-{}}}}}_i+{\ensuremath{{e^{{}-{}}}}}_j-{\ensuremath{{e^{{}-{}}}}}_i{\ensuremath{{e^{{}-{}}}}}_j\right) + \left(1-\Theta_{ij}\right){\ensuremath{{e^{{}+{}}}}}_i{\ensuremath{{e^{{}+{}}}}}_j\right] \label{e_errilas}\\ \mathbf{E} \sum_{(i,j)\in \mathcal{D}} \left|{\ensuremath{{\hat{\Theta}^U}}}-\Theta\right| = \sum_{(i,j)\in \mathcal{D}} \left[\Theta_{ij} {\ensuremath{{e^{{}-{}}}}}_i{\ensuremath{{e^{{}-{}}}}}_j + \left(1-\Theta_{ij}\right)\left({\ensuremath{{e^{{}+{}}}}}_i+{\ensuremath{{e^{{}+{}}}}}_j-{\ensuremath{{e^{{}+{}}}}}_i{\ensuremath{{e^{{}+{}}}}}_j\right)\right] \label{e_errulas}\end{gathered}$$ where $\mathcal{D}$ is the set of potential edges. As a simplifying assumption, let us take all informants to have the same error rates (hence ${\ensuremath{{e^{{}-{}}}}}_i={\ensuremath{{e^{{}-{}}}}}_j={\ensuremath{{e^{{}-{}}}}}$ and ${\ensuremath{{e^{{}+{}}}}}_i={\ensuremath{{e^{{}+{}}}}}_j={\ensuremath{{e^{{}+{}}}}}$ for all $i,j$). Let $M(\Theta)=\sum_{(i,j)\in \mathcal{D}} \Theta_{ij}$ be the number of edges in the true graph, and $N(\Theta)=\left|\mathcal{D}\right|-M(\Theta)$ the corresponding number of nulls. In this scenario, Eq. \[e\_errilas\] and \[e\_errulas\] reduce to $$\begin{gathered} \mathbf{E} \sum_{(i,j)\in \mathcal{D}} \left|{\ensuremath{{\hat{\Theta}^I}}}-\Theta\right| = M(\Theta) \left(2{\ensuremath{{e^{{}-{}}}}}-{\ensuremath{{e^{{}-{}}}}}^2\right) + N(\Theta){\ensuremath{{e^{{}+{}}}}}^2 \\ \mathbf{E} \sum_{(i,j)\in \mathcal{D}} \left|{\ensuremath{{\hat{\Theta}^U}}}-\Theta\right| = M(\Theta) {\ensuremath{{e^{{}-{}}}}}^2 + N(\Theta)\left(2{\ensuremath{{e^{{}+{}}}}}-{\ensuremath{{e^{{}+{}}}}}^2\right).\end{gathered}$$ Where the numbers of edges and nulls are similar (i.e., near density 0.5), the edgewise logic invoked above holds. But this is not always the case. To see why, let us consider the case in which the expected Hamming error under the intersection rule is greater than under the union rule, i.e. $$\begin{aligned} \left[\mathbf{E} \sum_{(i,j)\in \mathcal{D}} \left|{\ensuremath{{\hat{\Theta}^I}}}-\Theta\right|\right] &> \left[\mathbf{E} \sum_{(i,j)\in \mathcal{D}} \left|{\ensuremath{{\hat{\Theta}^U}}}-\Theta\right|\right]. \nonumber\\ \intertext{From the above, it follows that this condition corresponds to} M(\Theta) \left(2{\ensuremath{{e^{{}-{}}}}}-{\ensuremath{{e^{{}-{}}}}}^2\right) + N(\Theta){\ensuremath{{e^{{}+{}}}}}^2 &> M(\Theta) {\ensuremath{{e^{{}-{}}}}}^2 + N(\Theta)\left(2{\ensuremath{{e^{{}+{}}}}}-{\ensuremath{{e^{{}+{}}}}}^2\right), \nonumber \\ \intertext{and hence} M(\Theta) \left[{\ensuremath{{e^{{}-{}}}}}-{\ensuremath{{e^{{}-{}}}}}^2 \right] &> N(\Theta) \left[{\ensuremath{{e^{{}+{}}}}}-{\ensuremath{{e^{{}+{}}}}}^2 \right] \nonumber \\ \frac{M(\Theta)}{N(\Theta)} &> \frac{{\ensuremath{{e^{{}+{}}}}}(1-{\ensuremath{{e^{{}+{}}}}})}{{\ensuremath{{e^{{}-{}}}}}(1-{\ensuremath{{e^{{}-{}}}}})}, \label{e_ilascond}\end{aligned}$$ which implies that, for sparse graphs ($N(\Theta)\gg M(\Theta)$) the intersection LAS may outperform the union LAS even when ${\ensuremath{{e^{{}-{}}}}}>{\ensuremath{{e^{{}+{}}}}}$. Intuitively, this is is because the number of *false positive opportunities* is much larger than the number of false negative opportunities in a sparse graph—since the intersection LAS suppresses false positives, it eventually becomes favored at sufficiently low densities. This is illustrated graphically in Figure \[f\_ilas\_critden\], which shows the maximum density at which [${\hat{\Theta}^I}$]{}outperforms [${\hat{\Theta}^U}$]{}under the simplified scenario of Eq. \[e\_ilascond\]. As expected from the edgewise analysis, the impact of relative error rates is dominant on LAS performance for moderate densities; however, for densities below $\approx 0.1$, the intersection LAS outperforms the union LAS over a very wide range of informant error rates. Since densities this low (or lower) are extremely common in social network studies, this suggests that the intersection LAS may outperform the union LAS in many real-world settings. ![Maximum density at which ${\ensuremath{{\hat{\Theta}^I}}}$ has lower expected Hamming error than ${\ensuremath{{\hat{\Theta}^U}}}$, as a function of error rate. For sufficiently sparse graphs, [${\hat{\Theta}^I}$]{}can be superior even when false negative rates are high. (Note this pattern reverses when ${\ensuremath{{e^{{}-{}}}}}+{\ensuremath{{e^{{}+{}}}}}>1$, in which case individuals are perversely informative.) \[f\_ilas\_critden\]](ilas_critical_density.pdf){width="4.5in"} An Empirical Study of LAS Accuracy ================================== Although the above theoretical results are intuitive, they depend upon a number of simplifying assumptions. How does LAS accuracy play out in the real world? To address this, we conduct an empirical analysis of cognitive social structure data from several studies of two types of relations (friendship and advice-seeking ties) in organizations. We use statistical models based on the complete data (i.e., all informant reports on all edge variables) to estimate the underlying network structure, and then assess the relative accuracy of [${\hat{\Theta}^I}$]{}and [${\hat{\Theta}^U}$]{}based on these estimates. As we show, the empirical results from this analysis are consonant with our theoretical results regarding the superiority of the intersection LAS in sparse graphs. Data and Methods ---------------- To estimate the true network, we utilize a variant of the Bayesian Network Accuracy Model (BNAM) from @butts:sn:2003 with the efficiency-enhancing graph mixture priors of @butts:tr:2017, where we separately estimate the error rates of each informant self versus proxy ties. The foundations of the model are similar to that of cultural consensus theory [@romney.et.al:aa:1986], and more specifically the generalized Condorcet model described in @batchelder.romney:p:1988. As in these models, each informant is assumed to report on the state of each variable (here, each potential $i,j$ edge), erring independently with rates that depend upon the informant and the true state. In the BBNAM case, these are parameterized as informant-specific false positive and false negative rates (as described in section \[sec\_theory\]). Priors are placed on the error rates, and on the underlying graph structure; the former are here taken to be iid beta distributions, and the latter a continuous mixture of $U|man$ graphs of varying density and reciprocity. The resulting joint posterior provides estimates of both informant error rates and the true graph, given $Y$ (the array of informant reports). Formally, the model can be described as $$\begin{gathered} p(\gamma|\alpha_\gamma) = \mathrm{Dirichlet}(\gamma|\alpha_\gamma)\\ \Pr(\Theta|\gamma) = \prod_{{j,k} \in \mathcal{D}'} \left[\Theta_{jk}\Theta_{kj}\gamma_1 + \left(\Theta_{jk}(1-\Theta_{kj})+(1-\Theta_{jk})\Theta_{kj}\right)\gamma_2 + (1-\Theta_{jk})(1-\Theta_{kj})\gamma_3\right] \\ \begin{split} \Pr(Y|{\ensuremath{{e^{{}+{}}}}},{\ensuremath{{e^{{}-{}}}}},\Theta) =& \prod_{i=1}^N \prod_{(j,k)\in \mathcal{D}} \left[\Theta_{jk} \left[Y_{ijk} (1-{\ensuremath{{e^{{}-{}}}}}_{ijk}) + (1-Y_{ijk}){\ensuremath{{e^{{}-{}}}}}_{ijk}\right] \right.\\ &+ \left. \left(1-\Theta_{jk}\right)\left[Y_{ijk} {\ensuremath{{e^{{}+{}}}}}_{ijk} + (1-Y_{ijk})(1-{\ensuremath{{e^{{}+{}}}}}_{ijk})\right]\right] \end{split}\\\end{gathered}$$ where $\alpha_\gamma$ is a 3-vector of hyperparameters, $\gamma$ is a fully latent vector of prior dyad type frequencies, $\mathcal{D}'$ is the set of undirected dyads (in contrast to the set of directed dyads, $\mathcal{D}$), and other terms are as defined above. The assumption of homogenity of (respectively) self versus proxy report error rates gives us $$\begin{gathered} {\ensuremath{{e^{{}+{}}}}}_{ijk} =\begin{cases} {\ensuremath{{e^{{}+{}}}}}_{is} & \mathrm{if } i \in \{j,k\}\\ {\ensuremath{{e^{{}+{}}}}}_{ip} & \mathrm{otherwise} \end{cases}\\ {\ensuremath{{e^{{}-{}}}}}_{ijk} =\begin{cases} {\ensuremath{{e^{{}-{}}}}}_{is} & \mathrm{if } i \in \{j,k\}\\ {\ensuremath{{e^{{}-{}}}}}_{ip} & \mathrm{otherwise} \end{cases},\end{gathered}$$ where ${\ensuremath{{e^{{}+{}}}}}_{is},{\ensuremath{{e^{{}-{}}}}}_{is}$ are respectively $i$’s self-report error rates, and ${\ensuremath{{e^{{}+{}}}}}_{ip},{\ensuremath{{e^{{}-{}}}}}_{ip}$ are $i$’s proxy report error rates. We complete the model by placing independent Beta priors on these rates, giving $$\begin{gathered} p({\ensuremath{{e^{{}+{}}}}},{\ensuremath{{e^{{}-{}}}}}|\alpha_{{\ensuremath{{e^{{}+{}}}}}},\beta_{{\ensuremath{{e^{{}+{}}}}}},\alpha_{{\ensuremath{{e^{{}-{}}}}}},\beta_{{\ensuremath{{e^{{}-{}}}}}}) = \prod_{i=1}^n \left[ \mathrm{Beta}\left({\ensuremath{{e^{{}+{}}}}}_{is}|\alpha_{{\ensuremath{{e^{{}+{}}}}}_s},\beta_{{\ensuremath{{e^{{}+{}}}}}_s}\right) \mathrm{Beta}\left({\ensuremath{{e^{{}-{}}}}}_{is}|\alpha_{{\ensuremath{{e^{{}-{}}}}}_s},\beta_{{\ensuremath{{e^{{}-{}}}}}_s}\right) \right.\\ \left.\mathrm{Beta}\left({\ensuremath{{e^{{}+{}}}}}_{ip}|\alpha_{{\ensuremath{{e^{{}+{}}}}}_p},\beta_{{\ensuremath{{e^{{}+{}}}}}_p}\right) \mathrm{Beta}\left({\ensuremath{{e^{{}-{}}}}}_{ip}|\alpha_{{\ensuremath{{e^{{}-{}}}}}_p},\beta_{{\ensuremath{{e^{{}-{}}}}}_p}\right) \right].\end{gathered}$$ Samples from the joint posterior are obtained by a Metropolis within Gibbs algorithm, following the strategy of @butts:tr:2017 as described below. To compare LAS estimators, we employ data collected from eight networks in four studies: one of an entrepreneurial firm, two of technological firms, and one of a university group, hereby referred to respectively as “Silicon Systems,” “High Tech Managers,” “Italian University,” and “Pacific Distributors” [@krackhardt:sn:1987; @krackhardt:ch:1992; @casciaro:1998; @kilduff:2008]. Each of the four studies contains CSS data on advice-seeking and friendship relations. Silicon Systems (SS) was originally analyzed with 36 informants, but information on informants 13, 24, and 35 was missing for both friendship and advice networks, leaving a total of 33 informants.[^5] Italian University (IU) contains 25 informants. Pacific Distributors contains 47 informants. As organizations are common contexts for network studies, we regard these settings as broadly comparable to others commonly used by network researchers. We model both friendship and advice-seeking as directed relations, since both were collected using directed prompts (e.g. “to whom does X go for help and advice at work,” “whom does Y consider a friend”). Informant self-report and proxy report error rates for all eight networks are modeled as drawn a priori from iid Beta distributions as described above, with $\alpha$ and $\beta$ hyperparameters set to 1 and 11, respectively. As noted above, we employ the $U|man$ mixture prior for the network structure, with a Jeffreys hyperprior on the dyad type rates; this provides a minimally informative, “flat” prior with respect to density, reciprocity, and fine structure. The BNAM utilizes a MCMC procedure to jointly estimate the informant error rate parameters as well as the criterion graph. We take and discard 500 draws for our burn-in sample, and then sample additional 1000 draws from the joint posterior distribution in each case. We verify MCMC convergence via the $\hat{R}$ statistic. We employ the posterior network draws from the BNAM as our criterion, comparing the LAS intersection and LAS union against each draw; the distribution of error rates over posterior draws is thus a posterior estimate of the respective accuracies of [${\hat{\Theta}^I}$]{}and [${\hat{\Theta}^U}$]{}on each data set. The full BNAM estimate on the complete set of informant self and proxy reports represents the best available estimate of the total graph - and has been shown to be highly accurate in simulation studies [@butts:sn:2003; @butts:tr:2017] - providing a basis for assessing the much more information-limited LAS estimates. We measure the error of the LAS estimates versus the estimated criterion by the Hamming distance between the two graphs (i.e., the number of edge variables that would have to be changed to convert one into the other). The LAS estimator with lower posterior expected Hamming error is estimated to be more accurate for the case in question. Results ------- We begin our presentation of results by considering our informants’ estimated error rates (since these are expected to play a major role in LAS performance); LAS performance itself is then presented in section \[sec\_lasperf\]. ### Error Rates {#sec_error} We begin by taking posterior draws for the BNAM for each network, and examining the informant error rates; as shown in section \[sec\_formal\], these are expected to be strongly related to LAS performance. One immediate observation is that there are substantial differences in false negative rates for proxy vs. self-reports in all networks (Figure \[globaldiff\]). While there is considerable individual variability, individuals are on average substantially more likely to miss ties when reporting on others than when reporting on themselves. By turns, false positive rates are similar on average for self versus proxy reports, although there is a small but consistent tendency towards over-reporting of self-ties in all networks; the presence of a somewhat elongated right tail in each network also suggests that some individuals are especially prone to over-reporting for ties involving themselves vs. third party ties. This last result is in line with the conclusions by @kumbasar.et.al:ajs:1994, who found evidence that informants have a higher propensity to make false positive errors when describing their own ties. ![Distribution of differences between informants’ posterior mean error rates when reporting on their own ties versus reporting on ties between third parties. More positive numbers indicate higher error rates for self-report versus proxy reports; 0 indicates equal error rates. Means are shown by vertical lines. Study labels refer respectively to High-tech Managers (HM), Silicon Systems (SS), Pacific Distributors (PD), and Italian University (IU). The two relations measured were advice-seeking (AD) and friendship (FR). \[globaldiff\]](GlobalDiff.png){width="4in"} Figure \[globaldiff\] shows the differences between self and proxy reporting, but not absolute error levels; these are presented in Figure \[f\_fpfn\]. As LAS performance depends only on self-report error rates, we focus on those here. Overall, we see that the two error rates seem to be similar in most networks, with false negatives being slightly more prevalent on average. In one case (Pacific Distributors friendship) we see much higher false negative rates than false positive rates for most informants (a pattern typical of proxy report error rates), but the reverse is found in two cases (High-tech Manager advice and Italian University friendship). It is noteworthy that these outlying cases are associated neither with organizational context nor relation, suggesting that they stem from idiosyncratic factors. It is useful to contrast this situation with that for proxy reports, which are shown for the same networks in Figure \[f\_fpfn\_proxy\]. As one can observe in the proxy error reports, the false negative error rates dwarf the false positive error rates in magnitude. This is consistent with the general trend in global error rates estimated in @butts:sn:2003. As most reports in a CSS datum will be proxy reports, it stands to reason that the global error rates derived from @butts:sn:2003 would be closer to the proxy error rates that we display here. ![ Marginal posterior distributions of informant false positive (FP) and false negative (FN) self-report error rates for each of the 4 organizational settings: Silicon Systems (SS), High-tech Managers (HM), Italian University (IU), and Pacific Distributors (PD). Means indicated by vertical lines. \[f\_fpfn\]](EgoErrorRates.png){width="5in"} ![ Marginal posterior distributions of informant false positive (FP) and false negative (FN) proxy error rates for each of the 4 organizational settings: Silicon Systems (SS), High-tech Managers (HM), Italian University (IU), and Pacific Distributors (PD). Means indicated by vertical lines. \[f\_fpfn\_proxy\]](ProxyErrorRates.png){width="5in"} To the extent that false positive and false negative error rates are similar for self-report, our theoretical results suggest that the best-performing LAS will be governed by sparsity. To examine this question, we now turn to our empirical performance comparison. ### LAS Performance Comparison {#sec_lasperf} We present the Hamming error distribution of the two aggregation methods relative to the criterion in Figure \[LASPerformance\],s with the mean Hamming error of the two methods being presented in Table \[t\_hamerr\_sr\]. Although we focus on the performance of the LAS methods for aggregation, for completeness, we present the Hamming error distribution of the networks estimated from naive self-report (i.e., own report of outgoing ties) in Figure \[sec\_lasperf\]. We see that the LAS intersection consistently provides a more accurate estimate of the criterion graph relative to the LAS union. In addition, the posterior Hamming error distributions are fairly well-localized around their means. ![Posterior Hamming error distribution for the LAS Intersection and LAS Union methods relative to the criterion graph estimate (from the BNAM) for the various datasets, separated by environment and type of relation measured. The four environments were High-tech Managers (HM), Silicon Systems (SS), Pacific Distributors (PD), and Italian University (IU). The two relations measured were advice-seeking (AD) and friendship (FR). \[LASPerformance\]](LASPerformance.png){width="5in"} Dataset Union Intersection Self-Report --------- ------- -------------- ------------- HM-AD 166 75 124 HM-FR 106 35 72 SS-AD 204 93 141 SS-FR 173 87 121 PD-AD 693 392 583 PD-FR 271 167 232 IU-AD 91 27 58 IU-FR 155 46 101 : Hamming error of the LAS Union or LAS Intersection aggregation relative to the criterion graph (estimated as the BNAM central graph) for the various datasets, separated by environment and type of relation measured. The four environments were High-tech Managers (HM), Silicon Systems (SS), Pacific Distributors (PD), and Italian University (IU). Lower values indicate better performance. The two relations measured were advice-seeking (AD) and friendship (FR). \[t\_hamerr\_sr\] The definitive superiority of the LAS Intersection rule despite the relative similarity of self-report error rates suggests an origin in sparsity (i.e., having more opportunities to commit false positive vs. false negative errors). As Table \[t\_density\] shows, our networks are indeed fairly sparse, with estimated densities less than 0.3 in all cases. Thus, even in the Pacific Distributor friendship network, where the average false negative rate was more than twice as large as the average false positive rate, we see that there were nearly nine times as many opportunities for these false positives to occur as there were for false negatives. One hence obtains superior performance from suppressing the error that informants *have more chances to make,* as seen in Figure \[LASPerformance\]. Although extremely high ratios of false negative to false positive rates would lead to the opposite result, Table \[t\_density\] shows that these rate differences would need to be very large to overcome the opportunity effect. Dataset Density FP/FN Opportunities --------- --------- --------------------- HM-AD 0.286 2.50 HM-FR 0.119 7.40 SS-AD 0.123 7.13 SS-FR 0.121 7.26 IU-AD 0.096 9.42 IU-FR 0.157 5.37 PD-AD 0.257 2.89 PD-FR 0.112 7.93 : Posterior mean density estimates of the criterion network via the BNAM, separated by environment and type of relation measured. The four environments were High-tech Managers (HM), Silicon Systems (SS), Pacific Distributors (PD), and Italian University (IU). The two relations measured were advice-seeking (AD) and friendship (FR). All networks are sparse, providing from 2.5 to almost 9.5 times as many opportunities for false positive errors than false negative errors (third column). \[t\_density\] Discussion and Conclusion ========================= With this study we sought to determine which LAS method would reproduce an unknown criterion graph more accurately. Across all eight of the networks examined here, we found that the LAS Intersection outperformed the LAS Union. A cursory examination of the density of the graphs indicate that they are all relatively sparse. As the sparsity of the analyzed graphs serves to create more opportunities for false positives, this creates conditions that are more favorable for the LAS Intersection. Although this would be outweighed by a sufficiently high false negative rate (relative to the false positive rate), this disparity must become quite large to overcome the levels of sparsity observed in typical settings. Importantly, we do not find evidence of this disparity in self-report errors, implying that the LAS Union is unlikely to be superior except in networks for which the density approaches or exceeds 50%. Our findings are in line with Equation \[e\_ilascond\], which gives an approximate criterion for the conditions under which one or the other rule is expected to dominate. It should be noted that the results given here focus exclusively on Hamming error—i.e., the number of edge variables incorrectly inferred—and may not apply to all analyses. In particular, network measures that are differentially sensitive to false positive vs. false negative errors may have different optimal procedures. When in doubt, a simulation study should be performed prior to analysis. Although one LAS may be better than another in a given setting, this does not imply that the LAS is per se the best tool to use: indeed, as shown in Figure \[LASPerformance\], even the Intersection LAS shows high levels of Hamming error in all cases examined here. Where possible, we recommend obtaining as many measures of each edge variable as possible (either via complete CSS [@krackhardt:sn:1987] or arc sampling designs [@butts:sn:2003]) and subjecting them to model-based inference. In many cases, however, analysts must make due with only self-report data. In these settings, the Intersection LAS will often outperform the Union LAS, and we recommend the former over the latter so long as the networks in question are suitably sparse. [^1]: Department of Sociology, University of California, Irvine [^2]: Departments of Sociology, Statistics, and EECS, and Institute for Mathematical Behavioral Sciences, University of California, Irvine [^3]: This research was supported in part by ARO award W911NF-14-1-055. [^4]: @krackhardt:sn:1987 does not explicitly treat the LAS as a family of estimators per se, but employs them in a manner consistent with this interpretation. [^5]: The three non-responding informants and others’ perceptions of their ties were removed upon the release of this dataset to the public, and thus inference cannot be done to approximate the data we would have seen. We treat them as absent by design.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Granularity, anisotropy, local lattice distortions and their dependence on dopant concentration appear to be present in all cuprate superconductors, interwoven with the microscopic mechanisms responsible for superconductivity. Here we review anisotropy and penetration depth measurements to reassess the evidence for granularity, as revealed by the notorious rounded phase transition, the evidence for the three dimensional nature of superconductivity, uncovered by the doping dependence of transition temperature and anisotropy, and to reassess the relevance of the electron-lattice coupling, emerging from the oxygen isotope effects.' author: - 'H. Keller' - 'T. Schneider' title: 'Evidence for granularity, anisotropy and lattice distortions in cuprate superconductors and their implications' --- To appear in the proceedings of  Symmetry and Heterogeneity in High Temperature Superconductors, Erice-Sicily: 4-10 October 2003 Establishing and understanding the phase diagram of cuprate superconductors in the temperature - dopant concentration plane is one of the major challenges in condensed matter physics. Superconductivity is derived from the insulating and antiferromagnetic parent compounds by partial substitution of ions or by adding or removing oxygen. For instance La$_{2}$CuO$_{4}$ can be doped either by alkaline earth ions or oxygen to exhibit superconductivity. The empirical phase diagram of La$_{2-x}$Sr$_{x}$CuO$_{4}$ [@suzuki; @nakamura; @fukuzumi; @willemin; @kimura; @sasagawa; @hoferdis; @shibauchi; @panagopoulos] depicted in Fig.\[fig1\]a shows that after passing the so called underdoped limit $\left( x_{u}\approx 0.05\right) $, $T_{c}$ reaches its maximum value $T_{c}\left( x_{m}\right) $ at $x_{m}\approx 0.16$. With further increase of $x $, $T_{c}$ decreases and finally vanishes in the overdoped limit $x_{o}\approx 0.27$. This phase transition line is thought to be a generic property of cuprate superconductors [@tallon] and is well described by the empirical relation $$T_{c}\left( x\right) =T_{c}\left( x_{m}\right) \left( 1-2\left( \frac{x}{x_{m}}-1\right) ^{2}\right) =\frac{2T_{c}\left( x_{m}\right) }{x_{m}^{2}}\left( x-x_{u}\right) \left( x_{o}-x\right) , \label{eq1}$$ proposed by Presland *et al*.[@presland]. Approaching the endpoints along the $x$-axis, La$_{2-x}$Sr$_{x}$CuO$_{4}$ undergoes at zero temperature doping tuned quantum phase transitions. As their nature is concerned, resistivity measurements reveal a quantum superconductor to insulator (QSI) transition in the underdoped limit[@momono; @polen; @book; @klosters; @tshk; @tsphysB; @parks] and in the overdoped limit a quantum superconductor to normal state (QSN) transition[@momono]. ![(a) Variation of $T_{c}$ for La$_{2-x}$Sr$_{x}$CuO$_{4}$. Experimental data taken from [@suzuki; @nakamura; @fukuzumi; @willemin; @kimura; @sasagawa; @hoferdis; @shibauchi; @panagopoulos]. The solid line is Eq.(\[eq1\]) with $T_{c}\left( x_{m}\right) =39$K. (b) $\gamma _{T}$ versus $x$ for La$_{2-x}$Sr$_{x}$CuO$_{4}$. The squares are the experimental data for $\gamma _{T_{c}}$ [@suzuki; @nakamura; @willemin; @sasagawa; @hoferdis] and the triangles for $\gamma _{T=0}$ [@shibauchi; @panagopoulos]. The solid curve and dashed lines are Eq.(\[eq2\]) with $\gamma _{T_{c},0}=2$ and $\gamma _{T=0,0}=1.63$.[]{data-label="fig1"}](fig1a.EPS "fig:") ![(a) Variation of $T_{c}$ for La$_{2-x}$Sr$_{x}$CuO$_{4}$. Experimental data taken from [@suzuki; @nakamura; @fukuzumi; @willemin; @kimura; @sasagawa; @hoferdis; @shibauchi; @panagopoulos]. The solid line is Eq.(\[eq1\]) with $T_{c}\left( x_{m}\right) =39$K. (b) $\gamma _{T}$ versus $x$ for La$_{2-x}$Sr$_{x}$CuO$_{4}$. The squares are the experimental data for $\gamma _{T_{c}}$ [@suzuki; @nakamura; @willemin; @sasagawa; @hoferdis] and the triangles for $\gamma _{T=0}$ [@shibauchi; @panagopoulos]. The solid curve and dashed lines are Eq.(\[eq2\]) with $\gamma _{T_{c},0}=2$ and $\gamma _{T=0,0}=1.63$.[]{data-label="fig1"}](fig1b.EPS "fig:") Another essential experimental fact is the doping dependence of the anisotropy. In tetragonal cuprates it is defined as the ratio $\gamma =\xi _{ab}/\xi _{c}$ of the correlation lengths parallel $\left( \xi _{ab}\right) $ and perpendicular $\left( \xi _{c}\right) $ to CuO$_{2}$ layers ($ab$-planes). In the superconducting state it can also be expressed as the ratio $\gamma =\lambda _{c}/\lambda _{ab}$ of the London penetration depths due to supercurrents flowing perpendicular ($\lambda _{c}$ ) and parallel ($\lambda _{ab}$ ) to the $ab$-planes. Approaching a nonsuperconductor to superconductor transition $\xi $ diverges, while in a superconductor to nonsuperconductor transition $\lambda $ tends to infinity. In both cases, however, $\gamma $ remains finite as long as the system exhibits anisotropic but genuine 3D behavior. There are two limiting cases: $\gamma =1$ characterizes isotropic 3D- and $\gamma =\infty $ 2D-critical behavior. An instructive model where $\gamma $ can be varied continuously is the anisotropic 2D Ising model[@onsager]. When the coupling in the $y$ direction goes to zero, $\gamma =\xi _{x}/\xi _{y}$ becomes infinite, the model reduces to the 1D case, and $T_{c}$ vanishes. In the Ginzburg-Landau description of layered superconductors the anisotropy is related to the interlayer coupling. The weaker this coupling is, the larger $\gamma $ is. The limit $\gamma =\infty $ is attained when the bulk superconductor corresponds to a stack of independent slabs of thickness $d_{s}$. With respect to experimental work, a considerable amount of data is available on the chemical composition dependence of $\gamma $. At $T_{c}$ it can be inferred from resistivity ($\gamma =\xi _{ab}/\xi _{c}=\sqrt{\rho _{ab}/\rho _{c}}$) and magnetic torque measurements, while in the superconducting state it follows from magnetic torque and penetration depth ($\gamma =\lambda _{c}/\lambda _{ab}$) data. In Fig. \[fig1\]b we displayed the doping dependence of $1/\gamma _{T}$ evaluated at $T_{c}$ ($\gamma _{T_{c}}$) and $T=0$ ($\gamma _{T=0}$). As the dopant concentration is reduced, $\gamma _{T_{c}}$ and $\gamma _{T=0}$ increase systematically, and tend to diverge in the underdoped limit. Thus the temperature range where superconductivity occurs shrinks in the underdoped regime with increasing anisotropy. This competition between anisotropy and superconductivity raises serious doubts whether 2D mechanisms and models, corresponding to the limit $\gamma _{T}=\infty $, can explain the essential observations of superconductivity in the cuprates. From Fig. \[fig1\]b it is also seen that $\gamma _{T}\left( x\right) $ is well described by $$\gamma _{T}\left( x\right) =\frac{\gamma _{T,0}}{x-x_{u}}, \label{eq2}$$ where $\gamma _{T,0}$ is the quantum critical amplitude. Having also other cuprate families in mind, it is convenient to express the dopant concentration in terms of $T_{c}$. From Eqs.(\[eq1\]) and(\[eq2\]) we obtain the correlation between $T_{c}$ and $\gamma _{T}$: $$\frac{T_{c}\left( x\right) }{T_{c}\left( x_{m}\right) }=1-\left( \frac{\gamma _{T}\left( x_{m}\right) }{\gamma _{T}\left( x\right) }-1\right) ^{2},\ \ \gamma _{T}\left( x_{m}\right) =\frac{\gamma _{T,0}}{x_{m}-x_{u}} \label{eq3}$$ ![$T_{c}\left( x\right) /T_{c}\left( x_{m}\right) $ versus $\gamma _{T}\left( x_{m}\right) /$ $\gamma _{T}\left( x\right) $ for various cuprate families: La$_{2-x}$Sr$_{x}$CuO$_{4}$ ($\bullet $, $T_{c}\left( x_{m}\right) =37$K, $\gamma _{T_{c}}\left( x_{m}\right) =20$) [@suzuki; @nakamura; @willemin; @sasagawa; @hoferdis] , ($\bigcirc $, $T_{c}\left( x_{m}\right) =37$K, $\gamma _{T=0}\left( x_{m}\right) =14.9$) [@shibauchi; @panagopoulos], HgBa$_{2}$CuO$_{4+\delta }$ ($\blacktriangle $, $T_{c}\left( x_{m}\right) =95.6$K, $\gamma _{T_{c}}\left( x_{m}\right) =27$) [@hoferhg], Bi$_{2}$Sr$_{2}$CaCu$_{2}$O$_{8+\delta }$ ($\bigstar $, $T_{c}\left( x_{m}\right) =84.2$K, $\gamma _{T_{c}}\left( x_{m}\right) =133$) [@watauchi], YBa$_{2}$Cu$_{3}$O$_{7-\delta }$ ($\blacklozenge $, $T_{c}\left( x_{m}\right) =92.9$K, $\gamma _{T_{c}}\left( x_{m}\right) =8$) [@chien123], YBa$_{2}$(Cu$_{1-y}$Fe$_{y}$)$_{3}$O$_{7-\delta }$ ($\blacksquare $, $T_{c}\left( x_{m}\right) =92.5$K, $\gamma _{T_{c}}\left( x_{m}\right) =9$)[@chienfe], Y$_{1-y}$Pr$_{y}$Ba$_{2}$Cu$_{3}$O$_{7-\delta }$ ($\blacktriangledown $, $T_{c}\left( x_{m}\right) =91$K, $\gamma _{T_{c}}\left( x_{m}\right) =9.3$)[@chienpr], BiSr$_{2}$Ca$_{1-y}$Pr$_{y}$Cu$_{2}$O$_{8}$ ($\blacktriangleleft $, $T_{c}\left( x_{m}\right) =85.4$K, $\gamma _{T=0}\left( x_{m}\right) =94.3$)[@sun] and YBa$_{2}$(Cu$_{1-y}$ Zn$_{y}$)$_{3}$O$_{7-\delta }$ ($\blacktriangleright $, $T_{c}\left( x_{m}\right) =92.5$K, $\gamma _{T=0}\left( x_{m}\right) =9$)[@panagopzn]. The solid and dashed curves are Eq.(\[eq3\]), marking the flow from the maximum $T_{c}$ to QSI and QSN criticality, respectively.[]{data-label="fig2"}](fig2.EPS) Provided that this empirical correlation is not merely an artefact of La$_{2-x}$Sr$_{x}$CuO$_{4}$, it gives a universal perspective on the interplay of anisotropy and superconductivity, among the families of cuprates, characterized by $T_{c}\left( x_{m}\right) $ and $\gamma _{T}\left( x_{m}\right) $. For this reason it is essential to explore its generic validity. In practice, however, there are only a few additional compounds, including HgBa$_{2}$CuO$_{4+\delta }$[@hoferhg], for which the dopant concentration can be varied continuously throughout the entire doping range. It is well established, however, that the substitution of magnetic and nonmagnetic impurities depresses $T_{c}$ of cuprate superconductors very effectively[@xiao; @tarascon]. To compare the doping and substitution driven variations of the anisotropy, we depicted in Fig. \[fig2\] the plot $T_{c}\left( x\right) /T_{c}\left( x_{m}\right) $ versus $\gamma _{T}\left( x_{m}\right) /$ $\gamma _{T}\left( x\right) $ for a variety of cuprate families. The collapse of the data on the parabola, which is the empirical relation (\[eq3\]), reveals that this scaling form appears to be universal. Thus, given a family of cuprate superconductors, characterized by $T_{c}\left( x_{m}\right) $ and $\gamma _{T}\left( x_{m}\right) $, it gives a universal perspective on the interplay between anisotropy and superconductivity. Close to 2D-QSI criticality various properties are not independent but related by[@polen; @book; @klosters; @tshk; @tsphysB; @parks] $$T_{c}=\frac{\Phi _{0}^{2}R_{2}}{16\pi ^{3}k_{B}}\frac{d_{s}}{\lambda _{ab}^{2}\left( 0\right) }\propto \gamma _{T}^{-z}\propto \delta ^{z\overline{\nu }}, \label{eq4}$$ where $k_{B}$ is the Boltzmann constant, and $\Phi _{0}$ the elementary flux quantum. $\lambda _{ab}\left( 0\right) $ is the zero temperature in-plane penetration depth, $z$ is the dynamic critical exponent, $d_{s}$ the thickness of the sheets, and $\overline{\nu }$ the correlation length critical exponent of the 2D-QSI transitions. $\delta $ measures the distance from the critical point along the $x$ axis (see Fig.\[fig1\]a, and $R_{2}$ is a universal number. Since $T_{c}\propto d_{s}/\lambda _{ab}^{2}\left( 0\right) \propto n_{s}^{\Box }$, where $n_{s}^{\Box }$ is the aerial superfluid density, is a characteristic 2D property, it also applies to the onset of superfluidity in $^{4}$He films adsorbed on disordered substrates, where it is well confirmed[@crowell]. A great deal of experimental work has also been done in cuprates on the so called Uemura plot, revealing an empirical correlation between $T_{c}$ and $d_{s}/\lambda _{ab}^{2}\left( 0\right) $[@uemura]. Approaching 2D-QSI criticality, the data of a given family tends to fall on a straight line, consistent with Eq.(\[eq4\]). Differences in the slope reflect the family dependent value of $d_{s}$, the thickness of the sheets, becoming independent in the 2D limit[@book; @klosters; @tshk; @tsphysB; @parks]. The relevance of $d_{s}$ was also confirmed in terms of the relationship between the isotope effect on $T_{c}$ and $1/\lambda _{ab}^{2}$[@tshk; @tsiso]. Moreover, together with the scaling form (\[eq4\]) the empirical relation (\[eq1\]) implies 2D-QSI and 3D-QSN transitions with $z=1$, while the empirical relation for the anisotropy (Eqs.(\[eq2\]) and (\[eq3\])), require $\overline{\nu }=1$ at the 2D-QSI critical point. Thus, the empirical correlations point to a 2D-QSI transition with $z=1$ and $\overline{\nu }=1$.These estimates coincide with the theoretical prediction for a 2D disordered bosonic system with long-range Coulomb interactions, where $z=1$ and $\overline{\nu }\simeq 1$[@mpafisher; @ca; @herbutz1]. Here the loss of superfluidity is due to the localization of the pairs, which ultimately drives the transition. From the scaling relation (\[eq4\]) it is seen that measurements of the out of plane penetration depth of sufficiently underdoped systems allow to estimate the dynamic critical exponent $z$ directly, in terms of $T_{c}\propto \left( 1/\lambda _{c}^{2}\left( 0\right) \right) ^{z/\left( z+2\right) }$, which follows from Eq.(\[eq4\]) with $\gamma _{T}=\lambda _{c}\left( 0\right) /\lambda _{ab}\left( 0\right) $. In Fig.\[fig3\] we displayed the data of Hosseini[@hosseini] for heavily underdoped YBa$_{2}$Cu$_{3}$O$_{7-\delta }$ single crystals. The solid line is $T_{c}$ $=170\left( 1/\lambda _{c}^{2}\left( T=0\right) \right) ^{1/3}$ and uncovers the consistency with the 2D-QSI scaling relation $T_{c}\propto \left( 1/\lambda _{c}^{2}\left( 0\right) \right) ^{z/\left( z+2\right) }$ with $z=1$. ![$T_{c}$ versus $1/\lambda _{c}^{2}\left( T=0\right) $ for heavily underdoped YBa$_{2}$Cu$_{3}$O$_{7-\delta }$ single crystals, taken from Hosseini[@hosseini]. The solid line is $T_{c}$ $=170\left( 1/\lambda _{c}^{2}\left( T=0\right) \right) ^{1/3}$ and indicates the consistency with the 2D-QSI scaling relation $T_{c}\propto \left( 1/\lambda _{c}^{2}\left( 0\right) \right) ^{z/(z+2)}$ and $z=1$.[]{data-label="fig3"}](fig3.EPS) We have seen that the doping tuned flow to the 2D-QSI critical point is associated with a depression of $T_{c}$ and an enhancement of $\gamma _{T}$. It implies that whenever a QSI transition is approached, a no vanishing $T_{c}$ is inevitably associated with an anisotropic but 3D condensation mechanism, because $\gamma _{T}$ is finite for $T_{c}>0$ (see Figs.\[fig1\]b and \[fig2\]). This represents a serious problem for 2D models[@anderson] as candidates to explain superconductivity in the cuprates, and serves as a constraint on future work toward a complete understanding. Note that the vast majority of theoretical models focus on a single Cu-O plane, i.e., on the limit of zero intracell and intercell $c$-axis coupling. Since Eq.(\[eq4\]) is universal, it also implies that the changes $\Delta T_{c}$, $\Delta d_{s}$ and $\Delta \left( 1/\lambda _{ab}^{2}\left( T=0\right) \right) $, induced by pressure or isotope exchange are not independent, but related by $$\frac{\Delta T_{c}}{T_{c}}=\frac{\Delta d_{s}}{d_{s}}+\frac{\Delta \left( 1/\lambda _{ab}^{2}\left( 0\right) \right) }{\left( 1/\lambda _{ab}^{2}\left( 0\right) \right) }=\frac{\Delta d_{s}}{d_{s}}-2\frac{\Delta \left( \lambda _{ab}\left( 0\right) \right) }{\lambda _{ab}\left( 0\right) }. \label{eq5}$$ In particular, for the oxygen isotope effect ($^{16}$O vs. $^{18}$O) of a physical quantity $X$  the relative isotope shift is defined as $\Delta X/X=(^{18}X-^{16}X)/^{18}X$. In Fig.\[fig4\] we show the data for the oxygen isotope effect in La$_{2-x}$Sr$_{x}$CuO$_{4}$[@hofer214; @zhao1], Y$_{1-x}$Pr$_{x}$Ba$_{2}$Cu$_{3}$O$_{7-\delta }$[@zhao1; @khasanov123pr; @rksite] and YBa$_{2}$Cu$_{3}$O$_{7-\delta }$[@zhao1; @khasanov123f], extending from the underdoped to the optimally doped regime, in terms of $\Delta \left( \lambda _{ab}\left( 0\right) \right) /$ $\lambda _{ab}\left( 0\right) $ versus $\Delta T_{c}/T_{c}$. It is evident that there is a correlation between the isotope effect on $T_{c}$ and $\lambda _{ab}\left( 0\right) $ which appears to be universal for all cuprate families. Indeed, the solid line indicates the flow to the 2D-QSI transition and provides with Eq.(\[eq5\]) an estimate for the oxygen isotope effect on $d_{s}$, namely $\Delta d_{s}/d_{s}=3.3(4)\%$. Approaching optimum doping, this contribution renders the isotope effect on $T_{c}$ considerably smaller than that on $\lambda _{ab}\left( 0\right) $. As shown in Fig.\[fig5\], even in nearly optimally doped YBa$_{2}$Cu$_{3}$O$_{7-\delta }$, where $\Delta T_{c}/T_{c}=-0.26(5)\%$, a substantial isotope effect on the in-plane penetration depth, $\Delta \lambda _{ab}\left( 0\right) /\lambda _{ab}\left( 0\right) =-2.8(1.0)\%$, has been established by direct observation, using the novel low-energy muon-spin rotation technique[@khasanov123f]. Note that these findings have been obtained using various experimental techniques on powders, thin films and single crystals. ![Data for the oxygen isotope effect in underdoped La$_{2-x}$Sr$_{x}$CuO$_{4}$($\bigcirc $: x=0.15[@zhao1], $\blacktriangle $:x=0.08, 0.086 [@hofer214], Y$_{1-x}$Pr$_{x}$Ba$_{2}$Cu$_{3}$O$_{7-\delta }$ ($\bullet $: x=0, 0.2, 0.3, 0.4)[@zhao1; @khasanov123pr; @rksite] and YBa$_{2}$Cu$_{3}$O$_{7-\delta }$ ($\blacktriangledown $ [@zhao1], $\blacksquare $[@khasanov123f])  in terms of $\Delta \left( \lambda _{ab}\left( 0\right) \right) /$ $\lambda _{ab}\left( 0\right) $ versus -$\Delta T_{c}/T_{c}$. The solid line indicates the flow to 2D-QSI criticality and provides with Eq.(\[eq5\]) an estimate for the oxygen isotope effect on $d_{s}$, namely $\Delta d_{s}/d_{s}=3.3(4)\%$.[]{data-label="fig4"}](fig4.EPS) ![Magnetic field penetration profiles $B(z)$ for a $^{16}$O substituted (closed symbols) and a $^{18}$O substituted (open symbols) YBa$_{2}$Cu$_{3}$O$_{7-\delta }$-film measured in the Meissner state at 4 K and an external field of 9.2 mT, applied parallel to the surface of the film. The data are shown for implantation energies 3, 6, 10, 16, 22, and 29 keV starting from the surface of the sample. Solid curves are best fits to $B(z)=B_{0}\cosh[(t-z)/\lambda _{ab}]\cosh(t/\lambda _{ab})$. This is the form of the classical exponential field decay in the Meissner state $B(z)=B_{0}\exp(-z/\lambda _{ab})$, modified for a film with thickness $2t$ with flux penetrating from both sides. Taken from Khasanov *et al.*[@khasanov123f] []{data-label="fig5"}](fig5.EPS) Since upon oxygen isotope exchange the lattice parameters remain essentially unaffected[@raffa; @conder], the substantial isotope effect on the in-plane penetration depth uncovers the coupling between local lattice distortions and superfluidity and the failure of the Migdal-Eliashberg (ME) treatment of the electron-phonon interaction, predicting, $1/\lambda ^{2}\left( 0\right) $, to be independent of the ionic masses[@migdal]. Evidence for this coupling emerges from the oxygen isotope effect on $d_{s}$, the thickness of the superconducting sheets, upon isotope exchange, while the lattice parameters remain unaffected. Indeed, the relative shift, $\Delta d_{s}/d_{s}\approx 3.3(4)\%$, apparent in Fig.\[fig3\], implies local distortions of oxygen degrees of freedom, which do not modify the lattice parameters, and are coupled to the superfluid. Further evidence for this coupling emerges from the isotope effect on the granularity of the cuprates[@varenna; @tsrkhk]. Recently, it has been shown that the notorious rounding of the superconductor to normal state transition is fully consistent with a finite size effect, revealing that bulk cuprate superconductors break into nearly homogeneous superconducting grains of rather unique extent[@book; @varenna; @tsrkhk; @bled; @tsdc]. Even evidence for their surface and edge contributions to specific heat and penetration depth has been established[@tserice]. A characteristic feature of a finite size effect in the temperature dependence of the in-plane penetration depth $\lambda _{ab}$ is the occurrence of an inflection point giving rise to an extremum in $d\left( \lambda _{ab}^{2}\left( T=0\right) /\lambda _{ab}^{2}\left( T\right) \right) /dT$ at $T_{p}$. Here $\lambda _{ab}^{2}\left( T_{p}\right) $, $T_{p}$ and the length $L_{c}$ of the grains along the c-axis are related by[@varenna; @tsrkhk; @bled; @tsdc; @tserice] $$\frac{1}{\lambda _{ab}^{2}\left( T_{p}\right) }=\frac{16\pi ^{3}k_{B}T_{p}}{\Phi _{0}^{2}L_{c}}. \label{eq7}$$ Recently we explored the effect of oxygen isotope exchange in Y$_{1-x}$Pr$_{x}$Ba$_{2}$Cu$_{3}$O$_{7-\delta }$ on $L_{c}$ by means of in-plane penetration depth measurements[@tsrkhk]. Note that the shifts are not independent but according to Eq.(\[eq7\]) related by $$\frac{\Delta L_{c}}{L_{c}}=\frac{\Delta T_{p_{c}}}{T_{p_{c}}}+\frac{\Delta \lambda _{ab}^{2}\left( T_{p_{c}}\right) }{\lambda _{ab}^{2}\left( T_{p_{c}}\right) }. \label{eq8}$$ From the resulting estimates, listed in Table I, several observations emerge. First, $L_{c}$ increases systematically with reduced $T_{p_{c}}$. Second, $L_{c}$ grows with increasing $x$ and upon isotope exchange ($^{16}$O, $^{18}$O). Third, the relative shift of $T_{p_{c}}$ is very small. This reflects the fact that the change of $L_{c}$ is essentially due to the superfluid, probed in terms of $\lambda _{ab}^{2}$. Accordingly, $\Delta L_{c}/L_{c}\approx \Delta \lambda _{ab}^{2}/\lambda _{ab}^{2}$ for $x=0,\ 0.2 $ and $0.3$. ----------------------------------------------------------------------------------------------- x 0 0.2 0.3 ----------------------------------------------------------- ----------- ----------- ----------- $\Delta L_{c}/L_{c}$ 0.12(5) 0.13(6) 0.16(5) $\Delta T_{p_{c}}/T_{p_{c}}$ -0.000(2) -0.015(3) -0.021(5) $\Delta \lambda 0.11(5) 0.15(6) 0.15(5) _{ab}^{2}\left( T_{p_{c}}\right) /\lambda _{ab}^{2}\left( T_{p_{c}}\right) $ $^{16}T_{p_{c}}$(K) 89.0(1) 67.0(1) 52.1(1) $^{16}L_{c}$(A) 9.7(4) 14.2(7) 19.5(8) $^{18}T_{p_{c}}$(K) 89.0(1) 66.0(2) 51.0(2) $^{18}L_{c}$(A) 10.9(4) 16.0(7) 22.6(9) ----------------------------------------------------------------------------------------------- Table I: Finite size estimates for the relative changes of $L_{c}$, $T_{p_{c}}$ and $\lambda _{ab}^{2}\left( T_{p_{c}}\right) $ upon oxygen isotope exchange for Y$_{1-x}$Pr$_{x}$Ba$_{2}$Cu$_{3}$O$_{7-\delta }$[@tsrkhk]. To appreciate the implications of these estimates, we note again that for fixed Pr concentration the lattice parameters remain essentially unaffected [@conder; @raffa]. Accordingly, an electronic mechanism, without coupling to local lattice distortions implies $\Delta L_{c}=0$. On the contrary, a significant change of $L_{c}$ upon oxygen exchange requires local lattice distortions involving the oxygen lattice degrees of freedom and implies with Eq.(\[eq8\]) a coupling between these distortions and the superfluid. A glance to Table I shows that the relative change of the grains along the $c$-axis upon oxygen isotope exchange is significant, ranging from $12$ to $16\% $, while the relative change of the inflection point at $T_{p_{c}}$, or the transition temperature, is an order of magnitude smaller. For this reason the significant relative change of $L_{c}$ at fixed Pr concentration is accompanied by essentially the same relative change of $\lambda _{ab}^{2}$, which probes the superfluid. This uncovers unambiguously the existence and relevance of the coupling between the superfluid and lattice distortions, involving the oxygen lattice degrees of freedom. Furthermore the substantial isotope effect on the in-plane penetration depth at $T=T_{p_{c}}$ extends the evidence for the failure of the Migdal-Eliashberg (ME) theory of the electron-phonon interaction, predicting $1/\lambda ^{2}$ to be independent of the ionic masses[@migdal], to finite temperature. Although the majority opinion on the mechanism of superconductivity in the cuprates is that it occurs via a purely electronic mechanism involving spin excitations, and lattice degrees of freedom are supposed to be irrelevant, the relative isotope shifts $\Delta L_{c}/L_{c}\approx \Delta \lambda _{ab}^{2}/^{16}\lambda _{ab}^{2}\approx 14\%$ and $\Delta d_{s}/d_{s}\approx 3\%$ uncover clearly the existence and relevance of the coupling between the superfluid and local lattice distortions. Recent site-selective oxygen isotope ($^{16}$O/$^{18}$O) effect measurements of the in-plane penetration depth in Y$_{0.6}$Pr$_{0.4}$Ba$_{2}$Cu$_{3}$O$_{7-\delta }$, using the muon-spin rotation ($\mu $SR) technique show that the distortions arise from the oxygen sites within the CuO$_{2}$ planes (100 % within error bar)[@rksite]. Potential candidates are then the Cu-O bond-stretching-type modes showing a temperature dependence, which parallels that of the superconductive order parameter[@chung]. To summarize we observed remarkable consistency between the scaling properties of the experimental data for a variety of cuprates and those characterizing 2D-QSI transitions. The important implication there is that in cuprates a no vanishing transition temperature and superfluid density in the ground state are unalterably linked to a finite anisotropy. Furthermore, the oxygen isotope effect on the in-plane penetration depth and the spatial extent of the superconducting grains revealed the coupling between local lattice distortions and superfluidity, while the lattice parameters remain essentially unaffected. These findings raise serious doubts that 2D models [@anderson], neglecting granularity and local lattice distortions are potential candidates to explain superconductivity in cuprates. The author is grateful to D. Di Castro, R. Khasanov, K.A. Müller, and J. Roos for very useful comments and suggestions on the subject matter. This work was partially supported by the Swiss National Science Foundation and the NCCR program *Materials with Novel Electronic Properties* (MaNEP) sponsered by the Swiss National Science Foundation. [99]{} M. Suzuki and M.Hikita, Phys. Rev. B **44**, 249 (1991). Y. Nakamura and S. Uchida, Phys. Rev. Phys. Rev. B**47**, 8369 (1993). Y. Fukuzumi, K. Mizuhashi, K. Takenaka, and S. Uchida, Phys. Rev. Lett. **76**, 684 (1996). M. Willemin, C. Rossel, J. Hofer, H. Keller, and A. Revcolevschi, Phys. Rev. B **59**, 717 (1999). T. Kimura, K. Kishio, T. Kobayashi, Y. Nakayama, N. Motohira, K. Kitazawa, and K. Yamafuji, Physica C **192,** 247 (1992). T. Sasagawa, Y. Togawa, J. Shimoyama, A. Kapitulnik, K. Kitazawa, and K. Kishio, Phys. Rev. B **61**, 1610 (2000). J. Hofer, T. Schneider, J. M. Singer, M. Willemin, H. Keller, T. Sasagawa, K. Kishio, K. Conder, and J. Karpinski, Phys. Rev. B **62,** 631 (2000). T. Shibauchi, H. Kitano, K. Uchinokura, A. Maeda, T. Kimura, and K. Kishio, Phys. Rev. Lett. **72**, 2263 (1994). C. Panagopoulos, J. R. Cooper, T. Xiang, Y. S. Wang and C. W. Chu, Phys. Rev. B **61**, 3808 (2000). J. L. Tallon, C. Bernhard, H. Shaked, R. L. Hitterman, and J. D. Jorgensen, Phys. Rev. B **51**, 12911 (1995). M. R. Presland, J. L. Tallon, R. G. Buckley, R. S. Liu, and N. E. Flower, Physica C **176**, 95 (1991). N. Momono, M. Ido, T. Nakano, M. Oda, Y. Okajima, and K. Yamaya, Physica C **233**, 395 (1994). T. Schneider, Acta Physica Polonica A **91,** 203 (1997). T. Schneider and J. M. Singer, *Phase Transition Approach To High Temperature Superconductivity*, Imperial College Press, London, 2000. T. Schneider and J. M. Singer, J. of Superconductivity **13**, 789 (2000). T. Schneider and H. Keller, Phys. Rev. Lett. **86**, 4899 (2001). T. Schneider, Physica B **326**, 289 (2003). T. Schneider, cond-mat/0204236. L. Onsager, Phys. Rev. **65**, 117 (1944). J. Hofer, J. Karpinski, M. Willemin, G.I. Meijer, E.M. Kopnin, R. Molinski, H. Schwer, C. Rossel, and H. Keller, Physica C **297**, 103 (1998). G. Xiao, M. Z. Cieplak, J. Q. Xiao, and C. L. Chien, Phys. Rev. B **42**, 8752 (1990). J. M. Tarascon *et al.*, Phys. Rev. B **42**, 218 (1990). S. Watauchi, H. Ikuta, H. Kobayashi, J. Shimoyama, and K. Kishio, Phys. Rev. B **64**, 64520 (2001). T. R. Chien, W. R. Datars, B. W. Veal, A. P. Paulikas, P. Kostic, Chun Gu, and Y. Jiang, Physica C **229**, 273 (1994). T. R. Chien, W. R. Datars, M. D. Lan, J. Z. Liu, and R. N. Shelton, Phys. Rev. B **49**, 1342 (1994). T. R. Chien and W. R. DatarsJ. Z. Liu, M. D. Lan, and R. N. Shelton, Physica C **221**, 428 (1994). X. F. Sun, X. Zhao, X.-G. Li, and H. C. Ku, Phys. Rev. B **59**, 8978 (1999). C. Panagopoulos, J. R. Cooper, N. Athanassopoulou, and J. Chrosch, Phys. Rev. B **54**, 12721 (1996). P. A. Crowell, F. W. van Keuls, and J. R. Reppy, Phys. Rev. B **55**, 12620 (1997). Y. Uemura *et al.*, Phys. Rev. Lett. **62**, 2317 (1989). T. Schneider, Phys. Rev. B **67**, 134514 (2003). M. P. A. Fisher, G. Grinstein, and S. M. Girvin, Phys. Rev. Lett. **64**, 587 (1990. Min-Chul Cha, M. P. A. Fisher, M. Wallin, and A. P. Young, Phys. Rev. B **44**, 6883 (1991). I. F. Herbut, Phys. Rev. B **6**, 14723 (2000). A. Hosseini, The anisotropic microwave electrodynamics of YBCO, Ph.D. Thesis, University of British Columbia (2002). P. W. Anderson, P. A. Lee, M. Randeria, T. M. Rice, N. Trivedi, and F. C. Zhang, cond-mat/0311467. J. Hofer, K. Conder, T. Sasagawa, Guo-meng Zhao, M. Willemin, H. Keller, and K. Kishio, Phys. Rev. Lett. **84**, 4192 (2000). R. Khasanov, Studies of the oxygen-isotope effect on the magnetic field penetration depth in cuprate superconductors, Ph.D. Thesis, University of Zürich (2003). R. Khasanov, A. Shengelaya, K. Conder, E. Morenzoni, I. M. Savic, and H Keller, J. Phys.: Condens. Matter **15**, L17 (2003). R. Khasanov, A. Shengelaya, E. Morenzoni, M. Angst, K. Conder, I. M. Savic, D. Lampakis, E. Liarokapis, A. Tatsi, and H. Keller, Phys. Rev. B **68**, 220506(R) (2003). R. Khasanov, D. G. Eshchenko, H. Luetkens, E. Morenzoni, T. Prokscha, A. Suter, N. Garifianov, M. Mali, J. Roos, K. Conder, and H. Keller, Phys. Rev. Lett. (2004), in press. F. Raffa, T. Ohno, M. Mali, J. Roos, D. Brinkmann, K. Conder, and M. Eremin, Phys. Rev. Lett. **81**, 5912 (1998). K. Conder, Mater. Sci. Eng., R **32**, 41 (2001). A.B. Migdal, Zh. Eksp. Teor. Fiz. **34**, 1438 (1958) \[Sov.Phys. JETP **7**, 996 (1958)\]; G. M. Eliashberg, Zh. Eksp. Teor. Fiz. **38**, 966 (1960) \[Sov.Phys. JETP **11**, 696 (1960)\]. T. Schneider, cond-mat/0308595. T. Schneider, R. Khasanov, K. Conder, and H Keller, J. Phys. Condens: Matter **15**, L763 (2003). T. Schneider, Journal of Superconductivity, **17**, 41 (2004). T. Schneider and D. Di Castro, Phys. Rev. B **69**, 024502 (2004). T. Schneider, this volume. J.-H. Chung, T. Egami, R. J. Mc Queeney, M. Yethiraj, M. Arai, T. Yokoo, Y. Petrov, H. A. Mook, Y. Endoh, S. Tajima, C. Frost, and F. Dogan, Phys. Rev. B **67**, 014517 (2003).
{ "pile_set_name": "ArXiv" }
--- abstract: 'Solar energetic transients occurring in solar atmosphere are associated with catastrophic release of energy in the solar corona. These transients inject a part of their energy by various physical processes to the deeper, denser photospheric layer at which velocity and magnetic fields are measured using suitable spectral lines. Serious questions have been raised about the nature of the observed magnetic (and velocity) field changes associated with energetic transients as their measurements are expected to be affected by flare-induced line profile changes. In this paper, we shall discuss some recent progress on our understanding of the physical processes associated with such events.' author: - | Ashok Ambastha[^1] and Ram A. Maurya[^2]\ Udaipur Solar Observatory, Bari Road, P.O. Box 198, Udaipur 313 001\ date: 'Received — ; accepted —' title: Spectral Line Profile Changes Associated with Energetic Solar Transients --- \[firstpage\] Sun: activity – Sun: magnetic field – Sun: spectral profiles Introduction {#sec:intro} ============ Solar flares are sudden transient, localized eruptions of energy in the solar atmosphere. These events occur over a large range in space and time; lasting from a few seconds to several hours. Intensification of radiation over the entire electromagnetic spectrum, release of energetic particles and ejection of plasma clouds are some of the observed phenomena associated with flares. Flares produce short-lived waves traveling away from their site, such as, Moreton waves, and blast waves. More recently, flare-associated helioseismic effects are also reported, such as, compact acoustic sources, seismic waves, and amplification of acoustic or p-modes (Ambastha, Basu and Antia, 2003). It is generally accepted that solar transients,viz., flares and Coronal Mass Ejections (CMEs), derive their energy from stressed magnetic field. Therefore, magnetic stress or non-potentiality is expected to relax to a lower energy state after the release of excess energy from the the solar active region (AR). A wide variety of results are reported on the pre- and post-flare changes in magnetic field (Ambastha et al. ,1993; Kosovichev & Zharkova, 2001; Mathew & Ambastha, 2000; Wang et al., 2002; Sudol & Harvey, 2004). However, serious questions have also been raised about the nature of observed changes as the measurements are expected to be affected by flare-induced spectral line profile changes (Harvey, 1986; Qiu and Gary, 2003). We shall discuss here some observed phenomena occurring in the photospheric magnetic and velocity fields during large flares and interpret them with the help of the recently available spectral data. Observed Flare/CME related changes {#sec:using} ================================== To distinguish between the normal magnetic (velocity) field evolution of solar ARs, and the changes considered to be associated with flares and CMEs, high sensitivity magnetic (Doppler velocity) data are required at high-cadence and high spatial resolution. Such data are available from the ground-based Global Oscillation Network Group (GONG) since 1995, and the space-borne Solar Oscillations and Heliospheric Observatory (SOHO) during 1996-2009. The space-borne measurements are considered to be better as they are not affected by atmospheric degradations. More recently, very high quality data are being provided by the Solar Dynamics Observatory (SDO)-Helioseismic and Magnetic Imager (HMI) subsequent to its launch in 2010. Both abrupt and persistent changes in observed magnetic field parameters have been reported that last over a few minutes during the impulsive phase of many energetic solar transient events, i.e., flares and CMEs. However, an anomalous magnetic flux polarity reversal was first reported during the large X-class flare of 6 April 2001 using MDI magnetograms (Qiu and Gary, 2003). Rapidly moving transient features, with speeds ranging from 30 to 50 km-s$^{-1}$, were reported also in magnetic and Doppler images of AR NOAA 10486 during the super-flares X17/4B (2003 October 28) and X10/2B (2003 October 29) by Maurya & Ambastha (2009: hereafter MA09). These transient features seen in both the GONG and MDI data were also the sites of magnetic polarity reversals during the impulsive phase of the energetic flares (Figure 1). ![(a) MDI magnetogram taken around the peak phase of the 2003 October 28/11:07:03 UT. Solid and dotted profiles show the magnetic flux measured along lines RS (b) and PQ (c) at 10:59:03 UT and 11:07:03 UT, respectively. Polarity reversals are evident at points G and H around 11:07:03 UT, i.e., the peak phase of the flare. Else where along the two lines, magnetic flux remained nearly unchanged.\[f:one\]](imd28.ps "fig:"){width="4.0cm"} ![(a) MDI magnetogram taken around the peak phase of the 2003 October 28/11:07:03 UT. Solid and dotted profiles show the magnetic flux measured along lines RS (b) and PQ (c) at 10:59:03 UT and 11:07:03 UT, respectively. Polarity reversals are evident at points G and H around 11:07:03 UT, i.e., the peak phase of the flare. Else where along the two lines, magnetic flux remained nearly unchanged.\[f:one\]](mfluxG28.ps "fig:"){width="4.0cm"} ![(a) MDI magnetogram taken around the peak phase of the 2003 October 28/11:07:03 UT. Solid and dotted profiles show the magnetic flux measured along lines RS (b) and PQ (c) at 10:59:03 UT and 11:07:03 UT, respectively. Polarity reversals are evident at points G and H around 11:07:03 UT, i.e., the peak phase of the flare. Else where along the two lines, magnetic flux remained nearly unchanged.\[f:one\]](mfluxH28.ps "fig:"){width="4.0cm"} Such magnetic polarity sign reversals were first reported several years ago by Zirin & Tanaka (1981) in some large flares. They attributed it to the heating of lower atmosphere by the flare such that the core of the absorption line, used for measurement, turned to emission (Patterson, 1984). However, only a few cases of anomalous magnetic and Doppler transients driven by the flares have been reported in recent years. It appears that the transients may reach detectable levels only in very energetic flares. Both the flares in MA09 were extremely energetic white-light (WL) events. More recently, Kosovichev (2011) have reported “sunquake” sources in the regions of these features and suggested these to arise due to thermal and hydrodynamic effects of high-energy particles heating the lower atmosphere. One would like to ask whether these features are related to the line profile change (Ding et al. 2002; Qiu & Gary, 2003) or to other physical processes such as the impact of a shock wave (Zharkova & Zharkov, 2007) on the line-production region. Due to the non-availability of the required spectral data, the transients in the super-flares of 2003 (MA2009) were examined using only the imaged data of the flares at various wavelengths and inferences were drawn based on indirect methods. But the solution of the problem lies in analyzing spectral profile shapes before, during and after major flare events, not available thus far. Therefore, the question of spectral line profile changes in relation to the observed magnetic anomaly remained unresolved. Spectral line profile changes associated with flares ==================================================== Photospheric magnetic field measurements are sensitive to changes in the line profile during the intense heating phase of solar flares (Ding et al., 2002). They modeled the Ni [i]{} 6768 Å line profiles using radiative transfer to explain the reversals observed during flares. However, the sign reversals as observed by MDI and GONG are difficult to be explained by the sudden heating of the lower atmosphere. This is because the Ni [i]{} line (used in both MDI and GONG) is formed in the temperature minimum region and is rather stable against temperature changes (Bruls, 1993). ![(Top row): Ca [ii]{} H 3968 Å  filtergrams (observed by Hinode Solar Optical telescope) showing the spatio-temporal development of the two-ribbon X2.2 flare in AR NOAA 11158 on 2011 February 15. (Bottom row): Consecutive difference images of HMI magnetograms showing the enhanced “magnetic transientï” features associated with the flare during the same period as above along the dotted and dashed curves L1 and L2.\[fl:two\]](TimSeqImg-1.eps){width="11cm"} The non-LTE calculations for Ni [i]{} line have shown that this absorption line can turn into emission only by a large increase of electron density and not by thermal heating of the atmosphere by any other means. Qiu & Gary (2003), interestingly, noted that sign-reversals occurred near locations of strong magnetic field, cooler sunspot umbrae that are exactly co-aligned with HXR sources. They suggested that such non-thermal effects are most pronounced in cool atmosphere where continuum is maintained at low intensity level. This implies that the sign reversal anomaly may be associated with non-thermal electron beam precipitating to the lower atmosphere near cool sunspots. Photospheric transients as reported previously by MA09 were observed recently by SDO-HMI during the first X-class flare of the current solar cycle 24 on 2011 February 15 in AR NOAA 11158 as shown in Figure 2. (Detailed study of this event is reported by Maurya, Vemareddy & Ambastha 2012). Unlike the Ni [i]{} line used in MDI and GONG, the SDO-HMI uses Fe [i]{} 6173.3 Å. It observes the Sun in two circular polarizations at six wavelength positions ($\pm34.4, \pm103.2$ and $\pm172.0$mÅ) of the spectral line. Using these observations one can construct the line profile for any desired location of the full disk Sun (Martinez Oliveros et al., 2011). Therefore, it is now possible to analyze the line profile changes, if any, associated with energetic transients and its effects on the estimation of magnetic and Doppler velocity fields. ![Top: (a) Consecutive difference magnetogram, and (b) white-light images of the AR NOAA 11158 during peak phase (01:53:15 UT) of the X2.2 flare. Bottom: Line profiles at (c) a quiet location P1, and (d) a transient location P2 as marked in top panel. Fe [I]{} line profiles are plotted during pre- (01:33:45 UT) and peak- (01:53:15 UT) phases of the flare.\[f:three\]](LineProf.eps){width="11cm"} We found spatial association of the transients with the WLF kernels and HXR sources at the different phases of evolution of the X2.2 flare. Figure 3 shows the line profiles at a quiet location P$_1$ and at a location P$_2$ of the transient feature. Solid (dashed) curves represent the line profiles during the pre (peak) phase of the flare. We found that the line profile turned to emission at the wavelength $\Delta\lambda$=–34.4mÅ and became broader during the peak phase of the flare as compared to the pre-flare phase. No such change was found at the quiet, reference location P$_1$. For a quantitative study of the transient related changes in the Stokes profiles, we further analyzed the line profiles at several flaring and quiet locations of the AR NOAA 11158. The Stokes profiles [*I*]{}$\pm$[*U*]{} and [*I*]{}$\pm$[*Q*]{} at the transient location P2 were also examined in order to ascertain the results on the line shape changes during the large flare. We found that the magnetic flux and Doppler velocity values as estimated by HMI algorithms were affected by these spectral changes, resulting in the observed transient features. Summary and Conclusions {#sec:concl} ======================= It is evident from the spectral data available from SDO-HMI that the Fe [i]{} line profile changes occurred during the impulsive phase of the energetic X2.2 flare of 2011 February 15 in AR NOAA 11158. The observed changes could be related to various physical processes occurring during the flare as follows. The thermodynamic change during the impulsive phase of the flare can drastically change the height of line formation due to the moderate perturbation of temperature and density. This can also occur as a result of the large increase of electron density as shown earlier by the non-LTE calculations of Ding et al. (2002) for Ni [i]{} 6768 Å  line. They suggested that the non-thermal excitation and ionization by the penetrating electrons generate a higher electron density which enhances the continuum opacity, thereby pushing the formation height of the line upward. The precipitation of electrons and deposition of energy in the chromosphere, enhanced radiation in the hydrogen Paschen continuum gives rise to the line source function, leading to an increase of the line core emission relative to the far wing and continuum. We thus conclude that the anomalous magnetic and Doppler transients observed during the X2.2 flare of 2011 February 15 were related to the Fe [i]{} line profile changes. Both thermal and non-thermal physical processes operating during the flare may contribute to the spectral profile changes at different stages of the flare evolution. Therefore, the observed magnetic and velocity transients are essentially the observational signatures of these physical processes operating during the energetic event, and do not correspond to “real” changes in the magnetic (doppler) fields. Ambastha, A., Hagyard, M.  J., West, E.  A., 1993, Solar Phys., 148, 277. Ambastha, A., Basu, S., Antia, H. M., 2003, Solar Phys. 218, 151. Bruls, J. H. M. J., 1993, A & A, 269, 509. Ding, M.  D., Qiu,  J., Wang,  H., 2002, ApJ, 576, L83. Harvey, J., 1986, in Deinzer, W., Knolker, M., Voigt, H. H., eds., Small Scale Magnetic Flux Concentrations in the Solar Photosphere, Vandenhoeck & Ruprecht, Goettingen, p. 25. Kosovichev, A.  G., 2011, ApJ, 734, L15 Kosovichev, A.  G., Zharkova, V.  V., 2001, ApJ, 550, L105 Martinez Oliveros, J.  C., Couvidat, S., Schou, J., et al., 2011, Solar Phys., 269, 269 Mathew, S. K., Ambastha, A., 2000, Solar Phys. 197, 75. Maurya, R.  A., Ambastha, A., 2009, Solar Phys., 258, 31 Maurya, R.  A., Vemareddy P. , Ambastha, A., 2012, ApJ, 747:134 (doi: http://dx.doi.org/10.1088/0004-637X/747/2/134 ) Patterson, A., 1984, ApJ, 280, 884 Qiu, J., Gary, D.  E., 2003, ApJ, 599, 615 Sudol, J. J., Harvey, J. W., 2005, ApJ, 635, 647. Wang, H. et al. 2002, ApJ, 576, 497. Zharkova, V.  V., Zharkov, S.  I., 2007, ApJ, 664, 573 Zirin, H., Tanaka, K., 1981, ApJ, 250, 791. \[lastpage\] [^1]: email: `[email protected]` [^2]: Current address: Astronomy Program, Department of Physics and Astronomy, Seoul National University, Seoul 151-747, Korea., email: `[email protected]`
{ "pile_set_name": "ArXiv" }
--- abstract: 'This paper investigates the consensus problem of general linear multi-agent systems under the framework of optimization. A novel distributed receding horizon control (RHC) strategy for consensus is proposed. We show that the consensus protocol generated by the unconstrained distributed RHC can be expressed in an explicit form. Based on the resulting consensus protocol the necessary and sufficient conditions for ensuring consensus are developed. Furthermore, we specify more detailed consensus conditions for multi-agent system with general and one-dimensional linear dynamics depending on the difference Riccati equations (DREs), respectively. Finally, two case studies verify the proposed scheme and the corresponding theoretical results.' author: - 'Huiping Li,  Weisheng Yan [^1][^2]' bibliography: - 'ref.bib' title: 'Receding Horizon Control Based Consensus Scheme in General Linear Multi-agent Systems' --- sy-headinput.tex Consensus, receding horizon control (RHC), multi-agent systems, general linear systems, optimization, diagraph. INTRODUCTION {#sec_introduction} ============ In last two decades, the cooperative control of networked multi-agent systems has received a lot of attention due to its wide applications. In particular, the consensus problem is of significant importance, and has inspired much progress, e.g., [@Reza_04_consensus_TAC; @Moreau_05_TAC_consensus; @Wei_05_TAC_consensus]. In this paper, we are interested in solving the consensus problem of multi-agent systems from the distributed optimal control perspective. The multi-agent system under study is of fixed directed network topology and general linear time invariant (LTI) dynamics associated with each agent. The objective of this paper is to design a locally optimal consensus strategy for each agent, and further to investigate under what conditions the closed loop system can achieve consensus by the designed strategy. The optimality of control protocols brings many desired properties such as phase and gain margin, leading to robustness of the closed loop systems. The core difficulty of the cooperative optimal control for multi-agent systems lies in the fact that the centralized optimization problem cannot generally be distributed among agents, with few exceptions [@Cao_08_TSMC_consensus; @Movric_14_TAC_optimal_tracking]. As a result, the best way of circumventing the difficulty is to utilize the locally optimal control strategy and further combine it with regional information exchange scheme to address the system-level interaction and coupling, approximately achieving some global or cooperative behaviors. In the literature, one approach to the optimal cooperative control is the linear quadratic regulation (LQR) scheme. For example, the distributed LQR problem of multi-agent systems with LTI dynamics is studied in [@Borrelli_08_TAC_distributedLQR], showing that the overall stability can be guaranteed by appropriately designing the local LQR and using information exchange among network topology. The consensus problem with optimal Laplacian matrix for multi-agent systems of first-order dynamics is investigated in [@Cao_08_TSMC_consensus], where it is shown that the design of globally optimal Laplacian matrix can only be achieved by properly choosing the global cost function coupled with the network topology. Recently, the LQR-based consensus problem of multi-agent systems with LTI dynamics and fixed directed topology is addressed in [@Movric_14_TAC_optimal_tracking], indicating that the globally optimal consensus performance can be achieved by using locally optimal consensus protocol if and only if the overall performance index is selected in a special form depending on the graph structure. Another way of achieving the (sub-)optimal cooperative control of multi-agent systems is the distributed receding horizon control (RHC) strategy, also known as distributed model predictive control. Based on this approach, there have been many results developed for cooperative stabilization, formation control, and its applications. For example, the distributed RHC-based scheme for cooperative stabilization is proposed in [@Dunbar_06_Auto_DMPC; @Muller_12_DMPC_IJRNC], and the formation stabilization is addressed in [@Keviczky_06_auto_DMPC] and its application is reported in [@Keviczky_08_TCST_formation_MPC]. Furthermore, the robust distributed RHC problems that can be used for cooperative stabilization are studied in [@Richards_IJC_07_RDMPC] for linear systems with coupled constraints and in [@Li_TAC_14_RDRHC] for nonlinear systems. To further attack the unreliability of the communication networks, the cooperative stabilization problem of multi-agent nonlinear systems with communication delays are investigated in [@Li_SCL_DMPC_14; @Franco_08_TAC_DMPC; @Li_Auto_14_DRHC]. Note that all of these results use cost functions as Lyapunov functions to prove stability. Even though it is very desirable to achieve optimal consensus by distributed RHC scheme, there have been few results for the consensus problem of multi-agent systems due to the difficulty that the cost function may not be directly used as Lyapunov function. In [@Ferrari_09_TAC_MPC_consensus], Ferrari-Trecate [*et al.*]{} study the consensus problem of multi-agent systems of first-order and second-order dynamics, and the sufficient conditions for achieving consensus are developed by exploiting the geometry properties of the optimal path. Zhan [*et al.*]{} investigate the consensus problem of first-order sampled-data multi-agent systems in [@Zhan_13_auto_consensus], where state and control input information needs to be exchanged. Note that these two results are only focused on special type of linear systems, which is of limited use. In [@Johansson_08_auto_optimal_consensus] Johansson [*et al.*]{} propose to use the negotiation to reach the optimal consensus value by implementing the primal decomposition and incremental sub-gradient algorithm, but the effect of the network topology is not explicitly considered. It can seen that the receding horizon control -based consensus scheme for multi-agent systems with general LTI dynamics has not been solved, and the relationship between consensus and the interplay between the network topology and the RHC design is still unclear, which motivates this study. The main contribution of this paper is two-fold. - A novel distributed RHC strategy is proposed for designing the consensus protocol. In this strategy, each agent at each time instant only needs to obtain its neighbors’s state once via communication network, which is more efficient than the work in [@Li_TAC_14_RDRHC; @Dunbar_06_Auto_DMPC] (where the state and its predicted trajectory need to be transmitted) and [@Muller_12_DMPC_IJRNC; @Zhan_13_auto_consensus] (where the neighbors’ information needs to be exchanged for many times at each time instant). In addition, we show that the consensus protocol generated by the RHC is a feedback of the linear combination of each agent’s state and its neighbors’ states, and the feedback gains depend on a set of difference matrix equations. We believe our results partially extend the results in [@Ferrari_09_TAC_MPC_consensus; @Zhan_13_auto_consensus] to multi-agent systems with LTI dynamics. - Given the proposed distributed RHC strategy, a necessary and sufficient condition for ensuring consensus is developed. We show that the consensus can be reached if and only if the network topology contains a spanning tree and a simultaneous stabilization problem can be solved. Furthermore, more specifical sufficient consensus conditions depending on one Reccati difference equation for the multi-agent with LTI dynamics and one-dimensional linear dynamics are also developed, respectively. The remainder of this paper is organized as follows. Section \[Sec\_problem\_formulation\] introduces some well-known results from graph theory and formulates the problem to be studied. Section \[Sec\_consensus\_protocol\] presents the novel distributed RHC scheme, and develops a detailed consensus protocol. The necessary and sufficient conditions for enuring consensus are proposed in Section \[sec\_consensus\_analysis\], and more specifical sufficient consensus conditions for multi-agent systems with LTI dynamics and one-dimensional linear dynamics are also reported in this section. The case studies are demonstrated in Section \[Sec\_simulation\]. Finally, the conclusion remarks are summarized in Section \[Sec\_Conclusion\]. For the ease of presentation, the following notations are adopted in this paper. The symbol $\mathbb{R}$ represents the real space. For a real matrix $A$, its transposition and inverse (if the inverse exists) are denoted as “$A^{\rm T}$” and “$A^{-1}$”, respectively. If $A$ is a complex matrix, then the transposition is denoted by $A^{\rm H}$. Given a real (or complex) number $\lambda$, the absolute value (modulus) is defined as $|\lambda|$. Given a matrix (or a column vector) $X$ and another matrix $P$ with appropriate dimension, the $2$-induced norm (or the Euclidean norm) of $X$ is denoted by $\|X\|$ and the $P$-weighted norm of $X$ is denoted by $\|X\|_{P} \triangleq \sqrt{X^{\rm T}PX}$. Given matrix $Q$, $Q>0$ ($Q\geqslant0$) stands for the matrix $Q$ being positive definite (semi-positive definite). Define the column operation ${\rm col}\{x_1,x_2,\cdots,x_n\}$ as $[x_1^{\rm T},x_2^{\rm T},\cdots,x_n^{\rm T}]^{\rm T}$, where $x_1, x_2,\cdots,x_n$ are column vectors. $I_n$ stands for the identity matrix of dimension $n$, and $\textbf{1}_n$ represents an $n$-dimensional column vector $[1, \cdots, 1]^{\rm T}$. The symbol $\otimes$ stands for the Kronect product. Problem Formulation {#Sec_problem_formulation} =================== Consider a multi-agent system of $M$ linear agents. For each agent $i$, the dynamics is described as $$\label{equ_each_agent} x_i(k+1) = A x_i(k) + B u_i(k),$$ where $x_i(k+1) \in \mathbb{R}^{n}$ is the system state, and $u_i(k) \in \mathbb{R}^m$ is the control input of agent $i$. There exists a communication network among the $M$-agent system, and the network topology is described as a directional graph (diagraph) $\mathcal{G} \triangleq \{\mathcal{V},\mathcal{E},\mathcal{A}\}$. Here, $\mathcal{V} = \{i, i = 1,\cdots,M\}$ is the collection of the nodes of the digraph representing each agent $i$, $\mathcal{E} \subset \mathcal{V}\times \mathcal{V}$ denotes the edges of paired agents, and $\mathcal{A} =[a_{ij}]\in\mathbb{R}^{M\times M}$ is the adjacency matrix with $a_{ij}\geqslant 0$. If there is a connection from agent $i$ to $j$, then $a_{ij} = 1$; otherwise, $a_{ij} = 0$. We assume there is no self-circle in the diagraph $\mathcal{G}$, i.e., $a_{ii} = 0$. For each agent $i$, its neighbors are denoted by the agents from which it can obtain information, and the index set for agent $i$’s neighbors is denoted as $\mathcal{N}_i \triangleq \{j|(i,j)\subset \mathcal{E}\}$. The number of agents in $\mathcal{N}_i$ is denoted as $|\mathcal{N}_i|$. The in-degree of agent $i$ is denoted as $Deg_{in}(i) = \sum_{i=j}^M a_{ij}$, and the degree matrix is denoted as $\mathcal{D}={\rm diag}\{Deg(1),\cdots, Deg (M)\}$. The Laplacian matrix of $\mathcal{G}$ is denoted as $\mathcal{L} = \mathcal{D} - \mathcal{A}$. Arrange the eigenvalues of $\mathcal{L}$ as $|\lambda_1| \leqslant |\lambda_2| \leqslant \cdots \leqslant |\lambda_M|$. Assume that the diagraph $\mathcal{G}$ is fixed. Firstly, we recall some standing results from the graph theory [@Wei_05_TAC_consensus; @Reza_04_consensus_TAC; @You_11_TAC_consensus]. \[lem\_spanning\_tree\] The digraph $\mathcal{G}$ contains a spanning tree if and only if zero is a simple eigenvalue of the Laplacian matrix $\mathcal{L}$, i.e., $0 < |\lambda_2| \leqslant \cdots \leqslant |\lambda_M|$, and the corresponding right eigenvector is ${\rm \textbf{1}}$. For the linear system in (\[equ\_each\_agent\]), a standing assumption is made: The pair $[A,B]$ is controllable. We assume that at each time instant $k$, over the given communication network, agent $i$ can get state information $x_j(k)$, $j\in\mathcal{N}_i$ from its neighbors. The communication network is reliable and the information can be transmitted instantaneously without time consumption. [@You_11_TAC_consensus] The discrete-time multi-agent system in (\[equ\_each\_agent\]) with a given network topology $\mathcal{G}$, and under a distributed control protocol $u_i(k) = f(x_i(k), x_{-i}(k))$, is said to achieve consensus if, $$\lim_{k \rightarrow \infty } \|x_i(k) - x_j(k) \| = 0, j = 1, \cdots, M,$$ where $x_{-i}(k)$ are the collection of agent $i$’s neighbors’ states, i.e., $x_{-i}(k) \triangleq \{x_j,j\in\mathcal{N}_i\}$ and $f: \underbrace{\mathbb{R}^n \times \cdots\times \mathbb{R}^n}_{|\mathcal{N}_i| + 1}\rightarrow \mathbb{R}^m$. In this paper, we are interested in designing $u_i(k) = f(x_i(k), x_{-i}(k))$ using the distributed RHC strategy for each agent $i$ to achieve consensus. Distributed RHC based Consensus Protocol {#Sec_consensus_protocol} ======================================== This section first presents the distributed RHC based consensus strategy, and then develops the detailed consensus protocols for each agent. Distributed RHC Scheme ---------------------- For each agent $i$, we propose to utilize the following optimization problem to generate the consensus protocol. $$\begin{aligned} U_i^*(k) = \arg\min_{\hat{U}_i(k)} J_i(\hat{x}_i(k),\hat{U}_i(k),x_{-i}(k))\end{aligned}$$ subject to $$\begin{aligned} \hat{x}_i(k+n+1|k) = A \hat{x}_i(k+n|k) + B \hat{u}_i(k+n|k),\end{aligned}$$ where $n = 0, \cdots, N-1$. Here, $\hat{x}_i(k|k) = x_i(k)$, $\hat{U}_i(k)= \{ \hat{u}_i(k|k), \cdots, \hat{u}_i(k+N-1|k)\}$ and $J_i(\hat{x}_i(k),\hat{U}_i(k),x_{-i}(k)$ is defined as $$\begin{aligned} & J_i(\hat{x}_i(k),\hat{U}_i(k),x_{-i}(k))\nonumber\\ = &\sum_{n=0}^{N-1} \left(\|\hat{u}_i(k+n|k)\|_{R_i}^2 + \sum_{j\in \mathcal{N}_i}a_{ij}\|\hat{x}_i(k+n|k) - x_j(k)\|_{Q_i}^2\right)\nonumber\\ & + \sum_{j\in \mathcal{N}_i}a_{ij}\|\hat{x}_i(k+N|k) - x_j(k)\|_{Q_{iN}}^2,\end{aligned}$$ where $R_i>0$, $Q_i>0$ and $Q_{iN}>0$ are symmetric matrices, and $N>0$ is a positive integer which is called the prediction horizon. At each time instant $k$, each agent gets its neighbors’ state information $x_j(k)$, $j\in\mathcal{N}_i$, solves Problem 1 and then uses $u^*_i(k|k)$ as the control input to achieve consensus. In Problem 1, we use the term $\sum_{j\in \mathcal{N}_i}a_{ij}\|\hat{x}_i(k+n|k) - x_j(k)\|_{Q_i}^2$ to achieve consensus. Note that the consensus term can be rewritten as $|\mathcal{N}_i|\|\hat{x}_i(k+n|k) - ave(x_{-i}(k))\|_{Q_i}^2$, where $ave(x_{-i}(k))$ is the average of agent $i$’s neighboring states, which is consistent with the consensus term in [@Ferrari_09_TAC_MPC_consensus]. However, Problem 1 generates the first-order and second-order cases in [@Ferrari_09_TAC_MPC_consensus] to general linear systems. In addition, the solve of Problem 1 only requires the exchange of agent $i$ neighbors’ state for one time, which results in less communication load in comparison with these in [@Muller_12_DMPC_IJRNC; @Zhan_13_auto_consensus], where more information is needed to exchange for consensus. Specific Consensus Protocol --------------------------- This subsection shows that the consensus protocol generated by solving Problem 1 is can be expressed in an explicit form. By using the convex optimization approach, the result is reported in the following theorem. \[thm\_consensus\_law\] For the system in (\[equ\_each\_agent\]) with network topology $\mathcal{G}$ and the optimization problem 1, the optimal solution is given as: $$\hat{u}_i^*(k+n|k) = \sum_j^{M} a_{ij} [K_i(n) \hat{x}_i(k+n|k) + G_i(n) x_j(k)],$$ Here, $n = 0,\cdots,N-1$, $K_i(n) = -R^{-1}_i B^{\rm T} [I + \sum_{j=1}^{M}a_{ij} P_i(n+1)B R^{-1}_iB^{\rm T}]^{-1} P_i(n+1) A$, and $G_i(n) = -R^{-1}_i B^{\rm T} [I + \sum_{j=1}^{M}a_{ij} P_i(n+1)B R^{-1}_i B^{\rm T}]^{-1} \Delta_i(n+1)$, where $P_i(n+1)$ and $\Delta_i(n+1)$ satisfy the following matrix equations, respectively: $$\begin{aligned} P_i(n) = & A^{\rm T}[I + \sum_{j=1}^{M}a_{ij} P_i(n+1)B R^{-1}_i B^{\rm T}]^{-1}P_i(n+1)A + Q_i,\label{equ_Pn} \\ \Delta_i(n) =& A^{\rm T}[I + \sum_{j=1}^{M}a_{ij} P_i(n+1)B R^{-1}_i B^{\rm T}]^{-1}\Delta_i(n+1) -Q_i,\label{equ_Deltan}\end{aligned}$$ with the initial conditions $P_i(N) = Q_{iN}$ and $\Delta_i(N) = -Q_{iN}$. Furthermore, the consensus protocol for each agent $i$ is given as: $$\label{equ_consensus_law} u_i(k) = u_i^*(k|k).$$ By introducing the lagrange multipliers $\lambda_i(n+1) \in \mathbb{R}^n, n = 0, \cdots, N-1$, construct a lagrange function as $$\begin{aligned} L(\hat{U}_i(k),\hat{X}_i(k)) \triangleq & \frac{1}{2} J_i(\hat{x}_i(k),\hat{U}_i(k),x_{-i}(k)) \\ & + \sum_{n=0}^{N-1}\lambda^{\rm T}_i(n+1)[A\hat{x}_i(k+n|k)\nonumber \\ & + B \hat{u}_i(k+n|k) - \hat{x}_i(k+n+1|k)],\end{aligned}$$ where $\hat{X}_i(k) \triangleq \{\hat{x}_i(k|k),\cdots,\hat{x}_i(k+N-1|k)\}$. Since $Q_i>0, Q_{iN}>0$ and $R_i>0$, the objective function is strictly convex. According to the the Karush-Kuhn-Tucker (KKT) conditions [@Lu_12_optimization], there exists a unique global minima for Problem 1, which satisfies the following conditions: $$\begin{aligned} \frac{\partial L}{\partial \hat{x}_i^*(k+n|k)} = & \sum_{j=1}^{M} Q_i a_{ij}[\hat{x}_i^*(k+n|k)-x_j(k)]+ A^{\rm T} \lambda_i(n+1) - \lambda(n) = 0,\label{equ_KKT1} \\ \frac{\partial L}{\partial \hat{x}_i^*(k+N|k)} = & \sum_{j=1}^{M} Q_{iN} a_{ij}[\hat{x}_i^*(k+N|k)-x_j(k)]- \lambda_i(N) = 0,\label{equ_KKT2}\\ \frac{\partial L}{\partial \hat{u}_i^*(k+n|k)} & = R_i \hat{u}_i^*(k+n|k) + B^{\rm T} \lambda_i(n+1) = 0,\label{equ_KKT3}\\ A\hat{x}_i^*(k+n|k)& + B \hat{u}_i^*(k+n|k) - \hat{x}_i^*(k+n+1|k) = 0,\label{equ_KKT4}\end{aligned}$$ where $n = 0, \cdots, N-1$. In what follows, we first show an equation to evaluate $\lambda_i(n)$, by mathematical induction as follows: $$\label{equ_lamdai} \lambda_i(n) = \sum_{j=1}^{M} [P_i(n) \hat{x}_i^*(k+n|k) + \Delta_i(n) x_j(k)],$$ where, $ n = N, \cdots, 0$. Using (\[equ\_KKT3\]), we can get $$\label{equ_u_k} \hat{u}_i^*(k+n|k) = -R_i^{-1} B^{\rm T} \lambda_i(n+1).$$ When $n=N$, from (\[equ\_KKT2\]), we can obtain that $$\begin{aligned} \lambda_i(N) = & \sum_{j=1}^{M} a_{ij}Q_{iN}[\hat{x}_i^*(k+N|k)-x_j(k)]\nonumber\\ = & \sum_{j=1}^{M} a_{ij}[P_i(N)\hat{x}_i^*(k+N|k) + \Delta_i(N) x_j(k)].\end{aligned}$$ Assume that (\[equ\_lamdai\]) holds for some $n = l+1$ with $ l \leqslant N-1$. When $n = l$, by combining (\[equ\_KKT1\]), (\[equ\_KKT4\]) and (\[equ\_u\_k\]), we have $$\begin{aligned} \lambda_i(l) = & A^{\rm T} [I + \sum_{j=1}^{M}a_{ij} P_i(n+1)B R^{-1}_i B^{\rm T}]^{-1}\nonumber\\ & \times [\sum_{j=1}^{\rm M} a_{ij} (P_i(l+1) A\hat{x}_i^*(k+n|k)+ \Delta_i(l+1) x_j(k))]\nonumber\\ & + \sum_{j=1}^{M} a_{ij} Q_i(\hat{x}_i^*(k+n|k) - x_j(k))\nonumber\\ =& \sum_{j=1}^{M} a_{ij}[P_i(l) \hat{x}_i(k+n|k) + \Delta_i(l) x_j(k)].\end{aligned}$$ Therefore, using the mathematical induction, (\[equ\_lamdai\]) has been proven. Finally, the result in Theorem \[thm\_consensus\_law\] follows by plugging (\[equ\_lamdai\]) into (\[equ\_u\_k\]). This completes the proof. Note that $P_i(n)$ satisfies a modified Reccati difference equation in (\[equ\_Pn\]), which depends on the network topology. In order to decouple (\[equ\_Pn\]) from the network topology, we design $R_i$ such that $$\label{equ_design_R} R_1/|\mathcal{N}_1| = \cdots = R_M/|\mathcal{N}_M| = R,$$ where $R>0$. Using (\[equ\_design\_R\]), (\[equ\_Pn\]) and (\[equ\_Deltan\]) become $$\begin{aligned} P_i(n) = A^{\rm T}[I + P_i(n+1)B R^{-1} B^{\rm T}]^{-1}P_i(n+1)A + Q_i,\label{equ_RDE_Pn} \\ \Delta_i(n) = A^{\rm T}[I + P_i(n+1)B R^{-1} B^{\rm T}]^{-1}\Delta_i(n+1) -Q_i.\label{equ_RDE_Deltan}\end{aligned}$$ Therefore, $P_i(n)$ follows a standard RDE in (\[equ\_RDE\_Pn\]). Consensus Analysis {#sec_consensus_analysis} ================== This sections first develops a necessary and sufficient condition for achieving consensus of general linear agent systems, then further propose the sufficient conditions for multi-agent system of general LTI and one-dimensional dynamics, respectively. Necessary and Sufficient Condition ---------------------------------- The necessary and sufficient condition for reaching consensus is as follows. \[thm\_iff\] For the system in (\[equ\_each\_agent\]) with the network topology $\mathcal{G}$ and the consensus protocol in (\[equ\_consensus\_law\]), assume that $R_i$ is designed as in (\[equ\_design\_R\]). Then the consensus can be reached if and only if (a) $\mathcal{G}$ contains a spanning tree, and (b) $A-BK_i + (\lambda_i - 1)B G_i$ is stable for all $\lambda_i$, $i=2,\cdots, M$, where $\lambda_i$ are the nonzero eigenvalues of $\Gamma$ defined as $\Gamma = {\rm diag}(1/|\mathcal{N}_1|, \cdots, 1/|\mathcal{N}_M|) \mathcal{L}$, and $$\begin{aligned} K_i \triangleq & -R^{-1}B^{\rm T} [I + P_i(1) B R^{-1} B^{\rm T}]^{-1} P_i(1) A\\ G_i \triangleq & -R^{-1}B^{\rm T}[I + P_i(1) B R^{-1} B^{\rm T}]^{-1} \Delta_i(1).\end{aligned}$$ Before proving Theorem \[thm\_iff\], a lemma is first needed. \[lemma\_new\_graph\] If the graph $\mathcal{G}$ contains a spanning tree, then a new graph $\tilde{\mathcal{G}} \triangleq \{\mathcal{V}, \mathcal{E}, \tilde{\mathcal{A}}\}$ also contains a panning tree, with $\sum_{j=1}^{M} \tilde{a}_{ij} = 1$ for all $i$, and the corresponding Laplacian matrix is $\Gamma$, where $\tilde{\mathcal{A}} = {\rm diag}(1/|\mathcal{N}_1|, \cdots, 1/|\mathcal{N}_M|) \mathcal{A}$. The proof of lemma \[lemma\_new\_graph\] is straightforward, so it is omitted here. For each agent $i$, plugging the consensus law in (\[equ\_consensus\_law\]) into (\[equ\_each\_agent\]), and doing some algebraic operations, we have $$\begin{aligned} x_i(k+1) = A x_i(k) - \sum_{j=1}^{M}\tilde{a}_{ij}(B K_i x_i(k) + B G_i x_j(k)),\end{aligned}$$ where $\tilde{a}_{ij}$ is the element in $\tilde{\mathcal{A}}$. Define $\delta_i(k) = x_i(k) - x_1(k)$, and we have $$\begin{aligned} \delta_i(k) = (A-BK_i) \delta_i(k) - \sum_{j=1}^{M}(\tilde{a}_{ij}-\tilde{a}_{1j}) B G_i \delta_j(k),\end{aligned}$$ where $i = 2, \cdots, M$. By writing the above equations in an augmented form, one gets $$\begin{aligned} & \delta(k+1) = (I_{M-1}\otimes(A-BK_i))\delta(k)\nonumber\\ & - \left[\left( \begin{array}{ccc} \tilde{a}_{22}- \tilde{a}_{12} & \ldots & \tilde{a}_{2M}- \tilde{a}_{1M}\\ \vdots & \ddots & \vdots \\ \tilde{a}_{M2}- \tilde{a}_{12} & \ldots & \tilde{a}_{MM}- \tilde{a}_{1M} \\ \end{array} \right) \otimes (BG_i)\right]\delta(k),\end{aligned}$$ where $\delta(k) = {\rm col}(\delta_2(k), \cdots, \delta_M(k))$. The above equation can be further written as $$\begin{aligned} \label{equ_closed_A_delta} \delta(k+1) = A_\delta \delta(k)\end{aligned}$$ where $A_\delta = (I_{M-1}\otimes(A-BK_i))+ (\tilde{\mathcal{L}}_p + \textbf{1}_{M-1} \tilde{a}_1^{\rm T} \otimes B G_i) - (I_{M-1}\otimes B G_i)$, and $ \tilde{\mathcal{L}}_p = \left[ \begin{array}{cccc} 1 & -\tilde{a}_{22} & \ldots & \tilde{a}_{2M}\\ -\tilde{a}_{32} & 1 & \ldots & -\tilde{a}_{3M}\\ \vdots & \vdots & \ddots & \vdots \\ -\tilde{a}_{M2} & -\tilde{a}_{M3} & \ldots & 1\\ \end{array} \right]$ and $\tilde{a}_1 = [\tilde{a}_{12}, \cdots, \tilde{a}_{1M}]^{\rm T}$. From (\[equ\_closed\_A\_delta\]), it can be seen that the consensus can be reached if and only if all the eigenvalues of $A_\delta$ are in the unit circle. Next, we need to analyze the properties of the eigenvalues of $A_\delta$. According to Lemma \[lemma\_new\_graph\] and Lemma \[lem\_spanning\_tree\], $\Gamma$ has exactly one zero eigenvalue. Put the nonzero eigenvalues of $\Gamma$ in order as $|\lambda_2|\leqslant \cdots \leqslant |\lambda_M|$. Since $S^{-1} \Gamma S = \left[ \begin{array}{cc} 0 & -\tilde{a}_1^{\rm} \\ 0 & \tilde{\mathcal{L}}_p + \textbf{1}_{M-1} \tilde{a}_1^{\rm T} \\ \end{array} \right] $, the eigenvalues of $\tilde{\mathcal{L}}_p + \textbf{1}_{M-1} \tilde{a}_1^{\rm T}$ are $\lambda_2, \cdots, \lambda_M$. Therefore, there exists a nonsingular matrix $T$, such that $$\begin{aligned} T^{-1}(\tilde{\mathcal{L}}_p + \textbf{1}_{M-1} \tilde{a}_1^{\rm T} ) T = J = {\rm \diag}(J_1, \cdots, J_s),\end{aligned}$$ where $J_k$, $k=1,\cdots, s$ are upper triangular Jordan blocks and the principle diagonal elements are $\lambda_2, \cdots, \lambda_M$. As a result, we have $$\begin{aligned} & (T\otimes I_n)^{-1} A_\delta (T\otimes I_n) \nonumber \\ = & I_{M-1}\otimes(A-BK_i) + J\otimes B G_i - I_{M-1} B G_i.\end{aligned}$$ Thus, the eigenvalues of $A_\delta$ are $A-BK_i + (\lambda_i -1) B G_i$, for all $i = 2, \cdots, M$. That is, the consensus can be reached if and only if $A-BK_i + (\lambda_i -1) B G_i$ are stable. The proof is completed. Note that $K_i$ and $G_i$ depends on the matrix equations in (\[equ\_RDE\_Pn\]) and (\[equ\_RDE\_Deltan\]), respectively. In particular, the (\[equ\_RDE\_Deltan\]) is not a RDE, and thus might be of complex properties. In the following, we develop more detailed sufficient conditions to facilitate design. Sufficient Conditions for General Linear Systems ------------------------------------------------ In this section, we develop sufficient conditions for ensuring consensus, which are reported as follows. \[col\_sufficient\_LTI\] For the system in (\[equ\_each\_agent\]) with the network topology ${\mathcal{G}}$, assume that ${\mathcal{G}}$ contains a spanning tree, and $R_i$ is designed as in (\[equ\_design\_R\]). If $Q_{iN}$ and $Q_i$ are designed such that $$\begin{aligned} & Q_{iN} - P_i(N-1) >0, \label{equ_monotically_P}\\ & Q_i - \Xi_i -\|\lambda_j-1\|^2\nonumber\\ &\times \|I + B(R+B^{\rm T}PB)^{-1}B^{\rm T}\|_{[\bar{\rho}^{N-1}_i Q_{iN} + \sum_{l=1}^{N-2}\bar{\rho}_i^l Q_i]}^2 > 0,\label{equ_sufficient_Q}\end{aligned}$$ where $\Xi_i \triangleq \| B(R + B^{\rm T} P_i(1) B)^{-1}B^{\rm T} P_i(1) \bar{A}_i\|^2$, $\bar{A}_i = A - BK_i$, $\bar{\rho}_i = \rho(A^{\rm T}(I + P_i(1) B R^{-1}B)^{-1})$ and $\lambda_j$, $j=2, \cdots, M$, are the eigenvalues of $\tilde{\mathcal{L}}$. Before developing the proof of Theorem \[col\_sufficient\_LTI\], three lemmas are needed. \[lemma\_RDE\] For the RDE in (\[equ\_RDE\_Pn\]), if $Q_{iN} - P_i(N-1)>0$, then $ P_i(N)>\cdots, >P_i(1)$, and $\hat{u}_i^*(k+n|k) = BK_i(k+n|k)$ is a stabilizing controller law for the system in (\[equ\_each\_agent\]), $n = N-1, \cdots, 0$. The proof can be directly followed using Theorem 1 and Lemma 1 in [@Poubelle_88_TAC_FakeReccati] by considering $Q_i>0$. \[lemma\_special\_G\_i\] For the RDE in (\[equ\_RDE\_Pn\]) and all the eigenvalues $\lambda_2, \cdots, \lambda_M$ of $\tilde{\mathcal{L}}$, the following holds: $$\begin{aligned} \label{equ_temp_Pi(1)} & P_i(1) -(A- \lambda_i B K_i)^{\rm H} P_i(1)(A- \lambda_i B K_i) \nonumber\\ > & (1-\bar{\lambda}_i)P_i(1) + (\bar{\lambda}_i - 1)A^{\rm T}P_i(1)A + \bar{\lambda} Q_i,\end{aligned}$$ where $\bar{\lambda}_i = \lambda_i^{\rm H} + \lambda_i - \|\lambda_i\|^2$. According to [@Anderson_71_LOC], $K_i$ can also been written as $$\label{equ_K_i_newform} K_i = (R + B^{\rm T} P_i(1) B)^{-1}B^{\rm T} P_i(1) A.$$ Using (\[equ\_K\_i\_newform\]), we have $$\begin{aligned} & P_i(1) - (A- \lambda_i B K_i)^{\rm H} P_i(1)(A- \lambda_i B K_i) \nonumber\\ = & P_i(1) - A^{\rm T} P_i(1)A + (\lambda_i^{\rm H} + \lambda_i) A^{\rm T} P_i(1)\nonumber\\ &\times B (R+ B^{\rm T}P_i(1)B)^{-1}B^{\rm T} P_i(1)A \nonumber\\ & - \|\lambda_i\|^2 A^{\rm T} P_i(1) B (R+ B^{\rm T}P_i(1)B)^{-1}\nonumber \\ &\times B^{\rm T} P_i(1) B (R + B^{\rm T} P_i(1) B)^{-1} B^{\rm T} P_i(1) A\nonumber\\ \geqslant & P_i(1) - A^{\rm T} P_i(1)A \nonumber\\ & + \bar{\lambda_i} A^{\rm T} P_i(1) B (R+ B^{\rm T}P_i(1)B)^{-1}B^{\rm T} P_i(1)A.\end{aligned}$$ The DRE in (\[equ\_RDE\_Pn\]) can also be written as $$\begin{aligned} P_i(n) = & Q_i + A^{\rm T} P_i(n+1) A - A^{\rm T} P_i(n+1) B\nonumber\\ &\times (R+ B^{\rm T}P_i(n+1)B)^{-1}B^{\rm T} P_i(n+1)A.\end{aligned}$$ Therefore, we further have $$\begin{aligned} & P_i(1) - (A- \lambda_i B K_i)^{\rm H} P_i(1)(A- \lambda_i B K_i) \nonumber\\ \geqslant & P_i(1) - A^{\rm T} P_i(1)A + \bar{\lambda_i} (Q_i + A^{\rm T} P_i(1) - P_i(0)).\end{aligned}$$ According to Lemma \[lemma\_RDE\], it can be derived that $P_i(0) < P_i(1)$. As a result, (\[equ\_temp\_Pi(1)\]) follows. This completes the proof. The following lemma gives a bound of the sequence $\{\Delta_i(n)\}$. \[lemma\_delta\_auxiliary\] For any positive definite matrix $\Pi\in \mathbb{R}^{n\times n}$, the following holds: $$\|\Pi\|^2_{\Delta_i(n)} \leqslant \|\Pi\|^2_{\bar{\rho}^{N-n}_i Q_{iN} + \sum_{l=1}^{N-(n+1)}\bar{\rho}_i^l Q_i},$$ where $n = N, \cdots, 0$. Construct an auxiliary series $\{\tilde{\Delta}_i(n)\}$ as follows: $$\tilde{\Delta}_i(n) = \bar{\rho}_i \tilde{\Delta}_i(n+1) + Q_i,$$ where $n = N-1, \cdots, 0$, and $\tilde{\Delta}_i(N) = Q_{iN}$. Using Lemma \[lemma\_RDE\], it can be seen that $\tilde{\Delta}_i(n) + \Delta_i(n) \geqslant 0$. On the other hand, we have $$\tilde{\Delta}_i(n) = \bar{\rho}^{N-n}_i Q_{iN} + \sum_{l=1}^{N-(n+1)}\bar{\rho}_i^l Q_i.$$ Thus, the result in Lemma \[lemma\_delta\_auxiliary\] holds. *Proof of Theorem \[col\_sufficient\_LTI\]*: By direct calculation, we have $$\begin{aligned} & P_i(1) - [A-BK_i + (\lambda_i - 1) B G_i]^{H} \nonumber \\ & \times P_i(1)[A-BK_i + (\lambda_i -1)B G_i]\nonumber\\ = & P_i(1) - \bar{A}_i^{\rm T} P_i(1) \bar{A}_i -(\lambda_i-1)^{\rm H} \Delta_i(1)^{\rm T}\nonumber \\ & \times B(R + B^{\rm T}P_i(1)B)^{-1} B^{\rm T} P_i(1)\bar{A}_i\nonumber\\ & -(\lambda_i -1) \bar{A}_i^{\rm T}P_i(1)B(R + B^{\rm T}P_i(1)B)^{-1}B^{\rm T}\Delta_i(1)\nonumber\\ & -\|\lambda_i -1\|^2 \Delta_i(1)^{\rm T} B (R + B^{\rm T}P_i(1)B)^{-1}\nonumber \\ & \times B^{\rm T}P_i(1)B(R + B^{\rm T}P_i(1)B)^{-1} B^{\rm T} \Delta_i(1),\end{aligned}$$ where $G_i = (R+B^{\rm T}P_i(1)B)^{-1}B^{\rm T}\Delta_i(1)$ has been used. According to Lemma 5, we further have $$\begin{aligned} & P_i(1) - [A-BK_i + (\lambda_i - 1) B G_i]^{H} \nonumber \\ & \times P_i(1)[A-BK_i + (\lambda_i -1)B G_i]\nonumber\\ \geqslant & P_i(1) - \bar{A}_i^{\rm T} P_i(1) \bar{A}_i - \|\lambda_i -1\|^2 \Delta_i(1)^{\rm T}\nonumber \\ & \times (I + B(R+B^{\rm T}P_i(1)B)^{-1}B^{\rm T}) \Delta_i(1)- \Pi_i.\end{aligned}$$ Using Lemma \[lemma\_special\_G\_i\] with $\lambda_i = 1$, we get $$\begin{aligned} & P_i(1) - [A-BK_i + (\lambda_i - 1) B G_i]^{H} \nonumber \\ & \times P_i(1)[A-BK_i + (\lambda_i -1)B G_i]\nonumber\\ > & Q_i + \|\lambda_i -1\|^2 \Delta_i(1)^{\rm T}\nonumber \\ & \times (I + B(R+B^{\rm T}P_i(1)B)^{-1}B^{\rm T}) \Delta_i(1)- \Pi_i.\end{aligned}$$ Finally, applying the result in Lemma \[lemma\_delta\_auxiliary\] and using the condition in (\[equ\_sufficient\_Q\]), we obtain that $P_i(1) - [A-BK_i + (\lambda_i - 1) B G_i]^{H} P_i(1)[A-BK_i + (\lambda_i -1)B G_i]>0$. That is, $A-BK_i + (\lambda_i - 1) B G_i$ is stable for all $\lambda_2,\cdots, \lambda_M$. According to Theorem \[thm\_iff\], the consensus can be reached. This completes the proof. Sufficient Conditions for One-dimensional Systems ------------------------------------------------- For each agent $i$ of one-dimensional linear dynamics, the system equation becomes $$\label{equ_one_dimension_system} x_i(k+1) = a x_i(k) + b u_i (k).$$ For one-dimensional systems, the corresponding parameters $R, Q_i, Q_{iN}, P_i(n), \Delta_i(n)$ reduce to scalars $r, q_i, q_i^N, p_i(n),\delta_i(n)$. The result for ensuring consensus is reported in the following corollary. \[cor\_consensus\_1D\] For system in (\[equ\_one\_dimension\_system\]) with the network topology $\mathcal{G}$, assume that $r_i$ is designed as in (\[equ\_design\_R\]) and the $\mathcal{G}$ contains a panning tree. If $q_i$, $q_i^N$ and $r$ are designed such that $$\begin{aligned} & p_i(N) - p_i(N-1) >0,\label{equ_qi_qiN}\\ & |\alpha_i|^{N-1} q_i^N + \sum_{l = 1}^{N-2}|\alpha_i|^l q_i < \bar{\theta}_{min},\label{equ_qn_bound}\end{aligned}$$ then all the states of each agent can reach consensus. Here, $\alpha_i \triangleq \frac{r a}{r + b^2 p_i(1)}$, $\bar{\theta}_{min} = \min_{i}\{\theta_i\}$, $\theta_i = |ar|\frac{\sqrt{(1-a_{ci}^2)b_i^2 + (1-a_i)^2 a_{ci}^2}-(|1-a_i|a_c^2)}{(1-a_i)^2a_{ci}^2 + b_i^2}$, $a_{ci} = \frac{ar}{r + b^2 p_i(1)}$, $a_i = {\rm Re}(\lambda_i)$, $b_i = {\rm Imag}(\lambda_i)$, and $\lambda_i$ $i=2, \cdots, M$ are the eigenvalues of $\tilde{\mathcal{L}}$. By plugging $K_i = \frac{ab p_i(1)}{r + b^2 p_i(1)}$ and $G_i = \frac{b\delta_i(1)}{r + b^2 p_i(1)}$ into $A-B K_i + (\lambda_i -1) B G_i$, we have $$\begin{aligned} A-BK_i + (\lambda_i -1) B G_i = \frac{ar + (1-\lambda_i)b^2 \delta_i(1)}{r + b^2 p_i(1)}.\end{aligned}$$ To achieve consensus, we need to develop conditions to ensure that $|\frac{ar + (1-\lambda_i)b^2 \delta_i(1)}{r + b^2 p_i(1)}|<1$ for $\lambda_2, \cdots, \lambda_M$. To that need, we consider the term $T\triangleq |a r + (1-\lambda_i)b^2 \delta_i(1)|^2-|r + b^2 p_i(1)|^2$. Firstly, we derive an upper bound of $\delta_i(1)$. Define a sequence $\{\tilde{\delta}_i(n), n = N-1, \cdots, 0\}$ as $$\label{equ_tilde_delta} \tilde{\delta}_i(n) = \frac{ar\tilde{\delta}_i(n+1)}{r + b^2 p_i(1)} + q_i,$$ where $\tilde{\delta}_i(N) = q_{iN}$. Because of (\[equ\_qi\_qiN\]), Lemma \[lemma\_RDE\] can be used, implying $p_i(1)< \cdots < p_i(N)$. As a result, $|\tilde{\delta}_i(n)| \geqslant |\delta_i(n)|$. Therefore, a sufficient condition for guaranteeing $T<0$ is $$\begin{aligned} \label{equ_1D_inequality} (|ar| + |1-a_i| b^2 |\tilde{\delta}_i(1)|)^2 + b_i^2 b^2 |\tilde{\delta}_i(1)|^2 -(r + b^2 p_i(1))^2 < 0.\end{aligned}$$ By some algebraic operations, (\[equ\_1D\_inequality\]) reduces to $$\begin{aligned} \label{equ_1D_inequality_second} (|c_i|^2 + |d_i|^2) (\tilde{\delta}_i(1))^2 + 2 |a_c| |c_i| |\tilde{\delta}_i(1)| + |a_c|^2 - 1 <0,\end{aligned}$$ where $c_i = \frac{|1-a_i|b^2 |a_c|}{ar}$ and $d_i = \frac{b^2 b_i}{ar}$. On the other hand, from (\[equ\_tilde\_delta\]), we can obtain that $|\tilde{\delta}_i(1)|\leqslant |\alpha_i|^{N-1} q_{iN} + \sum_{l = 1}^{N-2}|\alpha_i|^l q_i$. Using the condition in (\[equ\_qn\_bound\]), we have $|\tilde{\delta}_i(1)|\leqslant \bar{\theta}_{min}$, which ensures (\[equ\_1D\_inequality\_second\]) holds. This completes the proof. Simulation Studies {#Sec_simulation} ================== In this section, two examples are provided to verify the theoretical results. One-dimensional Case -------------------- Consider a multi-agent system with $5$ agents, and the model for each agent is given as $$x_i(k+1) = 2 x_i(k) + u_i(k).$$ The adjacency matrix $\mathcal{A}$ of $\mathcal{G}$ is given as $\mathcal{A} = \left[ \begin{array}{ccccc} 0 & 1 & 0 & 0 & 1 \\ 0 & 0 & 1 & 0 & 0 \\ 0 & 0 & 0 & 1 & 0 \\ 1 & 0 & 1 & 0 & 0 \\ 1 & 1 & 1 & 0 & 0 \\ \end{array} \right] $. The Laplacian matrix $\tilde{\mathcal{L}}$ of $\tilde{\mathcal{G}}$ can be figured out as $\tilde{\mathcal{L}} = \left[ \begin{array}{ccccc} 1 & -\frac{1}{2} & 0 & 0 & -\frac{1}{2} \\ 0 & 1 & -1 & 0 & 0 \\ 0 & 0 & 1 & -1 & 0 \\ -\frac{1}{2} & 0 & -\frac{1}{2} & 1 & 0 \\ -\frac{1}{3} & -\frac{1}{3} & -\frac{1}{3} & 0 & 1 \\ \end{array} \right]$. It can be seen that $\mathcal{G}$ contains a panning tree. The parameters $q_i$, $q_{iN}$ are designed as $q_i=2$, and $q_{iN}=6$, $i=1,\cdots, 5$. The parameter $r$ is designed as $r = 1$, and thus, the corresponding $r_i$, $i=1,\cdots, 5$, are $r_1 = \frac{1}{2}$, $r_2 = 1$, $r_3 = 1$, $r_4 = \frac{1}{2}$, and $r_5 = \frac{1}{3}$. The horizon is chosen as $N=3$. By checking the condition in (\[equ\_qi\_qiN\]), we have $p_i(3)-p_i(2) = 0.5714>0$. Furthermore, $\bar{\theta}_{min}$ is calculated as $3.3391$, and the term $|\alpha_i|^{2} q_{iN} + |\alpha_i| q_i = 1.2186$. Thus, (\[equ\_qn\_bound\]) verifies. According to Corollary \[cor\_consensus\_1D\], the multi-agent system will achieve consensus. We plot the simulated trajectories of the five agents in Fig. \[fig\_1d\_state\], showing that the system can reach consensus. ![Trajectories for 5 agents with 1D system dynamics.[]{data-label="fig_1d_state"}](1D_state.eps){width="50.00000%"} General Linear Systems Case --------------------------- Consider a multi-agent system with $3$ agents, and each of a general LTI dynamics. The system matrices are $A = \left[ \begin{array}{cc} 2 & 0 \\ 1.2 & -1 \\ \end{array} \right]$, $B = \left[\begin{array}{c} 1 \\ 1 \\ \end{array} \right]$. The diagraph $\mathcal{G}$ contains a spanning tree, and it adjacency matrix $\mathcal{A}$ is $\mathcal{A} = \left[ \begin{array}{ccc} 0 & 1 & 1 \\ 0 & 0 & 1 \\ 1 & 1 & 0 \\ \end{array} \right]$. Thus, the corresponding matrix $\tilde{\mathcal{L}} = \left[ \begin{array}{ccc} 1 & -\frac{1}{2} & -\frac{1}{2} \\ 0 & 1 & -1 \\ -\frac{1}{2} & -\frac{1}{2} & 1 \\ \end{array} \right] $. The matrices $Q_i$ and $Q_{iN}$ are designed as $Q_i = {\rm \diag}(2, 2)$ and $Q_{iN} = {\rm \diag}(15, 20)$, $i=1, 2, 3$. The matrices $R_i$ are designed as $R_1 = \frac{1}{2}$, $R_2 = 1$ and $R_3 = \frac{1}{2}$, respectively, which satisfies (\[equ\_design\_R\]) with $R = 1$. The prediction horizon is chosen as $N = 10$. Calculate $P_i(10) - P_i(9) = \left[ \begin{array}{cc} 5.2 & -6 \\ -6 & 9.111 \\ \end{array} \right]>0 $. Thus, the condition in (\[equ\_monotically\_P\]) is verified. The eigenvalues of $\tilde{\mathcal{L}}$ are $\lambda_1 = 0$, $\lambda_2 = 1.5 + j0.5$, and $\lambda_3 = 1.5 - j0.5$. When $\lambda_2 = 1.5 \pm j0.5$, the left-hand side of (\[equ\_sufficient\_Q\]) is $[0.7045 -0.0621; -0.0621 0.7089]$, which is positive definite. Therefore, the condition in (\[equ\_sufficient\_Q\]) is satisfied. The simulated system trajectories are plotted in Fig. \[fig\_2d\_state1\] and Fig. \[fig\_2d\_state2\], respectively. From these two figures, it can be seen that the consensus is reached. ![Trajectories of state 1.[]{data-label="fig_2d_state1"}](2D_state1.eps){width="50.00000%"} ![Trajectories of state 2.[]{data-label="fig_2d_state2"}](2D_state2.eps){width="50.00000%"} CONCLUSIONS {#Sec_Conclusion} =========== In this paper, we have proposed a novel consensus scheme by using the distributed RHC for general LTI multi-agent systems. The necessary and sufficient conditions for ensuring consensus have been developed. Furthermore, we have developed more easily solvable conditions for multi-agent systems of general LTI and one-dimensional system dynamics, respectively. The developed theoretical results have been verified by two numerical studies. [^1]: This work was supported by the start-up research fund of the Northwestern Polytechnical University(NPU); the basic research foundation of NPU with grant no. GEKY 1004. [^2]: The authors are with the Department of Automation, School of Marine Science and Technology, Northwestern Polytechnical University, Xi’an, 710072, China (e-mail: [email protected];[email protected]).
{ "pile_set_name": "ArXiv" }
--- author: - | Carmine Abate^1^ Roberto Blanco^1^ Deepak Garg^2^ Cătălin Hriţcu^1^ Marco Patrignani^3,4^ Jérémy Thibault^1^\ ^1^Inria Paris ^2^MPI-SWS ^3^Stanford University ^4^CISPA Helmholz Center for Information Security\ - Carmine Abate - Roberto Blanco - Deepak Garg - Cătălin Hriţcu - Marco Patrignani - Jérémy Thibault bibliography: - 'local.bib' - 'mp.bib' - 'safe.bib' subtitle: Exploring Robust Property Preservation for Secure Compilation title: - | **[Journey Beyond Full Abstraction]{}\ Exploring Robust Property Preservation for Secure Compilation** - Journey Beyond Full Abstraction --- =1 input[texdirectives.tex]{} Robustly Preserving Hyperproperties {#sec:hyperprop} =================================== So far, we have studied the robust preservation of trace properties, which are properties of [*individual*]{} traces of a program. In this section we generalize this to [*hyperproperties*]{}, which are properties of [*multiple*]{} traces of a program [@ClarksonS10]. A well-known hyperproperty is noninterference, which has many variants [[@GoguenM82; @McLean92; @ZdancewicM03; @AskarovHSS08; @SabelfeldM03][@GoguenM82; @McLean92; @ZdancewicM03; @AskarovHSS08]]{}, but which usually requires considering two traces of a program that differ on secret inputs. Another hyperproperty is bounded mean response time over all executions. We study robust preservation of many subclasses of hyperproperties: all hyperproperties (), subset-closed hyperproperties (), hypersafety and $K$-hypersafety (), and hyperliveness (). These criteria are in the red area in . Robust Hyperproperty Preservation () {#sec:rhp} ------------------------------------ While trace properties are sets of traces, hyperproperties are sets of sets of traces [@ClarksonS10]. We call the set of traces of a whole program $W$ the [*behavior*]{} of $W$: $\behav{W} = \{ t ~|~ W \sem t \}$. A hyperproperty is a set of allowed behaviors. Program $W$ satisfies hyperproperty $H$ if the behavior of $W$ is a member of $H$, $\behav{W} \in H$, or, equivalently, $\{ t ~|~ W \sem t \} \in H$. Contrast this to $W$ satisfying trace property $\pi$, which holds if the behavior of $W$ is a subset of the set $\pi$, $\behav{W} \subseteq \pi$, or, equivalently, $\forall t.~ W \sem t \Rightarrow t \in \pi$. So while a trace property determines whether each individual trace of a program should be allowed or not, a hyperproperty determines whether the set of traces of a program, its behavior, should be allowed or not. For instance, the trace property $\pi_{123} = \{t_1, t_2, t_3\}$ is satisfied by programs with behaviors such as $\{t_1\}$, $\{t_2\}$, $\{t_2, t_3\}$, and $\{t_1, t_2, t_3\}$, but a program with behavior $\{t_1, t_4\}$ does not satisfy $\pi_{123}$. A hyperproperty like $H_{1+23} = \{ \{t_1\}, \{t_2, t_3\}\}$ is satisfied only by programs with behavior $\{t_1\}$ or with behavior $\{t_2, t_3\}$. A program with behavior $\{t_2\}$ does not satisfy $H_{1+23}$, so hyperproperties can express that if some traces ($t_2$) are possible then some other traces ($t_3$) should also be possible. A program with behavior $\{t_1,t_2,t_3\}$ also does not satisfy $H_{1+23}$, so hyperproperties can express that if some traces ($t_2$ and $t_3$) are possible then some other traces ($t_1$) should not be possible. Finally, trace properties can be easily lifted to hyperproperties: A trace property $\pi$ becomes the hyperproperty $[\pi] = 2^{\pi}$, the powerset of $\pi$. We say that a partial program $P$ [*robustly satisfies*]{} a hyperproperty $H$ if it satisfies $H$ for any context $C$. Given this we define as the preservation of robust satisfaction of arbitrary hyperproperties: $$\begin{aligned} \rhp:\quad \forall H \in 2^{2^\ii{Trace}}.~\forall\src{P}.~ & (\forall\src{C_S} \ldotp \src{\behav{C_S\hole{P}}} \in H) \Rightarrow \\& (\forall \trg{C_T} \ldotp \trg{\behav{C_T\hole{{{\ensuremath{\bl{\left.\src{P}\right\downarrow}}}}}}} \in H)\end{aligned}$$ The equivalent () property-free characterization of is not very surprising: $$\begin{aligned} \pf{\rhp}:\quad&\ \forall\src{P}.~ \forall\trg{C_T}.~ \exists \src{C_S}.~ \trg{\behav{C_T\hole{{{\ensuremath{\bl{\left.\src{P}\right\downarrow}}}}}}} = \src{\behav{C_S\hole{P}}} \\ \pf{\rhp}:\quad&\ \forall\src{P}.~ \forall\trg{C_T}.~ \exists \src{C_S}.~ \forall t.~ \trg{C_T\hole{{{\ensuremath{\bl{\left.\src{P}\right\downarrow}}}}} \sem} t \iff \src{C_S\hole{P}\sem} t\end{aligned}$$ This requires that, for every partial program $\src{P}$ and target context $\trg{C_T}$, there is a (back-translated) source context $\src{C_S}$ that perfectly preserves the set of traces of $\trg{C_T\hole{{{\ensuremath{\bl{\left.\src{P}\right\downarrow}}}}}}$ when linked to $\src{P}$. There are two differences from : (1) the $\exists \src{C_S}$ and $\forall t$ quantifiers are swapped, so the back-translated $\src{C_S}$ must work for all traces $t$, and (2) the implication in ($\Rightarrow$) became a two-way implication in ($\iff$), so compilation has to perfectly preserve the set of traces. In particular the compiler cannot refine behavior (remove traces), it cannot implement nondeterministic scheduling via a deterministic scheduler. In the following subsections we study restrictions of to various subclasses of hyperproperties. To prevent duplication we define $\rhp(X)$ to be the robust satisfaction of a class $X$ of hyperproperties (so above is simply $\rhp(2^{2^\ii{Trace}})$): $$\begin{aligned} \rhp(X):\quad \forall H \in X.~\forall\src{P}.~ &(\forall\src{C_S} \ldotp \src{\behav{C_S\hole{P}}} \in H) \Rightarrow\\ & (\forall \trg{C_T} \ldotp \trg{\behav{C_T\hole{{{\ensuremath{\bl{\left.\src{P}\right\downarrow}}}}}}} \in H)\end{aligned}$$ Robust Subset-Closed Hyperproperty Preservation () {#sec:rschp} -------------------------------------------------- If one restricts robust preservation to only subset-closed hyperproperties then refinement of behavior is again allowed. A hyperproperty $H$ is subset-closed, written $H {\in} \ii{SC}$, if for any two behaviors $b_1$ and $b_2$ so that$b_1 {\subseteq} b_2$, if $b_2 {\in} H$ then $b_1 {\in} H$. For instance, the lifting $[\pi]$ of any trace property $\pi$ is subset-closed, but the hyperproperty $H_{1+23}$ above is not. It can be made subset-closed by allowing all smaller behaviors: $H_{1+23}^\ii{SC} = \{\emptyset, \{t_1\}, \{t_2\}, \{t_3\}, \{t_2, t_3\}\}$ is subset-closed. [*Robust Subset-Closed Hyperproperty Preservation*]{} () is simply defined as $\rhp(\ii{SC})$. The equivalent () property-free characterization of simply gives up the $\Leftarrow$ direction of : $$\begin{multlined} \pf{\rschp}:\quad \forall\src{P}.~ \forall\trg{C_T}.~ \exists \src{C_S}.~ \forall t.~ \trg{C_T\hole{{{\ensuremath{\bl{\left.\src{P}\right\downarrow}}}}}} \mathrel{\trg{\sem}} t \Rightarrow \src{C_S\hole{P}} \mathrel{\src{\sem}} t \end{multlined}$$ The most interesting subclass of subset-closed hyperproperties is hypersafety, which we discuss in the next sub-sectionnext. The appendix ()also studies the subclass of $K$-subset-closed hyperproperties [@mastroeni2018verifying], which can be seen as generalizing $K$-hypersafety below. Robust Hypersafety Preservation () {#sec:rhsp} ---------------------------------- \[sec:rnip\] Hypersafety is a generalization of safety that is very important in practice, since several important notions of noninterference are hypersafety, such as termination-insensitive noninterference [@SabelfeldS01; @Fenton74; @AskarovHSS08], observational determinism [@McLean92; @ZdancewicM03; @Roscoe95], and nonmalleable information flow [@CecchettiMA17]. According to Alpern and Schneider [@AlpernS85], the “bad thing” that a safety property disallows must be [*finitely observable*]{} and [*irremediable*]{}. For safety the “bad thing” is a finite trace prefix that cannot be extended to any trace satisfying the safety property. For hypersafety, @ClarksonS10 generalize the “bad thing” to a finite set of finite trace prefixes that they call an [*observation*]{}, drawn from the set $\ii{Obs} = 2^\ii{FinPref}_\ii{Fin}$, which denotes the set of all finite subsets of finite prefixes. They then lift the prefix relation to sets: an observation $o \in \ii{Obs}$ is a prefix of a behavior $b \in 2^\ii{Trace}$, written $o {\leq} b$, if $\forall m \in o.~ \exists t \in b.~ m {\leq} t$. Finally, they define hypersafety analogously to safety, but the domains involved include an extra level of sets: $$\ii{Hypersafety} {\triangleq} \{ H ~|~ \forall b {\not\in} H.~ (\exists o {\in} \ii{Obs}.~ o {\leq} b \wedge (\forall b' {\geq} o.~ b' {\not\in} H)) \}$$ Here the “bad thing” is an observation $o$ that cannot be extended to a behavior $b'$ satisfying the hypersafety property $H$. We use this to define [*Robust Hypersafety Preservation*]{} () as $\rhp(\ii{Hypersafety})$ and propose the following equivalent () characterization for it: $$\begin{multlined} \pf{\rhsp}:~ \forall\src{P}.~ \forall\trg{C_T}.~ \forall o \in \ii{Obs}.~\\ \quad\qquad o \leq \trg{\behav{C_T\hole{{{\ensuremath{\bl{\left.\src{P}\right\downarrow}}}}}}} \Rightarrow \exists \src{C_S}.~ o \leq \src{\behav{C_S\hole{P}}} \end{multlined}$$ This says that to prove one needs to be able to back-translate a partial program $\src{P}$, a context $\trg{C_T}$, and a prefix $o$ of the behavior of $\trg{C_T\hole{{{\ensuremath{\bl{\left.\src{P}\right\downarrow}}}}}}$, to a source context $\src{C_S}$ so that the behavior of $\src{C_S\hole{P}}$ extends $o$. It is possible to use the finite set of finite executions corresponding to observation $o$ to drive this back-translation (as we do in ). For hypersafety the involved observations are finite sets but their cardinality is otherwise unrestricted. In practice though, most hypersafety properties can be falsified by very small sets: counterexamples to termination-insensitive noninterference [@SabelfeldS01; @Fenton74; @AskarovHSS08] and observational determinism [@McLean92; @ZdancewicM03; @Roscoe95] are observations containing 2 finite prefixes, while counterexamples to nonmalleable information flow [@CecchettiMA17] are observations containing 4 finite prefixes. To account for this, @ClarksonS10 introduce $K$-hypersafety as a restriction of hypersafety to observations of a fixed cardinality $K$. Given $\ii{Obs}_K = 2^\ii{FinPref}_\ii{Fin(K)}$, the set of observations with cardinality $K$, all definitions and results above can be ported to $K$-hypersafety by simply replacing $\ii{Obs}$ with $\ii{Obs}_K$. Specifically, we denote by $\criteria{\rkhsp}{rkhsp}$ the criterion $\rhp(\ii{\mbox{$K$-}Hypersafety})$. The set of lifted safety properties, $\{[\pi] ~|~ \pi \in \ii{Safety} \}$, is precisely the same as $1$-hypersafety, since the counterexample for them is a single finite prefix. For a more interesting example, termination-insensitive noninterference () [@AskarovHSS08; @SabelfeldS01; @Fenton74] can be defined as follows in our setting: $$\begin{aligned} \ii{TINI} \triangleq \{ b ~|~ \forall t_1~t_2 {\in} b.~ &(t_1~\ii{terminating} \land t_2~\ii{terminating} ~\\ &\land \ii{pub\text{-}inputs}(t_1){=}\ii{pub\text{-}inputs}(t_2)) \\ &\Rightarrow \ii{pub\text{-}events}(t_1){=}\ii{pub\text{-}events}(t_2) \}\end{aligned}$$ This definition requires that trace events are either inputs or outputs, each of them associated with a security level: public or secret. ensures that for any two terminating traces of the program behavior for which the two sequences of public inputs are the same, the two sequences of public events—inputs and outputs—are also the same. is $2$-hypersafety, since $b \not\in \ii{TINI}$ implies that there exist finite traces $t_1$ and $t_2$ that agree on the public inputs but not on all public events, so we can simply take $o = \{ t_1, t_2 \}$. Since the traces in $o$ are already terminated, any extension $b'$ of $o$ can only add extra traces, $\{ t_1, t_2 \} \subseteq b'$, so $b' \not\in \ii{TINI}$ as needed to conclude that is in $2$-hypersafety. In , we write [*Robust Termination-Insensitive Noninterference Preservation*]{} () for $\rhp(\{\ii{TINI}\})$. Separation Between Properties and Hyperproperties {#sec:hypersafety-separations} ------------------------------------------------- Enforcing is strictly more demanding than enforcing . Because even (robust 2-hypersafety preservation) implies , a compilation chain satisfying has to make sure that a target-level context cannot infer more information about the internal data of ${{\ensuremath{\bl{\left.\src{P}\right\downarrow}}}}$ than a source context could infer about the data of $\src{P}$. By contrast, a compilation chain can allow arbitrary [*reads*]{} of ${{\ensuremath{\bl{\left.\src{P}\right\downarrow}}}}$’s internal data, even if $\src{P}$’s data is private at the source level. Intuitively, for proving , the source context produced by back-translation can guess any secret ${{\ensuremath{\bl{\left.\src{P}\right\downarrow}}}}$ receives in the [ *single*]{} considered execution, but for the single source context needs to work for [*two*]{} different executions, potentially with two different secrets, so guessing is no longer an option. We use this idea to prove a more general separation result $\rtp \not\Rightarrow \rtinip$, by exhibiting a toy compilation chain in which private variables are readable in the target language, but not in source. \[thm:rtp-doesnt-imply-rtinip\] $\rtp \not\Rightarrow \rtinip$ This implies a strict separation between all criteria based on hyperproperties (the red area in , having as the bottom) and all the ones based on trace properties (the yellow area in having as the top). Using a more complex counterexample involving a system of $K$ linear equations, we have also shown that, for any $K$, robust preservation of $K$-hypersafety, does not imply robust preservation of $(K{+}1)$-hypersafety. \[thm:rkhsp-doesnt-imple-rkkhsp\] $\forall K.~ \rkhsp \not\Rightarrow \rkkhsp$ Where Is Robust Hyperliveness Preservation? {#sec:hyperliveness} ------------------------------------------- [*Robust Hyperliveness Preservation*]{} () does not appear in , because it is provably equivalent to (or, equivalently, ). We define as $\rhp(\ii{Hyperliveness})$ for the following standard definition of  [@ClarksonS10]: $$\ii{Hyperliveness} \triangleq \{ H ~|~ \forall o \in \ii{Obs}.~ \exists b {\geq} o.~ b \in H \}$$ The proof that implies () involves showing that $\{ b ~|~ b {\not=} \trg{\behav{C_T\hole{{{\ensuremath{\bl{\left.\src{P}\right\downarrow}}}}}}} \}$, the hyperproperty allowing all behaviors other than $\trg{\behav{C_T\hole{{{\ensuremath{\bl{\left.\src{P}\right\downarrow}}}}}}}$, is hyperliveness. Another way to obtain this result though is from the fact that, as in previous models [@AlpernS85], each hyperproperty can be decomposed as the intersection of two hyperliveness properties. This collapse of [*preserving*]{} hyperliveness and [*preserving*]{} all hyperproperties happens irrespective of the adversarial contexts. Related Work {#sec:related} ============ #### Full Abstraction was originally used as a criterion for secure compilation in the seminal work of @Abadi99 and has since received a lot of attention [@MarcosSurvey]. @Abadi99 and, later, @Kennedy06 identified failures of full abstraction in the Java to JVM and C\# to CIL compilers, some of which were fixed, but also others for which fixing was deemed too costly compared to the perceived practical security gain. @AbadiFG02 proved full abstraction of secure channel implementations using cryptography, but to prevent network traffic attacks they had to introduce noise in their translation, which in practice would consume network bandwidth. Ahmed  [[@AhmedB11; @Ahmed15; @NewBA16; @AhmedB08][@AhmedB11; @Ahmed15; @NewBA16]]{} proved the full abstraction of type-preserving compiler passes for simple functional languages. @AbadiP12 and @JagadeesanPRR11 expressed the protection provided by address space layout randomization as a probabilistic variant of full abstraction. @FournetSCDSL13 devised a fully abstract compiler from a subset of ML to JavaScript. Patrignani  [[@PatrignaniASJCP15; @LarmuseauPC15; @PatrignaniDP16][@PatrignaniASJCP15]]{} studied fully abstract compilation to machine code, starting from single modules written in simple, idealized object-oriented and functional languages and targeting a hardware isolation mechanism similar to Intel’s SGX [@sgx]. Until recently, most formal work on secure interoperability with linked target code was focused only on fully abstract compilation. The goal of our work is to explore a diverse set of secure compilation criteria, some of them formally stronger than (the interesting direction of) full abstraction at least in various determinate settings, and thus potentially harder to achieve and prove, some of them apparently easier to achieve and prove than full abstraction, but most of them not directly comparable to full abstraction. This exploration clarifies the trade-off between security guarantees and efficient enforcement for secure compilation: On one extreme, robustly preserves only trace properties, but does not require enforcing confidentiality; on the other extreme, robustly preserving relational properties gives very strong guarantees, but requires enforcing that both the private data and the code of a program remain hidden from the context, which is often much harder to achieve. The best criterion to apply depends on the application domain, but our framework can be used to address interesting design questions such as the following: [*(1) What secure compilation criterion, when violated, would the developers of practical compilers be willing to fix at least in principle?*]{} The work of @Kennedy06 indicates that fully abstract compilation is not such a good answer to this question, and we wonder whether or could be better answers. [*(2) What secure compilation criterion would the translations of @AbadiFG02 still satisfy if they did not introduce (inefficient) noise to prevent network traffic analysis?*]{} @AbadiFG02 explicitly leave this problem open in their paper, and we believe one answer could be , since it does not require preserving any confidentiality. We also hope that our work can help eliminate common misconceptions about the security guarantees provided (or not) by full abstraction. For instance, @FournetSCDSL13 illustrate the difficulty of achieving security for JavaScript code using a simple example policy that (1) restricts message sending to only correct URLs and (2) prevents leaking certain secret data. Then they go on to prove full abstraction apparently in the hope of preventing contexts from violating such policies. However, part (1) of this policy is a safety property and part (2) is hypersafety, and as we showed in fully abstract compilation does not imply the robust preservation of such properties. In contrast, proving would directly imply this, without putting any artificial restrictions on code introspection, which are unnecessarily required by full abstraction. Unfortunately, this is not the only work in the literature that uses full abstraction even when it is not the right hammer. Our exploration also forced us to challenge the assumptions and design decisions of prior work. This is most visible in our attempt to use as generic and realistic a trace model as possible. To start, this meant moving away from the standard assumption in the hyperproperties literature [@ClarksonS10] that all traces are infinite, and switching instead to a trace model inspired by CompCert’s [@Leroy09] with both terminating and non-terminating traces, and where non-terminating traces can be finite but not finitely observable (to account for silent divergence). This more realistic model required us to find a class of trace properties to replace liveness. #### Development of Two pieces of concurrent work have examined more carefully how to attain and prove one of the weakest of our criteria, (). @PatrignaniG18 show for compilers from simple sequential and concurrent languages to capabilities [@WatsonWNMACDDGL15]. They observe that if the source language has a verification system for robust safety and compilation is limited to verified programs, then can be established without directly resorting to back-translation. (This observation has also been made independently by Dave Swasey in private communication to us.) @AbateABEFHLPST18 aim at devising secure compilation chains for protecting mutually distrustful components written in an unsafe language like C. They show that by moving away from the full abstraction variant used in earlier work [@JuglaretHAEP16] to a variant of our criterion from , they can support a more realistic model of dynamic component compromise, while at the same time obtaining a criterion that is easier to achieve and prove than full abstraction. #### Hypersafety Preservation The high-level idea of specifying secure compilation as the preservation of properties and hyperproperties goes back to the work of @PatrignaniG17. However, that work’s technical development is limited to one criterion—the preservation of finite prefixes of program traces by compilation. Superficially, this is similar to one of our criteria, , but there are several differences even from . First, @PatrignaniG17 do not consider adversarial contexts explicitly. This might suffice for their setting of closed reactive programs, where traces are inherently fully abstract (so considering the adversarial context is irrelevant), but not in general. Second, they are interested in designing a criterion that accommodates specific fail-safe like mechanisms for low-level enforcement, so the preservation of hypersafety properties is not perfect, and one has to show, for every relevant property, that the criterion is meaningful. However, @PatrignaniG17 consider translations of trace symbols induced by compilation, an extension that would also be interesting for our criteria (as shown for in ). #### Proof techniques @NewBA16 present a back-translation technique based on a universal type embedding in the source for the purpose of proving full abstraction of translations from typed to untyped languages. In we adapted the same technique to show for a simple translation from a statically typed to a dynamically typed language with first-order functions and input-outputI/O. @DevriesePP16 show that even when a precise universal type does not exist in the source, one can use an approximate embedding that only works for a certain number of execution steps. They illustrate such an approximate back-translation by proving full abstraction for a compiler from the simply-typed to the untyped $\lambda$-calculus. introduced a “trace-based” back-translation technique. They were interested in proving full abstraction for so-called trace semantics. This technique was then adapted to show full abstraction of compilation chains to low-level target languages [[@PatrignaniASJCP15; @PatrignaniC15; @PatrignaniDP16; @AgtenSJP12][@PatrignaniASJCP15]]{}. In , we showed how these trace-based techniques can be extended to prove all the criteria below in , which includes robust preservation of safety, of noninterference, and in a determinate setting also of observational equivalence. While many other proof techniques have been previously proposed [[@AbadiFG02; @AhmedB11; @AbadiP12; @FournetSCDSL13; @JagadeesanPRR11; @AhmedB08][@AbadiFG02; @AhmedB11; @AbadiP12; @FournetSCDSL13; @JagadeesanPRR11]]{}, proofs of full abstraction remain notoriously difficult, even for simple translations, with apparently simple conjectures surviving for decades before being finally settled [@DevriesePP18]. It will be interesting to investigate which existing full abstraction techniques can be repurposed to show the stronger criteria from . For instance, it will be interesting to determine the strongest criterion from for which an approximate back-translation [@DevriesePP16] can be used. #### Source-level verification of robust satisfaction While this paper studies the *preservation* of robust properties in compilation chains, formally verifying that a partial source program robustly satisfies a specification is a challenging problem in itselftoo. So far, most of the research has focused on techniques for proving observational equivalence [[@JeffreyR05; @DelauneH17; @JeffreyR05b; @AbadiBF18; @ChevalKR18][@JeffreyR05; @DelauneH17]]{} or trace equivalence [@BaeldeDH17; @ChevalCD13]. Robust satisfaction of trace properties has been model checked for systems modeled by nondeterministic Moore machines and properties specified by branching temporal logic [@KupfermanV99]. Robust safety, the robust satisfaction of safety properties, was studied for the analysis of security protocols [[@GordonJ04; @Backes:Hritcu:Maffei:11; @Backes:Hritcu:Maffei:08bb][@GordonJ04]]{}, and more recently for compositional verification [@SwaseyGD17]. Verifying the [*robust*]{} satisfaction of relational hyperproperties beyond observational equivalence and trace equivalence seems to be an open research problem. For addressing it, one can hopefully take inspiration in extensions of relational Hoare logic [@benton04relational] for dealing with cryptographic adversaries represented as procedures parameterized by oracles [@BartheDGKSS13]. #### Other Kinds of Secure Compilation In this paper we investigated the various kinds of security guarantees one can obtain from a compilation chain that protects the compiled program against linked adversarial low-level code. While this is an instance of [*secure compilation*]{} [@dagstuhl-sc-2018], this emerging area is much broader. Since there are many ways in which a compilation chain can be “more secure”, there are also many different notions of secure compilation, with different security goals and attacker models. A class secure compilation chains is aimed at providing a “safer” semantics for unsafe low-level languages like C and C++, for instance ensuring memory safety [@NagarakatteMZ15; @cheri_asplos2015; @CheckedC]. Other secure compilation work is targeted at closing down side-channels: for instance by preserving the secret independence guarantees of the source code [@BartheGL18], or making sure that the code erasing secrets is not simply optimized away by the unaware compilers [@SimonCA18; @BessonDJ18; @DSilvaPS15; @DengN18]. Closer to our work is the work on building compartmentalizing compilation chains [@AbateABEFHLPST18; @BessonBDJW19; @WatsonWNMACDDGL15; @GudkaWACDLMNR15] for unsafe languages like C and C++. In particular, as mentioned above, @AbateABEFHLPST18 have recently showed how can be extended to express the security guarantees obtained by protecting mutually distrustful components against each other. #### Relation to provably correct compilers The closest related work is on “compositional compiler correctness”, but that work doesn’t do security as it assumes that the linked target-level context is well-behaved, for instance it behaves like a source-level context. Conclusion and Future Work {#sec:conclusion} ========================== \[sec:future\] This paper proposes a foundation for secure interoperability with linked target code by exploring many different criteria based on robust property preservation (). Yet the road to building [*practical*]{} secure compilation chains achieving any of these criteria remains long and challenging. Even for , scaling up to realistic programming languages and efficiently enforcing protection of the compiled program without restrictions on the linked context is challenging [@AbateABEFHLPST18; @PatrignaniG18]. For the problem is even harder, because one also needs to protect the secrecy of the program’s data, which is especially challenging in a realistic model in which the context can observe side-channels like timing. Here, an -like property might be the best one can hope for in practice. In this paper we assumed for simplicity that traces are exactly the same in both the source and target language, and while this assumption is currently true for other work like CompCert [@Leroy09] as well, it is a restriction nonetheless. We plan to lift this restriction in the future. In the appendix (), we illustrate how to lift this restriction for the , and in the future we hope to generalize all other criteria in a similar way. This paper made the simplifying assumption that the source and the target languages have the same trace model, and while this assumption is currently true for CompCert [@Leroy09], it is a restriction nonetheless. Fortunately, the criteria of this paper can be easily extended to take a relation between source and target traces as an extra parameter of the compilation chain. It is also easy to automatically lift this relation on traces to a relation on sets of traces, sets of sets of traces, etc. What is less obvious is whether this automatic lifting is what one always wants, and more importantly whether the users of a secure compilation chain will understand this relation between the properties they reason about at the source languages and the ones they get in the target language. We plan to investigate this in the future, building on recent work [@PatrignaniG17]. Finally, the relation between the criteria of and fully abstract compilation requires some further investigation. This paper identifies the sufficient conditions under which follows from some of these criteria, which gives us a one-way relation to full abstraction in the cases in which observational equivalence coincides with trace equivalence [@Engelfriet85; @ChevalCD13]. Even under this assumption, what is less clear is whether there are any sufficient conditions for full abstraction to imply any of the criteria of . The separation result of shows that compiler correctness, even when reasonably compositional (satisfying and ), is not enough. Yet the fact that fully abstract compilers often provide both the necessary enforcement mechanisms and the proof techniques to achieve even the highest criterion in (as illustrated in ) suggests that there is more to full abstraction than currently meets the eye. One possible additional assumption could be the “fail-safe behavior” of @PatrignaniG17, which should at least intuitively remove our counterexample from , but which seems much harder to formally state in our setting compared to their closed reactive programs setting. Nonetheless, investigating whether any extra assumptions make full abstraction imply our criteria is interesting future work. We are grateful to Akram El-Korashy, Arthur Azevedo de Amorim, Ştefan Ciobâcă, Dominique Devriese, Guido Martínez, Marco Stronati, Dave Swasey, Éric Tanter, and the anonymous reviewers for their valuable feedback and in many cases also for participating in various discussions. This work was in part supported by the [European Research CouncilERC](https://erc.europa.eu) under ERC Starting Grant SECOMP (715753), by the German Federal Ministry of Education and Research (BMBF) through funding for the CISPA-Stanford Center for Cybersecurity (FKZ: 13N1S0762), and by DARPA grant SSITH/HOPE (FA8650-15-C-7558). Proof Techniques for $\pf{\rrhp}_{\bowtie}$ and $\pf{\rfrxp}_{\bowtie}$ {#sec:proofs-appendix} ======================================================================= This section presents the formal details of . As explained in the main paper, we use two different proof techniques, one that is “context-based”, and the other “trace-based”, to prove two different security criteria for the same compilation chain. We argue that one of these techniques, the trace-based one, while less powerful, still gives us an interesting criterion, and should be more generic, as it relies less on the details of the languages. #### A remark on the security criteria used in this section In the languages used in this example, not all programs and contexts can be linked together. In order for it to be the case, they have to satisfy some interfacing constraints. Here, these constraints are the existence of functions called but not defined by the context, and, in the source language, also agreement on the types of these functions. We introduce the operators $\src{\bowtie}$ and $\trg{\bowtie}$ to represent these constraints. For instance, $\src{P\bowtie C}$ means that $\src{C}$ is linkable with $\src{P}$. We prove two () and (), named $\pf{\rrhp}_{\bowtie}$ () and $\pf{\rfrxp}_{\bowtie}$ (), that take into account these linkability predicates.
{ "pile_set_name": "ArXiv" }
--- abstract: 'The Bloch theorem is a powerful theorem stating that the expectation value of the U(1) current operator averaged over the entire space vanishes in large quantum systems. The theorem applies to the ground state and to the thermal equilibrium at a finite temperature, irrespective of the details of the Hamiltonian as far as all terms in the Hamiltonian are finite ranged. In this work we present a simple yet rigorous proof for general lattice models. For large but finite systems, we find that both the discussion and the conclusion are sensitive to the boundary condition one assumes: under the periodic boundary condition, one can only prove that the current expectation value is inversely proportional to the linear dimension of the system, while the current expectation value completely vanishes before taking the thermodynamic limit when the open boundary condition is imposed. We also provide simple tight-binding models that clarify the limitation of the theorem in dimensions higher than one.' author: - Haruki Watanabe title: A proof of the Bloch theorem for lattice models --- Introduction ============ The Bloch theorem [@PhysRev.75.502] states that the equilibrium state of a thermodynamically large system, in general, does not support non-vanishing expectation value of the averaged current density of any conserved U(1) charge, regardless of the details of the Hamiltonian such as the form of interactions or the size of the excitation gap. Despite its wide applications, the proof of the theorem in the existing literature is mostly for specific continuum models [@PhysRev.75.502; @doi:10.1143/JPSJ.65.3254; @PhysRevD.92.085011]. There are also proofs for lattice models [@PhysRevB.78.144404; @Tada2016; @Bachmann] but the setting considered in these works are not fully general. For example, Ref. [@PhysRevB.78.144404] is for a concrete spin model with a translation symmetry and the assumption of their discussion is unclear. Ref. [@Tada2016] assumes an extended-Hubbard type Hamiltonian and their definition of the current operator heavily relies on the specific form of the kinetic term. Finally, Ref. [@Bachmann] assumes the uniqueness of the ground state with non-vanishing excitation gap. In this work, we revisit the proof and clarify several confusing points about the Bloch theorem. We summarize the assumption and the statement of the theorem under the periodic boundary condition in Sec. \[sec:setup\] and give a proof for general models defined on a one-dimensional lattice in Sec. \[sec:proof\]. We discuss the theorem under the open boundary condition in Sec. \[sec:open\]. Finally we clarify the limitation of the theorem in higher dimensions in Sec. \[sec:highd\]. Bloch theorem under periodic boundary condition =============================================== Setup and statement {#sec:setup} ------------------- Let us consider a quantum many-body system defined on a one dimensional lattice. We impose the periodic boundary condition with system size $L$. The Hamiltonian $\hat{H}$ of the system can be very general. It may contain arbitrary hopping matrices and interactions as far as each term in the Hamiltonian is short-ranged (i.e., the size of its support is finite and does not scale with $L$) and respects the U(1) symmetry we discuss shortly. In particular, we *do not* put any restriction on the translation symmetry, the ground state degeneracy, or the excitation gap. Hence the result is applicable not only to periodic lattice with arbitrary number of sub-lattices but also, for example, to quasi-crystals or disordered systems. To simplify the notation we set the lattice constant to be $1$ and denote lattice sites by $x\in\{1,2,\cdots,L\}$. We assume that the Hamiltonian $\hat{H}$ commutes with the particle number operator $$\hat{N}=\sum_{x=1}^L\hat{n}_x.$$ Here, $\hat{n}_x$ is the local charge density operator at site $x$. We assume that density operators at different sites commute, $[\hat{n}_x,\hat{n}_{x'}]=0$. The U(1) symmetry implies the conservation law: $$i[\hat{H},\hat{n}_x]+\hat{j}_{x+\frac{1}{2}}-\hat{j}_{x-\frac{1}{2}}=0,\label{eq:cons}$$ where $\hat{j}_{x+\frac{1}{2}}$ is the local U(1) current operator that measures the net charge transfer across the ‘seam’ in between $x$ and $x+1$ \[Fig. \[fig\] (a)\]. We present the precise definition of $\hat{j}_{x+\frac{1}{2}}$ in Sec. \[subsec:current\]. With this setting, the Bloch theorem states that the ground state expectation value of the local current operator vanishes in the limit of the large system size: $$\lim_{L\rightarrow\infty}\langle\text{GS}|\hat{j}_{x+\frac{1}{2}}|\text{GS}\rangle=0.\label{statement1}$$ Here $|\text{GS}\rangle$ is the ground state of $\hat{H}$ with the energy eigenvalue $E_{\text{GS}}$. When there is a ground state degeneracy we arbitrary pick one state. The current conservation law in Eq. , together with $\hat{H}|\text{GS}\rangle=E_{\text{GS}}|\text{GS}\rangle$, implies $$\langle\text{GS}|\hat{j}_{x+\frac{1}{2}}|\text{GS}\rangle=\langle\text{GS}|\hat{j}_{x-\frac{1}{2}}|\text{GS}\rangle\label{constantj}$$ for all $x=1,2,\cdots,L$, meaning that the expectation value is independent of the position. Therefore, we can equally state the Bloch theorem in terms of the averaged current operator $$\begin{aligned} &\hat{\bar{j}}\equiv\frac{1}{L}\sum_{x=1}^L\hat{j}_{x+\frac{1}{2}},\label{eq:ac}\\ &\lim_{L\rightarrow\infty}\langle\text{GS}|\hat{\bar{j}}|\text{GS}\rangle=0.\label{statement12}\end{aligned}$$ This statement can be directly generalized to a finite temperature $T>0$ [@doi:10.1143/JPSJ.65.3254; @Tada2016] described by the Gibbs state (we set $k_B=1$): $$\begin{aligned} &\hat{\rho}_0\equiv \frac{1}{Z}e^{-\hat{H}/T},\quad Z\equiv\text{tr}\big(e^{-\hat{H}/T}\big),\label{Gibbs}\\ &\lim_{L\rightarrow\infty}\text{tr}\big(\hat{\rho}_0\hat{j}_{x+\frac{1}{2}}\big)=\lim_{L\rightarrow\infty}\text{tr}\big(\hat{\rho}_0\hat{\bar{j}}\big)=0.\label{statement2}\end{aligned}$$ Proof {#sec:proof} ----- ### Variational principle {#variation} Our proof of the theorem makes use of the twist operator introduced by Ref. [@Lieb1961], which reads $$\hat{U}_m\equiv e^{\frac{2\pi im}{L}\sum_{x=1}^Lx\hat{n}_x},\quad m\in\mathbb{Z}. \label{twistop}$$ This unitary operator is consistent with the periodic boundary condition since replacing $x$ with $x+L$ in the exponent does not affect $\hat{U}_m$ as $e^{2\pi im\hat{N}}=1$. The key observation of the proof is the following Taylor expansion in the power series of $L^{-1}$, which we show in Sec. \[subsec:current\]: $$\hat{U}_m^\dagger\hat{H}\hat{U}_m=\hat{H}+2\pi m \hat{\bar{j}}+O(L^{-1}).\label{eq:Taylor}$$ Taking the ground state expectation value of this equation, we find the following relation for the energy expectation value of the variational state $|\Phi_m\rangle\equiv\hat{U}_m|\text{GS}\rangle$: $$\langle\Phi_m|\hat{H}|\Phi_m\rangle=E_{\text{GS}}+2\pi m \langle \text{GS}|\hat{\bar{j}}|\text{GS}\rangle+O(L^{-1}).$$ Suppose first that $\langle \text{GS}|\hat{\bar{j}}|\text{GS}\rangle>0$. Then we find that $\langle\Phi_m|\hat{H}|\Phi_m\rangle$ with $m<0$ is lower than the ground state energy for a large $L$, which contradicts with the variational principle. If $\langle \text{GS}|\hat{\bar{j}}|\text{GS}\rangle<0$, $|\Phi_m\rangle$ with $m>0$ does the same job. Hence, $\langle \text{GS}|\hat{\bar{j}}|\text{GS}\rangle$ cannot remain nonzero as $L\rightarrow\infty$ and must be smaller than or equal to $O(L^{-1})$. The is the proof of Eq. , which also gives Eq.  with the help of Eq. . The above variational argument is common among the majority of proofs in the literature [@PhysRev.75.502; @PhysRev.75.502; @doi:10.1143/JPSJ.65.3254; @PhysRevD.92.085011; @PhysRevB.78.144404]. ### Finite temperature The proof of the Bloch theorem for a finite temperature is almost identical to that for the ground state. Given the Hamiltonian $\hat{H}$ and a density operator $\hat{\rho}$, in general, the free energy at $T>0$ is given by $$F(\hat{\rho})=\text{tr}\big(\hat{\rho}\hat{H}+T\hat{\rho}\ln \hat{\rho}\big).$$ This is minimized by the Gibbs state in Eq.  with the minimum value $F(\hat{\rho}_0)=-k_BT\ln Z$ [@PhysRev.106.620; @Sagawa]. Using Eq. , we find $$\begin{aligned} F(\hat{U}_m\hat{\rho}_0\hat{U}_m^\dagger)&=\text{tr}\left[\hat{\rho}_0(\hat{U}_m^\dagger\hat{H}\hat{U}_m)+T\hat{\rho}_0\ln \hat{\rho}_0\right]\notag\\ &=F(\hat{\rho}_0)-2\pi m\,\text{tr}\big(\hat{\rho}_0\hat{\bar{j}}\big)+O(L^{-1}).\end{aligned}$$ If the magnitude of the current expectation value is bigger than $O(L^{-1})$, we get $F(\hat{\rho}_m)<F(\hat{\rho}_0)$ for either $m=\pm1$, which is a contradiction. ### Local current operator {#subsec:current} ![\[fig\] (a) Twisted boundary condition with the U(1) phase $e^{-i\theta_{\bar{x}}}$ at the seam $\bar{x}=x+\frac{1}{2}$. (b) Generalized twisted boundary condition with a seam at every $\bar{x}$ for $x=1,2,\cdots,L$. ](fig1.pdf){width="1.0\columnwidth"} It remains to verify Eq. . This requires a precise formulation of the local current operator. To this end, let us temporary introduce the *twisted* boundary condition. We place the position of the ‘seam’ to be somewhere in between $x$ and $x+1$, which we denote by $\bar{x}\equiv x+\frac{1}{2}$ \[Fig. \[fig\] (a)\]. Let $\theta_{\bar{x}}$ be the angle of the twist. Later we will set $\theta_{\bar{x}}=0$, as, after all, we are interested in the original system under the periodic boundary condition. The Hamiltonian $\hat{H}^{\theta_{\bar{x}}}$ under the twisted boundary condition has $\theta_{\bar{x}}$-dependence localized around the seam. This is because every term in the original Hamiltonian $\hat{H}$ that goes across the seam acquires a phase $e^{i\ell_{\bar{x}}\theta_{\bar{x}}}$. For example, a hopping term $tc_{x+1}^\dagger c_x+\text{h.c.}$ becomes $te^{-i\theta_{\bar{x}}}c_{x+1}^\dagger c_x+\text{h.c.}$, while a pair-hopping term $t\hat{c}_{x+1,\uparrow}^\dagger\hat{c}_{x+1,\downarrow}^\dagger\hat{c}_{x,\downarrow}\hat{c}_{x,\uparrow}+\text{h.c.}$ becomes $te^{-2i\theta_{\bar{x}}}\hat{c}_{x+1,\uparrow}^\dagger\hat{c}_{x+1,\downarrow}^\dagger\hat{c}_{x,\downarrow}\hat{c}_{x,\uparrow}+\text{h.c.}$ More generally, a term in $\hat{H}$ is multiplied by the factor $e^{i(n_{\bar{x}}^a-n_{\bar{x}}^c)\theta_{\bar{x}}}$ where $n_{\bar{x}}^a$ ($n_{\bar{x}}^c$) is the number of annihilation (creation) operators in the right side of the seam in the term. Other terms in $\hat{H}$ that reside either one side of the seam remain unchanged. The local current operator across the seam under the *periodic* boundary condition is given by $$\hat{j}_{\bar{x}}\equiv\partial_{\theta_{\bar{x}}}\hat{H}^{\theta_{\bar{x}}}\big|_{\theta_{\bar{x}}=0}.\label{eq:lc}$$ The current operator defined this way satisfies the conservation law in Eq. . To see this explicitly, let us introduce a seam for every $\bar{x}=x+\frac{1}{2}$ ($x=1,2,\cdots,L$) and denote the twisted Hamiltonian by $\hat{H}^{(\theta_{\bar{1}},\theta_{\bar{2}},\cdots,\theta_{\bar{L}})}$ \[Fig. \[fig\] (b)\]. It satisfies $$\begin{aligned} &\hat{H}=\hat{H}^{(\theta_{\bar{1}},\theta_{\bar{2}},\cdots,\theta_{\bar{L}})}\big|_{\theta_{\bar{1}}=\theta_{\bar{2}}=\cdots=\theta_{\bar{L}}=0},\\ &\hat{j}_{\bar{x}}=\partial_{\theta_{\bar{x}}}\hat{H}^{(\theta_{\bar{1}},\theta_{\bar{2}},\cdots,\theta_{\bar{L}})}\big|_{\theta_{\bar{1}}=\theta_{\bar{2}}=\cdots=\theta_{\bar{L}}=0}\end{aligned}$$ and $$\begin{aligned} &e^{i\epsilon \hat{n}_{x}}\hat{H}^{(\theta_{\bar{1}},\theta_{\bar{2}},\cdots,\theta_{\bar{L}})}e^{-i\epsilon \hat{n}_{x}}\notag\\ &=\hat{H}^{(\theta_{\bar{1}},\cdots,\theta_{\bar{x}-2},\theta_{\bar{x}-1}-\epsilon,\theta_{\bar{x}}+\epsilon,\theta_{\bar{x}+1},\cdots,\theta_{\bar{L}})}.\label{eq:Htr}\end{aligned}$$ This relation implies that $\theta_{\bar{x}}$ can be identified with the background U(1) gauge field $A_{\bar{x}}$. When Eq.  for $\theta_{\bar{1}}=\theta_{\bar{2}}=\cdots=\theta_{\bar{L}}=0$ is expanded in the power series of $\epsilon$, the $O(\epsilon)$-term reproduces the conservation law . It also follows by using Eq.  repeatedly that $$\hat{U}_m^\dagger\hat{H}\hat{U}_m=\hat{H}^{(\frac{2\pi m}{L},\cdots,\frac{2\pi m}{L})}.\label{v1}$$ The Taylor series of the right-hand side reads $$\hat{H}^{(\frac{2\pi m}{L},\cdots,\frac{2\pi m}{L})}=\sum_{\ell=0}^\infty\tfrac{1}{\ell!}(\tfrac{2\pi m}{L})^\ell\hat{H}^{(\ell)},$$ where $\hat{H}^{(\ell)}$ ($\ell=0,1,2,\cdots$) is defined by $$\sum_{x_1,x_2,\cdots,x_\ell=1}^L\partial_{\theta_{\bar{x}_1}}\partial_{\theta_{\bar{x}_2}}\cdots\partial_{\theta_{\bar{x}_\ell}}\hat{H}^{(\theta_{\bar{1}},\cdots,\theta_{\bar{L}})}\Big|_{\theta_{\bar{1}}=\cdots=\theta_{\bar{L}}=0}.\label{v2}$$ For example, $\hat{H}^{(0)}=\hat{H}$ and $$\hat{H}^{(1)}=\sum_{x=1}^L\partial_{\theta_{\bar{x}}}\hat{H}^{(\theta_{\bar{1}},\cdots,\theta_{\bar{L}})}\Big|_{\theta_{\bar{1}}=\cdots=\theta_{\bar{L}}=0}=L\hat{\bar{j}}.\label{v3}$$ For short-ranged Hamiltonians, each $\hat{H}^{(\ell)}$ is at most the order of $L$ at least for $\ell=O(L^0)$. Eqs. – altogether verify Eq.  and the proof is completed. Discussions {#sec:extention} ----------- Here let us make some remarks on the theorem. ### Relation to the Lieb-Schultz-Mattis theorem The conclusion in Sec. \[variation\] implies that the variational state $|\Phi_m\rangle=\hat{U}_m|\text{GS}\rangle$ is a low-energy state whose excitation energy $|\langle\Phi_m|\hat{H}|\Phi_m\rangle-E_{\text{GS}}|$ is bounded by $O(L^{-1})$. Further assuming the translation symmetry $\hat{T}_1$ with $\hat{T}_1\hat{n}_x\hat{T}_1^\dagger=\hat{n}_{x+1}$, we find [@Lieb1961; @Affleck1986; @PhysRevLett.79.1110] $$\hat{T}_1\hat{U}_m\hat{T}_1^\dagger=\hat{U}_me^{-2\pi m i\frac{\hat{N}}{L}}.$$ Suppose that the ground state $|\text{GS}\rangle$ is an eigenstate of $\hat{T}_1$ and $\hat{N}$ and that the filling fraction $\nu\equiv \langle\text{GS}|\hat{N}|\text{GS}\rangle/L$ is not an integer. Then the variational state $|\Phi_m\rangle$ and the ground state $|\text{GS}\rangle$ have inequivalent eigenvalues of $\hat{T}_1$ and hence are orthogonal to each other. This implies the Lieb-Schultz-Mattis theorem for translation invariant one-dimensional systems which suggests the presence of either a ground state degeneracy or a gapless excitation when $\nu\notin\mathbb{Z}$ [@Lieb1961; @Affleck1986; @PhysRevLett.79.1110]. Note that we did not assume any additional symmetry such as the spatial inversion or the time-reversal symmetry unlike the original argument [@Lieb1961; @Affleck1986; @PhysRevLett.79.1110]. In the context of the Lieb-Schultz-Mattis theorem, it appears that Ref. [@Koma2000] removed such an assumption for first time by using the variational argument, but as we have seen here this logic was old and can be traced back to the original work of the Bloch theorem [@PhysRev.75.502]. ![\[fig2\] (a) The band structure of the tight-binding model for $L=12$ and $\theta_0=0.9\pi/L$. Blue (white) dots represents occupied (unoccupied) states at $\mu=-2t\cos(\frac{\pi N}{L})$ with $N=5$. (b) The current expectation value of the ground state as a function of $\theta_0$ for $L=12$ and $N=5$. ](fig2.pdf){width="1.0\columnwidth"} ### Persistent current in a finite system {#subsec:persistent} The Bloch theorem allows a persistent current of the order $O(L^{-1})$ in a finite system. For a later purpose, let us consider a concrete tight-binding model with the nearest neighbor hopping $t>0$. $$\hat{H}=-te^{-i\theta_0}\sum_{x=1}^L\hat{c}_{x+1}^\dagger \hat{c}_x+\text{h.c.}$$ We introduced a phase $e^{-i\theta_0}$ to break the time-reversal symmetry. Introducing the Fourier transform $\hat{c}_{k}^\dagger=\frac{1}{\sqrt{L}}\sum_{x=1}^L\hat{c}_{x}^\dagger e^{i k x}$ for $k=\frac{2\pi q}{L}$ ($q=1,2,\cdots L$), we find [@PhysRevB.37.6050] $$\hat{H}=\sum_k \varepsilon_k \hat{c}_k^\dagger \hat{c}_k,\quad\hat{\bar{j}}=\frac{1}{L}\sum_k\partial_k\varepsilon_k \hat{c}_k^\dagger \hat{c}_k$$ with $\varepsilon_k=-2t\cos(k+\theta_0)$ \[Fig. \[fig2\] (a)\]. For concreteness, let us set the Fermi energy to be $\mu=-2t\cos(\frac{\pi N}{L})$ for an odd number of particles $N$, and consider the Fermi sea $|\text{GS}\rangle=\prod_{k, \varepsilon_k<\mu}c_k^\dagger|0\rangle$ \[Fig. \[fig2\] (a)\]. The current expectation value $\langle \text{GS}|\hat{\bar{j}}|\text{GS}\rangle=\frac{1}{L}\sum_{k, \varepsilon_k<\mu}\partial_k\varepsilon_k$ shows the periodicity in $\theta_0$ with the period $2\pi/L$ \[Fig. \[fig2\] (b)\]. Its maximum value is given by $$j_{\text{max}}\equiv\lim_{\theta_0\uparrow\frac{\pi}{L}}\langle \text{GS}|\hat{\bar{j}}|\text{GS}\rangle=\tfrac{2t}{L}\sin(\tfrac{\pi N}{L})=O(L^{-1}). \label{eq:jmax}$$ These results are consistent with the previous studies, for example, in Ref. [@PhysRevB.37.6050]. ### Extensions #### Continuum models Continuum models can be treated simply by replacing $\sum_{x=1}^L$ with $\int_0^Ldx$, for example. For continuum models, the Noether theorem provides the definition of the conserved U(1) current. The key relation Eq.  remains unchanged. #### Long-range interactions The assumption on the range of hopping matrices and interactions can be slightly relaxed. They are not necessarily strictly finite-ranged. For example, any term $\hat{o}$ that does not depend on $\theta_{\bar{x}}$ (i.e. $\hat{U}_m^\dagger\hat{o}\hat{U}_m=\hat{o}$) can be safely added. This class includes the density-density interactions such as the Coulomb interaction among electrons. Terms with a long-range tail that depend on $\theta_{\bar{x}}$ are also allowed as long as the order estimate of the series expansion in Eq.  is preserved. In addition to exponentially decaying terms, power-low decaying terms with a large enough exponent can also be added. The minimum value of the power depends on the detailed form of the term. For instance, in the case of the hopping term $$t\sum_{n=1}^Ln^{-\alpha}\sum_{x=1}^L\hat{c}_{x+n}^\dagger \hat{c}_x+\text{h.c.},$$ the power $\alpha$ must be greater than $3$. \[When $2<\alpha\leq3$, the Bloch theorem still holds but the upper bound of the current expectation value decays slower than $O(L^{-1})$.\] #### Other conserved current The argument in this work coherently applies to any conserved current density associated with an internal U(1) symmetry. For example, when the $z$-component of the total spin is conserved in a spin model with spin $S$ ($=1/2, 1, 3/2. \cdots$) on each site, we can set $\hat{n}_x=\hat{S}_x^z-S$ [@PhysRevLett.78.1984] in our discussion above to prove the absence of the equilibrium spin current. However, our argument is not applicable, for example, to the energy current density as there does not exist the corresponding twist operator. Recently, a completely new argument for the energy current has been developed in Ref. [@Anton]. Bloch theorem under open boundary condition {#sec:open} =========================================== When we impose the open boundary condition instead of the periodic boundary condition, we can actually prove a stronger statement: $$\langle\text{GS}|\hat{\bar{j}}|\text{GS}\rangle=\text{tr}\big(\hat{\rho}_0\hat{\bar{j}}\big)=0.\label{openbc}$$ Unlike the case with the periodic boundary condition, we do not have to take the large $L$ limit. To see this, note that the position operator $$\hat{P}=\frac{1}{L}\sum_{x=1}^Lx\hat{n}_x$$ is well defined under the open boundary condition. This operator is the one in the exponent of the twist operator  and is also known as the polarization operator [@PhysRevX.8.021065]. Using Eq.  repeatedly, we find $$[i\hat{H},\hat{P}]=-\hat{j}_{L+\frac{1}{2}}+\frac{1}{L}\sum_{x=1}^L\hat{j}_{x-\frac{1}{2}}.$$ Because the boundary is open and the current cannot leak out from or flow into the system, we have $\hat{j}_{L+\frac{1}{2}}=\hat{j}_{\frac{1}{2}}=0$ so that $$[i\hat{H},\hat{P}]=\hat{\bar{j}},$$ where $\hat{\bar{j}}$ is the averaged current operator given in Eq. . This is the many-body version of the relation $\hat{v}=i[\hat{H},\hat{x}]$ in the single-particle quantum mechanics. Eq.  follows immediately by combining this expression of $\hat{\bar{j}}$ with $\hat{H}|\text{GS}\rangle=E_{\text{GS}}|\text{GS}\rangle$ and $[\hat{H},\hat{\rho}_0]=0$. Higher dimensions {#sec:highd} ================= Models in higher dimensions can be treated by reducing them to one dimension either by compactifying all other directions by the periodic boundary condition or by taking a finite-width strip with the open boundary condition [@PhysRevD.92.085011; @Tada2016; @Bachmann; @PhysRevB.37.6050]. All quantities then contain an additional summation over transverse directions. We still impose the periodic boundary condition in $x$. Below we provide more details for two spatial dimensions. Two dimensions -------------- We define the current operator in the same way as we did in Sec. \[subsec:current\]. When the twisted boundary condition is introduced at $\bar{x}$, all terms in the Hamiltonian across the seam \[the black line in Fig. \[fig3\] (a)\] acquires the phase $e^{i\ell_{\bar{x}}\theta_{\bar{x}}}$ as before. $$\partial_{\theta_{\bar{x}}}\hat{H}^{\theta_{\bar{x}}}\big|_{\theta_{\bar{x}}=0}=\sum_{y=1}^{L_y}\hat{j}_{(\bar{x},y)}^x.$$ Correspondingly, Eqs. , and become $$\begin{aligned} &\hat{\bar{j}}^x\equiv\frac{1}{L_xL_y}\sum_{x=1}^{L_x}\sum_{y=1}^{L_y}\hat{j}_{(\bar{x},y)}^x,\\ &\hat{U}_m\equiv e^{\frac{2\pi im}{L_x}\sum_{x=1}^{L_x}\sum_{y=1}^{L_y}x\hat{n}_{(x,y)}},\\ &\hat{U}_m^\dagger\hat{H}\hat{U}_m=\hat{H}+2\pi m L_y\hat{\bar{j}}^x+O(L_x^{-1}L_y).\end{aligned}$$ In this case, the Bloch theorem states that the expectation value of the averaged current density vanishes in the ground state or in the thermal equilibrium at $T>0$: $$\lim_{L_x\rightarrow\infty}\langle\text{GS}|\hat{\bar{j}}^x|\text{GS}\rangle=\lim_{L_x\rightarrow\infty}\text{tr}\big(\hat{\rho}_0\hat{\bar{j}}^x\big)=0.$$ It is important to note, however, that both of the following quantities may have a non-vanishing expectation value in the limit of large $L_x$. 1. The total current integrated over transverse directions ($L_y\hat{\bar{j}}^x$ in two dimensions). 2. The local current density ($\hat{j}_{(\bar{x},y)}^x$ in two dimensions). ![\[fig3\] (a) Two dimensional counterpart of Fig. \[fig\]. (b) Example of an insulator with a nonzero bulk current. (c) Example of an insulator with a nonzero edge current. ](fig3.pdf){width="1.0\columnwidth"} Bulk current ------------ An example of the first case is given by the $L_y$ copies of decoupled 1D chains described by the tight-binding model in Sec. \[subsec:persistent\]. As each chain supports a persistent current of $O(L_x^{-1})$, we find $$L_yj_{\text{max}}^x=\tfrac{L_y}{L_x}2t\sin(\tfrac{\pi N}{L_xL_y})=O(L_y/L_x).$$ For example, when $L_y=L_x$, this is an $O(1)$ quantity which does not vanish in the large $L_x$ limit. Edge current ------------ To provide an example of the second case, let us consider a two-dimensional periodic array of decoupled 1D rings and impose the open boundary condition in $y$ \[Fig. \[fig3\] (b)\]. Each 1D ring is formed by the tight-binding model considered in Sec. \[subsec:persistent\] with $L=4$ and $N=1$. Every ring supports a loop current $$j_{\text{loop}}=\tfrac{t}{2}\sin\phi_0,\quad \phi_0\in(-\tfrac{\pi}{4},\tfrac{\pi}{4}).$$ This is an $O(1)$ quantity, independent of $L_x$ or $L_y$. In the bulk, contributions from neighboring loops cancel and the local current density vanishes after proper coarse-graining [@jackson1999]. However, at the edge, there is a residual contribution that flows along the edge as illustrated in Fig. \[fig3\] (b) implying the nonzero expectation value of $\hat{j}_{(\bar{x},y)}^x$ at the edge. This is nothing but the magnetization current $\vec{\nabla}\times\vec{m}$ originating from the orbital magnetization $\vec{m}$ produced by the loop currents [@jackson1999]. Of course, a similar situation occurs for Chern insulators but our model is simpler as gapless chiral edge modes are absent. Conclusion ========== In this work we clarified the actual assumption of the Bloch theorem, which was unclear in the discussions in the literature and is turns out to be just the U(1) symmetry and the locality of the Hamiltonian. We gave a simple proof for the most general version of the theorem. Since the theorem holds regardless of the details of the states, as far as they are in the ground state or in a thermal equilibrium, the same upper bound on the persistent current applies, for example, to superconductors [@Tada2016]. We also clarified the difference of the statement under periodic and open boundary conditions: in the periodic case the current density can be the order of $L^{-1}$, in contrast to the case with the open boundary condition where the current expectation value identically vanishes without taking the thermodynamic limit. Finally we discussed a few illuminating tight-binding models in which (i) the net current flow integrated over the transverse direction or (ii) the local current density near the boundary does not vanish even in the limit of the large system size. These non-vanishing current expectation values do not contradict with the general theorem. I would like to thank S. Bachmann, A. Kapustin, and L. Trifunovic for useful discussions on this topic. In particular, I learned the generalization to the Gibbs state from the authors of Ref. [@Bachmann] and I am indebted to their email correspondence. I also thank T. Momoi for informing us of Ref. [@PhysRevB.78.144404]. This work is supported by JST PRESTO Grant No. JPMJPR18LA. [17]{}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 [****, ()](\doibase 10.1103/PhysRev.75.502) [****,  ()](\doibase 10.1143/JPSJ.65.3254) [****,  ()](\doibase 10.1103/PhysRevD.92.085011) [****,  ()](\doibase 10.1103/PhysRevB.78.144404) [****,  ()](\doibase 10.1007/s10955-016-1629-2) @noop [ ()]{} [****,  ()](\doibase 10.1016/0003-4916(61)90115-4) [****, ()](\doibase 10.1103/PhysRev.106.620) “,” in [**](\doibase 10.1142/9789814425193_0003), pp.  [****, ()](\doibase 10.1007/BF00400304) [****,  ()](\doibase 10.1103/PhysRevLett.79.1110) [****,  ()](\doibase 10.1023/A:1018604925491) [****,  ()](\doibase 10.1103/PhysRevB.37.6050) [****,  ()](\doibase 10.1103/PhysRevLett.78.1984) @noop [ ()]{} [****,  ()](\doibase 10.1103/PhysRevX.8.021065) [**](http://cdsweb.cern.ch/record/490457),  ed. (, , )
{ "pile_set_name": "ArXiv" }
--- abstract: 'We perform an analytical and numerical study of the phase transitions in three-dimensional $Z(N)$ lattice gauge theories at finite temperature for $N>4$ exploiting equivalence of these models with a generalized version of the two-dimensional vector Potts models in the limit of vanishing spatial coupling. In this limit the Polyakov loops play the role of $Z(N)$ spins. The effective couplings of these two-dimensional spin models are calculated explicitly. It is argued that the effective spin models have two phase transitions of BKT type. This is confirmed by large-scale Monte Carlo simulations. Using a cluster algorithm we locate the position of the critical points and study the critical behavior across both phase transitions in details. In particular, we determine various critical indices, compute the helicity modulus, the average action and the specific heat. A scaling formula for the critical points with $N$ is proposed.' --- [**Phase transitions in strongly coupled three-dimensional $Z(N)$ lattice gauge theories at finite temperature**]{} 0.3cm O. Borisenko$^{1\dagger}$, V. Chelnokov$^{1*}$, G. Cortese$^{2\dagger\dagger}$, R. Fiore$^{3\P}$, M. Gravina$^{4\ddagger}$, A. Papa$^{3\P}$, I. Surzhikov$^{1**}$ 0.6cm ${}^1$ *Bogolyubov Institute for Theoretical Physics,* *National Academy of Sciences of Ukraine,* *03680 Kiev, Ukraine* 0.2cm ${}^2$ *Instituto de Física Teórica UAM/CSIC,* *Cantoblanco, E-28049 Madrid, Spain* *and Departamento de Física Teórica,* *Universidad de Zaragoza, E-50009 Zaragoza, Spain* 0.2cm ${}^3$ *Dipartimento di Fisica, Università della Calabria,* *and Istituto Nazionale di Fisica Nucleare, Gruppo collegato di Cosenza* *I-87036 Arcavacata di Rende, Cosenza, Italy* 0.2cm ${}^4$ *Department of Physics, University of Cyprus, P.O. Box 20357, Nicosia, Cyprus* 0.6cm ------------------------------------------------------------------------ : $^\[email protected],   $^*[email protected],   $^{\dagger\dagger}[email protected],   $^{\P}$fiore, papa  @cs.infn.it, $^{\ddagger}[email protected],  $^{**}$i\_van\[email protected] Introduction ============ The phase structure of three-dimensional ($3d$) pure $Z(N)$ lattice gauge theories (LGTs) has been the subject of an intensive study for more than three decades. It is well known by now that the zero-temperature models possess a single phase transition which disappears in the limit $N\to\infty$ [@bhanot]. Thus, the $U(1)$ LGT has a single confined phase in agreement with theoretical results [@3d_u1]. The deconfinement phase transition at finite temperature is well understood and studied for $N=2,3$. These models belong to the universality class of $2d$ $Z(N)$ spin models and exhibit a second order phase transition in agreement with the Svetitsky-Yaffe conjecture [@svetitsky]. Much less is known about the finite-temperature deconfinement transition when $N>4$. The Svetitsky-Yaffe conjecture is known to connect critical properties of $3d$ $Z(N)$ LGTs with the corresponding properties of $2d$ spin models, if they share the same global symmetry of the action. It is widely expected, and in many cases proved by either analytical or numerical methods, that some $2d$ $Z(N>4)$ spin models (like the vector Potts model) have two phase transitions of infinite order, known as the Berezinskii-Kosterlitz-Thouless (BKT) phase transitions. Then, according to the conjecture, the phase transitions in some $3d$ $Z(N>4)$ gauge models at finite temperature could exhibit two phase transitions as well. Moreover, if the correlation length diverges when approaching the critical point, these transitions should be of the BKT type and belong to the universality class of the corresponding $2d$ $Z(N)$ spin models. The BKT phase transition is known to take place in a variety of $2d$ systems: certain spin models, $2d$ Coulomb gas, sine-Gordon model, Solid-on-Solid model, etc. The most elaborated case is the $2d$ $XY$ model [@berezin; @kosterlitz1; @kosterlitz2]. There are several indications that this type of phase transition is not a rare phenomenon in gauge models at finite temperature - one can argue that in some $3d$ lattice gauge models the deconfinement phase transition is of BKT type as well. A well known example is the deconfinement phase transition in the $U(1)$ LGT. Indeed, certain analytical [@svetitsky; @parga; @lat_07] as well as numerical results [@3du1ft; @3du1full] unambiguously indicate the BKT nature of the phase transition[^1]. Many details of the critical behavior of $2d$ $Z(N)$ spin models are well known – see the review in Ref. [@Wu]. The $Z(N)$ spin model in the Villain formulation has been studied analytically in Refs. [@elitzur; @savit; @kogut; @nienhuis; @kadanoff; @Cardy]. It was shown that the model has at least two phase transitions when $N\geq 5$. The intermediate phase is a massless phase with power-like decay of the correlation function. The critical index $\eta$ has been estimated both from the renormalization group (RG) approach of the Kosterlitz-Thouless type and from the weak coupling series for the susceptibility. It turns out that $\eta(\beta^{(1)}_{\rm c})=1/4$ at the transition point from the strong coupling (high-temperature) phase to the massless phase, [*i.e.*]{} the behavior is similar to that of the $XY$ model. At the transition point $\beta^{(2)}_{\rm c}$ from the massless phase to the ordered low-temperature phase one has $\eta(\beta^{(2)}_{\rm c})=4/N^2$. A rigorous proof that the BKT phase transition does take place, and so that the massless phase exists, has been constructed in Ref. [@rigbkt] for both Villain and standard formulations of the vector Potts model. Universality properties of vector Potts models were studied via Monte Carlo simulations in Ref. [@cluster2d] for $N=6,8,12$ and in Refs. [@lat_10; @2dzn; @lat_11; @2dzn7_17] for $N=5,7,17$. Results for the critical indices $\eta$ and $\nu$ agree well with the analytical predictions obtained from the Villain formulation of the model. We expect that $3d$ $Z(N)$ gauge models at finite temperature exhibit the same critical properties as $Z(N)$ spin models in two dimensions. In particular, gauge models with $N>4$ may possess two phase transitions of the BKT type. On the basis of the Svetitsky-Yaffe conjecture the critical behavior of the gauge model in this case is governed by the $2d$ $Z(N)$ spin model. In particular, one expects the following values of critical indices: $\nu = 1/2,\ \eta = 1/4$ at the first transition and $\nu = 1/2,\ \eta = 4/N^2$ at the second transition. To the best of our knowledge this scenario was not verified in the literature by either analytical or numerical means. The main goal of the present paper is to fill this gap and to study the nature of deconfining phase transitions in $3d$ $Z(N)$ LGTs. The fact that the BKT transition has infinite order makes it hard to study its properties using analytical methods. In most of the cases studied one uses a renormalization group (RG) technique like in Ref. [@elitzur]. Unfortunately, there are no direct ways to generalize transformations of Ref. [@elitzur], leading to RG equations, to $3d$ $Z(N)$ LGTs except for the limiting case $N\to\infty$. To study the phase structure of these models we need numerical simulations. Here, however, another problem appears related to the very slow, logarithmic convergence to the thermodynamic limit in the vicinity of the BKT transition. It is thus necessary to use both large-scale simulations and combine them with the finite-size scaling methods. For a full finite-temperature gauge model this is an ambitious program, especially if one wants to study several values of $N$. We have therefore decided to utilize an approach developed by some of the present authors in Refs. [@3du1ft; @3du1full] and to divide the whole investigation into two steps. The finite-temperature model is formulated on anisotropic lattice with different spatial and temporal couplings. Here, following [@3du1ft], as a first step, we study the limit of vanishing spatial coupling. Since this approximation does not affect the global symmetry properties of the model, we hope that in this limit the model belongs to the same universality class as the full theory. It is well known that for both three- and four-dimensional gauge models at finite temperature this is indeed the case, at least for $N=2,3,4$. Furthermore, in this limit the gauge model can be mapped onto a generalized $Z(N)$ spin model in $2d$. Thus, this investigation sheds some light on the details of the critical behavior of the general Potts model, [*i.e.*]{} beyond the vector Potts model, which is usually considered in the literature. In the present paper we study the phase transitions in models with $N=5,7$ in great details and in addition we locate critical points of the models with $N=9,13$. Our computations are performed on lattices with temporal extent $N_t=2,4$ and with spatial size in the range $L\in [128-2048]$. This paper is organized as follows. In Section 2 we formulate our model and study some of its properties analytically. In particular, we establish the exact relation with a generalized $2d$ $Z(N)$ spin model and discuss some of the RG predictions regarding the critical behavior. Also, we give simple analytical estimates for the critical couplings. In Section 3 we present the setup of Monte Carlo simulations, define the observables used in this work and present the numerical results of simulations for $N=5,7$. In Section 4 we present results for the critical points in the models with $N=9,13$ and discuss the scaling of the critical points with $N$. Our conclusions and perspectives are given in Section 5. Analytical considerations ========================= Relation of the $3d$ $Z(N)$ LGT to a generalized $2d$ $Z(N)$ vector model ------------------------------------------------------------------------- We work on a $3d$ lattice $\Lambda = L^2\times N_t$ with spatial extension $L$ and temporal extension $N_t$; $\vec{x}=(x_0,x_1,x_2)$, where $x_0\in [0,N_t-1]$ and $x_1,x_2\in [0,L-1]$ denote the sites of the lattice and $e_n$, $n=0,1,2$, denotes a unit vector in the $n$-th direction. Periodic boundary conditions (BC) on gauge fields are imposed in all directions. The notations $p_t$ ($p_s$) stand for the temporal (spatial) plaquettes, $l_t$ ($l_s$) for the temporal (spatial) links. We introduce conventional plaquette angles $s(p)$ as $$s(p) \ = \ s_n(x) + s_m(x+e_n) - s_n(x+e_m) - s_m(x) \ . \label{plaqangle}$$ The $3d$ $Z(N)$ gauge theory on an anisotropic lattice can generally be defined as $$Z(\Lambda ;\beta_t,\beta_s;N) \ = \ \prod_{l\in \Lambda} \left ( \frac{1}{N} \sum_{s(l)=0}^{N-1} \right ) \ \prod_{p_s} Q(s(p_s)) \ \prod_{p_t} Q(s(p_t)) \ . \label{PTdef}$$ The most general $Z(N)$-invariant Boltzmann weight with $N-1$ different couplings is $$Q(s) \ = \ \exp \left [ \sum_{k=1}^{N-1} \beta_p(k) \cos\frac{2\pi k}{N}s \right ] \ . \label{Qpgen}$$ The Wilson action corresponds to the choice $\beta_p(1)=\beta_p$, $\beta_p(k)=0, k=2,...,N-1$. The $U(1)$ gauge model is defined as the limit $N\to\infty$ of the above expressions. To study the phase structure of $3d$ $Z(N)$ LGTs in the strong coupling limit ($\beta_s = 0$) one can map the gauge model to a generalized $2d$ spin $Z(N)$ model with the action $$\label{modaction} S \ =\ \sum_{x}\ \sum_{n=1}^2 \sum_{k = 1}^{N-1} \ \beta_k \ \cos \left( \frac{2 \pi k}{N} \left(s(x) - s(x+e_n) \right) \right) \ .$$ The effective coupling constants $\beta_k$ are derived from the coupling constant $\beta_t \equiv \beta$ of the $Z(N)$ LGT, using the following equation (the Wilson action is used for the gauge model): $$\beta_k \ =\ \frac{1}{N} \sum_{p = 0}^{N - 1} \ln(Q_p) \cos \left(\frac{2 \pi p k}{N} \right) \ , \label{couplings}$$ where $$\begin{aligned} \label{Qk} Q_k&\ =\ &\sum_{p = 0}^{N - 1} \ \left(\frac{B_p}{B_0}\right)^{N_t} \ \cos \left(\frac{2 \pi p k}{N} \right) \ , \\ B_k&\ =\ &\sum_{p = 0}^{N - 1} \exp \left[ \beta \cos \left(\frac{2 \pi p}{N} \right) \right ] \cos \left(\frac{2 \pi p k}{N} \right) \ . \label{couplings_coeff}\end{aligned}$$ These equations can be easily obtained in a few steps following similar computations for $3D$ $SU(N)$ model in the same limit $\beta_s=0$ (see, for example, Ref. [@eff_action] and references therein): - Fourier expansion of the original Boltzmann weight; - Integration over spatial gauge fields; this leads to an effective $2d$ model for the Polyakov loops with the Boltzmann weight $Q(s(x)-s(x+e_n))$ defined in (\[Qk\]); - Exponentiation and re-expansion in a new Fourier series. Renormalization group and critical behavior ------------------------------------------- To gain some information on the critical behavior of the gauge model at $\beta_s=0$, one can perform the RG study of the theory. Let us consider the $2d$ model obtained after integration over spatial gauge fields, $$Z(\Lambda ;\beta; N) \ = \ \prod_{x\in \Lambda} \left ( \frac{1}{N} \sum_{s(x)=0}^{N-1} \right ) \ \prod_{l} \ \left [ \sum_{k = 0}^{N - 1} \ \left(B_k \right )^{N_t} \ \cos \left(\frac{2 \pi k}{N} (s(x) - s(x+e_n)) \right) \right ] \ . \label{PF_orig}$$ The coefficients $B_k\equiv B_k(\beta)$ are given in (\[couplings\_coeff\]) and can be represented as $$B_k \ = \ \sum_{r=-\infty}^{\infty} \ I_{Nr+k}(\beta) \ . \label{Bk_exp}$$ Here, $I_k(x)$ is the modified Bessel function. The spin variables $s(x)$ can be associated with the Polyakov loops of the original model. The RG equations can be obtained only for the Villain formulation of the model. Replacing the Bessel function with its asymptotics and using the Poisson summation formula, one finds $$\begin{aligned} Z(\Lambda ;\beta; N) \ &=& \ \prod_{x\in \Lambda} \left ( \frac{1}{N} \sum_{s(x)=0}^{N-1} \right ) \\ &\times& \prod_{l} \left [ \sum_{m=-\infty}^{\infty} \exp \left [ -\frac{1}{2}\beta_{\rm eff} \left ( \frac{2\pi }{N}(s(x) - s(x+e_n))+2\pi m \right )^2 \right ] \right ] \nonumber \ . \label{PF_villain}\end{aligned}$$ This is nothing but the Villain formulation of $2d$ $Z(N)$ vector model with $\beta_{\rm eff}$ defined as $$\beta_{\rm eff} \ = \ \beta/N_t \ . \label{b_eff}$$ The RG equations for the model (\[PF\_villain\]) have been constructed in [@elitzur]. Their analysis has been performed by us in a recent work [@2dzn7_17]. Therefore, all the conclusions as to the critical behavior remain valid in the present case. We shortly list the main results below: 1. The critical RG trajectories in the planes $(\beta_{\rm eff},y)$ and $(\beta_{\rm eff},z)$, where $y$, $z$ are the activities of the vortex configurations, coincide with those of the $2d$ $Z(N)$ vector spin model; 2. The critical index $\nu$ has been computed numerically from the RG equations for a variety of $N$. It takes on the value 1/2 for all $N$ considered, in particular for $N=5,7$; 3. The calculation of the two-point correlation function reveals that the index $\eta$ is equal to 1/4 at the first transition point, while $\eta=4/N^2$ at the second critical point [@elitzur]. The same behavior is valid for the correlation of the Polyakov loops in our model; 4. In [@2dzn7_17] we have calculated the dependence of the critical points on $N$. These scaling formulae are expected to hold in the present case for any fixed $N_t$ and will be the subject of the discussion in Section 4. Estimation of the critical points --------------------------------- The location of the critical points in $3d$ $Z(N)$ LGTs for $N>4$ is unknown. Therefore, before presenting numerical results, it is instructive to give some simple analytical predictions for the critical values of $\beta^{\rm crit}$ for different values $N_t$. Such a prediction could serve then as the starting point for the numerical search of the critical points. Following [@3du1ft], such critical values can be easily estimated if one knows $\beta^{\rm crit}$ for $N_t=1$. Since the model with $N_t=1$ coincides with the $2d$ $Z(N)$ model, approximate critical points for other values of $N_t$ can be computed from the equality $$\frac{B_1}{B_0}(\beta^{\rm crit}) \ = \ \left [ \frac{B_1}{B_0}(x)\right]^{N_t} \ , \label{crpointest}$$ where $x$ on the right-hand side denotes the unknown critical point. Solving the last equation numerically one finds $x$. As for the critical points $\beta^{\rm crit}$ at $N_t=1$, we use our previous estimates from [@2dzn]. As it will be seen below, the predicted values are in a reasonable agreement with the numerical results. Table \[tbl:bpc\_pred\] summarizes approximate analytical predictions for the critical couplings at the two transitions, denoted by $\beta_{\rm c}^{(1)}$ and $\beta_{\rm c}^{(2)}$, respectively; the last two columns show Monte Carlo values. $N$ $N_t$ $\beta_{\rm c}^{(1)}$ $\beta_{\rm c}^{(2)}$ ${\beta_{\rm c}^{(1)}}_{\rm MC}$ ${\beta_{\rm c}^{(2)}}_{\rm MC}$ ----- ------- ----------------------- ----------------------- ---------------------------------- ---------------------------------- 5 1 - - 1.051(1) 1.105(1) 5 2 1.8393 1.9057 1.87(1) 1.940(7) 5 4 2.7761 2.8515 2.813(3) 2.898(4) 7 1 - - 1.111(1) 1.88(8) 7 2 1.9861 3.1303 2.031(7) 3.366(7) 7 4 3.2995 4.9669 3.406(8) 5.158(7) : Values of $\beta_{\rm c}^{(1)}$ and $\beta_{\rm c}^{(2)}$ expected for $N_t = 1,\ 2,\ 4$ in $Z(N)$ with $N = 5,\ 7$. \[tbl:bpc\_pred\] Numerical results ================= Setup of the Monte Carlo simulation ----------------------------------- To study the phase transitions we used the cluster algorithm described in [@2dzn]. The model is studied on a square $L \times L$ lattice $\Lambda$ with periodic BC. Simulations were carried on for $N_t = 2,\ 4$, but could easily be done also for other values of $N_t$, since the parameter $N_t$ appears only in the definition of the couplings (\[couplings\]). As original action of the gauge model we used the conventional Wilson action. For each Monte Carlo run the typical number of generated configurations was $10^6$, the first $5\times10^4$ of them being discarded to ensure thermalization. Measurements were taken after 10 updatings and the uncertainty on primary observables was estimated by the jackknife method combined with binning. We considered the following observables: - complex magnetization $M_L = |M_L| e^{i \psi}$, $$\label{complex_magnetization} M_L \ =\ \sum_{x \in \Lambda} \exp \left( \frac{2 \pi i}{N} s(x) \right) \;;$$ - population $S_L$, $$\label{population} S_L \ =\ \frac{N}{N - 1} \left(\frac{\max_{i = 0, N - 1} n_i} {L^2} - \frac{1}{N} \right)\;,$$ where $n_i$ is number of $s(x)$ equal to $i$; - real part of the rotated magnetization $M_R = |M_L| \cos(N \psi)$ and normalized rotated magnetization $m_\psi = \cos(N \psi)$; - susceptibilities of $M_L$, $S_L$ and $M_R$: $\chi_L^{(M)}$, $\chi_L^{(S)}$, $\chi_L^{(M_R)}$, $$\label{susceptibilities} \chi_L^{(\mathbf\cdot)} \ =\ L^2 \left(\left< \mathbf\cdot^2 \right> - \left< \mathbf\cdot \right>^2 \right)\;;$$ - Binder cumulants $U_L^{(M)}$ and $B_4^{(M_R)}$, $$\begin{aligned} U_L^{(M)}&\ =\ &1 - \frac{\left\langle \left| M_L \right| ^ 4 \right\rangle}{3 \left\langle \left| M_L \right| ^ 2 \right\rangle^2}\;, \nonumber \\ \label{binderU} B_4^{(M_R)}&\ =\ & \frac{\left\langle \left| M_R - \left\langle M_R \right\rangle \right| ^ 4 \right\rangle} {\left\langle \left| M_R - \left\langle M_R \right\rangle \right| ^ 2 \right\rangle ^ 2 }\;; \label{binderBMR}\end{aligned}$$ - helicity modulus $\Upsilon$, $$\Upsilon \ =\ \left \langle e \right \rangle - L^2 \beta \left \langle s^2 \right \rangle \ , \label{helicity_modulus}$$ where $$e \equiv \frac{1}{L^{2}} \sum_{ { \left \langle ij \right \rangle }_{x}} \cos{\left(\theta_{i}-\theta_{j} \right)} \ ,\ s\equiv\frac{1}{L^{2}} \sum_{ {\left \langle ij \right \rangle}_{x}} \sin{(\theta_{i} -\theta_{j})}\ ,\ \theta_i \equiv \frac{2\pi}{N} s(i)$$ and the notation $ {\left \langle ij \right \rangle}_{x} $ means nearest-neighbors spins in the $x$-direction. Determination of the critical couplings --------------------------------------- A clear indication of the three-phase structure emerges from the inspection of the scatter plot of the complex magnetization $M_L$ at different values of $\beta$: as we move from low to high $\beta$, we observe the transition from a disordered phase (uniform distribution around zero) through an intermediate phase (ring distribution) up to the ordered phase ($N$ isolated spots), as Fig. \[fig:scatter\] shows for the case of $Z(5)$ on a $16^2\times 2$ lattice. ![Scatter plot of the complex magnetization $M_L$ at $\beta$=1.6, 1.9 and 2.5 in $Z(5)$ on a $16^2\times 2$ lattice.[]{data-label="fig:scatter"}](plot16.eps "fig:"){width="32.00000%"} ![Scatter plot of the complex magnetization $M_L$ at $\beta$=1.6, 1.9 and 2.5 in $Z(5)$ on a $16^2\times 2$ lattice.[]{data-label="fig:scatter"}](plot19.eps "fig:"){width="32.00000%"} ![Scatter plot of the complex magnetization $M_L$ at $\beta$=1.6, 1.9 and 2.5 in $Z(5)$ on a $16^2\times 2$ lattice.[]{data-label="fig:scatter"}](plot25.eps "fig:"){width="32.00000%"} The first and most important numerical task is to determine the value of the two critical couplings in the thermodynamic limit, $\beta_c^{(1)}$ and $\beta_c^{(2)}$, that separate the three phases. To this aim we have adopted several methods, which we list here: - Methods for the determination of $\beta_{\rm c}^{(1)}$: \(a) locate the position $\beta_{\rm pc}^{(1)}(L)$ of the peak of the susceptibility $\chi_L^{(M)}$ of the complex magnetization $|M_L|$ on lattices with various spatial size and find $\beta_{\rm c}^{(1)}$ by a fit with the following scaling function, dictated by the essential scaling: $$\beta^{(1)}_{\rm pc}=\beta^{(1)}_{\rm c} +\frac{A}{(\ln L + B)^{\frac{1}{\nu}}}\quad , \label{b_pc}$$ taking $\nu$ equal to 1/2; \(b) estimate the crossing point of the curves giving the behavior of the Binder cumulant $U_L^{(M)}$ versus $\beta$ on lattices with different spatial size $L$ or, alternatively, search for the value of $\beta_{\rm c}^{(1)}$ which optimizes the overlap of these curves when they are plotted against $(\beta-\beta_{\rm c}^{(1)})(\ln L)^{1/\nu}$, with $\nu$ fixed at 1/2; \(c) consider the helicity modulus $\Upsilon$ near the phase transition and define $\beta_{\rm pc}^{(1)}(L)$ as the value of $\beta$ such that $\eta(\beta) \equiv 1/(2 \pi \beta \Upsilon)=1/4$ on the lattice with spatial size $L$ [@YO91], then find $\beta_{\rm c}^{(1)}$ by a fit with the scaling function $$\beta^{(1)}_{\rm pc}=\beta^{(1)}_{\rm c} + \frac{A}{\ln L + B}\;, \label{helicity_scaling}$$ valid under the assumption that the phase transition belongs to the $XY$ universality class. - Methods for the determination of $\beta_{\rm c}^{(2)}$: \(d) same as the method (a) using instead the susceptibility $\chi_L^{(S)}$ of the population $S_L$; \(e) same as the method (b) using instead simultaneously the Binder cumulant $B_4^{(M_R)}$ and the order parameter $m_\psi$. As an illustration of the methods (a) and (d), we show in Figs. \[chi\_ML\_Z7\_Nt2\] and \[chi\_SL\_Z7\_Nt2\] the behavior of $M_L$ (and of its susceptibility) and that of $S_L$ (and of its susceptibility) versus $\beta$ in $Z(7)$ on lattices with $N_t=2$ and $L$ ranging from 128 to 2048. In Table \[peaks\] we summarize all the values of $\beta_{\rm pc}^{(1)}(L)$ and $\beta_{\rm pc}^{(2)}(L)$ found in this work for the application of methods (a) and (d) in $Z(N)$ with $N=5,7$ for $N_t=2,4$. ![Behavior of $M_L$ and of its susceptibility versus $\beta$ in $Z(7)$ on lattices with $N_t=2$ and $L$ ranging from 128 to 2048.[]{data-label="chi_ML_Z7_Nt2"}](ML_vs_beta.eps "fig:"){width="45.00000%"} ![Behavior of $M_L$ and of its susceptibility versus $\beta$ in $Z(7)$ on lattices with $N_t=2$ and $L$ ranging from 128 to 2048.[]{data-label="chi_ML_Z7_Nt2"}](CHIML_vs_beta.eps "fig:"){width="45.00000%"} ![Behavior of $S_L$ and of its susceptibility versus $\beta$ in $Z(7)$ on lattices with $N_t=2$ and $L$ ranging from 128 to 2048.[]{data-label="chi_SL_Z7_Nt2"}](SL_vs_beta.eps "fig:"){width="45.00000%"} ![Behavior of $S_L$ and of its susceptibility versus $\beta$ in $Z(7)$ on lattices with $N_t=2$ and $L$ ranging from 128 to 2048.[]{data-label="chi_SL_Z7_Nt2"}](CHISL_vs_beta.eps "fig:"){width="45.00000%"} $N$ $N_t$ $L$ $\beta_{\rm pc}^{(1)}$ $\beta_{\rm pc}^{(2)}$ ----- ------- ------ ------------------------ ------------------------ 128 1.7656(1) 1.9773(5) 192 1.7808(1) 1.9740(3) 256 1.7910(1) 1.9713(4) 5 2 384 1.80124(7) 1.9689(5) 512 1.80774(6) 1.9658(3) 768 1.81540(5) 1.9626(2) 1024 1.82013(4) 1.9614(4) 2048 1.83001(5) 16 2.4913(9) 32 2.5928(6) 64 2.6538(5) 128 2.692(1) 2.9376(8) 5 4 192 2.7131(8) 2.934(2) 256 2.7226(7) 2.928(1) 384 2.7357(7) 2.927(2) 512 2.921(2) 768 2.920(3) 1024 2.917(2) 128 1.8644(4) 3.461(2) 192 1.8873(2) 3.454(1) 256 1.9011(2) 3.443(2) 7 2 384 1.9184(2) 3.433(1) 512 1.9289(2) 3.427(2) 768 1.9421(2) 3.420(2) 1024 1.9504(5) 3.416(2) 128 3.14(1) 5.257(1) 192 3.18(1) 5.242(2) 256 3.20(1) 5.2371(7) 7 4 384 3.2220(2) 5.223(2) 512 3.2392(1) 5.216(2) 768 3.26024(6) 5.209(2) 1024 3.2737(3) 5.198(2) : Summary of all the determinations of $\beta_{\rm pc}^{(1)}$ and $\beta_{\rm pc}^{(2)}$ in $Z(N)$ on lattices with size $L^2\times N_t$. \[peaks\] As an illustration of the method (b), we show in Fig. \[UL\_Z7\_Nt2\] the behavior of $U^{(M)}_L$ versus $\beta$ in $Z(7)$ on lattices with $N_t=2$ and $L$ ranging from 128 to 2048. Similarly, as an illustration of the method (e), we show in Figs. \[B4\_Z5\_Nt4\] and \[mpsi\_Z5\_Nt4\] the behavior of $B_4^{(M_R)}$ and of $m_\psi$ versus $\beta$ in $Z(5)$ on lattices with $N_t=4$ and $L$ ranging from 128 to 2048. ![Binder cumulant $U_L^{(M)}$ as function of $\beta$ (left) and of $(\beta-\beta_c) (\ln L)^{1/\nu}$ (right) in $Z(7)$ on lattices with $N_t=2$ and $L$ ranging from 128 to 2048.[]{data-label="UL_Z7_Nt2"}](UML_vs_beta.eps "fig:"){width="45.00000%"} ![Binder cumulant $U_L^{(M)}$ as function of $\beta$ (left) and of $(\beta-\beta_c) (\ln L)^{1/\nu}$ (right) in $Z(7)$ on lattices with $N_t=2$ and $L$ ranging from 128 to 2048.[]{data-label="UL_Z7_Nt2"}](UML_vs_logL.eps "fig:"){width="45.00000%"} ![Binder cumulant $B_4^{(M_R)}$ as a function of $\beta$ (left) and of $(\beta-\beta_{\rm c}) (\ln L)^{1/\nu}$ (right) in $Z(5)$ on lattices with $N_t=4$ and $L$ ranging from 128 to 2048.[]{data-label="B4_Z5_Nt4"}](_Z5_NT4_B4MR_vs_beta.eps "fig:"){width="45.00000%"} ![Binder cumulant $B_4^{(M_R)}$ as a function of $\beta$ (left) and of $(\beta-\beta_{\rm c}) (\ln L)^{1/\nu}$ (right) in $Z(5)$ on lattices with $N_t=4$ and $L$ ranging from 128 to 2048.[]{data-label="B4_Z5_Nt4"}](_Z5_NT4_B4MR_vs_logL.eps "fig:"){width="45.00000%"} ![$m_{\psi}$ as a function of $\beta$ (left) and of $(\beta-\beta_{\rm c}) (\ln L)^{1/\nu}$ (right) in $Z(5)$ on lattices with $N_t=4$ and $L$ ranging from 128 to 2048.[]{data-label="mpsi_Z5_Nt4"}](_Z5_NT4_MPSI_vs_beta.eps "fig:"){width="45.00000%"} ![$m_{\psi}$ as a function of $\beta$ (left) and of $(\beta-\beta_{\rm c}) (\ln L)^{1/\nu}$ (right) in $Z(5)$ on lattices with $N_t=4$ and $L$ ranging from 128 to 2048.[]{data-label="mpsi_Z5_Nt4"}](_Z5_NT4_MPSI_vs_logL.eps "fig:"){width="45.00000%"} As an illustration of the method (c), we show in Fig. \[helicity\] the behavior of the helicity modulus $\Upsilon$ versus $\beta$ along with the line $\Upsilon = 1/(2\pi\beta\eta),\ \eta = 1/4$, describing pseudocritical points, in $Z(5)$ and in $Z(7)$ on lattices with $N_t=2,4$ and several values of the spatial extension $L$. For larger values of $L$ only the points around the intersection were simulated. ![Helicity modulus as a function of $\beta$ in $Z(5)$ with $N_t=2$ (top left), in $Z(5)$ with $N_t=4$ (top right), in $Z(7)$ with $N_t=2$ (bottom left) and in $Z(7)$ with $N_t=4$ (bottom right).[]{data-label="helicity"}](_Z5_NT2_HELICITY.eps "fig:"){width="45.00000%"} ![Helicity modulus as a function of $\beta$ in $Z(5)$ with $N_t=2$ (top left), in $Z(5)$ with $N_t=4$ (top right), in $Z(7)$ with $N_t=2$ (bottom left) and in $Z(7)$ with $N_t=4$ (bottom right).[]{data-label="helicity"}](_Z5_NT4_HELICITY.eps "fig:"){width="45.00000%"} ![Helicity modulus as a function of $\beta$ in $Z(5)$ with $N_t=2$ (top left), in $Z(5)$ with $N_t=4$ (top right), in $Z(7)$ with $N_t=2$ (bottom left) and in $Z(7)$ with $N_t=4$ (bottom right).[]{data-label="helicity"}](_Z7_NT2_HELICITY.eps "fig:"){width="45.00000%"} ![Helicity modulus as a function of $\beta$ in $Z(5)$ with $N_t=2$ (top left), in $Z(5)$ with $N_t=4$ (top right), in $Z(7)$ with $N_t=2$ (bottom left) and in $Z(7)$ with $N_t=4$ (bottom right).[]{data-label="helicity"}](_Z7_NT4_HELICITY.eps "fig:"){width="45.00000%"} Finally, we report in Table \[crit\_betas\] the determinations of the critical couplings $\beta_{\rm c}^{(1)}$ and $\beta_{\rm c}^{(2)}$ in $Z(N)$ with $N$=5 and 7 for $N_t$=2 and 4, specifying the adopted method. $N$ $N_t$ $\beta_{\rm c}^{(1)}$ method $\beta_{\rm c}^{(2)}$ method ----- ------- ----------------------- -------- ----------------------- -------- 1.878(2) a 1.91(6) d 5 2 1.87(1) b 1.940(7) e 1.8801(7) c 2.832(3) a 2.88(3) d 5 4 2.813(3) b 2.898(4) e 2.829(1) c 2.070(2) a 3.31(9) d 7 2 2.031(7) b 3.366(7) e 2.069(1) c 3.47(1) a 5.0(2) d 7 4 3.406(8) b 5.158(7) e 3.4782(8) c : Summary of the determinations of $\beta_{\rm c}^{(1)}$ and and $\beta_{\rm c}^{(2)}$ in $Z(N)$ on lattices with size $L^2\times N_t$. The fourth (sixth) column gives the method adopted to find $\beta_{\rm c}^{(1)}$ ($\beta_{\rm c}^{(2)}$). \[crit\_betas\] Determination of critical indices at the two transitions {#sec:betas} -------------------------------------------------------- Once critical couplings have been estimated, we are able to extract some critical indices and check the hyperscaling relation. We start the discussion from the first transition. According to the standard finite-size scaling (FSS) theory, the equilibrium magnetization $|M_{L}|$ at criticality should obey the relation $|M_{L}| \sim L^{-\beta / \nu}$, if the spatial extension $L$ of the lattice is large enough [^2]. Therefore, we fit data of $|M_L|$ at $\beta^{(1)}_{\rm c}$ on all lattices with size $L$ not smaller than a given $L_{\rm min}$ with the scaling law $$|M_{L}|=A L^{-\beta/\nu} \ln^r L\;, \label{magn_fss}$$ where a non-zero value for $r$ takes into account the possibility of logarithmic corrections [@Kenna-Irving; @Hasenbusch]. The FSS behavior of the susceptibility $\chi^{(M)}_L$ is given by $\chi^{(M)}_L\sim L^{\gamma/ \nu}$, where $\gamma/\nu=2-\eta$ and $\eta$ is the magnetic critical index. Therefore we fit data of $\chi^{(M)}_L$ at $\beta^{(1)}_{\rm c}$ on all lattices with size $L$ not smaller than a given $L_{\rm min}$ according to the scaling law $$\chi^{(M)}_{L}= A L^{\gamma/\nu} \ln^r L\;. \label{chiM_fss}$$ As the value of the critical coupling $\beta^{(1)}_{\rm c}$ we use the central value of the determination from the method (b) (see Table \[crit\_betas\]). The results of the fits are summarized in Tables \[indices\_Z5\_Nt2\], \[indices\_Z5\_Nt4\], \[indices\_Z7\_Nt2\], \[indices\_Z7\_Nt4\], for the cases of $Z(5)$ with $N_t$=2, $Z(5)$ with $N_t$=4, $Z(7)$ with $N_t$=2 and $Z(7)$ with $N_t$=4, respectively. The reference value for the index $\eta$ at this transition is 1/4, whereas the hyperscaling relation to be fulfilled is $\gamma/\nu+2\beta/\nu=d=2$. $L_{\min}$ $\beta/\nu$ $r_{\beta/\nu}$ $\chi^2_{\beta/\nu}$ $\gamma/\nu$ $r_{\gamma/\nu}$ $\chi^2_{\gamma/\nu}$ $d = 2\beta/\nu + \gamma/\nu$ $\eta = 2-\gamma/\nu$ ------------ ------------- ----------------- ---------------------- -------------- ------------------ ----------------------- ------------------------------- ----------------------- 16 0.12040(7) 0$^\ast$ 4.43 1.692(2) 0$^\ast$ 52.74 1.933(2) 0.308(2) 0.1228(5) 0.013(3) 2.43 1.97(1) $-$1.44(6) 0.90 2.22(1) 0.03(1) 1.668(2) 0.125$^\ast$ 61.75 1.908(2) 0.332(2) 32 0.12047(8) 0$^\ast$ 4.49 1.712(2) 0$^\ast$ 22.49 1.953(2) 0.288(2) 0.1239(7) 0.019(4) 2.12 1.97(2) $-$1.4(1) 0.99 2.22(2) 0.03(2) 1.690(2) 0.125$^\ast$ 26.40 1.931(2) 0.310(2) 64 0.1206(1) 0$^\ast$ 4.30 1.729(3) 0$^\ast$ 6.93 1.970(3) 0.271(3) 0.126(1) 0.034(7) 1.51 1.93(3) $-$1.2(2) 0.73 2.18(3) 0.07(3) 1.708(3) 0.125$^\ast$ 8.30 1.949(3) 0.292(3) 128 0.1208(1) 0$^\ast$ 2.43 1.740(3) 0$^\ast$ 3.00 1.982(3) 0.260(3) 0.126(2) 0.03(1) 1.72 1.94(9) $-$1.3(5) 0.84 2.19(9) 0.06(9) 1.720(3) 0.125$^\ast$ 3.48 1.962(3) 0.280(3) 192 0.1210(1) 0$^\ast$ 2.09 1.745(4) 0$^\ast$ 1.97 1.987(4) 0.255(4) 0.125(3) 0.03(2) 2.06 1.91(4) $-$1.1(2) 0.97 2.16(4) 0.09(4) 1.726(4) 0.125$^\ast$ 2.25 1.968(4) 0.274(4) 256 0.1211(2) 0$^\ast$ 2.19 1.752(5) 0$^\ast$ 1.44 1.994(5) 0.248(5) 0.124(4) 0.02(3) 2.56 1.906(5) $-$1.02(3) 1.19 2.15(1) 0.094(5) 1.733(5) 0.125$^\ast$ 1.56 1.975(5) 0.267(5) 384 0.1212(2) 0$^\ast$ 2.44 1.758(6) 0$^\ast$ 1.02 2.000(6) 0.242(6) 0.1224(2) 0.0079(9) 3.27 1.836(6) $-$0.53(3) 1.37 2.080(6) 0.164(6) 1.740(6) 0.125$^\ast$ 1.04 1.982(6) 0.260(6) 512 0.1215(3) 0$^\ast$ 2.18 1.762(7) 0$^\ast$ 1.06 2.005(8) 0.238(7) 0.1021(2) $-$0.135(1) 1.21 1.840(6) $-$0.54(3) 1.88 2.044(7) 0.160(6) 1.744(7) 0.125$^\ast$ 1.02 1.987(8) 0.256(7) 768 0.1209(4) 0$^\ast$ 1.71 1.76(1) 0$^\ast$ 1.49 2.00(1) 0.24(1) 0.1220(3) 0.008(1) 3.52 1.834(8) $-$0.54(3) 3.24 2.078(8) 0.166(8) 1.74(1) 0.125$^\ast$ 1.47 1.98(1) 0.26(1) 1024 0.1209(6) 0$^\ast$ 3.41 1.76(2) 0$^\ast$ 2.93 2.00(2) 0.24(2) 1.74(2) 0.125$^\ast$ 2.89 1.98(2) 0.26(2) : Critical indices $\beta/\nu$ and $\gamma/\nu$ for the first transition in $Z(5)$ with $N_t$ = 2, determined by the fits given in Eqs. (\[magn\_fss\]) and (\[chiM\_fss\]) on the complex magnetization $M_L$ and its susceptibility $\chi_L^{(M)}$ at $\beta_{\rm c}^{(1)}= 1.87$ for different choices of the minimum lattice size $L_{\rm min}$ (an asterisk indicates a fixed parameter). The $\chi^2$ of the two fits, given in the columns four and seven, is the reduced one. \[indices\_Z5\_Nt2\] $L_{\min}$ $\beta/\nu$ $r_{\beta/\nu}$ $\chi^2_{\beta/\nu}$ $\gamma/\nu$ $r_{\gamma/\nu}$ $\chi^2_{\gamma/\nu}$ $d = 2 \beta/\nu + \gamma/\nu$ $\eta = 2-\gamma/\nu$ ------------ ------------- ----------------- ---------------------- -------------- ------------------ ----------------------- -------------------------------- ----------------------- 128 0.1215(2) 0$^\ast$ 0.45 1.730(5) 0$^\ast$ 1.23 1.973(5) 0.270(5) 0.126(4) 0.02(2) 0.30 1.808(6) $-$0.46(3) 1.36 2.06(1) 0.192(6) 1.709(5) 0.125$^\ast$ 1.32 1.952(5) 0.291(5) 192 0.1217(3) 0$^\ast$ 0.25 1.734(6) 0$^\ast$ 1.20 1.977(7) 0.266(6) 0.1228(3) 0.007(1) 0.32 1.707(7) 0.17(3) 1.59 1.952(7) 0.293(7) 0.25 1.714(6) 0.125$^\ast$ 1.19 1.957(7) 0.286(6) 256 0.1218(3) 0$^\ast$ 0.30 1.740(8) 0$^\ast$ 1.07 1.983(8) 0.260(8) 0.1229(3) 0.007(1) 0.47 1.62(3) 0.7(2) 0.88 1.87(3) 0.38(3) 1.720(8) 0.125$^\ast$ 0.98 1.963(8) 0.280(8) 384 0.1217(5) 0$^\ast$ 0.44 1.73(1) 0$^\ast$ 0.24 1.97(1) 0.27(1) 0.1228(4) 0.007(2) 0.86 1.785(9) $-$0.38(4) 0.62 2.03(1) 0.215(9) 1.71(1) 0.125$^\ast$ 0.22 1.95(1) 0.29(1) 512 0.1220(7) 0$^\ast$ 0.49 1.72(2) 0$^\ast$ 0.05 1.96(2) 0.28(2) 1.70(2) 0.125$^\ast$ 0.05 1.94(2) 0.30(2) : The same as Table \[indices\_Z5\_Nt2\] for $Z(5)$ with $N_t=4$, determined at $\beta_c^{(1)} = 2.813$. \[indices\_Z5\_Nt4\] $L_{\min}$ $\beta/\nu$ $r_{\beta/\nu}$ $\chi^2_{\beta/\nu}$ $\gamma/\nu$ $r_{\gamma/\nu}$ $\chi^2_{\gamma/\nu}$ $d = 2 \beta/\nu + \gamma/\nu$ $\eta = 2- \gamma/\nu$ ------------ ------------- ----------------- ---------------------- -------------- ------------------ ----------------------- -------------------------------- ------------------------ 128 0.12862(9) 0$^\ast$ 11.95 1.768(2) 0$^\ast$ 0.93 2.025(2) 0.232(2) 0.141(1) 0.078(9) 0.93 1.763(2) 0.030(3) 1.15 2.045(5) 0.237(2) 1.747(2) 0.125$^\ast$ 1.18 2.004(2) 0.253(2) 192 0.1290(1) 0$^\ast$ 7.73 1.767(3) 0$^\ast$ 1.09 2.025(3) 0.233(3) 0.143(2) 0.09(1) 0.87 1.89(2) $-$0.8(1) 0.33 2.17(3) 0.11(2) 1.748(3) 0.125$^\ast$ 1.36 2.006(3) 0.252(3) 256 0.1294(1) 0$^\ast$ 5.84 1.771(4) 0$^\ast$ 0.83 2.029(4) 0.229(4) 0.145(3) 0.10(2) 0.95 1.87(3) $-$0.6(2) 0.44 2.16(3) 0.13(3) 1.752(4) 0.125$^\ast$ 1.00 2.010(4) 0.248(4) 384 0.1297(2) 0$^\ast$ 4.96 1.775(5) 0$^\ast$ 0.58 2.034(5) 0.225(5) 0.1517(2) 0.1480(8) 0.27 1.865(5) $-$0.61(2) 0.55 2.168(5) 0.135(5) 1.757(5) 0.125$^\ast$ 0.64 2.016(5) 0.243(5) 512 0.1303(2) 0$^\ast$ 2.43 1.775(6) 0$^\ast$ 0.77 2.035(7) 0.225(6) 0.1504(2) 0.1389(9) 0.38 1.94(6) $-$1.2(4) 0.32 2.25(6) 0.06(6) 1.757(6) 0.125$^\ast$ 0.86 2.017(7) 0.243(6) 768 0.1309(4) 0$^\ast$ 0.76 1.784(9) 0$^\ast$ 0.01 2.05(1) 0.216(9) 0.1308(2) -0.001(1) 1.54 1.875(6) $-$0.65(3) 0.00 2.137(7) 0.125(6) 1.767(9) 0.125$^\ast$ 0.01 2.03(1) 0.233(9) 1024 0.1311(5) 0$^\ast$ 1.36 1.79(1) 0$^\ast$ 0.00 2.05(2) 0.21(1) 1.77(1) 0.125$^\ast$ 0.01 2.03(2) 0.23(1) : The same as Table \[indices\_Z5\_Nt2\] for $Z(7)$ with $N_t=2$, determined at $\beta_c^{(1)} = 2.031$. \[indices\_Z7\_Nt2\] $L_{\min}$ $\beta/\nu$ $r_{\beta/\nu}$ $\chi^2_{\beta/\nu}$ $\gamma/\nu$ $r_{\gamma/\nu}$ $\chi^2_{\gamma/\nu}$ $d = 2 \beta/\nu + \gamma/\nu$ $\eta = 2-\gamma/\nu$ ------------ ------------- ----------------- ---------------------- -------------- ------------------ ----------------------- -------------------------------- ----------------------- 128 0.1294(1) 0$^\ast$ 4.64 1.772(3) 0$^\ast$ 1.72 2.031(4) 0.228(3) 0.141(2) 0.07(1) 0.54 1.766(4) 0.04(2) 2.16 2.049(9) 0.234(4) 1.751(3) 0.125$^\ast$ 1.76 2.010(4) 0.249(3) 192 0.1298(2) 0$^\ast$ 3.61 1.771(5) 0$^\ast$ 2.12 2.031(5) 0.229(5) 0.1456(2) 0.0961(8) 0.20 1.85(8) $-$0.5(5) 2.64 2.14(8) 0.15(8) 1.751(5) 0.125$^\ast$ 2.20 2.010(5) 0.249(5) 256 0.1302(2) 0$^\ast$ 1.43 1.778(6) 0$^\ast$ 1.84 2.038(6) 0.222(6) 0.1432(2) 0.0811(9) 0.19 1.74(3) 0.2(2) 2.58 2.03(3) 0.26(3) 1.758(6) 0.125$^\ast$ 1.77 2.018(6) 0.242(6) 384 0.1307(3) 0$^\ast$ 0.37 1.773(9) 0$^\ast$ 2.48 2.03(1) 0.227(9) 0.1303(2) $-$0.003(1) 0.78 1.871(7) $-$0.63(3) 5.32 2.132(7) 0.129(7) 1.754(9) 0.125$^\ast$ 2.45 2.01(1) 0.246(9) 512 0.1310(5) 0$^\ast$ 0.08 1.78(1) 0$^\ast$ 4.86 2.04(1) 0.22(1) 1.76(1) 0.125$^\ast$ 4.77 2.02(1) 0.24(1) : The same as Table \[indices\_Z5\_Nt2\] for $Z(7)$ with $N_t=4$, determined at $\beta_c^{(1)} = 3.406$. \[indices\_Z7\_Nt4\] The procedure for the determination of the critical indices at the second transition is similar to the one for the first transition, with the difference that the fit with the scaling laws Eqs. (\[magn\_fss\]) and (\[chiM\_fss\]) is to be applied to data of the rotated magnetization, $M_R$, and of its susceptibility, $\chi_L^{(M_R)}$, respectively. As the value of the critical coupling $\beta^{(2)}_{\rm c}$ we use the one determined from the method (e) (see Table \[crit\_betas\]). The results of the fits are summarized in Tables \[indices\_Z5\_Nt2\_2\], \[indices\_Z5\_Nt4\_2\], \[indices\_Z7\_Nt2\_2\], \[indices\_Z7\_Nt4\_2\], for the cases of $Z(5)$ with $N_t$=2, $Z(5)$ with $N_t$=4, $Z(7)$ with $N_t$=2 and $Z(7)$ with $N_t$=4, respectively. The reference value for the index $\eta$ at this transition is $4/N^2$, [*i.e.*]{} $\eta=0.16$ for $N=5$ and $\eta=0.0816..$ for $N=7$, whereas the hyperscaling relation to be fulfilled is $\gamma/\nu+2\beta/\nu=d =2$. A general comment is that in many of the cases we investigated both $d$ and $\eta$ at the two critical points slightly differ from the expected values, though these differences cancel to a large extent if we define $\eta$ as $2 \beta/\nu$. $L_{\min}$ $\beta/\nu$ $r_{\beta/\nu}$ $\chi^2_{\beta/\nu}$ $\gamma/\nu$ $r_{\gamma/\nu}$ $\chi^2_{\gamma/\nu}$ $d = 2 \beta/\nu + \gamma/\nu$ $\eta = 2 - \gamma/\nu$ ------------ ------------- ----------------- ---------------------- -------------- ------------------ ----------------------- -------------------------------- ------------------------- 128 0.037(8) 0$^\ast$ 2.22 1.918(2) 0$^\ast$ 0.52 1.99(2) 0.082(2) $-$0.27(3) $-$1.8(2) 1.07 1.900(3) 0.11(1) 0.53 1.36(7) 0.100(3) 192 0.03(1) 0$^\ast$ 2.10 1.918(3) 0$^\ast$ 0.64 1.97(2) 0.082(3) $-$0.25(4) $-$1.6(2) 1.52 1.821(3) 0.59(1) 0.31 1.33(8) 0.179(3) 256 0.01(1) 0$^\ast$ 2.27 1.916(4) 0$^\ast$ 0.64 1.94(3) 0.084(4) $-$0.20(5) $-$1.3(3) 2.38 1.900(3) 0.10(1) 0.82 1.5(1) 0.100(3) 384 $-$0.02(2) 0$^\ast$ 1.77 1.913(6) 0$^\ast$ 0.62 1.88(5) 0.087(6) $-$0.01(2) 0.03(6) 3.56 1.87(3) 0.3(2) 1.06 1.85(6) 0.13(3) 512 $-$0.05(3) 0$^\ast$ 1.37 1.912(8) 0$^\ast$ 1.21 1.81(7) 0.088(8) : Critical indices $\beta/\nu$ and $\gamma/\nu$ for the second transition in $Z(5)$ with $N_t$ = 2, determined by the fits given in Eqs. (\[magn\_fss\]) and (\[chiM\_fss\]) on the rotated magnetization $M_R$ and its susceptibility $\chi_L^{(M_R)}$ at $\beta_{\rm c}^{(2)}= 1.940$ for different choices of the minimum lattice size $L_{\rm min}$ (an asterisk indicates a fixed parameter). \[indices\_Z5\_Nt2\_2\] $L_{\min}$ $\beta/\nu$ $r_{\beta/\nu}$ $\chi^2_{\beta/\nu}$ $\gamma/\nu$ $r_{\gamma/\nu}$ $\chi^2_{\gamma/\nu}$ $d = 2 \beta/\nu + \gamma/\nu$ $\eta = 2 - \gamma/\nu$ ------------ ------------- ----------------- ---------------------- -------------- ------------------ ----------------------- -------------------------------- ------------------------- 128 0.17(1) 0$^\ast$ 1.52 1.850(3) 0$^\ast$ 0.07 2.20(3) 0.150(3) $-$0.28(5) $-$2.6(3) 0.30 1.847(4) 0.02(2) 0.08 1.3(1) 0.153(4) 192 0.15(1) 0$^\ast$ 0.40 1.849(4) 0$^\ast$ 0.07 2.15(3) 0.151(4) $-$0.1(1) $-$1.5(6) 0.25 1.847(4) 0.01(2) 0.09 1.6(2) 0.153(4) 256 0.15(2) 0$^\ast$ 0.51 1.849(5) 0$^\ast$ 0.07 2.14(4) 0.151(5) $-$0.09(9) $-$1.5(5) 0.36 1.846(5) 0.01(2) 0.10 1.7(2) 0.154(5) 384 0.12(3) 0$^\ast$ 0.24 1.847(8) 0$^\ast$ 0.07 2.09(7) 0.153(8) 0.10(2) $-$0.15(9) 0.45 1.845(6) 0.01(2) 0.14 2.04(5) 0.155(6) 512 0.10(5) 0$^\ast$ 0.21 1.85(1) 0$^\ast$ 0.11 2.1(1) 0.15(1) : The same as Table \[indices\_Z5\_Nt2\_2\] for $Z(5)$ with $N_t=4$, determined at $\beta_c^{(2)} =2.898$. \[indices\_Z5\_Nt4\_2\] $L_{\min}$ $\beta/\nu$ $r_{\beta/\nu}$ $\chi^2_{\beta/\nu}$ $\gamma/\nu$ $r_{\gamma/\nu}$ $\chi^2_{\gamma/\nu}$ $d = 2 \beta/\nu + \gamma/\nu$ $\eta = 2 - \gamma/\nu$ ------------ ------------- ----------------- ---------------------- -------------- ------------------ ----------------------- -------------------------------- ------------------------- 128 0.034(6) 0$^\ast$ 2.33 1.921(2) 0$^\ast$ 0.45 1.99(1) 0.079(2) $-$0.27(5) $-$1.9(3) 0.34 1.89(3) 0.2(2) 0.27 1.3(1) 0.11(3) 192 0.018(7) 0$^\ast$ 0.72 1.919(2) 0$^\ast$ 0.23 1.96(2) 0.081(2) $-$0.19(4) $-$1.3(3) 0.26 1.903(2) 0.10(1) 0.26 1.53(9) 0.097(2) 256 0.010(9) 0$^\ast$ 0.26 1.919(2) 0$^\ast$ 0.28 1.94(2) 0.081(2) $-$0.12(7) $-$0.8(4) 0.18 1.904(3) 0.10(1) 0.28 1.7(1) 0.096(3) 384 0(1) 0$^\ast$ 11.54 1.919(3) 0$^\ast$ 0.35 2(2) 0.081(3) $-$0.12(6) $-$0.9(4) 0.21 1.904(3) 0.10(1) 0.37 1.7(1) 0.096(3) 512 0.00(2) 0$^\ast$ 0.25 1.917(4) 0$^\ast$ 0.21 1.92(3) 0.083(4) 0.00(1) 0.00(5) 0.38 1.903(3) 0.10(1) 0.29 1.91(3) 0.097(3) 768 $-$0.01(2) 0$^\ast$ 0.11 1.916(6) 0$^\ast$ 0.29 1.90(5) 0.084(6) $-$0.01(1) 0.02(6) 0.22 1.903(4) 0.10(2) 0.56 1.89(3) 0.097(4) 1024 $-$0.00(3) 0$^\ast$ 0.14 1.918(8) 0$^\ast$ 0.53 1.91(7) 0.082(8) : The same as Table \[indices\_Z5\_Nt2\_2\] for $Z(7)$ with $N_t=2$, determined at $\beta_c^{(2)} = 3.366$. \[indices\_Z7\_Nt2\_2\] $L_{\min}$ $\beta/\nu$ $r_{\beta/\nu}$ $\chi^2_{\beta/\nu}$ $\gamma/\nu$ $r_{\gamma/\nu}$ $\chi^2_{\gamma/\nu}$ $d = 2 \beta/\nu + \gamma/\nu$ $\eta = 2 - \gamma/\nu$ ------------ ------------- ----------------- ---------------------- -------------- ------------------ ----------------------- -------------------------------- ------------------------- 128 0.037(8) 0$^\ast$ 2.22 1.918(2) 0$^\ast$ 0.52 1.99(2) 0.082(2) $-$0.27(3) $-$1.8(2) 1.07 1.900(3) 0.11(1) 0.53 1.36(7) 0.100(3) 192 0.03(1) 0$^\ast$ 2.10 1.918(3) 0$^\ast$ 0.64 1.97(2) 0.082(3) $-$0.25(4) $-$1.6(2) 1.52 1.821(3) 0.59(1) 0.31 1.33(8) 0.179(3) 256 0.01(1) 0$^\ast$ 2.27 1.916(4) 0$^\ast$ 0.64 1.94(3) 0.084(4) $-$0.20(5) $-$1.3(3) 2.38 1.900(3) 0.10(1) 0.82 1.5(1) 0.100(3) 384 $-$0.02(2) 0$^\ast$ 1.77 1.913(6) 0$^\ast$ 0.62 1.88(5) 0.087(6) $-$0.01(2) 0.03(6) 3.56 1.87(3) 0.3(2) 1.06 1.85(6) 0.13(3) 512 $-$0.05(3) 0$^\ast$ 1.37 1.912(8) 0$^\ast$ 1.21 1.81(7) 0.088(8) : The same as Table \[indices\_Z5\_Nt2\_2\] for $Z(7)$ with $N_t=4$, determined at $\beta_c^{(2)} = 5.158$. \[indices\_Z7\_Nt4\_2\] There is an independent method to determine the critical exponent $\eta$, which does not rely on the prior knowledge of the critical coupling, but is based on the construction of a suitable universal quantity [@Loison99; @2dzn]. The idea is to plot $\chi_{L}^{(M_{R})}L^{\eta-2}$ versus $B_{4}^{(M_{R})}$ and to look for the value of $\eta$ which optimizes the overlap of curves from different volumes. This method is illustrated in Fig. \[CHIvsB4z7\_2\_nt4\] for the case of $Z(7)$ with $N_t=4$: for $\eta=0.25$ the overlap is optimal in the lower branch of the curves, corresponding to the region of the first transition, while per $\eta=0.0816\simeq 4/7^2$ the overlap is optimal in the upper branch. Another option is to plot $M_{R}L^{\eta/2}$ versus $m_\psi$, which leads to overlapping curves for $\eta$ fixed at the value of the second phase transition, as illustrated in Fig. \[MRLvMPSIz7\_2\] for the case of $Z(7)$ with $N_t=2$ and $N_t=4$. ![Correlation between $\chi_{L}^{(M_{R})}L^{\eta-2}$ and the Binder cumulant $B_{4}^{(M_{R})}$ in $Z(7)$ with $N_t=4$ for $\eta=0.25$ (left) and for $\eta=0.0816$ (right) on lattices with $L$ ranging from 128 to 1024.[]{data-label="CHIvsB4z7_2_nt4"}](_Z7_NT4_CHIL_vs_B4MR1.eps "fig:"){width="45.00000%"} ![Correlation between $\chi_{L}^{(M_{R})}L^{\eta-2}$ and the Binder cumulant $B_{4}^{(M_{R})}$ in $Z(7)$ with $N_t=4$ for $\eta=0.25$ (left) and for $\eta=0.0816$ (right) on lattices with $L$ ranging from 128 to 1024.[]{data-label="CHIvsB4z7_2_nt4"}](_Z7_NT4_CHIL_vs_B4MR2.eps "fig:"){width="45.00000%"} ![Correlation between $M_{R}L^{\eta/2}$ and $m_\psi$ in $Z(7)$ with $N_t=2$ (left) and in $Z(7)$ with $N_t=4$ (right) for $\eta=0.0816$ on lattices with various values of $L$.[]{data-label="MRLvMPSIz7_2"}](MRL_vs_MPSI.eps "fig:"){width="45.00000%"} ![Correlation between $M_{R}L^{\eta/2}$ and $m_\psi$ in $Z(7)$ with $N_t=2$ (left) and in $Z(7)$ with $N_t=4$ (right) for $\eta=0.0816$ on lattices with various values of $L$.[]{data-label="MRLvMPSIz7_2"}](_Z7_NT4_MRL_vs_MPSI.eps "fig:"){width="45.00000%"} ![Mean link energy (left) and specific heat (right) versus $\beta$ in $Z(7)$ with $N_t = 4$ on lattices with $L$ ranging from 128 to 1024.[]{data-label="fig:spheat_Z7_NT4"}](Z7_NT4_ENERGY.eps "fig:"){width="45.00000%"} ![Mean link energy (left) and specific heat (right) versus $\beta$ in $Z(7)$ with $N_t = 4$ on lattices with $L$ ranging from 128 to 1024.[]{data-label="fig:spheat_Z7_NT4"}](Z7_NT4_SPHEAT.eps "fig:"){width="45.00000%"} Other checks of the nature of the phase transitions --------------------------------------------------- To produce further evidence in favor of the fact that the phase transitions investigated so far are both of infinite order, we have calculated the mean link energy $E$ and the specific heat $C$ around the transitions in $Z(5)$ and in $Z(7)$ with $N_t$=2 and $N_t=4$ (see Fig. \[fig:spheat\_Z7\_NT4\] for the case of $Z(7)$ with $N_t=4$, for example). In all cases the dependence of $E$ and $C$ on $\beta$ is continuous, so that first and second order transitions are ruled out. Behavior with $N$ of the critical couplings =========================================== The results of this work and those available in the literature allow us to make some considerations about the behavior with $N$ of the critical couplings $\beta_{\rm c}^{(1)}$ and $\beta_{\rm c}^{(2)}$. Examining our data for $\beta_{\rm c}^{(1,2)}$, one concludes that, for a fixed $N_t$, - $\beta_{\rm c}^{(1)}$ converges to the $XY$ value very fast, like $\exp(-a N^2)$  , - $\beta_{\rm c}^{(2)}$ diverges like $N^2$ . To better see the dependence on $N$ of the critical couplings, we have found also the critical coupling $\beta_{\rm c}^{(1,2)}$ in $Z(9)$ and $Z(13)$ with $N_t=2$ and $N_t=4$, using the method (b) for $\beta_{\rm c}^{(1)}$ and the method (e) for $\beta_{\rm c}^{(2)}$ (see Section \[sec:betas\] for the details of these methods). In Table \[Ndep\] we summarize the present knowledge about the position of the critical points for $3d$ $Z(N)$ gauge models at $\beta_s=0$. $N$ $N_t$ $\beta_{\rm c}^{(1)}$ $\beta_{\rm c}^{(2)}$ Reference ---------- ------- ----------------------- ----------------------- ----------- 5 2 1.87(1) 1.940(7) this work 7 2 2.031(7) 3.366(7) this work 9 2 2.04(3) 5.38(4) this work 13 2 2.02(1) 10.815(8) this work $\infty$ 2 – $\infty$ 5 4 2.813(3) 2.898(4) this work 7 4 3.406(8) 5.158(7) this work 9 4 3.50(1) 8.28(1) this work 13 4 3.490(6) 16.94(2) this work $\infty$ 4 3.42(1) $\infty$ [@3du1ft] : Summary of the known values of the critical couplings $\beta_{\rm c}^{(1)}$ and $\beta_{\rm c}^{(2)}$ in $Z(N)$ with $N_t=2,4$. \[Ndep\] One should expect that the $3d$ $Z(N)$ gauge models at $\beta_s=0$ satisfy the scaling predicted by RG, probably up to ${\cal{O}}(N)$ corrections. One could try therefore to fit the available Monte Carlo data for $\beta_{\rm c}^{(1,2)}$ with formulae predicted by RG and modified to account for such corrections. We find that the critical couplings for the first transition are well reproduced by the function $$\beta_{\rm c}^{(1)} = A - B N^2 \exp {\left( - \frac{N^2}{C} \right)}\;,$$ for suitable values of the parameters $A$, $B$ and $C$, both for $N_t=2$ and $N_t=4$. The critical couplings for the second transition are well reproduced instead by the following functions: $$\beta_{\rm c}^{(2)} = A N^2 + B N + C$$ and $$\beta_{\rm c}^{(2)} = \frac{A}{1-\cos\left(\frac{2 \pi}{N}\right)} + B N + C\;,$$ for suitable values of the parameters $A$, $B$ and $C$, both for $N_t=2$ and $N_t=4$. The last formula has been suggested in Ref. [@bhanot] in the context of the zero-temperature theory. ![Fits of critical values for $\beta_{\rm c}^{(2)}$ as a function of $N$ for $N_t=2$ (left) and $N_t=4$ (right).[]{data-label="fig:scaling2"}](dep_on_N_nt2_4.eps "fig:"){width="45.00000%"} ![Fits of critical values for $\beta_{\rm c}^{(2)}$ as a function of $N$ for $N_t=2$ (left) and $N_t=4$ (right).[]{data-label="fig:scaling2"}](dep_on_N_nt4_4.eps "fig:"){width="45.00000%"} As an illustration of our fits, Fig. \[fig:scaling2\] shows the dependence of $\beta_{\rm c}^{(2)}$ on $N$ (parameter $B=0$ in fitting formulas). One concludes from these plots that to distinguish between two scalings one should probably have data for smaller values of $N=2,3,4$ at $\beta_s=0$. Conclusions =========== In this paper we have studied the $3d$ $Z(N)$ gauge theory at the finite temperature in the strong coupling region $\beta_s=0$. This study was based on the exact relation of this model to a generalized $2d$ $Z(N)$ spin model. In Section 2 we established the exact relation between couplings of these two models, described qualitatively some of the RG predictions for effective model and gave analytical estimations of the critical couplings. The main, numerical part of the work has been devoted to the localization of the critical couplings and to the computation of the critical indices: - We have determined numerically the two critical couplings of $Z(N=5,7,9,13)$ LGTs and given estimates of the critical indices $\eta$ at both transitions. For the first time we have a clear indication that for all $N\geq 5$ the scenario of three phases is realized: a disordered phase at small $\beta_t$, a massless or BKT one at intermediate values of $\beta_t$ and an ordered phase, occurring at larger and larger values of $\beta_t$ as $N$ increases. This matches perfectly with the $N\to\infty$ limit, [*i.e.*]{} the finite-temperature $3d$ $U(1)$ LGT (at $\beta_s=0$), where the ordered phase is absent; - We have found that the values of the critical index $\eta$ at the two transitions are compatible with the theoretical expectations; in order to reproduce the expected value of $\eta$, we had to take into account logarithmic corrections with the exponent $r$ fixed to 0.125; - The index $\nu$ also appears to be compatible with the value $1/2$, in agreement with RG predictions. Results listed above present further evidence that finite-temperature $3d$ $Z(N)$ LGTs for $N>4$ undergo two phase transitions of the BKT type. Moreover, this model belongs to the universality class of the $2d$ $Z(N)$ spin model, at least in the strong coupling limit $\beta_s=0$. Considering the determinations of the critical couplings as a function of $N$, we have conjectured the approximate scaling for $\beta_c^{(1,2)}(N)$. ![Phase structure of the general $Z(5)$ spin model (see text for explanation).[]{data-label="fig:phasez5"}](phase.eps){width="50.00000%"} Finally, the study performed here allows to improve our knowledge of the phase diagrams of the generalized $2d$ $Z(N)$ spin models. As an example, we plot in Fig. \[fig:phasez5\] the general phase diagram for $N=5$ in the $(t_1, t_2)$-plane, where $t_i = (B_i/B_0)^{N_t}$, and $B_i$ are defined in (\[couplings\_coeff\]). Here, the line $AB$ is self-dual line, SPM corresponds to the standard Potts model, VPM to the vector model. The SPM undergoes a first order phase transition with a critical point occurring on the self-dual line. The line of the first order phase transition terminates at the point 2. Its approximate position was computed in Ref. [@Domany]. Shown are also the locations of the critical points for the VPM ($N_t=1$) and for $Z(5)$ LGT with $N_t=2,4$. The parametric curves for different $N_t$ lie very close to each other, so we cannot trace a sufficiently big part of the curve while changing $N_t$. On the other hand it shows that already the model with $N_t=4$ presents a very good approximation to the finite-temperature limit. Indeed, the parametric curve for $N_t=8$ is almost indistinguishable from the curve with $N_t=4$. Acknowledgments =============== The work of O.B. was supported by the Program of Fundamental Research of the Department of Physics and Astronomy of NAS, Ukraine. The work of G.C. and M.G. was supported in part by the European Union under ITN STRONGnet (grant PITN-GA-2009-238353). G.C. thanks Matteo Giordano for useful discussions. [4]{} G. Bhanot and M. Creutz, Phys. Rev. D [**21**]{} (1980) 2892. A. Polyakov, Nucl. Phys. B [**120**]{} (1977) 429; T. Banks, J. Kogut, R. Myerson, Nucl. Phys. B [**121**]{} (1977) 493; M. Göpfert, G. Mack, Commun. Math. Phys. [**81**]{} (1981) 97. B. Svetitsky, L. Yaffe, Nucl. Phys. B [**210**]{} (1982) 423. V. Berezinskii, Sov. Phys. JETP [**32**]{} (1971) 493. J. Kosterlitz, D. Thouless, J. Phys. C [**6**]{} (1973) 1181. J. Kosterlitz, J. Phys. C [**7**]{} (1974) 1046. N. Parga, Phys. Lett. B [**107**]{} (1981) 442. O. Borisenko, [*Critical behaviour of 3D U(1) LGT at finite temperature*]{}, PoS LAT [**2007**]{} (2007) 170. O. Borisenko, M. Gravina, A. Papa, J. Stat. Mech. [**2008**]{} (2008) P08009. O. Borisenko, R. Fiore, M. Gravina, A. Papa, J. Stat. Mech. [**2010**]{} (2010) P04015. F.Y. Wu, Rev. Mod. Phys. [**54**]{} (1982) 235. S. Elitzur, R.B. Pearson, J. Shigemitsu, Phys. Rev. D [**19**]{} (1979) 3698. M. B. Einhorn, R. Savit, and E. Rabinovici, Nucl. Phys. B [**170**]{} (1980) 16. C.J. Hamer, J.B. Kogut, Phys. Rev. B [**22**]{} (1980) 3378. B. Nienhuis, J. Statist. Phys. [**34**]{} (1984) 731. L.P. Kadanoff, J. Phys. A [**11**]{} (1978) 1399. J.L. Cardy, J. Phys. A [**13**]{} (1980) 1507. J. Fröhlich, T. Spencer, Commun. Math. Phys. [**81**]{} (1981) 527. Y. Tomita, Y. Okabe, Phys. Rev. B [**65**]{} (2002) 184405. O. Borisenko, G. Cortese, R. Fiore, M. Gravina, A. Papa, [*Critical properties of the two-dimensional Z(5) vector model*]{}, PoS LATTICE [**2010**]{} (2010) 274 \[arXiv:1101.0512 \[hep-lat\]\]. O. Borisenko, G. Cortese, R. Fiore, M. Gravina and A. Papa, Phys. Rev. E [**83**]{} (2011) 041120. O. Borisenko, G. Cortese, R. Fiore, M. Gravina, A. Papa, [*Critical properties of 2D Z(N) models for $N>4$*]{}, PoS LATTICE [**2011**]{} (2011) 304 \[arXiv:1110.6385 \[hep-lat\]\]. O. Borisenko, V. Chelnokov, G. Cortese, R. Fiore, M. Gravina, A. Papa, Phys. Rev. E [**85**]{} (2012) 021114. M. Billò, M. Caselle, A. D’Adda, Int. J. Mod. Phys. A [**12**]{} (1997) 5753. A. Yamagata and I. Ono, J. Phys. [**A24**]{} (1991) 265. R. Kenna and A. C. Irving, Nucl. Phys. B [**485**]{} (1997) 583. M. Hasenbusch, J. Phys. A [**38**]{} (2005) 5869. D. Loison, J. Phys.: Condens. Matter [**11**]{} (1999) L401. E. Domany, D. Mukamel, A. Schwimmer, J. Phys. A [**13**]{} (1980) L311. [^1]: It should be noted, however, that the numerical results of [@3du1full] point to a critical index $\eta$ larger than its $XY$ value by almost a factor of 2 for $N_t=8$. Therefore, the question of the universality remains open for this model. [^2]: The symbol $\beta$ here denotes a critical index and not, obviously, the coupling of the theory. In spite of this inconvenient notation, we are confident that no confusion will arise, since it will be always clear from the context which $\beta$ is to be referred to.
{ "pile_set_name": "ArXiv" }
--- abstract: 'A generalization of Chern-Simons gauge theory is formulated in any dimension and arbitrary gauge group where gauge fields and gauge parameters are differential forms of any degree. The quaternion algebra structure of this formulation is shown to be equivalent to a three $\mathbb{Z}_2$-gradings structure, thus clarifying the quaternion role in the previous formulation.' --- EPHOU 16-014\ Sept. 2016 \ [Quaternion based generalization of Chern-Simons theories in arbitrary dimensions ]{}\ [Alessandro D’Adda[^1], Noboru Kawamoto[^2], Naoki Shimode[^3],\ [and]{} Takuya Tsukioka[^4]]{}\ [ *INFN Sezione di Torino, and\ Dipartimento di Fisica Teorica, Universita di Torino,\ I-10125 Torino, Italy\ Department of Physics, Hokkaido University,\ Sapporo 060-0810, Japan\ School of Education, Bukkyo University,\ Kyoto 603-8301, Japan*]{} [**1  Introduction**]{} A formulation of gauge theory in terms of differential forms has the advantage that it automatically generates a general coordinate invariant formulation since an explicit metric dependence does not appear. One interesting approach along this line is the applications of Chern-Simons action to 3-dimensional gravity [@Witten-3dgravity]. In the formulation of the standard gauge theory only 1-form gauge fields and 0-form gauge parameters play a role as differential forms. It is natural to ask if one can formulate gauge theories in terms of all the degrees of differential forms. A positive answer was given by one of authors (N.K.) and Watabiki many years back with a graded Lie algebra setting [@kawa-wata]. In this paper we focus on the generalization of the Chern-Simons action to arbitrary dimensions with arbitrary degrees of differential forms as gauge fields and parameters for Lie algebra setting and clarify the origin of the quaternion structure which was discovered in the original formulation [@kawa-wata]. In the present formulation the introduction of graded Lie algebra is not required. [**2  The origin of the quaternion structure for a three grading gauge system**]{} When we consider standard Abelian gauge theory with differential forms we identify gauge field as one-form and gauge parameter as zero-form. In this gauge system $\mathbb{Z}_2$-grading structure of even-form and odd-form is present. If we define $\Lambda_+$ as a set of even forms and $\Lambda_-$ as a set of odd forms, we have \_+ ’\_+=’\_+ \_+\_+, \_+ \_-=\_- \_+\_-, \_- ’\_-=-’\_- \_-\_+, \[z2grading\] where $\lambda_+, \lambda'_+ \in \Lambda_+,~ \lambda_-, \lambda'_- \in \Lambda_-$ and $\wedge$ is a wedge product. Fermionic and bosonic fields have similar $\mathbb{Z}_2$-grading structure with an obvious correspondence. Let us consider two types of fields $\Phi_{(a,b,c)}$ and ${\mathcal F}_{(a,b,c)}$ which have a three $\mathbb{Z}_2$-grading structure $(a,b,c)$ with $a,b,c, =0$ or 1. For simplicity we assume that these fields have Abelian nature. Then we introduce two types of commuting structure with respect to the three gradings: \_[(a,b,c)]{} ’\_[(a’,b’,c’)]{} &=& (-1)\^[aa’ +bb’+cc’]{} ’\_[(a’,b’,c’)]{} \_[(a,b,c)]{}, \[grading1\]\ [F]{}\_[(a,b,c)]{} [F]{}’\_[(a’,b’,c’)]{} &=& (-1)\^[(a+b+c)(a’+b’+c’)]{} [F]{}’\_[(a’,b’,c’)]{} [F]{}\_[(a,b,c)]{}. \[grading2\] In (\[grading1\]) the three gradings are independent whereas in (\[grading2\]) there is one global grading corresponding $a+b+c$. Let us introduce an object ${\bf q}(a,b,c)$ satisfying the following commuting structure: (a,b,c) [**q**]{}(a’,b’,c’) = (-1)\^[aa’+bb’+cc’ +(a+b+c)(a’+b’+c’)]{} [**q**]{}(a’,b’,c’) [**q**]{}(a,b,c). \[quaternion-grading\] Then we have the following commuting relation: (\_[(a,b,c)]{}[**q**]{}(a,b,c))&& (’\_[(a’,b’,c’)]{}[**q**]{}(a’,b’,c’))\ &=&(-1)\^[(a+b+c)(a’+b’+c’)]{} (’\_[(a’,b’,c’)]{}[**q**]{}(a’,b’,c’)) (\_[(a,b,c)]{} [**q**]{}(a,b,c)), \[product-grading\] where we assume $\Phi_{(a,b,c)}$ and ${\bf q}(a,b,c)$ are not interacting and thus commuting: \_[(a,b,c)]{} [**q**]{}(a,b,c) = [**q**]{}(a,b,c) \_[(a,b,c)]{}. \[q-commuting\] We recognize now that a field of the ${\mathcal F}$-type, namely with a single global grading, can be written in terms of a field $\Phi_{(a,b,c)}$ with three distinct gradings as \_[(a,b,c)]{} =\_[(a,b,c)]{} [**q**]{}(a,b,c). \[q-valued-field\] In order to define a product of the ${\mathcal F}$-fields, we have to define a product in the ${\bf q}(a,b,c)$ space that satisfies (\[quaternion-grading\]). This is given by (a,b,c)[**q**]{}(a’,b’,c’) = (-1)\^[aa’+bb’+cc’+a’b+b’c+c’a]{} [**q**]{}(a+a’,b+b’,c+c’), \[quaternion-closure\] where the sums are defined modulo 2. This product is associative. There are eight possible ${\bf q}(a,b,c)$’s for $a,b,c = 0,1$. However with respect to their commuting structure (\[quaternion-grading\]) and with respect to the product (\[quaternion-closure\]) there are only four independent ${\bf q}(a,b,c)$. In fact the sign factors in (\[quaternion-grading\]) and (\[quaternion-closure\]) are invariant for $(a,b,c) \rightarrow (a+1,b+1,c+1)$ and the same for $(a',b',c')$. The units ${\bf q}(a,b,c)$ can then be identified in pairs and renamed as: $$\renewcommand{\arraystretch}{1.4} \begin{array}{rcl} {\bf q}(1,1,1) = {\bf q}(0,0,0) \equiv {\bf 1}, &\quad& {\bf q}(1,0,0) = {\bf q}(0,1,1) \equiv {\bf i}, \\ {\bf q}(0,1,0) = {\bf q}(1,0,1) \equiv {\bf j}, &\quad& {\bf q}(0,0,1) = {\bf q}(1,1,0) \equiv {\bf k}. \end{array} \renewcommand{\arraystretch}{1} \label{quat-def}$$ It is easy to recognize now that ${\bf 1,i,j,k}$ satisfy the quaternion algebra $$\renewcommand{\arraystretch}{1.4} \begin{array}{c} {\bf i}^2 = {\bf j}^2 = {\bf k}^2 = -{\bf 1}, \\ {\bf i}{\bf j} = -{\bf j}{\bf i} = {\bf k}, \ \ {\bf j}{\bf k} = -{\bf k}{\bf j} = {\bf i}, \ \ {\bf k}{\bf i} = -{\bf i}{\bf k} = {\bf j}. \end{array} \renewcommand{\arraystretch}{1} \label{quat-com-rel}$$ It is important to notice that eq.(\[quaternion-closure\]) defines the quaternion algebra in an unconventional way, different from its standard mathematical introduction, and that it links in an unexpected way the quaternion algebra to the existence of three independent gradings. We can introduce now two types of ${\mathcal F}$-fields, corresponding respectively to $a+b+c$ odd and even: &=&(1,1 ,1) \_[(1,1,1)]{} + [**q**]{}(1,0,0) \_[(1,0,0)]{} + [**q**]{}(0,1,0) \_[(0,1,0)]{} + [**q**]{}(0,0,1) \_[(0,0,1)]{}\ &=&\_[(1,1,1)]{} + [**i**]{} \_[(1,0,0)]{} + [**j**]{} \_[(0,1,0)]{} + [**k**]{} \_[(0,0,1)]{}, \[gene-field\]\ [V]{} &=&(0,0,0) \_[(0,0,0)]{} + [**q**]{}(0,1,1) \_[(0,1,1)]{} + [**q**]{}(1,0,1) \_[(1,0,1)]{} + [**q**]{}(1,1,0) \_[(1,1,0)]{}\ &=&\_[(0,0,0)]{} + [**i**]{} \_[(0,1,1)]{} + [**j**]{}\_[(1,0,1)]{} + [**k**]{} \_[(1,1,0)]{}. \[gene-para\] Assuming that the fields $\Phi_{(a,b,c)}$ are Abelian ${\mathcal A}$ and ${\mathcal V}$ are odd and even elements in a $\mathbb{Z}_2$ commuting algebra: ’=-[A]{}’[A]{},     = [V]{}[A]{},    ’ = [V]{}’[V]{}, \[gene-z2commuting\] where ${\mathcal A}'$ and ${\mathcal V}'$ are defined by $\Phi'(a',b',c')$. It also follows immediately from (\[gene-field\]), (\[gene-para\]) and the multiplication rules of quaternions that if we denote by ${\bf\Lambda}_-$ and ${\bf\Lambda}_+$ the sets of fields respectively of the ${\mathcal A}$-type and ${\mathcal V}$-type then ’=-[A]{}’[A]{} \_+,    = [V]{}[A]{} \_-,   ’ = [V]{}’[V]{} \_+. \[gene-z2grading\] In conclusion ${\mathcal A}$ and ${\mathcal V}$ are anticommuting and commuting quaternionic fields whose component fields $\Phi_{(a,b,c)}$ possess three independent $\mathbb{Z}_2$-gradings whose physical meaning and interpretation may be different. We shall use them in what follows to formulate gauge theories with higher differential forms. The importance of the different sign choice in (\[grading1\]) and (\[grading2\]) was noticed in [@Deligne] for two- and three-gradings formulations of supersymmetric gauge theories. [**3  Higher form gauge fields and non-Abelian extension**]{} We shall consider fields $\Phi_{(a,b,c)}(x)$ with space-time dependence so it is natural to associate one of the gradings, -conventionally the second one labeled by $b$-, as representing the grading of even and odd differential forms in space-time. We shall also take the first grading, labeled by $a$, as some sort of fermion-boson grading (whose exact nature will be briefly discussed at the end) while the third grading will be left unspecified and denoted by the suffix $c(=0,1)$. We introduce then the following notations: $$\renewcommand{\arraystretch}{1.4} \begin{array}{rcl} \Phi_{(0,0,c)}(x) \!\!\!\!&\equiv&\!\!\!\! \hat{A}_c(x)={\hbox{direct sum of bosonic even forms}}, \\ \Phi_{(0,1,c)}(x) \!\!\!\!&\equiv&\!\!\!\! A_c(x) ={\hbox{direct sum of bosonic odd forms}}, \\ \Phi_{(1,0,c)}(x) \!\!\!\!&\equiv&\!\!\!\! \hat{\psi}_c(x)={\hbox{direct sum of fermionic even forms}}, \\ \Phi_{(1,1,c)}(x) \!\!\!\!&\equiv&\!\!\!\! \psi_c(x) = {\hbox{direct sum of fermionic odd forms}}. \end{array} \renewcommand{\arraystretch}{1} \label{grading-notation}$$ Here and in the following we shall denote bosons and fermions, respectively, with Roman and Greek letters while even forms and odd forms are differentiated by hat and non-hat. We introduce the exterior derivative operator $d=dx^\mu\partial_\mu$ as an ${\mathcal A}$-type operator since it has grading; $(0,1,0)$: Q= [**q**]{}(0,1,0) d =[**j**]{} d. \[gene-d-op\] A more general form of $Q$ will be briefly discussed at the end of the paper. Extending from Abelian to non-Abelian gauge field theory, we identify the following ${\mathcal A}$ and ${\mathcal V}$ as generalized gauge field and gauge parameter respectively: $$\renewcommand{\arraystretch}{1.4} \begin{array}{rcl} {\mathcal A} \!\!\!\!&=&\!\!\!\! {\mathcal A}^B T^B = ({\bf 1} {\psi}_1^B + {\bf i} \hat{\psi}_0^B + {\bf j} {A}_0^B + {\bf k} \hat{A}_1^B)T^B, \\ {\mathcal V}\!\!\!\!&=&\!\!\!\! {\mathcal V}^B T^B = ({\bf 1} \hat{a}_0^B + {\bf i} {a}_1^B + {\bf j} \hat{\alpha}_1^B + {\bf k} {\alpha}_0^B) T^B, \end{array} \renewcommand{\arraystretch}{1} \label{gene-nonAbelian}$$ where $T^B$’s are generators of Lie algebra. It is important to realize at this stage that ${\mathcal A}^B$ and ${\mathcal V}^B$ satisfy the same grading structure as ${\mathcal A}$-type and ${\mathcal V}$-type in (\[gene-z2grading\]). Hereafter all products such as ${\mathcal A}^A{{\mathcal A}'}^B,~{\mathcal A}^A{\mathcal V}^B,~ {{\mathcal V}'}^A{\mathcal V}^B \cdots$ are understood as wedge products. It is important to remark at this stage that ${\mathcal A}$, ${\mathcal V}$ and $Q$ have the same formal properties as respectively one form gauge fields, zero form gauge parameters and the differential operator $d$ in gauge theories. So any gauge theory that can be entirely written in terms of forms, without use of the Hodge operator, like Chern-Simons theory and Einstein gravity, admits a generalization where all fields and parameters are replaced by the quaternionic analogues ${\mathcal A}$ and ${\mathcal V}$ and gauge invariance in terms of the generalized parameters is automatically preserved. In this paper we shall concentrate on the generalized Chern-Simons theory, which was already studied in a similar framework in [@kawa-wata]. [**4  Generalized Chern-Simons actions in arbitrary dimensions**]{} We can now construct a dimension independent formulation of generalized Chern-Simons action: S = (Q[A]{} + \^3 ) = [**1**]{}S\^1 + [**i**]{}S\^i + [**j**]{}S\^j + [**k**]{}S\^k, \[GCSA\] where the $\Tr$ is taken on a representation of the Lie algebra. Notice that ${\mathcal A}$ contains now forms of arbitrary degree and so the integral is not restricted to be on a three dimensional space-time but can be formulated in any dimensions. This generalized Chern-Simons action is invariant under the following generalized gauge transformation: = Q[V]{} +\[[A]{},[V]{}\] =[**1**]{}\_1 + [**i**]{}\_0 + [**j**]{}A\_0 + [**k**]{}\_1. \[GGT\] The proof of the gauge invariance of the generalized Chern-Simons action can be derived from the following properties of the generalized gauge fields and parameters: 1. $Q^2 =0$, 2. $\{\overrightarrow{Q},{\bf \lambda_-} \} = Q{\bf \lambda_-}, ~~~ [\overrightarrow{Q},{\bf \lambda_+}] = Q{\bf \lambda_+},$ 3. $\Tr({\bf \lambda_+}{\bf \lambda'_+}) = \Tr({\bf \lambda'_+}{\bf \lambda_+}), ~~~ \Tr({\bf \lambda_-}{\bf \lambda_+}) = \Tr({\bf \lambda_+}{\bf \lambda_-}), ~~~ \Tr({\bf \lambda_-}{\bf \lambda'_-}) =- \Tr({\bf \lambda'_-}{\bf \lambda_-})$, where $\lambda_-,\lambda'_-$ and $\lambda_+,\lambda'_+$ are, respectively, ${\mathcal A}$-type and ${\mathcal V}$-type fields and parameters in (\[gene-nonAbelian\]). Here $\{\ ,\ \}$ and $[\ , \ ]$ are anti-commutator and commutator, respectively. $\overrightarrow{Q}={\bf j}\overrightarrow{d}$ is an exterior derivative operating on the right. In this quaternion valued formulation each sector of quaternion coefficients are equivalent in (\[GCSA\]) and (\[GGT\]) since quaternions commute with fields and interact only among themselves. We can thus derive the following four types of generalized Chern-Simons actions: S\^1&=& ,\ S\^i &=& ,\ S\^j &=& ,\ S\^k &=& , which are invariant under the following generalized gauge transformations: $$\renewcommand{\arraystretch}{1.4} \begin{array}{rcl} \delta A_0 \!\!\!\!&=&\!\!\!\!d \hat{a}_0 +[A_0, \hat{a}_0] +\{\hat{A}_1,a_1\} +[\psi_1,\hat{\alpha}_1] -\{\hat{\psi}_0,\alpha_0\} , \\ \delta \hat{A}_1\!\!\!\!&=&\!\!\!\! -d a_1 -\{A_0,a_1\} + [\hat{A}_1, \hat{a}_0] + [\psi_1,\alpha_0] +\{\hat{\psi}_0,\hat{\alpha}_1\} , \\ \delta\psi_1\!\!\!\!&=&\!\!\!\! -d \hat{\alpha}_1 -[A_0,\hat{\alpha}_1] -[\hat{A}_1,\alpha_0] +[\psi_1, \hat{a}_0] - [\hat{\psi}_0, a_1] , \\ \delta\hat{\psi}_0 \!\!\!\!&=&\!\!\!\! d \alpha_0 +\{A_0,\alpha_0\} -\{\hat{A}_1,\hat{\alpha}_1\} +[\psi_1,a_1] + [\hat{\psi}_0,\hat{a}_0]. \end{array} \label{GGTs} \renewcommand{\arraystretch}{1}$$ In these generalized gauge transformations we find that commutators and anti-commutators are mixed. It is, however, important to realize that the generators of the Lie algebra appear only in commutators and thus are algebraically closed. For example $\{A_0,\alpha_0\} =A_0^B \alpha_0^C[T^B,T^C]$, $\{\hat{\psi}_0,\hat{\alpha}_1\} = \hat{\psi}_0^B\hat{\alpha}_1^C[T^B,T^C]$,   $\{\hat{A}_1,a_1\} = \hat{A}_1^B a_1^C[T^B,T^C], \ \cdots$ and so on. Notice that the standard 3-form Chern-Simons action and the standard gauge transformation are included respectively in the first two terms of $S^j$ and of $\delta A_0$. It is important to recognize that the integrand of the generalized Chern-Simons action has the ${\mathcal A}$-type nature given in (\[gene-field\]) with respect to the quaternion structure. $S^j$ is thus bosonic odd-dimensional action and $S^k$ is bosonic even-dimensional action. $S^1$ and $S^i$ are, respectively, fermionic odd-dimensional and fermionic even-dimensional actions, whose physical interpretation is not yet clear at this moment. Notice that the structure of $S^j$ is different from the one of the other actions $S^1,S^i$ and $S^k$ which have instead a similar field structure. The origin of this difference and similarity comes respectively from the special choice of the exterior derivative operator as a ${\bf j}$-component quaternion; $Q={\bf j}d$ and from the permutation invariance in the quaternion space. [**5  Connection with graded Lie algebra formulation**]{} The quaternion structure in formulating higher form gauge theory was discovered long time ago by one of authors (N.K.) and Watabiki [@kawa-wata]. In fact the generalized Chern-Simons actions (\[GCSAs\]) and the gauge transformations (\[GGTs\]) were already given in [@kawa-wata] in a graded Lie algebra framework. At the time the origin of quaternions in formulating the generalized gauge theory was not clear. In the present paper the origin of the quaternion is clarified and is based on the 3-grading structure. Due to this clarification a Lie algebra formulation instead of graded Lie algebra formulation has been successfully realized. Here we show how these two formulations are related in the current context. In order that a product of algebra valued fields and parameters of ${\mathcal A}$-type and ${\mathcal V}$-type again belongs to ${\mathcal A}$-type or ${\mathcal V}$-type field or parameter, the products should be defined by the following graded commutator: ’ {[A]{},[A]{}’ }, \[[A]{},[V]{} \],’ \[[V]{},[V]{}’ \], \[closure-alg\] where ${\mathcal A}, {\mathcal A}'$ and ${\mathcal V}, {\mathcal V}'$ are (graded) Lie algebra valued ${\mathcal A}$- and ${\mathcal V}$-type fields and/or parameters, respectively. In [@kawa-wata] the necessity of algebraic closure for the product was noticed but not clearly stated while this point was stressed in [@Schwarz]. Let us now consider the case where we do not introduce the third grading denoted by the suffix $c=(0,1)$, but introduce two types of generators $T^B$ and $\Sigma^\alpha$ which close as a graded Lie algebra: = f\^[BC]{}\_D T\^D,     \[T\^B,\^\] = g\^[B]{}\_\^,    {\^,\^} = h\^\_B T\^B, \[GLA\] where $f^{BC}_D,g^{B\alpha}_\beta$ and $h^{\alpha \beta}_B$ are the structure constants of the graded Lie algebra. If we define ${\mathcal A}$-type and ${\mathcal V}$-type graded Lie algebra valued fields and/or parameters as: $$\renewcommand{\arraystretch}{1.4} \begin{array}{rcl} {\mathcal A}\!\!\!\!&=&\!\!\!\! {\bf 1}\psi^\alpha\Sigma^\alpha + {\bf i}\hat{\psi}^BT^B +{\bf j}A^BT^B +{\bf k}\hat{A}^\alpha\Sigma^\alpha \in {\bf \Lambda}'_-, \\ {\mathcal V}\!\!\!\!&=&\!\!\!\! {\bf 1}\hat{a}^BT^B + {\bf i}a^\alpha\Sigma^\alpha +{\bf j}\hat{\alpha}^\alpha\Sigma^\alpha +{\bf k}\alpha^BT^B \in {\bf \Lambda}'_+, \end{array} \label{defAV-GLA}$$ where ${\bf \Lambda}'_-$ and ${\bf \Lambda}'_+$ denote the set of ${\mathcal A}$-type and ${\mathcal V}$-type fields and/or parameters, respectively. We can then show ’ ’\_+, ’\_-, ’’\_+, \[closure-prod-GLA\] which are the extended version of the closure relation (\[gene-z2grading\]) for graded Lie algebra. In order to obtain the gauge invariance of the generalized Chern-Simons action (\[GCSA\]) for the graded Lie algebra framework, it is necessary the item 3. in section 4 to be fulfilled for graded Lie algebra counterparts: \[[V]{},[V]{}’\] = [Str]{} \[[A]{},[V]{}\] = [Str]{}{[A]{},[A]{}’} =0. \[Str-def\] All these conditions can be satisfied if super trace (Str) satisfies the following relations for graded generators: \[T\^B,T\^C\] = [Str]{}\[T\^B,\^\] = [Str]{}{\^,\^}=0. \[Str-relations\] In particular the odd generators $\Sigma^\alpha$’s anti-commute within the super trace. This result shows that the 3rd grading with Lie algebra formulation is equivalent to the graded Lie algebra formulation without 3rd grading but with the super trace relations (\[Str-relations\]). In the earlier generalized gauge theory formulation, the introduction of graded Lie algebra is necessary to formulate odd dimensional generalized Chern-Simons actions including all degrees of differential forms equally successful as for the even dimensional case [@kawa-wata]. In other words the 3rd grading of the current formulation is mandatory for a equal footing treatment of generalized Chern-Simons actions for both even and odd dimensions. Thus introduction of three grading, equivalently the quaternion structure, is very important for the dimension independent treatment of the generalized gauge theory. As an application of the above graded Lie algebra formulation, Clifford algebra formulation which satisfies loosened version of the condition (\[closure-alg\]); a product of $T^B$ and $\Sigma^\alpha$ generators close within product, was investigated to formulate conformal gravity in two and four dimensions [@kawa-wata2]. [**6  D-dimensional generalized Chern-Simons actions**]{} It is important to realize that in the generalized Chern-Simons action the generalized gauge invariance is valid order by order in the form degree. In other words 0-form, 1-form, 2-form, 3-form, 4-form,$\cdots$ sectors of generalized Chern-Simons actions are separately invariant under the generalized gauge transformations of the gauge fields. To derive explicit forms of the generalized Chern-Simons actions with third grading $(c=0,1)$ for Lie algebra setting we introduce the following notations to clarify the differential form degrees: $$\renewcommand{\arraystretch}{1.4} \begin{array}{rcl} {\mathcal A} \!\!\!\!&=&\!\!\!\! {\bf 1} {\psi}_1 + {\bf i} \hat{\psi}_0 + {\bf j} {A}_0 + {\bf k} \hat{A}_1 \\ &=&\!\!\!\!{\bf 1}(\psi_1^{(1)} + \psi_1^{(3)} +\cdots ) + {\bf i} (\psi_0^{(0)} + \psi_0^{(2)} +\psi_0^{(4)} +\cdots ) \\ & &\!\!\!\! +{\bf j}(\omega_0^{(1)} + \Omega_0^{(3)} + \cdots ) +{\bf k} (\phi_1^{(0)} + B_1^{(2)} + H_1^{(4)}+ \cdots ), \\ {\mathcal V}\!\!\!\! &=&\!\!\!\! {\bf 1} \hat{a}_0 + {\bf i} {a}_1 + {\bf j} \hat{\alpha}_1 + {\bf k} {\alpha}_0 \\ &=&\!\!\!\! {\bf 1}(v_0^{(0)} + b_0^{(2)} + h_0^{(4)} +\cdots ) +{\bf i}(u_1^{(1)} + U_1^{(3)} +\cdots ) \\ & &\!\!\!\! +{\bf j}(\alpha_1^{(0)} + \alpha_1^{(2)} +\alpha_1^{(4)} \cdots ) + {\bf k}(\alpha_0^{(1)} + \alpha_0^{(3)} + \cdots ), \end{array} \label{GGFP-compo}$$ where $\phi_1^{(0)}, \omega_0^{(1)}, B_1^{(2)}, \Omega_0^{(3)}, H_1^{(4)}, \cdots$ are bosonic gauge fields of 0-, 1-, 2-, 3-, 4-form, $\cdots$ with the form degree in the parentheses. Similarly $v_0^{(0)}, u_1^{(1)}, b_0^{(2)}, U_1^{(3)}, h_0^{(4)}, \cdots$ are bosonic gauge parameters of 0-, 1-, 2-, 3-, 4-form, $\cdots$, respectively. Fermionic gauge fields and parameters are simply denoted by $\psi$ and $\alpha$ and the suffix “$i$” in $(i)$ denotes differential form degree. Hereafter we omit the differential form degree assignments for bosonic gauge fields and parameters for simplicity. Component expressions of the bosonic generalized Chern-Simons actions in 0-, 1-, 2-, 3-, 4-dimensions are given by S\_0\^k&=& ,\ S\_1\^j&=&,\ S\_2\^k &=&,\ S\_3\^j&=& ,\ S\_4\^k &=& , where $S^j_{(2k+1)}$ and $S^k_{(2k)}$ for $k=0,1,2, \cdots$ are, respectively, $(2k+1)$-dimensional and $(2k)$-dimensional actions. These generalized Chern-Simons actions are invariant under the following generalized gauge transformations for bosons: $$\renewcommand{\arraystretch}{1.4} \begin{array}{rcl} \delta\phi_1\!\!\!\!&=&\!\!\!\! [\phi_1,v_0] + \{\hat{\psi}_0^{(0)},\hat{\alpha}_1^{(0)}\} , \\ \delta\omega_0\!\!\!\!&=&\!\!\!\! dv_0 + [\omega_0,v_0] +\{\phi_1,u_1\} +[\psi_1^{(1)},\hat{\alpha}_1^{(0)}] -\{\hat{\psi}_0^{(0)},\alpha_0^{(1)}\} , \\ \delta B_1\!\!\!\!&=&\!\!\!\! -du_1-\{\omega_0,u_1\}+[\phi_1,b_0]+[B_1,v_0] +\{\hat{\psi}_0^{(0)}, \hat{\alpha}_1^{(2)}\} + [\psi_1^{(1)},\alpha_0^{(1)}] +\{\hat{\psi}_0^{(2)},\hat{\alpha}_1^{(0)}\} , \\ \delta\Omega_0\!\!\!\!&=&\!\!\!\! db_0+[\omega_0,b_0] +\{\phi_1,U_1\} + \{B_1,u_1\} +[\Omega_0,v_0] \\ &&\!\!\!\! -\{\hat{\psi}_0^{(0)},\alpha_0^{(3)}\}+[\psi_1^{(1)},\hat{\alpha}_1^{(2)}] -\{\hat{\psi}_0^{(2)},\alpha_0^{(1)}\}+[\psi_1^{(3)},\hat{\alpha}_1^{(0)}], \\ \delta H_1\!\!\!\!&=&\!\!\!\! -dU_1 -\{\omega_0,U_1\} +[\phi_1,h_0] + [B_1,b_0] - \{ \Omega_0,u_1\} +[H_1,v_0] \\ &&\!\!\!\! +\{\hat{\psi}_0^{(0)},\alpha_1^{(4)}\} + [\psi_1^{(1)},\alpha_0^{(3)}] +\{\hat{\psi}_0^{(2)},\hat{\alpha}_1^{(2)}\} + [\psi_1^{(3)},\alpha_0^{(1)}] +\{\hat{\psi}_0^{(4)},\hat{\alpha}_1^{(0)}\}, \end{array} \label{GGT-compo-boso}$$ and for fermions: $$\renewcommand{\arraystretch}{1.4} \begin{array}{rcl} \delta\hat{\psi}_0^{(0)} \!\!\!\!&=&\!\!\!\! [\hat{\psi}_0^{(0)},v_0] - \{\phi_1,\hat{\alpha}_1^{(0)}\} , \\ \delta\psi_1^{(1)} \!\!\!\!&=&\!\!\!\! -d \hat{\alpha}_1^{(0)} - [\omega_0,\hat{\alpha}_1^{(0)}] -[\phi_1,\alpha_0^{(1)}] - [\hat{\psi}_0^{(0)},u_1] + [\psi_1^{(1)},v_0] , \\ \delta\hat{\psi}_0^{(2)} \!\!\!\!&=&\!\!\!\! d \alpha_0^{(1)} + \{\omega_0,\alpha_0^{(1)}\} -\{\phi_1,\hat{\alpha}_1^{(2)}\} - \{B_1,\hat{\alpha}_1^{(0)}\} +[\hat{\psi}_0^{(0)},b_0] + [\psi_1^{(1)},u_1] +[\hat{\psi}_0^{(2)},v_0] , \\ \delta \psi_1^{(3)} \!\!\!\!&=&\!\!\!\! -d \hat{\alpha}_1^{(2)} -[\omega_0,\hat{\alpha}_1^{(2)}] -[\phi_1,\alpha_0^{(3)}] - [B_1,\alpha_0^{(1)}] - [\Omega_0,\hat{\alpha}_1^{(0)}] \\ &&\!\!\!\! -[\hat{\psi}_0^{(0)},U_1] +[\psi_1^{(1)},b_0] -[\hat{\psi}_0^{(2)},u_1] +[\psi_1^{(3)},v_0] , \\ \delta \hat{\psi}_0^{(4)} \!\!\!\!&=&\!\!\!\! d \alpha_0^{(3)} +\{\omega_0,\alpha_0^{(3)}\} - \{\phi_1,\alpha_1^{(4)}\} - \{B_1,\hat{\alpha}_1^{(2)}\} + \{\Omega_0,\alpha_0^{(1)}\} -\{H_1,\hat{\alpha}_1^{(0)}\} \\ &&\!\!\!\! +[\hat{\psi}_0^{(0)},h_0] + [\psi_1^{(1)},U_1] + [\hat{\psi}_0^{(2)},b_0] +[\psi_1^{(3)},u_1] + [\hat{\psi}_0^{(4)},v_0]. \end{array} \label{GGT-compo-fer}$$ It is possible to derive component expressions of fermionic generalized Chern-Simons actions in 0-, 1-, 2-, 3-, 4-dimensions. We do not have an interpretation of the actions corresponding to $S^1$ and $S^i$ in (\[GCSAs\]). [**7   2-grading gauge system**]{} We have given a formulation of higher form gauge theory in terms of the 3-grading structure with the quaternion algebra accommodating the generalized gauge system. We may wonder what happens if we consider only a 2-grading structure. It turns out that the quaternion structure is kept to classify ${\mathcal A}$-type and ${\mathcal V}$-type generalized gauge fields and parameters. In the current assignment of the 3-grading structure, a 2-grading formulation can be derived by simply setting fermionic gauge fields and parameters to be zero: = 0, =0. \[2-grading-sys\] In this setting a bosonic version of generalized Chern-Simons actions can be obtained with all the higher degrees of forms included. As one can see the leading terms of the actions in each dimensions have a typical $BF$-type structure except for 0 dimension. Let us specifically consider here the 3-, and 4-dimensional generalized Chern-Simons actions without fermions, which include the standard Chern-Simons action in 3 dimensions and $BF$ action in 4 dimensions, respectively: $$\renewcommand{\arraystretch}{2.0} \begin{array}{rcl} S_3^j(\psi=0)\!\!\!\!&=&\!\!\!\! \displaystyle \int\Tr\bigg[-\frac{1}{2}\omega_0d\omega_0 -\frac{1}{3}\omega_0^3 +\phi_1(dB_1 + [\omega_0,B_1]) -\Omega_0\phi_1^2 \bigg], \\ S_4^k(\psi=0) \!\!\!\!&=&\!\!\!\! \displaystyle\int\Tr\bigg[ -B_1(d\omega_0+\omega_0^2 ) -\phi_1(d \Omega_0+\{\omega_0,\Omega_0\}+B_1^2) - H_1 \phi_1^2 \bigg], \end{array} \label{3-4-GCSAs}$$ which are invariant under the following generalized gauge transformations: $$\renewcommand{\arraystretch}{1.4} \begin{array}{rcl} \delta\phi_1\!\!\!\!&=&\!\!\!\! [\phi_1,v_0], \\ \delta\omega_0\!\!\!\!&=&\!\!\!\! dv_0 + [\omega_0,v_0] +\{\phi_1,u_1\}, \\ \delta B_1\!\!\!\!&=&\!\!\!\! -du_1-\{\omega_0,u_1\}+[\phi_1,b_0]+[B_1,v_0], \\ \delta\Omega_0\!\!\!\!&=&\!\!\!\! db_0+[\omega_0,b_0] +\{\phi_1,U_1\} + \{B_1,u_1\} +[\Omega_0,v_0], \\ \delta H_1\!\!\!\!&=&\!\!\!\! -dU_1 -\{\omega_0,U_1\} +[\phi_1,h_0] + [B_1,b_0] - \{\Omega_0,u_1\} +[H_1,v_0]. \end{array} \label{GGT-compo-no-ferm}$$ One can easily see that these generalized actions are natural generalizations of the standard Chern-Simons action in 3 and 4 dimensions that include all the degrees of differential form. The 4-dimensional generalized Chern-Simons action includes the $BF$ action as a leading term. We can interpret this formulation as a generalization of the Chern-Simons action to arbitrary dimensions and thus we call these actions generalized Chern-Simons actions. It should be noted here that in the generalized gauge transformations (\[GGT-compo-no-ferm\]) commutators and anti-commutators are mixed. However after taking into account the odd-form nature and the odd-grading nature of the suffix “1”, all the anti-commutators turn into commutators so that algebra is closed within the Lie algebra as it was already noted after Eqs. (\[GGTs\]). It is important to realize that all the even-form generalized gauge fields in (\[3-4-GCSAs\]) and all the odd-form generalized gauge parameters in (\[GGT-compo-no-ferm\]) carry the suffix $1$. The even-form gauge fields and the odd-form gauge parameters have a hidden grading structure, whose nature is not completely determined and could also be interpreted as a grading of fermionic nature. [**8  Equations of motions for generalized Chern-Simons actions**]{} The equations of motion for the generalized Chern-Simons actions can be derived and are given by the vanishing condition for the generalized curvature: $${\mathcal F} = Q{\mathcal A}+{\mathcal A}^2 ={\bf 1}{\mathcal F}^1+{\bf i}{\mathcal F}^i+{\bf j}{\mathcal F}^j + {\bf k}{\mathcal F}^k = 0,$$ with $$\renewcommand{\arraystretch}{1.4} \begin{array}{rcl} {\mathcal F}^1\!\!\!\!&=&\!\!\!\! -dA_0 - A_0^2 - \hat{A}_1^2+ \psi_1^2 - \hat{\psi}_0^2 = 0, \\ {\mathcal F}^i \!\!\!\!&=&\!\!\!\! d\hat{A}_1 + [A_0, \hat{A}_1] + \{\psi_1,\hat{\psi}_0\}= 0, \\ {\mathcal F}^j \!\!\!\!&=&\!\!\!\! d\psi_1 + \{A_0, \psi_1 \} + [\hat{A}_1, \hat{\psi}_0] = 0, \\ {\mathcal F}^k \!\!\!\!&=&\!\!\!\! -d\hat{\psi}_0 -[A_0,\hat{\psi}_0] + \{\hat{A}_1,\psi_1\}=0. \end{array} \label{GEMs}$$ Notice that the first line of (\[GEMs\]) is a generalization of the Maurer-Cartan equation and it admits the solution = [G]{}\^[-1]{} Q [G]{}, with = e\^[i[V]{}]{}, where ${\mathcal V}$ is a generalized parameter as in Eq. (\[gene-nonAbelian\]). Component expressions for the equations of motion can be derived by equating to zero separately the sectors of the generalized curvature with different form degree. Bosonic equations of motion for 0-, 1-, 2-, 3-, 4-form sectors are given by: $$\renewcommand{\arraystretch}{1.4} \begin{array}{rcl} -\phi_1^2 -(\hat{\psi}_0^{(0)})^2 \!\!\!\!&=&\!\!\!\! 0, \\ d\phi_1 +[\omega_0,\phi_1] +\{\psi_1^{(1)},\hat{\psi}_0^{(0)}\} \!\!\!\!&=&\!\!\!\! 0, \\ -d\omega_0 -\omega_0^2 - \{\phi_1,B_1\} + (\psi_1^{(1)})^2 -\{\hat{\psi}_0^{(0)},\hat{\psi}_0^{(2)}\} \!\!\!\!&=&\!\!\!\! 0, \\ dB_1 + [\omega_0,B_1] + [\Omega_0,\phi_1] + \{\psi_1^{(1)},\hat{\psi}_0^{(2)}\} + \{\psi_1^{(3)},\hat{\psi}_0^{(0)}\} \!\!\!\!&=&\!\!\!\! 0, \\ -d\Omega_0 -\{\omega_0,\Omega_0\} -\{\phi_1,H_1\} - B_1^2 + \{\psi_1^{(1)},\psi_1^{(3)}\} - \{\hat{\psi}_0^{(0)},\hat{\psi}_0^{(4)}\} -(\hat{\psi}_0^{(2)})^2 \!\!\!\!&=&\!\!\!\! 0. \end{array} \label{GEM-compo-boson}$$ And corresponding fermionic equations of motion are $$\renewcommand{\arraystretch}{1.4} \begin{array}{rcl} [\phi_1, \hat{\psi}_0^{(0)}] \!\!\!\!&=&\!\!\!\!0, \\ -d\hat{\psi}_0^{(0)} -[\omega_0,\hat{\psi}_0^{(0)}] + \{\phi_1,\psi_1^{(1)}\} \!\!\!\!&=&\!\!\!\!0, \\ d\psi_1^{(1)} + \{\omega_0,\psi_1^{(1)}\} +[\phi_1,\hat{\psi}_0^{(2)}] + [B_1,\hat{\psi}_0^{(0)}] \!\!\!\!&=&\!\!\!\! 0, \\ -d\hat{\psi}_0^{(2)} - [\omega_0,\hat{\psi}_0^{(2)}] +\{\phi_1,\psi_1^{(3)}\} +\{B_1,\psi_1^{(1)}\} -[\Omega_0,\hat{\psi}_0^{(0)}] \!\!\!\!&=&\!\!\!\! 0, \\ d\psi_1^{(3)} +\{\omega_0,\psi_1^{(3)}\} + [\phi_1,\hat{\psi}_0^{(4)}] +[B_1,\hat{\psi}_0^{(2)}] +\{\Omega_0,\psi_1^{(1)}\} +[H_1,\hat{\psi}_0^{(0)}] \!\!\!\!&=&\!\!\!\! 0. \end{array} \label{GEM-compo-ferm}$$ The equations of motion of 3- and 4-dimensional generalized Chern-Simons actions in (\[3-4-GCSAs\]) are simply given by setting $\psi=0$ in (\[GEM-compo-boson\]). It is interesting to realize that these generalized equations of motion have the generalized gauge invariance. [**9  Generalized topological relations**]{} A generalized gauge theory version of the second Chern character can be defined and it is related to the generalized Chern-Simons action in analogy with the standard gauge theory: ([F]{}\^2) = Q( [A]{}Q[A]{} +\^3 ) \^1 + [**i**]{}[G]{}\^i + [**j**]{}[G]{}\^j + [**k**]{}[G]{}\^k , \[top-rel\] where ${\mathcal F}$, ${\mathcal A}$ and $Q$ are defined (\[GEMs\]), (\[GGFP-compo\]) and (\[gene-d-op\]), respectively. From the sector-wise equivalence for the quaternion in (\[top-rel\]) we can obtain the following new topological relations: $$\renewcommand{\arraystretch}{2.0} \begin{array}{rcl} {\mathcal G}^1 \!\!\!\!&\equiv&\!\!\!\! \displaystyle\int \Tr\left(({\mathcal F}^1)^2 - ({\mathcal F}^i)^2 - ({\mathcal F}^j)^2 - ({\mathcal F}^k)^2\right) = -2\int\Tr\, d{\mathcal L}_{GCS}^j, \\ {\mathcal G}^i \!\!\!\!&\equiv&\!\!\!\! \displaystyle\int\Tr\left( \{{\mathcal F}^1, {\mathcal F}^i\} +[{\mathcal F}^j,{\mathcal F}^k] \right) = 2\int\Tr\, d{\mathcal L}_{GCS}^k, \\ {\mathcal G}^j \!\!\!\!&\equiv&\!\!\!\! \displaystyle\int\Tr\left( \{{\mathcal F}^1, {\mathcal F}^j\} +[{\mathcal F}^k,{\mathcal F}^i] \right) = 2\int\Tr\, d{\mathcal L}_{GCS}^1, \\ {\mathcal G}^k \!\!\!\!&\equiv&\!\!\!\! \displaystyle \int\Tr\left( \{{\mathcal F}^1, {\mathcal F}^k\} +[{\mathcal F}^i,{\mathcal F}^j] \right) = -2\int\Tr\, d{\mathcal L}_{GCS}^i, \end{array} \label{top-rel-compo}$$ where ${\mathcal F}^A\, (A=1,i,j,k)$ are given in (\[GEMs\]) and ${\mathcal L}_{GCS}^A \, (A=1,i,j,k)$ are given by the following relations: S\^A =\_[GCS]{}\^A, (A=1,i,j,k), with $S^A$ given in (\[GCSAs\]). It should be recognized here that $\Tr({\mathcal F}^2)$ and correspondingly the right hand side in (\[top-rel\]) have the ${\mathcal V}$-type quaternion structure. Therefore ${\mathcal G}^1$ and ${\mathcal G}^i$ are bosonic even- and odd-form sectors while ${\mathcal G}^j$ and ${\mathcal G}^k$ are fermionic even- and odd-form sectors respectively. Reduction to 2-grading formulation which includes all degrees of differential forms for bosonic gauge fields with hidden grading for even-forms, can be easily obtained by simply setting $\psi=0$ in ${\mathcal G}^1$ and ${\mathcal G}^i$ while ${\mathcal G}^j$ and ${\mathcal G}^k$ disappear. For a given space-time dimension the expressions of ${\mathcal G}^A$ can be derived by extracting in (\[top-rel-compo\]) the sector of the corresponding from degree. For example 4-dimensional counter part of the generalized topological relation without fermions is given by \^1\_4 &=&\ &=& d, \[topo-rel-compo\] where the last term is the exterior derivative times the 3-dimensional generalized Chern-Simons action in (\[3-4-GCSAs\]). It should be noted that 0-, 2-, 4-form gauge fields $\phi_1,~B_1$ and $H_1$ carry a hidden $\mathbb{Z}_2$-grading. Notice also that some nontrivial cancellations occur in going from the second to the third term due to the hidden grading. It is interesting to realize that the topological Yang-Mills action is included as the first term in the action in (\[topo-rel-compo\]). [**10  Generalized differential operators**]{} In the component expressions for the actions given in (\[GCSA-compo\]) we have specifically assigned the first and the second grading as (fermion, boson) and (even form, odd form) so that the second grading is related to the space-time structure of the generalized fields and parameters. There is some freedom in the interpretation of gradings. A physical meaning of the third grading is unspecified in this paper. This assignment of gradings automatically generates an asymmetric nature of exterior derivative operator with respect to the quaternion structure as in (\[gene-d-op\]). This special assignment is the origin of the difference between the action $S^j$ and the rest of the actions and the similarity of $S^1,S^i$ and $S^k$, as is already pointed out. We have introduced fermionic fields which have an anti-symmetric space-time tensor structure. There is no spinor introduced. One may wonder what kind of fermions they are. It was shown that these fermions can be identified as ghost fields of quantized generalized Chern-Simons, which turned out to be an infinitely reducible gauge system [@GCS-quant]. It was noticed that the generalized differential operator can be extended to: Q=-[**i**]{}s + [**j**]{}d, \[extended-Q1\] where $s$ is a fermionic zero form and can be interpreted as a BRST operator. The generalized Chern-Simons action (\[GCSA\]) has the gauge invariance with the generalized gauge parameters which turn into ghosts after quantization. It is surprising to find that ${\mathcal V}$-type gauge parameters turn into ghosts as ${\mathcal A}$-type gauge fields due to the change into alternative grading for fermion and boson. This structure is realized as a BRST transformation by the addition of $-{\bf i}s$ term for the $Q$-operator in (\[extended-Q1\]). It is fundamentally reflected from the fact that the generalized gauge transformation is infinitely reducible [@GCS-quant]. The quantization of the generalized Chern-Simons action[@GCS-quant] was completed by BV and BFV formalism [@BV; @BFV]. BV formalism is the special realization of Q-P manifold, [*à la*]{} AKSZ formalism [@AKSZ]. The connection of the current formulation of the generalized gauge theory and the AKSZ formalism is of a great interest. It is expected that the topological particle field theory formulation which was discovered to be equivalent to the quaternion formulation of the generalized gauge theory plays an important role [@TPFT]. As a generalization of the $Q$ operator (\[extended-Q1\]) we propose the most symmetric differential operator: Q=-[**1**]{}sd-[**i**]{}s + [**j**]{}d +[**k**]{}, \[extended-Q2\] where $\sigma$ is an exterior derivative operator corresponding to the 3rd grading. This $Q$ operator satisfies the item 1. and 2. in section 4 and thus define a more general and symmetric generalized gauge theory. [**11   Conclusions and discussions**]{} The generalized Chern-Simons gauge theory to accommodate all the degrees of differential forms as Lie algebra valued gauge fields and parameters is explicitly formulated for arbitrary even and odd dimensions. The role of quaternion in the formulation is clarified as an alternative presentation for three grading structure of the generalized gauge fields and parameters. The connection of the Lie algebra formulation with the original graded Lie algebra formulation is also clarified. In considering a gauge system of three physically independent gradings, it is natural to assume that these gradings commute each other. For example the differential form grading of space-time and the quantized ghost grading have no physical connection in general. Thus we assume that these gradings are commutative as in (\[grading1\]). In this physical system the quaternion accompanied formulation of the generalized gauge theory is necessary. On the other hand this formulation can be equivalently described by the formulation of the total three gradings of (\[grading2\]). There has been recently a growing interest in higher form gauge theory as a way of finding a unified formulation of extended objects such as D-branes [@string-related1; @string-related2; @string-related3] and of getting a deeper understanding of duality related prescriptions for gauge theories [@Seiberg]. We consider that the present formulation as well as its original version [@kawa-wata] have possible applications in this new context and also in the older formulations of topological field theories [@BF], Chern-Simons related gravities [@Witten-3dgravity; @kawa-wata2; @CSgf-ext], and higher spin formulations [@higher-spin]. It is also known that the standard Chern-Simons and BF actions lead simplicial gravities of Ponzano-Regge type in 3 dimensions and 15-$j$ topological gravity type in 4 dimensions [@PR-G; @BF-G]. One may wonder what is the role of the extra differential forms other than the ones constructing gravity background. The generalized gauge theory formulation suggests a hope that we may be able to formulate a simplicial gravity theory with matter introduced naturally on the simplicial lattice by simplex-form correspondence [@NK-R]. Finally we would like to point out that physical identification of the unknown third grading would be very important for finding the generalized gauge symmetry. [99]{} E. Witten, Nucl. Phys. [**B311**]{} (1988) 46. N. Kawamoto and Y. Watabiki,  Commun. Math. Phys. [**144**]{} (1992) 641; Mod. Phys. Lett. [**A7**]{} (1992) 1137. P. Deligne and D. Freed, in Appendix “Sign manifesto” in “Supersolutions” of “Quantum Fields and Strings, A course for mathematicians”, 2 vols. Amer. Math. Soc. Providence 1999, . M.V. Movshev and A. Schwarz, Proc. Symp. Pure Math. [**90**]{} (2015) 327,\ . N. Kawamoto and Y. Watabiki, Nucl. Phys. [**B396**]{} (1993) 326; Phys. Rev. [**D45**]{} (1992) 605. N. Kawamoto, K. Suehiro, T. Tsukioka, and H. Umetsu, Nucl. Phys. [**B532**]{} (1998) 429, . I.A. Batalin and G.A. Vilkovisky, Phys. Lett. [**102B**]{} (1981) 27, Phys. Rev. [**D28**]{} (1983) 2567. E.S. Fradkin and G.A. Vilkovisky, Phys. Lett. [**55B**]{} (1975) 224.\ I.A. Batalin and E.S. Fradkin, Phys. Lett. [**122B**]{} (1983) 157, Phys. Lett. [**128B**]{} (1983) 303. A.Y. Alexandrov, M. Kontsevich, A. Schwartz, and O. Zaboronsky, Int. J. Mod. Phys. [**A12**]{} (1997) 1405, . N. Kawamoto and Y. Watabiki, Commun. Math. Phys. [**148**]{} (1992) 169. T. Takayanagi, S. Terashima, and T. Uesugi, JHEP [**0003**]{} (2001) 019,\ .\ T. Asakawa, S. Sugimoto, and S. Terashima, JHEP [**0203**]{} (2002) 034,\ . J. Bagger and N. Lambert, Phys. Rev.[**D75**]{} (2007) 045020, ; Phys. Rev. [**D77**]{} (2008) 065008, .\ A. Gustavsson, Nucl. Phys. [**B811**]{} (2009) 66, . J.C. Baez and J. Huerta, Gen. Rel. Grav. [**43**]{} (2011) 2335, . T. Banks and N. Seiberg, Phys. Rev. [**D83**]{} (2011) 084019, .\ D. Gaiotto, A. Kapustin, N. Seiberg, and B. Willett, JHEP [**1502**]{} (2015) 172,\ . D. Birmingham, M. Blau, M. Rakowski, and G. Thompson, Phys. Rep. [**209**]{} (1991) 129. S.W. MacDowell and F. Mansouri, Phys. Rev. Lett. [**38**]{} (1977) 739, \[Erratum-ibid. [**38**]{}, (1977) 1376\].\ K.S. Stelle and P. C. West, Phys. Rev. [**D21**]{}, (1980) 1466.\ A. D’Adda, Conference:C80-07-21. p.268-299 Proceedings. M.A. Vasiliev, Nucl. Phys. [**B616**]{} (2001), \[Erratum-ibid. [**652**]{}, (2003) 407\],\ . M. Caselle, A. D’Adda, and L. Magnea, Phys. Lett. [**232B**]{} (1989) 457.\ N. Kawamoto, H.B. Nielsen, and N. Sato, Nucl. Phys. [**B555**]{} (1999) 625,\ . N. Kawamoto, N. Sato, and Y. Uchida, Nucl. Phys. [**B574**]{} (2000) 809,\ . N. Kawamoto, Prog. Theor. Phys. Suppl. [**134**]{} (1999) 629, . [^1]: `[email protected]` [^2]: `[email protected]` [^3]: `[email protected]` [^4]: `[email protected]`
{ "pile_set_name": "ArXiv" }
--- abstract: 'The possibility of performing the experimental research in the field of fundamental physics based on the unique instrument – European X-ray Free Electron Laser (E-XFEL) is considered in this paper. The calculations of the reaction $\gamma + \gamma \to e^{+} + e^{-}$ cross section for gamma quanta with $E \sim~(1-100)\,\mbox{GeV}$ energy with X-ray photons are performed. The possibility of experimental registration of reaction $\gamma + \gamma \to e^{+} + e^{-}$ product are reviewed. Also, the optical depth of the interaction between gamma-rays with the E-XFEL’s photon pulses is estimated. Astrophysical applications are discussed.' address: | $^1$ Ioffe Institute, Saint Petersburg, Russia\ $^2$ DESY, Hamburg, Germany author: - 'A N Popov$^{1}$, S V Bobashev$^{1}$, N O Bezverkhnii$^{1}$, A A Sorokin$^{1,2}$' title: 'On the possibility of research the photon-photon interaction at the European X-ray Free Electron Laser – European XFEL ' --- Introduction ============ The scattering of two photons is a well-known quantum process which exhibits the non-linear nature of quantized electromagnetic fields. Although light-by-light scattering was predicted almost 100 years ago its direct experimentally observation still remains to be discovered.Until today, several ways of directly detected this phenomenon in the laboratory have been proposed. For example, Compton-backscattered photons against laser photons [@7], collisions of photons from microwave waveguides [@8] or high-power lasers [@9; @10; @11], and at photon colliders [@12; @13; @holeraum]. In [@LHCPHPH] using computer simulations, the authors show that lead-lead collisions provide the opportunity for seeing a scattering of quasireal photons in the LHC. Direct observation of photon-photon scattering is worth the effort because of any deviation from the predicted could be evidence of new physics, e.g., the existence of supersymmetry. In this paper, we discuss the possibilities of conducting experimental research in the field of fundamental physics based on the E-XFEL. The E-XFEL is the international research project which involved 12 states. It was officially launched on September 1, 2017. Compared with modern a synchrotron light sources, the peak brightness of the E-XFEL is more than 8 orders of magnitude higher, the radiation has a high degree of transverse coherence and pulse duration is about 10 fs. The particularities of the E-XFEL’s radiation make it possible to conduct groundbreaking research in fundamental physics. Table 1 shows the X-ray beam parameters of one of the three channels of the E-XFEL. [@\*[4]{}[l]{}]{} Parameter&\ Energy $e^-$ (GeV) & $17.5$ & $17.5$ & $10.0$\ Wavelength (nm) & $0.4$ & $1.6$ & $6.4$\ Eneregy $\gamma$ (keV) & $3.1$ & $0.8$ & $0.2$\ Peak power (GW) & $80$ & $130$ & $135$\ Average power (W) & $260$ & $420$ & $580$\ FWHM ($\mu m$) & $60$ & $70$ & $95$\ Beam divergence ($\mu rad$) & $3.4$ & $11.4$ & $27$\ Coherence time (fs) & $0.34$ & $0.88$ & $1.9$\ Spectral width (%) & $0.2$ & $0.3$ & $0.73$\ Pulse duration (fs) & $100$ & $100$ & $100$\ Photons in impulse & $1.6 \cdot 10^{13}$ & $1.0 \cdot 10^{14}$ & $4.3 \cdot 10^{14}$\ Average flux (phot/s) & $5.2 \cdot 10^{17}$ & $3.4 \cdot 10^{18}$ & $1.4 \cdot 10^{19}$\ Peak brilliance & $2.0 \cdot 10^{33}$ & $5.0 \cdot 10^{32}$ & $0.6 \cdot 10^{32}$\ Average brilliance & $6.4 \cdot 10^{24}$ & $1.6 \cdot 10^{24}$ & $2.0 \cdot 10^{24}$\ ![\[fig\_structure\] Comparison of gamma photon sources according to data from [@PhysRevX.7]. The “Gold target” is the bremsstrahlung gamma-ray source from [@holeraum].](impulse_structure.eps){width="14pc"} ![\[fig\_structure\] Comparison of gamma photon sources according to data from [@PhysRevX.7]. The “Gold target” is the bremsstrahlung gamma-ray source from [@holeraum].](Sources.eps){width="16pc"} Experimental proposal and possible gamma-ray sources ==================================================== The experimental proposal in simplified form is as follows. A high-energy photon beam, generated via the gamma-ray source is fired into a vacuum channel where it intersects with the E-XFEL’s x-ray photon beam. The scheme takes advantage of the high photon densities that are characteristic of the radiation fields produced by E-XFEL. The vacuum channel is surrounded by multi-layer charged particle detectors for detecting elector-positron pairs. The detectors are protected from external radiation and cosmic rays. One of the central problems of this proposal is the gamma-ray source. In the work [@holeraum], it is proposed to use bremsstrahlung radiation of ultrarelativistic electrons with an energy of about GeV passing through a gold target to create a gamma-ray beam. Electrons and positrons emitted from the back surface of a gold target deviate from the interaction region using a magnetic field. Another way is proposed in a recent paper [@PhysRevX.7]. Electromagnetic dipole cascades can become one of the possible sources of gamma. The paper demonstrated the possibility of creating a stable directional source of photons of GeV energies. Using 3D QED particle-in-cell modeling and analytical evaluations, the authors show that it is possible to attain a peak power of 10 PW. When this becomes possible, there will be a real possibility to carry out this kind of experiments. Fig. 2 shows the approximate location of possible gamma-ray sources on a brilliance-energy map. The optical depth estimation ============================ The temporal structure of the E-XFEL laser, shown in Fig. 1, is a sequence of trains consisting of 2700 femtosecond pulses. The total repetition rate of the trains is 10 Hz. For one train, the optical depth becomes equal to $5.4 \cdot 10^{-4}$, that is, when the gamma-quantum crosses 100 trains, the optical depth becomes equal to 1. The cross section of interaction of gamma quantum to thermal photon with producing electron-positron pair is [@Gould1967_cross_section]: $$\sigma = \frac{\pi}{2} \, r_{e}^{2} \, (1 - v^{2}) \, \left( (3 - v^{4}) \, {\mathrm{ln}} \left( \frac{1+v}{1-v} \right) - 2\, v \, (2 - v^{2}) \right) \, h(s)$$ where $r_{e} = \frac{e^{2}}{m c^{2}}$ is classical electron radius, $m$ is mass of electron, $h(s)$ is Heaviside function ($h(s) = 1$ at $s>0$ and $h(s) = 0$ at $s<0$), $$v = \sqrt{ 1 - 1/s \, } \mbox{\ \ and \ \ } s = \frac{ E \epsilon }{ m^{2} c^{4} } ( 1 - \cos\Psi ) , \label{eqn_1}$$ $E$ is energy of gamma quantum, $\epsilon$ is energy of thermal photon, radiated by intracluster gas, $\Psi$ is angle between its impulses. Consider a situation where photons move towards each other, i. e. $\theta = \pi$. the X-ray beam in XFEL bunch reaches 100 photons on square $10^{-16} \mbox{ cm}^2$ for impulse in order 10 fs. It remains to make a closer study of number of photons are per area, equal to area of the reaction cross section. Put in (\[eqn\_1\]) the product of the energies $\epsilon E$ equal to $1.2 m_e^2 c^4$. Then (for $\theta = \pi$) $\sigma \approx 2 \cdot 10^{-25} \mbox{ cm}^2$. This means that the optical depth $\tau \approx 2 \cdot 10^{-7}$ is collected for one pulse. The magnitude of the cross section depends on the angle of interaction (Fig. 3). Figure 4 shows the dependence of the reaction cross section on the energy of the oncoming gamma photon. ![\[fig\_v6\_L\] The dependence of cross-section $\sigma$ on gamma photon energy with three values of X-ray photon energy. ](angle.eps){width="16pc"} ![\[fig\_v6\_L\] The dependence of cross-section $\sigma$ on gamma photon energy with three values of X-ray photon energy. ](energy.eps){width="16pc"} Astrophysical applications ========================== Another reason to research this problem are numerous astrophysical applications. For example, two-photon pair-production processes are one of the mechanisms of electron-positron plasma generation in the popular outer gap model of magnetospheric emission from pulsars [@MonthlyNotices]. Another example is gamma-ray emission from cosmological sources. The observed gamma-ray spectrum of them contains information about physical conditions and composition of matter in the Universe at various stages of its evolution. In this case, it is very important to take into account the factors leading to spectral distortions as absorbtion due to the interaction with background photons. The process of absorption of gamma rays with the production of electron-positron pairs in the collision of them with cosmological background photons [@ruffini] makes the most important contribution to distortion of spectrum in the gamma-ray range. For gamma quanta with energies $E \sim 100 \mbox{ GeV} \div 100 \mbox{ TeV}$, this is mainly interaction with clusters of galaxies, which are the largest gravitationally bound objects in the Universe. The noticeable part of the baryon component of the cluster is a rarefied ($n \sim 10^{- 3} \div 10^{-2} \mbox{cm}^{-3}$) hot ($T \sim 1 \div 10 \mbox {keV}$) intracluster gas [@ruffini; @vikhlinin2006]. Moreover, due to the bremsstrahlung of electrons of this gas, clusters are also transformed into huge photon reservoirs [@lea]. Therefore, in addition to the components of the intergalactic background, distortions in the spectrum of distant cosmological sources of gamma radiation can also be caused by the interaction of gamma-rays with photons radiated by gas in cluster of galaxies. The authors of [@popov] have found out that the interaction with the photons of the bremsstrahlung of gas in clusters of galaxies makes a small contribution to the optical depth, compared with absorption by microwave, infrared, visible and radio photons, for almost all gamma-ray energies. However, for energies $E \sim 1-100 \, \mbox{GeV}$, the scattering effect on photons of the bremsstrahlung galaxies can dominate and amounts to $\tau \approx 10^{-5}$. Conclusions =========== In conclusion, two-photon pair-production processes should easily be observable at this kind experiment on E-XFEL. This experiment can be considered both as a test of the standard model and as a simulation of astrophysical processes. It is a good possibility to research the absorption of gamma radiation in the laboratory. However, the implementation of the experiment described in this article is a task with many problems: 1. The real source of gamma-ray. 2. The photon interaction geometry. 3. Special aspects of the interaction of photons with coherent E-XFEL’s radiation. These problems require further study and will be considered in the further papers. We sincerely thank D. P. Barsukov for help, comments and useful discussions. References {#references .unnumbered} ========== [99]{} Karnig O.Mikaelian [ *Phys. Lett. B*]{}, [**115**]{}, 267-269 (1982) Eriksson, Daniel and Brodin, Gert and Marklund, Mattias and Stenflo, Lennart [ *Phys. Rev. A*]{}, [**70**]{}, 013808 (2004) F. Moulin, D. Bernard, F. Amiranoff [ *F. Z. Phys. C - Particles and Fields*]{}, [**72**]{}, 607 (1996) Bernard, D., Moulin, F., Amiranoff, F. et al. [ *Eur. Phys. J. D*]{}, [**10**]{}, 141 (2000) E. Lundström, G. Brodin, J. Lundin, M. Marklund, R. Bingham, J. Collier, J. T. Mendonça, and P. Norreys [ *Phys. Rev. Lett.*]{}, [**96**]{}, 083602 (2006) S. Chattopadhyay, A.M. Sessler, Nuclear Instruments and Methods in Physics Research Section A, [**355**]{}, 19-41 (1995) G. Jikia and A. Tkabladze [ *Physics Letters B*]{}, [**323**]{}, 453-458 (1994) O. J. Pike, F. Mackenroth, E. G. Hill and S. J. Rose [ *Nature Photonics*]{}, [**8**]{}, 434–436 (2014) David d’Enterria, and Gustavo G. da Silveira, [ *Phys. Rev. Lett.*]{} [**11**]{}, 080405 (2013) A. Gonoskov, A. Bashinov, S. Bastrakov, et al. [*Phys. Rev. X*]{} [**7**]{}, 041003 (2017) Gould R J and Schreder G P 1967 [*Phys. Rev.* ]{} [**155**]{} 1404-7 Daniele Viganò, Diego F. Torres, Kouichi Hirotani, Martín E. Pessah [ *Monthly Notices of the Royal Astronomical Society*]{}. [**447**]{}, 2631–2648 (2015) Ruffini R., Vereshchagin G. V., Xue, S. S. [ *Astrophysics and Space Science*]{}. [**361**]{}, id 82 (2016) A. Vikhlinin, A. Kravtsov, et al. [ *Astrophys. J.*]{} [**640**]{}, 691 (2006) R. A. Sunyaev, Ya. B. Zeldovich [ *Comm. Astrophys. Sp. Phys*]{} [**4**]{}, 173 (1972) S. M. Lea, J. Silk, E. Kellogg, S. Murray [ *Astrophys. J.*]{} [**184**]{}, L105 (1973) A. N. Popov, D. P. Barsukov, A. V. Ivanchik [ *PAZh*]{}, [**44**]{}, 633 (2018)
{ "pile_set_name": "ArXiv" }
--- abstract: 'We develop a theory of amorphous interfaces in glass-forming liquids. We show that the statistical properties of these surfaces, which separate regions characterized by different amorphous arrangements of particles, coincide with the ones of domain walls in the random field Ising model. A major consequence of our results is that super-cooled liquids are characterized by two different static lengths: the point-to-set $\xi_{PS}$ which is a measure of the spatial extent of cooperative rearranging regions and the wandering length $\xi_\perp$ which is related to the fluctuations of their shape. We find that $\xi_\perp$ grows when approaching the glass transition but slower than $\xi_{PS}$. The wandering length increases as $s_c^{-1/2}$, where $s_c$ is the configurational entropy. Our results strengthen the relationship with the random field Ising model found in recent works. They are in agreement with previous numerical studies of amorphous interfaces and provide a theoretical framework for explaining numerical and experimental findings on pinned particle systems and static lengths in glass-forming liquids.' author: - Giulio Biroli Chiara Cammarota bibliography: - 'bib.bib' title: 'Fluctuations and Shape of Cooperative Rearranging Regions in Glass-Forming Liquids' --- Introduction ============ Developing a theory of the glass transition remains one of the most fundamental challenge of statistical physics and condensed matter. The interest in this problem actually goes well beyond the physics of molecular super-cooled liquids. The reason is that glassy behaviour is ubiquitous; it appears in a large variety of contexts: from physical systems like colloids and granular material to central problems in other branches of sciences like computer science, economics and biology. Recent years have witnessed important and substantial progress in its understanding. Several theoretical approaches have grown in importance and in the level of detailed predictions and explanations [@birgar13]. In particular, the Random First Order Transition (RFOT) theory originally introduced by Kirkpatrick, Thirumalai and Wolynes [@kithwo89] has been boosted from new theoretical ideas and techniques [@mezpar99; @boubir04; @wollub12] and innovative simulation studies [@cagrve07; @bibcgv08; @parisi09; @sautar10; @chchta12; @berkob12; @homare12]. The initial idea of Kirkpatrick, Thirumalai and Wolynes that super-cooled liquids are in a mosaic state, a kind of micro phase separated state in which the number of possible phases is huge, has been made concrete and testable in analytical computations and numerical simulations. We understand now how to define and—measure—the spatial extent of amorphous order, [*i.e.*]{} the length-scale over which particles (or molecules) in super-cooled liquids are arranged in an ordered, even though apparently chaotic, fashion. Numerical simulations have shown that this length $\xi_{PS}$, called point-to-set, grows upon super-cooling and plays an important role in the static and dynamical behaviours. In this work we unveil the existence of a second static length-scale, that together with $\xi_{PS}$ is central to the physics of super-cooled liquids and rules the relaxation within the RFOT picture.\ The main physical ingredient of RFOT are the surface tension $\Upsilon$ and the configurational entropy density $s_c(T)$. The former is a measure of the extra free-energy cost paid when two different amorphous phases are in contact through a common surface. The latter quantifies the multiplicity of amorphous phases in which the liquid can freeze. The mosaic state results from the competition between the configurational entropy gain due to local fluctuations between all possible amorphous states (all different type of “tiles” of the mosaic) and the surface energy loss due to the mismatch at the boundary between two amorphous states. A lot of analytical and numerical works have been devoted to characterise the spatial extent of the “tiles” of the mosaic, also called cooperative rearranging regions (CRR). Very few investigations have instead focused on their [*interfaces*]{}, for which no clear picture has arisen yet. Analytical studies based on Kac-models describe these interfaces as flat [@frasem11; @zarfra10], whereas numerical works instead suggest that these interfaces wander similarly to domain walls in disordered magnets [@bibcgv08; @cacggv09a; @cacggv09b]. Their width has been directly measured in recent numerical simulations and showed to grow mildly when temperature is lowered [@parisi09; @berkob12; @korobe12; @sckobp02; @grtcgv13; @hobekr14]. The physical reason for this growth is still to be elucidated though. Basic questions remain unanswered: How much do these interfaces fluctuate? How do their fluctuations depend on temperature, in particular do interfaces become rougher or flatter approaching the glass transition? How does their width compare with the other characteristic length, the point-to-set length, which measures the spatial extent of CRRs? In order to answer all these questions, fully characterise the real space structure of super-cooled liquids and the mosaic state advocated by RFOT theory it is crucial to develop a complete theory of fluctuating interfaces between amorphous states. This task is particularly timely since amorphous interfaces have started to be directly probed in recent numerical simulations on pinned particle systems [[[@parisi09; @berkob12; @korobe12; @sckobp02; @grtcgv13; @hobekr14]]{}]{}, and the first experimental results obtained by using optical traps in colloidal liquids have just come out [@gonags14].\ In this work we develop such a theory, obtain detailed predictions and provide explanations for previous numerical findings. We show that interfaces are rough because pinned by self-induced disorder and that they are characterised by wandering exponents identical to the ones of domain walls in the Random Field Ising Model (RFIM), thus strengthening the relationship between super-cooled liquids and the RFIM found in [[[@stewol08; @frparr11; @cambir13; @frapar13; @bicatt14]]{}]{}. [[These results allow us to establish that super-cooled liquids are characterized by two different static lengths, $\xi_{PS}$ and $\xi_\perp$, which measure respectively the spatial extent and the fluctuations of the shape of CRRs. Their scaling with respect to $s_c$ is different: $\xi_{PS}\propto s_c(T)^{-1/(3-\theta)}$ with $\theta>1$, whereas $\xi_\perp\propto s_c(T)^{-1/2}$. Therefore, when $s_c$ decreases approaching the glass transition, boundary fluctuations grow but less strongly than the linear size. The resulting shape of CRRs is shown pictorially in Fig. 1.]{}]{} [[In order to study amorphous interfaces we adapt the theoretical protocol that has been used to characterize the size of CRRs [@boubir04]. We focus on equilibrated configurations, $\cal C'$, constrained to have a high overlap behind an infinite plane with a reference equilibrium configuration, $\cal C$. This can be operativly realised by taking an equilibrium configuration, pinning all particles behind a plane and resampling the configuration of the remaining free particles. The overlap field between $\cal C$ and $\cal C'$ is bound to be high close to the plane and to reach a low value, characteristic of bulk behavior, far from it. We define the [*amorphous interface*]{} for a given $\cal C'$ as the surface separating the high overlap region from the low overlap one.]{}]{} [[Within RFOT theory, the physics behind the formation of amorphous interfaces is similar to wetting [@fisher84]: on the one hand it is favourable for the system to change amorphous state (or CRRs) beyond the plane because this allows a net gain in configurational entropy $s_c$. On the other hand, this leads to a free-energy loss due to the $\Upsilon$. However, since the loss term scales as the surface, whereas the gain term scales as the volume, the former cannot counterbalance the latter and the drop in the overlap field is always favourable. Understanding how this takes place and how the resulting amorphous interfaces fluctuate is one of the main aim of this work.]{}]{} [[An important remark on the procedure proposed above is that although $\xi_{PS}$ is finite, it holds $\xi_{PS}\gg\xi_\perp$ when $s_c$ is small, as we shall show. Thus, the curvature of CRRs is negligible on the scale over which interfaces fluctuate. On this scale the approximation of considering that the average profile of interfaces is flat and infinite, i.e. taking the overlap high behind an infinite plane, is justified. For simplicity, we shall first restrict our study to this case and take later into account the finite extension of CRRs. In the following we only focus on the temperature regime below the mode coupling transition temperature, $T_{MCT}$, where high and low overlap states are well-defined and it makes sense to use the concept of configurational entropy. We shall come back in the conclusion to the regime close to $T_{MCT}$.]{}]{}\ As noticed in several works, see [@frapar95; @krzzdb11; @cambir13] and refs therein, the overlap plays the role of an order parameter and the configurational entropy acts like a field favoring the low overlap state. In consequence, the situation is similar to the case of the ferromagnetic Ising model in [[a field, $H$,]{}]{} below the critical temperature. The counterparts of high and low overlaps are positive and negative magnetisations, whereas the configurational entropy plays the role of [[a negative field $H$]{}]{}. This analogy has been proved to be instructive in understanding the physics of glass-forming liquids [@krzzdb11; @cambir13]. In consequence, we start our analysis by discussing the physical picture one obtains from it. Whithin this framework, pinning particles behind a wall amounts to forcing all spins behind a wall to point up. In this way, one induces an interface between the positively magnetized region close to the wall and the negative magnetized region favored by $H$ far from the wall. It is well known that in this case the effective Hamiltonian for the interface position $h({\mathbf x})$ ($h$ is the distance between the interface and the plane at position ${\mathbf x}$ in the $d-1$ planar dimensional space) reads [[[@dikrwa80]]{}]{}: $${\mathcal H}[h(\mathbf x)]=\int d{\mathbf x}\left[ \sigma \frac{\nabla h^2}{2}+Hh(\mathbf x)\right] \label{Hising}$$ where $\sigma$ is the surface-energy cost. The statistics of the interface is obtained by integrating over all interface configurations weighted by their corresponding Boltzmann weight with the constraint $h{{(\mathbf x)}}\ge 0$. At zero temperature ${\mathcal H}$ is minimized by choosing $h{{(\mathbf x)}}=0$ for all $\mathbf x$, i.e. the interface is flat and stuck on the plane. For finite temperatures the interface fluctuate to gain entropy. In two dimensions the interface is a line and the functional integral can be mapped into a quantum mechanical problem that can be solved exactly [@fisher84]. One finds that is entropically favorable for the interface to wander over a length $\xi_{\perp}\propto H^{-1/3}$ perpendicular to the plane and a length $\xi_{\parallel}\propto H^{-2/3}$ parallel to the plane. In three dimensions the wandering is logarithmic only, and in four and higher dimensions the interface is flat [@fisher89]. In conclusion, the analogy with the ferromagnetic Ising model suggests that 3D amorphous interfaces are essentially almost flat and $\xi_\perp$ diverges logarithmically with $s_c$. A previous analysis based on Kac models also lead to a similar conclusion: interfaces are flat and characterized by $\xi_\perp \propto -\ln s_c$ (for energetic reasons) [@frasem11; @zarfra10]. However, a crucial physical ingredient has not been taken into account yet: self-induced quenched disorder. The specific reference configuration naturally introduces quenched randomness in the problem, which plays a very important role in the physics of super-cooled liquids as already shown in [@stewol08; @frparr11; @cambir13; @frapar13; @bicatt14]. Note that in a super-cooled liquid there is no frozen-in disorder: it is the configuration from which the system has to escape in order to flow that plays the role of $\mathcal{C}$, i.e. of self-induced disorder. As it is known for random manifolds in random environments, disorder leads to a huge enhancement of the wandering of the interface, so large that thermal fluctuations become completely irrelevant. It is reasonable to expect that a similar phenomenon could also take place for amorphous interfaces. Indeed, by using replica field theory we show that the Hamiltonian governing the long-wavelength fluctuations of amorphous interfaces is given by (\[Hising\]) plus a random potential term $\int d \mathbf x\, V_R\left(h(\mathbf x),\mathbf x\right)$, whose statistical properties are the same ones found for interfaces in the Random Field Ising Model. The scaling theory of the the RFIM then allow us to work out the behavior of the wandering length $\xi_\perp$, which we find to diverge as the square root of $1/s_c$ in three dimensions. In the following we derive the mapping to the RFIM. The resulting effective action for amorphous interfaces and the corresponding scaling theory for $\xi_\perp$ are presented in the next section. ![Cartoon of a cooperative rearranging region: the linear spatial extent is of the order $\xi_{PS}$ whereas the external shape is rough and fluctuating over the length-scale $\xi_\perp$.[]{data-label="CRR"}](Droplets1.pdf){width=".3\textwidth"} Derivation and Mapping to the RFIM ================================== The starting point of our analysis is considering the statistical field theory for the overlap field $p(z,{\mathbf x})$, which measures the similarity between two equilibrium configurations: the first is free whereas the second is constrained to coincide with the first behind a plane ($z$ is the distance from the plane and ${\mathbf x}$ are the coordinates along the plane). This boundary condition leads naturally to the existence of an interface, whose position along the $z$ axis, $h(\mathbf x)$, corresponds to the region in space where $p(z,{\mathbf x})$ jumps from the high value enforced by the constraint close to the plane to the low value favored by the configurational entropy. Our aim in the following is to obtain the effective field theory on $h(\mathbf x)$ starting from the one on $p(z,{\mathbf x})$. Following previous works that derived effective interface Hamiltonian [@dikrwa80] we assume that the most relevant configurations of $p(z,{\mathbf x})$ are the ones corresponding to a single interface positioned in $h(\mathbf x)$. This is natural since having more than one interface is unlikely. In consequence, henceforth we only focus on configurations $p^h(z,{\mathbf x})=q_{EA}$ for $0<z<h({\mathbf x})$ and zero for $z>h({\mathbf x})$, where $q_{EA}$ is the typical overlap of two configurations in the same amorphous state (it is associated with the local Debye-Waller factor characterizing molecular motion in the glass-forming liquid). We neglect the smoothness of the decrease from one to $q_{EA}$ just after the wall and from $q_{EA}$ to zero at $z\simeq h({\mathbf x})$. Both simplifications are inessential to establish the effective field theory of $h({\mathbf x})$, as we shall discuss later. The effective Hamiltonian ${\mathcal H}_R[h(\mathbf x)]$ is obtained evaluating the action for the overlap field for $p=p^h(z,{\mathbf x})$: $${\mathcal H}_R[h(\mathbf x)]={\cal S}[p^h(z,\mathbf x)|{\cal C}]$$ where we have made explicit the dependence on the reference equilibrium configuration ${\cal C}$ that introduces the quenched disorder. In order to show that ${\mathcal H}_R[h(\mathbf x)]$ coincides with (\[Hising\]) plus a random potential term, $\int d \mathbf x \,V_R\left(h(\mathbf x),\mathbf x\right)$, whose variance is the one characteristic of domain walls in the RFIM, we compute the average and the variance of ${\mathcal H}_R[h(\mathbf x)]$. We shall show that the former is equal to ${\mathcal H}[h(\mathbf x)]$ and the latter, $\overline{\left(({\mathcal H}_R[h_1(\mathbf x)]-{\mathcal H}[h_1(\mathbf x)])-({\mathcal H}_R[h_2(\mathbf x)]-{\mathcal H}[h_2(\mathbf x)])\right)^2}$ is [[proportional]{}]{} to the volume, ${\mathcal V_{h_1,h_2}}$, of the space embedded by the two interfaces described by $h_1(\mathbf x)$ and $h_2(\mathbf x)$. This is indeed the result expected for the RFIM case, where the correlator of the random potential [[is]{}]{} $\overline{V_R\left(h_1(\mathbf x),\mathbf x\right)V_R\left(h_2(\mathbf x),\mathbf x\right)} {{\propto}} |h_1(\mathbf x_1)-h_2(\mathbf x_2)|\delta(\mathbf x_1-\mathbf x_2)$ [@wiese03].\ The computation of the cumulants of ${\mathcal H}_R[h(\mathbf x)]$ is performed by introducing [[$n$]{}]{} different copies (or real replicas) of the system in presence of the same “disorder” $\cal C$ and averaging the replicated system over $\cal C$. Following Ref. [[[@bicatt14]]{}]{} we define the action of the replicated system by the identity $$\exp\left(-{\cal S}_r[\{p_a\}]\right)=\overline{\exp\left(-\sum_a {\cal S}[p_a(h,\mathbf x)|{\cal C}]\right)}^{\cal C}$$ [[where $a\in[0,n]$]{}]{}. The action ${\cal S}_r[\{p_a\}]$ generates all the cumulants of ${\cal S}[p(z,\mathbf x)|{\cal C}]$ through the equation: $${\cal S}_r[\{p_a\}]=\sum_a S_1[p_a]-\frac{1}{2}\sum_{a,b}S_2[p_a,p_b]+\frac{1}{3!}\sum_{a,b,c,}S_3[p_a,p_b,p_c]\dots$$ where $$S_1[p]=\overline{{\cal S}[p|{\cal C}]}^{\cal C} \,\,,\,\,S_2[p_1,p_2]=\overline{{\cal S}[p_1|{\cal C}]{\cal S}[p_2|{\cal C}]}^{\cal C}-\overline{{\cal S}[p_1|{\cal C}]}^{\cal C}\overline{{\cal S}[p_2|{\cal C}]}^{\cal C}$$ and so on.\ In order to extract $S_1[p]$ from ${\cal S}_r[\{p_a\}]$ one considers all replicas equal, i.e. $p_a=p\,\forall a$, and pick from ${\cal S}_r[\{p_a\}]$ the term linear in $n$ since $S_2, S_3,…$ are respectively of the order $n^2,n^3$ etc.. Similarly, in order to obtain $S_2[p_1,p_2]$ one subdivides all replicas in two groups, such that $p_a$ is equal to $p_1$ and $p_2$ for replicas respectively belonging to the first and the second group. In this case $S_2[p_1,p_2]$ is simply the part of the action proportional to $n_1n_2$ and can be therefore easily selected in the limit $n_1,n_2 \rightarrow 0$. The technical procedure to follow in order to compute $ {\cal S}_r[\{p_a\}]$ was derived in [@bicatt14]. In the following we just quote the final result: $ {\cal S}_r[\{p_a\}]$ is obtained as the free energy of the replicated field theory for $n+1$ copies $\alpha\in[0,n]$ of the system, in which one fixes the overlaps $q_{\alpha0}$ with the [[reference]{}]{} configuration [[labelled by]{}]{} $\beta=0$ to be equal to $p_a$ and integrates out all the others. As in Ref. [@bicatt14] the integration is performed by saddle-point (a more careful evaluation of the functional integral is not expected to give rise to any qualitative change). We use as action of the replicated field theory the Landau’s one proposed in Ref. [[[@dzscwo09]]{}]{}: $$\begin{aligned} {\cal S}[\{q_{\alpha\beta}\}]=\frac{E_0}{k_BT}\int_{z\mathbf x}\left\{\frac{c}{2}\sum_{\alpha\neq \beta}\left(\partial_{z\mathbf x}q_{\alpha\beta}(z,\mathbf x)\right)^2+\sum_{\alpha\neq \beta}v(q_{\alpha\beta})\right. \nonumber \\ \left. -\frac{u}{3}\sum_{\alpha\neq\beta\neq\gamma}q_{\alpha\beta}(z,\mathbf x)q_{\beta\gamma}(z,\mathbf x)q_{\gamma\alpha}(z,\mathbf x)\right\}\end{aligned}$$ where it has been defined $v(q)=tq^2/2-(u+w)q^3/3+yq^4/4$, the main temperature dependence is in $t\sim k_B(T-T_0)/E_0$, $E_0$ is the liquid’s energy scale, and $T_0$ is a constant. From previous analyses we know that $v(q)+uq^3/3$ develops a secondary minimum below $T_{MCT}$ in correspondence of the Edwards-Anderson overlap value $q_{EA}$. The height of this minimum is the configurational entropy $s_c(T)$.\ In order to compute the first cumulant we take $p_a=p^h(z,\mathbf x)$. Because of the resulting replica symmetry we solve the saddle point equations assuming $q_{ab}=q \ \ \forall a\neq b$ and $n\rightarrow 0$. We first consider the case of a flat interface, i.e. $h(x)=h$, $p^h(z,\mathbf x)=p(z)=q_{EA}\theta(h-z)$ where the latter is the Heaviside function. In this case $q$ only depends on $z$ and its saddle point equation reads: $$\label{eqq} c\partial^2_z q(z)-v'(q(z))-2uq(z)^2+uq_{EA}^2\theta(h-z)=0$$ By numerically solving this equation[^1] we found that an interface profile for $p^h$ induces a similar interface profile $q^h(z)=I(z-h)$, where $I$ is constant until $z=h$ and then decreases rapidly to zero, see Fig. 1. This physically makes sense since replicas that are forced to have a high overlap with a reference configuration until a distance $h$ from the wall are expected to also have a high cross-overlap. By plugging these profiles into the replicated action and focusing on the term proportional to $n$ one naturally finds two contributions, one that is proportional to the volume between the wall and the interface and another one, independent on the interface position, that scales as the surface[^2]: $$\begin{aligned} \label{h1} \overline{{\mathcal H}_R[h]}&\simeq&\int d{\mathbf{x}}\{ h[v(q_{EA})+uq_{EA}^3/3]+\Sigma\}\nonumber\\ &=&\int d{\mathbf{x}} s_c h+\Sigma L^{d-1}\,\,,\end{aligned}$$ where $\Sigma$ is the cost per unit surface of creating an interface and $L^{d-1}$ the wall surface. In order to compute $\Sigma$ correctly one should take into account an optimized and smooth form of the interface along the $z$ direction. We do not need to worry about this complication since the term linear in $h$, the one we are interested in, is independent of it. Let us now also include in the analysis long-wavelength fluctuations of $h(x)$. Simple arguments show that in this case $q(z,{\mathbf{x}})$ has to follow “adiabatically” the profile of $p^h(z,\mathbf x)=q_{EA}\theta(h({\mathbf{x}})-z)$, i.e. $q(z,{\mathbf{x}})=I(z-h({\mathbf{x}}))$ up to sub-leading corrections in gradients of $h({\mathbf{x}})$. By plugging this expression in the action one finally obtains that $\overline{{\mathcal H}_R[h]}$ at large length-scales and for $h\gg1$ is precisely equal to ${{\mathcal H}[h]}$ defined in eq. \[Hising\], with $\sigma=c$ and $H=s_c$, plus a constant term equal to $\Sigma$ times the wall surface.\ Having obtained the first part of our technical results, we now turn to the study of the fluctuations of ${\mathcal H}_R[h]$. As discussed above, we have to consider two groups of replicas having an overlap profile with the reference configuration $p^{h_1}$ and $p^{h_2}$ respectively. As before we start by focusing on flat interfaces positioned at $h_1$ and $h_2$. Without loss of generality we will consider $h_1<h_2$. By writing the saddle-point equation on $q_{ab}$ for $n_1,n_2 \rightarrow 0$ one finds that the overlap between replicas of the same group satisfies equation (\[eqq\]) where the role of $p^h$ is played by $p^{h_1}$ and $p^{h_2}$ respectively. The overlap $q_{12}$ between replicas of different groups satisfies the equation: $$\begin{aligned} c\Delta q_{12}&=&v'(q_{12})-uq_{EA}^2\theta(h_1-z)\theta(h_2-z)\nonumber\\ &&+uq_{12}(q(z-h_1)+q(z-h_2)) \ . %-u(q_{12}-p_1)(q_{12}-p_2)\end{aligned}$$ Our numerical solutions [^3] show that $q_{12}$ assumes a profile very similar to $p^{h_1}$, i.e. the interface profile closer to the wall: $q_{12}$ is equal to $q_{EA}$ for $z<h_1$ and has a sharp drop to zero just after, see Fig.2. By plugging all the overlap profiles in the replicated action one finds that the second cumulant $S_2[p_1,p_2]$ reads $$\begin{aligned} S_2&=&-2\int_{z\mathbf x} \left\{ 2v(q_{12})-2up^{h_1}p^{h_2}q_{12}\right. \nonumber \\ &&\left.+uq_{12}^2(q^{h_1}+q^{h_2})\right\} \ .\end{aligned}$$ Using this result and in the limit $h_1\gg1$, $h_2\gg1$, $h_1-h_2\gg1$ we obtain that the variance of the fluctuations $\delta {\cal S}[p|{\cal C}]={\cal S}[p|{\cal C}]-\overline{{\cal S}[p|{\cal C}]}^{\cal C}$ reads: $$\begin{aligned} \label{s2} \overline{\left(\delta {\cal S}[p^{h_1}|{\cal C}]-\delta {\cal S}[p^{h_2}|{\cal C}]\right)^2}^{\cal C}&=&\\ &&\hspace{-2.8cm}=S_2[p^{h_1},p^{h_1}]+S_2[p^{h_2},p^{h_2}]-2S_2[p^{h_1},p^{h_2}] \nonumber \\ &&\hspace{-2.8cm}=4{\cal V}_{h_1,h_2} \left(\dfrac{u}{3}q_{EA}^3-s_c\right)\nonumber\end{aligned}$$ [[where ${\cal V}_{h_1,h_2}$ is the volume embedded by the surfaces $h_1$ and $h_2$]{}]{}. As discussed previously, if one considers long wave-length fluctuations of $h_1({\mathbf{x}}),h_2({\mathbf{x}})$, the overlaps $q^{h_1}, q^{h_2}, q_{12}$ follow adiabatically the solutions obtained for flat profiles. By plugging the corresponding solutions into the action one finds that $S_2$ acquires an extra contribution due to the gradient terms which can be neglected at leading order since scales as the area of the interfaces and not as the volume ${\cal V}_{h_1,h_2}$. In consequence, result (\[s2\]) also holds for non-flat interfaces. Effective Hamiltonian and Scaling Theory ======================================== We now collect all previous results and write down the effective Hamiltonian governing the long-wavelength fluctuations of amorphous interfaces. It is a random functional of $h(\mathbf x)$ whose average and variance are the ones computed previously. We do not have computed higher cumulants but these are not expected to be relevant (in a renormalization group sense). Hence, for simplicity we shall take them equal to zero in the following. The final model for amorphous interfaces reads: $${\mathcal H}_R[h(\mathbf x)]=\int d{\mathbf x}\left[ c \frac{\nabla h^2}{2}+s_ch(\mathbf x)\right] +\int d \mathbf x \,V_R\left(h(\mathbf x),\mathbf x\right)+\Sigma L^{d-1} \label{Hint}$$ where $V_R(h(\mathbf x))$ is a random Gaussian potential with zero mean and $\Sigma$ is independent of the shape of the interface profile. The variance of $V_R(h(\mathbf x))$ reads $$\begin{aligned} \overline{V_R\left(h_1(\mathbf x_1),\mathbf x_1\right)V_R\left(h_2(\mathbf x_2),\mathbf x_2\right)}=\\ &\hspace{-2.8cm}=\left(\dfrac{u}{3}q_{EA}^3-s_c\right)|h_1(\mathbf x_1)-h_2(\mathbf x_2)|\delta(\mathbf x_1-\mathbf x_2)\nonumber\end{aligned}$$ Note that the variance is positive, as it should, since in the regime we are interested in, i.e. $T$ close to $T_K$, the configurational entropy $s_c$ is small[^4].\ The model we ended up is identical to the one describing domain walls in the RFIM in presence of an external field. We can therefore use previous insights developed in this case, in particular the scaling theory of [@fisher89], to work out the behavior of amorphous interfaces. The fluctuations of the interfaces are determined by the balance between two competing mechanisms. An interface closer to the wall leads to a gain of configurational entropy and hence to an [*effective attractive interaction*]{} $W_a(\ell)$. In fact the transition from high to the low overlap state at a distance $\ell$ from the wall leads to a free energy density gain (per unit surface) equal to $W_a(\ell)=W_a(0)+s_c \ell$. On the other hand, because of the random field disorder the interface wanders over increasingly large length-scales in order to find an optimized configuration that goes through favourable energetic regions, as it is known for the RFIM. Forcing the interface to wander no more than a distance $\ell$ from the wall induces a constraint and hence to a less optimized configuration, i.e. to a higher energy. As discussed in [@fisher89] this produces an [*effective repulsive potential*]{} between wall and interface (per unit surface) equal to $W_r(\ell)=W_r(0)+\frac{b}{\ell^\tau}$. The balance between these two mechanisms sets the value of the typical distance of the amorphous interface from the wall: $\xi_\perp\propto s_c^{-1/(\tau+1)}$. On length-scales smaller than $\xi_\perp$, the effective attraction due to the configurational entropy can be neglected and the interface fluctuations are similar to the ones of a free interface [@fisher89]. Thus, moving along the plane one encounters over length-scales $\xi_\parallel\propto \xi_\perp^{1/\zeta}$ independent transverse fluctuations of the interface of the order $\xi_\perp$ ($\zeta$ is the roughness exponent of free RFIM interfaces).\ For the RFIM the Imry-Ma argument, validated by Functional Renormalization Group analysis and numerical simulations, gives $\zeta=(5-d)/3$ and $\tau=2/\zeta-2=(2d-4)/(5-d)$ [@fisher89; @wiese03]. This leads to a width of amorphous interfaces scaling as $$\xi_\perp\propto s_c^{-(5-d)/(d+1)}$$ In consequence, we finally find that amorphous interfaces wander in three dimensions over a length $\xi_\perp\propto s_c^{-1/2}$. ![Overlap profiles $p^{h_1}(z)$, $p^{h_2}(z)$, $q^{h_1}(z)$, $q^{h_2}(z)$, $q_{12}(z)$ for $h_1=10, h_2=30$ and $s_c=0$ (the unit of length is $\sqrt{c}$). Note that as explained in the text $q^{h_1}(z)=I(z-h_1)$. For the Landau action considered in the text $q_{EA}=1$ when $s_c=0$. Similar results are obtained for a generic small value of $s_c$.[]{data-label="profile"}](out.pdf){width=".5\textwidth"} Physical consequences, predictions and comparison to numerical experiments ========================================================================== Until now we have considered amorphous interfaces between regions of high overlap and low overlap separated by an infinite plane. This can be achieved by pinning particles behind a wall as discussed in the introduction (more on this later one). However, in a bulk glass-forming liquid the length which characterizes the spatial extent of amorphous order and the linear size of the CRRs is finite. It is called point-to-set and within RFOT scales as $\xi_{PS}\propto s_c^{-1/(d-\theta)}$ [@kithwo89]. The growth of $\xi_{PS}$ is due to a mechanism completely different from the one associated to $\xi_\perp$ and, accordingly, the growth law is different. In three dimensions, within a Kac-like (instanton) approach one finds $\theta=\left. d-1\right.=2$ [@franz05; @dzscwo09], scaling arguments by Kirkpatrick, Thirumalai and Wolynes suggest $\theta=\left.\frac d 2\right.= \frac 3 2$ [@kithwo89], whereas some numerical results seem to indicate $\theta=2$ [@cacggv09a; @cacggv09b]. Although a conclusive result on the value of $\theta$ is still missing, all indications point forward a value of $\theta$ such that $1/(d-\theta)$, is larger than $1/2$ in three dimensions. In consequence, we do find as anticipated that CRRs, i.e. the regions over which the system is coherently in one amorphous state, are separated by interfaces that are rough but fluctuate less than the typical size of the regions, see Fig. 1 for a pictorial representation. This remains true for higher dimensions. Interestingly, for $d=2$ the exponents of $\xi_{PS}$ and $\xi_{\perp}$ become equal ($d-1=d/2=(5-d)/(d+1)=1$) possibly indicating that $d=2$ is the lower critical dimension for the glass transition, as also suggested by other arguments [[[@cambir13; @frapar13]]{}]{}.\ The existence of two different static length-scales governing the physics of super-cooled liquids, $\xi_{PS}$ and $\xi_\perp$, is a major fact to take into account in understanding the outcomes of simulations and experiments in glass-forming liquids, in particular when probing static correlations. Our findings make clear that lengths extracted by different ways of pinning, in particular the wall geometry versus the spherical geometry, probe [*different physical lengths*]{}. This naturally provides an explanation for the difference found in numerical simulations between these two cases. In fact it was shown that the lengths probed using these two pinning geometries grow in a different way, more mildly in the former case as indeed expected from our results [@korobe12; @grtcgv13]. Moreover, the decay of the overlap at the center of the cavity was found to be quite different from the decay of the overlap from the wall. In the first case, it becomes sharper at lower temperature (it can be fitted by an increasingly more compressed exponential) [@bibcgv08; @homare12; @grtcgv13], whereas in the latter the form remains unchanged and exponential-like [@berkob12; @grtcgv13]. These differences have a natural explanation within the physical picture arising from our theory in which two growing static lengths, $\xi_\perp$ and $\xi_{PS}$, intervene. Since the ratio $\xi_\perp/\xi_{PS}$ decreases by lowering the temperature, the CRRs are better and better defined on the scale $\xi_{PS}$ and therefore the decay at the center of the cavity becomes indeed sharper and sharper at lower temperature. Instead, the decay of the overlap from the wall is governed by the length-scale $\xi_\perp$ only (the point-to-set length does not play any role). In this case it is natural to expect scaling with respect to $\xi_\perp$, as also shown for manifolds in random media, and hence a decay that does not change form, in particular does not become sharper by lowering the temperature. Note that another case in which taking into account the existence of two different static length-scales is crucial to explain numerical data has been discussed recently in [@macagr14]: in order to rationalize the finite size scaling of the specific heat for pinned systems in a cavity geometry one needs to consider both $\xi_\perp$ and $\xi_{PS}$. To test our scaling predictions it would be worth pushing further numerical simulations to obtain the dependence $\xi_\perp$ on $s_c$ for realistic model of super-cooled liquid. The RFIM character of amorphous interfaces that follows from our theory is already well supported by the numerical results of [@cacggv09a; @cacggv09b] which found a roughening exponent $\zeta\simeq 0.62-0.75$ and energy fluctuations scaling as $\ell ^{2\zeta}$. These two results compare extremely well with our predictions, which also lead to the same scaling of energy fluctuations and $\zeta=2/3\simeq 0.66$. Conclusion ========== In this work we showed that amorphous interfaces are rough in three dimensions and we obtained the scaling with the configurational entropy of the length-scale over which they wander. Their statistical properties are identical to the ones of domain walls in random ferromagnets, a fact that strengthens even more the relationship between the physics of supercooled liquids and of the RFIM discussed in [@stewol08; @frparr11; @cambir13; @frapar13; @bicatt14]. One of our major results is that there are two different static length-scales governing the physics of super-cooled liquids: the point-to set length $\xi_{PS}$, related to the spatial extent of CRRs, and the wandering length $\xi_\perp$ related to the fluctuations of their external shape. Our findings are in good agreement with previous numerical results, some of which were considered contradictory but find a natural explanation within our theory. We focused on the regime below $T_{MCT}$ where CRRs are well-formed and configurational entropy and interfaces are meaningful concepts. Approaching $T_{MCT}$ we expect $\Upsilon$, and hence $c$, to decrease. This makes fluctuations more favorable. At a certain point, when they become so large that the long-wavelength theory with a simple square gradient term is not suitable anymore, the description of the interface we used might break down. We suspect that close to $T_{MCT}$ this leads to different scaling forms and is associated to the fractal, or stringy, nature of CRRs found in [[[@stscwo06]]{}]{}.\ Finally, in view of the recent studies of the glass transition in high spatial dimensions [[[@eavrei09; @chikpz11; @chikpz12; @chcjpz13; @chkupa14]]{}]{} it is interesting to remark that our theory predicts a highly non-trivial dimensional dependence. In particular we find $\xi_\perp\propto s_c^{-(5-d)/(1+d)}$ and, hence, an upper critical dimension $d_u=5$. In higher dimensions amorphous interfaces are flat, $\xi_\perp$ does not increase and, hence, only one static growing length-scale accompanies the glass transition. This is a striking change in the nature of the glass transition that would be worth testing numerically.\ In conclusion, the predictions obtained in this work provide a full characterization of the shape of cooperative rearranging regions in super-cooled liquids. They are instrumental in interpreting, understanding and devising new numerical simulations and experiments on static correlation in glass-forming liquids and clarify differences and relationships between the plethora of static lengths studied in recent years. An issue worth studying further, that we leave for future work, is whether in cases more accessible to experimental investigations, such as free surfaces instead of amorphous walls, the length $\xi_\perp$ can be probed [@gruebel10]. We thank T. Giamarchi, G. Parisi, G. Tarjus and M. Tarzia for very useful discussions. We acknowledge support from the ERC grants NPRGGLASS (GB) and CRIPHERASY (CC) (no. 247328). [^1]: In the numerical integration the derivative of $q$ at the interface is chosen to optimize the action (actually maximize for reasons related to the $n\rightarrow0$ limit). [^2]: In eq. (\[h1\]) we only consider the leading term in $h$. The first correction to this behavior dependent on $h$ is of the form $\exp(-Kh)$ where $K$ is a constant. This term, in absence of quenched disorder, pushes the interface slightly away from the plane and leads to the result $\xi_\perp\propto -\ln s_c$ obtained in [@frasem11; @zarfra10]. It can be neglected in the following since self-induced disorder leads to much stronger fluctuations and a more rapid increase of $\xi_\perp$ with $s_c$. [^3]: In the numerical integration the derivative of $q_{12}$ at the interface is chosen to optimize the action (actually minimize for reasons related to the $n\rightarrow0$ limit). [^4]: It is possible to show that the disorder is short-ranged in $\mathbf x_1-\mathbf x_2$, thus for simplicity we have considered it $\delta$-correlated. Introducing explicitly finite range correlations would not alter our conclusions.
{ "pile_set_name": "ArXiv" }
--- abstract: 'The present status of kaon photo- and electroproduction on the nucleon is briefly reviewed. Some current important issues in this field are discussed.' address: | Departemen Fisika, FMIPA, Universitas Indonesia Depok 16426, Indonesia\ [email protected] author: - 'T. MART' title: PROGRESS AND ISSUES IN THE ELECTROMAGNETIC PRODUCTION OF KAON ON THE NUCLEON --- Introduction {#sec:intro} ============ Compared to the case of pions the electromagnetic productions of kaons are less understood. This is due to the fact that the strangeness quantum number is explicitly present in the final state of the process. Although the reactions involving kaons are slightly more complicated, the additional degrees of freedom created by the strangeness can give information not available from nucleons and pions. For instance, due to the conservation of the strangeness, the production of kaons is always accompanied by the creation of hyperons, which can be used to deeply explore the structure of the nucleus since they are not blocked by the Pauli principle. On the other hand, this elementary process provides an important input for the calculation of hypernuleus photo- or electroproduction. Along with the fact that some $N^*$ resonances that were predicted in quark models have only noticeable branching ratios into the $K \Lambda$ channel, kaon photo- and electroproduction clearly have drawn many attentions for more than five decades. Theoretical Models ================== The earliest attempt to theoretically explain kaon photo- and electroproduction was proposed by Kawaguchi and Moravcsik more than 50 years ago[@masasaki]. Interestingly, all of the six possible isospin channels were already considered by utilizing only three Feynman diagrams of the Born terms, though hitherto no experimental data were available. The results were obviously very modest and, in fact, the cross sections for the $K^0\Lambda$ and $K^0\Sigma^0$ channels were predicted to be zero since neither nucleon anomalous magnetic moments nor resonance contributions were taken into account. Since that time considerable efforts were devoted to explain the appearing experimental data, although only Thom who tried to seriously fit the data to an isobar model. After that, there were a number of works using isobar model[@deo], dispersion relation[@nelipa; @pickering], multipoles analysis[@schorsch], and Regge approach[@levyph]. After the mid-seventies, the interest in this field was temporarily dormant, mainly due to the lack of experimental facilities. The interest in kaon photo- and electroproduction was revived by the constructions of modern accelerators such as those in JLAB, MAMI, ELSA, and others laboratories. The work of Adelseck, Bennhold, and Wright started the new era of phenomenological models in $K^+\Lambda$ photoproduction[@abw]. This work was refined with the inclusion of more data and electroproduction process[@adel2] and extended to all isospin channels[@terry1]. A chiral quark model with less parameters has been also put forward[@zpli] in an attempt to recover the low energy theorem. The results of this model are modest and, in fact, the $K^0\Sigma^+$ cross section is predicted to be larger than that for $K^+\Sigma^0$ by a factor of approximately two, since contributions from the seagull and $s$-channel resonance have opposite signs. Modern calculations of kaon photoproduction exploit chiral perturbation theory[@chpt] and coupled channels analysis[@coupled-channels; @coupled]. In the higher energy regions, the Regge[@Guidal:1997hy] or hybrid models[@hybrid] (a combination of the Feynman diagrammatic technique and Regge formalism) turn out to be more appropriate. Nevertheless, for practical use such as for nuclear applications and other phenomenological studies, the single channel isobaric analysis is still proven to be powerful[@terry1; @Mart:1999ed; @single_channel]. Experimental Data {#sec:data} ================= Historically, experimental data of kaon photo- and electroproduction can be divided into two categories, old data (published before 1980) and new data (published after 1990). A list of references for old data is given, e.g., in Refs.[@AS90; @saphir98], whereas the corresponding kinematical coverage for the photoproduction is shown in the left panel of Fig. \[fig:data\]. Surprisingly, the old data cover also forward angle regions, the case which is very difficult to achieve with the presently available technologies, as obviously shown in the right panel of Fig. \[fig:data\]. As discussed in the following section, this region is in fact very decisive for constraining the models and for nuclear applications. Nevertheless, the data quality is in general poor and, as a consequence, the cross sections data do not show resonance structures. There were few data available for the recoiled $\Lambda$ polarization and two data points on the target asymmetry. Kaon photoproduction started to become more interesting after the publication of SAPHIR data[@saphir98] on the $K^+\Lambda$ and $K^+\Sigma^0$ channels in 1998, since the corresponding error bars allow for an identification of resonance structures, especially in the $K^+\Lambda$ channels. Further analysis of SAPHIR data[@glander04] as well as the new CLAS data[@bradford06] confirmed these structures. Note that in both SAPHIR and CLAS data sets, there are plenty of data on the recoil polarization. Together with the new data from the LEPS[@sumihama06; @hicks07] and GRAAL[@lleres07] collaborations, these data provide a strong constraint on the proliferation of phenomenological models. In the $K\Sigma$ channels, new data have also appeared from SAPHIR collaboration[@lawall05] ($K^0\Sigma^+$), and from the LEPS collaboration[@kohri06] ($K^+\Sigma^-$). The ability to reverse the direction of photon helicity in the SPRING8 has led to the measurements of the photon asymmetry $\Sigma$ in the $K^+\Lambda$, $K^+\Sigma^0$, as well as $K^+\Sigma^-$ channels[@sumihama06; @kohri06]. Since the polarization of the recoiled $\Lambda$ can be obtained without any additional apparatus, measurements of the double polarization observables $C_x$ and $C_z$ (or $O_x$ and $O_z$) are possible, provided that the polarized photon beams are available. The first measurement of these observables was performed by the CLAS collaboration[@Bradford:2006ba] and because the available beams at JLAB are circularly polarized, the corresponding observables are $C_x$ and $C_z$. Very recently, by utilizing the linearly polarized photon the GRAAL collaboration has been able to measure the $O_x$ and $O_z$ observables[@Lleres:2008em] from threshold up to $E_\gamma=1.5$ GeV. The electroproduction experiments with high statistics have been also performed at JLAB[@mohring03; @carman03; @ambrozewicz]. It is reported that the longitudinal and transverse components of the cross section can be nicely separated[@mohring03; @ambrozewicz], whereas the transferred polarization in the $\vec{e}p\to e'K^+\vec{\Lambda}$ process has been measured[@carman03]. Current Important Issues in Kaon Photoproduction ================================================ The problem of data discrepancy ------------------------------- In spite of their unprecedented high statistics, the new CLAS[@bradford06] and SAPHIR[@glander04] data reveal a lack of mutual consistency in the forward and backward regions. This problem can be summarized by the total cross section shown in Fig. \[fig:tot\], where it is obvious that fitting a phenomenological model to both data sets simultaneously results in a model which is inconsistent to all data sets. As shown by Ref.[@mart06] this problem also hinders a reliable extraction of the resonance parameters, since the use of SAPHIR and CLAS data, individually or simultaneously, leads to quite different resonance parameters. Therefore, the presently available data do not allow for a firm conclusion on the extracted “missing resonances”. By using four different isobar models Ref.[@Bydzovsky:2006wy] studied the statistical properties of both CLAS and SAPHIR data. It is found that the SAPHIR data are coherently shifted down with respect to the CLAS and LEPS[@sumihama06] data, especially at forward kaon angles. A global scaling factor of 15% is required to remove this discrepancy. The phenomenon also implies that the LEPS data are more consistent with the CLAS data than the SAPHIR ones. Interestingly, at forward angles both CLAS and LEPS data sets can be described reasonably well within the isobaric model without hadronic form factors. This can be understood from the energy distribution of the cross sections at this kinematics, as will be discussed in the following subsections. The statistical study recommend that a combination of CLAS, LEPS, and old[^1] data sets in the fitting data base is the more preferred choice, while a combination of CLAS and SAPHIR data should not be assumed in fixing parameters of models, especially at forward angles. Despite considerable efforts to alleviate this problem, it is still difficult to determine which data set should be used to obtain a reliable phenomenological model. The reason is that in all analyses the experimental data are fitted by adjusting a set of free parameters, while the precise values of these parameters are not well known. Furthermore, the extracted parameters are not unique and also sensitive to the number of resonances used in a model. In view of this, Ref.[@Mart:2008ik] proposed to use other quantities which can be predicted by the models and can be directly compared with the results from other measurements or model predictions. One of the possible quantities is the contribution of the $\gamma p\to K^+\Lambda$ channel to the Gerasimov-Drell-Hearn (GDH) sum rule [@gdh], which relates the anomalous magnetic moment of the proton $\kappa_p$ to the total photoabsorption cross sections $\sigma_{\rm TT'}$, i.e. $\kappa_p^2 = (m_p^2/\pi^2\alpha) \int_{0}^{\infty}dE_\gamma\,\sigma_{\rm TT'}/ E_\gamma$. By looking at the energy distributions of the $\sigma_{\rm TT'}$ obtained from models that fit to CLAS and SAPHIR data shown in the left panel of Fig. \[fig:sigTT\], it is obvious that the two data sets lead to different contribution to $\kappa_p$. It is then concluded that contribution from the model that fits to the SAPHIR data sets is more preferred[@Mart:2008ik]. Nevertheless, experimental measurement of the $\sigma_{\rm TT'}$ is clearly required to confirm this claim. Since the SAPHIR detector has been dismantled, measurements of the $\sigma_{\rm TT'}$ by the CLAS or MAMI collaboration seem to be the only choice. The effects of $C_x$ and $C_z$ data ----------------------------------- The beam-recoil polarization observables data, $C_x$ and $C_z$, published by the CLAS collaboration[@Bradford:2006ba] recently, indicate that the $\Lambda$ polarization is predominantly in the direction of the spin of the incoming photon. It is interesting that recent analyses found that these data seems to be difficult to explain. In Ref. [@Mart:2008ik] it is reported that the inclusion of these data reduces the complicated structure of the total cross sections $\sigma_{\rm TT'}$ as shown in the right panel of Fig. \[fig:sigTT\], which indicates that the $C_x$ and $C_z$ data select only certain important resonances, i.e. the $S_{11}(1650)$, $P_{11}(1710)$, $P_{13}(1720)$, and $P_{13}(1900)$ resonances. This finding corroborates the result of a recent coupled channels analysis. [@Anisovich:2007bq] The second peak in the cross section ------------------------------------ The first version of SAPHIR data on $\gamma p\to K^+ \Lambda$ released in 1994 did not give any hint about resonance structures in the cross section, since the statistics was very limited[@saphir_old1]. However, by collecting more statistics the second version of SAPHIR data[@saphir98] released in 1998 display two distinct peaks in the total cross section at $W\approx 1700$ MeV and 1900 MeV. More accurate data published by the SAPHIR[@glander04] and CLAS[@bradford06] collaborations 5 and 8 years later, respectively, confirm the existence of the two peaks (see Fig.\[fig:tot\]). By comparing the extracted resonance widths and branching fractions with the predictions of a constituent quark model[@cqm], Ref.[@Mart:1999ed] concluded that the second peak corresponds to the $D_{13}(1895)$ nucleon resonance that is missing from the PDG list[@pdg98]. Although many subsequent analyses used this spin 3/2 nucleon resonance in their models, there was a suggestion[@Janssen:2001pe] that this peak could come from other resonances such as $S_{11}$, $P_{11}$ or $P_{13}$. However, the possibility of using these resonances, instead of the $D_{13}(1895)$, has been also discussed in Ref.[@Mart:1999ed], but it was ruled out after comparing the extracted widths and branching fractions with the predictions of quark model. Further study using a multipoles approach[@mart06] found that, in spite of their differences, both SAPHIR[@glander04] and CLAS[@bradford06] data sets indicate that the peak originates from the $D_{13}$ resonance with a mass between $1911-1936$ MeV. The necessity of the $D_{13}$ in the analyses of $K^+\Lambda$ photoproduction with a mass of around 1900 MeV has been also corroborated by recent coupled channels analyses[@coupled-channels]. Note that in these analyses the recoil polarization data have been also taken into account. It is quite interesting, however, that a recent effective field theory analysis suggests that the peak should correspond to the $P_{11}$ state which is a bound state of $K\bar KN$ with a mixture of $a_0(980)N$ and $f_0(980)N$ components[@MartinezTorres:2009cw]. To confirm this idea, two experiments are proposed, i.e. first to exclude contributions of the spin 3/2 states and second to find an indication of the $K\bar KN$ bound state. On the other hand, a preliminary study[@nelson09] performed by constraining the free parameters in a multipole analysis[@mart06] to the PDG values indicates that the peak originates from the contributions of both $S_{11}$ and $P_{11}$ states with masses around 1920 MeV. In this study it is found that the contribution of the $S_{11}$ state is stronger than that of the $P_{11}$ and the result does not depend on which data set is being used in the fitting database. The problem at forward angles ----------------------------- Since the dominant contributions to the total cross section shown in Fig. \[fig:tot\] come from the forward angles data, the largest discrepancies between the CLAS and SAPHIR data obviously show up at the forward angles. Unfortunately, the available phenomenological models vary wildly at these kinematical regions. This situation is clearly shown in Fig. \[fig:forward\], in which we can see that for $0^\circ\le\theta_K\le 30^\circ$ our best knowledge on kaon photoproduction cannot tell us about the actual values of differential cross sections. Meanwhile, theoretical predictions of the hypernuclear photo- or electroproduction rely heavily on the elementary operator extracted from kaon photo- or electroproduction data at very forward angles. As an example, in the electromagnetic production of the hypertriton, the magnitude of differential cross section is only realistic for experimental measurement at $0^\circ\le\theta_K\le 20^\circ$, whereas at $\theta_K\approx 25^\circ$ the cross section is practically zero[@Mart:2008gq]. In view of this, accurate and reliable experimental data as well as theoretical formulation of kaon photoproduction in the forward region are the first condition toward a reliable prediction of the hypernuclear photoproduction cross section. The controversy of hadronic form factors ---------------------------------------- It has been well known that the background terms become divergent at high energies. Therefore, the use of hadronic form factors to reduce the background contribution is desired. It is also known that the use of this form factor leads to the violation of gauge invariance of the amplitude. As a result, several methods have been put forward in the last decades to overcome this problem. Two of the well known prescriptions are due to Ohta [@ohta89] and Haberzettl [@hh97g]. As can be seen from the left panel of Fig. \[fig:had\_ff\], the latter is superior since it can provide a reasonable description of experimental data, i.e. producing relatively much smaller $\chi^2$, using values for the leading couplings constants close to the SU(3) values. In Ohta’s method this is not possible due to the absence of a hadronic form factor in the electric current contribution. Since then, most of the phenomenological models include hadronic form factors in their background terms. On the other hand, Ref.[@Janssen:2001pe] proposed to use certain hyperon resonances, the $S_{01}(1800)$ and $P_{01}(1810)$, for counterbalancing the strength of the Born terms. This is achieved through a destructive interference between these $u$-channel resonances and the Born terms, thereby reducing the strength to a realistic level. However, the method is not so convincing for the huge number of presently available experimental data and, on the other hand, the use of hadronic form factor is more flexible and simultaneously takes into account the fact that nucleons are composite objects, and not point-like. Recently, Ref.[@Bydzovsky:2006wy] has demonstrated that the behavior of forward-angle cross sections as a function of the energy for models that include the hadronic form factors is far from realistic. This finding is depicted in the right panel of Fig. \[fig:had\_ff\], where the hadronic form factors drastically suppress the cross sections at $W\ge 2$ GeV. Although no data are available for comparison at this kinematics, such a strong damping has not been observed in the experimental data near the forward angles. As can be seen in Figs. 4 and 6 of Ref.[@mart06] differential cross sections for $\theta_K=18^\circ-37^\circ$ tend to be flat up to $W=2.5$ GeV. Therefore, the concept of hadronic form factors in meson photoproduction needs to be revisited in the future. Other isospin channels ---------------------- By considering the conservation of quantum numbers, there are six possible reaction channels for kaon photoproduction on the nucleon, three on the proton and three on the neutron. Except for the $\gamma p\to K^+\Lambda$ channel, these possible reactions are shown in Fig. \[fig:iso\], where the calculations obtained from Kaon-Maid are compared with the available data. Note that there are few data points available just recently for the $\gamma n\to K^+\Sigma^-$ differential cross section released by the LEPS collaboration[@kohri06]. Thus, there are no data for the neutral kaon photoproduction on the neutron, $\gamma n\to K^0\Lambda$ and $\gamma n\to K^0\Sigma^0$, although from Fig. \[fig:iso\] it is clear that the data can severely constrain the model, especially in the $\gamma n\to K^0\Lambda$ channel. Furthermore, the investigation of the $YN$ potentials through the $\gamma d\to KYN$ reaction requires the information from these two channels[@salam06]. We note that important progress in this direction has been indicated by the newly published $\gamma d\to K^0KY$ data by the Tohoku’s group[@Tsu08]. It is also important to note that the $K^0$ photoproduction excludes the $t$-channel in the Born terms. Thus, by comparing photoproductions of $K^+$ and $K^0$, one can study the influence of the kaon propagator in kaon photoproduction. Kaon electroproduction ---------------------- It is widely known that the electroproduction process can give more information not available from the photoproduction one. One of the important information is the electromagnetic structure (form factor) of the kaon, which is hiding in the longitudinal cross section ($d\sigma_{\rm L}/d\Omega$). In contrast to the pion case, where the mass of pion is much smaller and therefore the $t$-channel can dominate the process, so that an independent extraction is possible[@Mart:2008sw], the extraction of kaon electromagnetic form factor requires a reliable phenomenological model. Consequently, accurate measurements of $d\sigma_{\rm L}/d\Omega$ are required for this purpose. However, as seen from the lower panels of Fig. \[fig:electro\], the presently available data still do not allow for the extraction of this form factor. Furthermore, to suppress contaminations from nucleon resonances, measurements with $W\ge 2.4$ GeV are recommended. Summary and Conclusion ====================== The recent progress and some important issues in photo- and electroproduction of kaon on the nucleon have been briefly presented. It is quite apparent that more experimental and theoretical works are needed to settle the present problems in this field. Nevertheless, with the modern concept of accelerator and spectrometer technologies available at JLAB, MAMI, ELSA, SPRING8, and other laboratories, we are quite optimistic that these problems can be solved in the near future. Acknowledgments {#acknowledgments .unnumbered} =============== The author acknowledges the support from the Universitas Indonesia. [9]{} M. Kawaguchi and M. J. Moravcsik, Phys. Rev. [**107**]{}, 563 (1957). B. B. Deo and A. K. Bisoi, [*Phys. Rev. D*]{} [**9**]{}, 288 (1974). N. F. Nelipa, Nucl. Phys. [**82**]{}, 680 (1966). A. R. Pickering, [*Nucl. Phys.*]{} [**B66**]{}, 493 (1973). W. Schorsch, J. Tietge, and W. Weilnböck, [*Nucl. Phys.*]{} [**B25**]{}, 179 (1970). N. Levy, W. Majerotto, and B. J. Read, [*Nucl. Phys.*]{} [**B55**]{}, 493 (1973);[**B55**]{}, 513 (1973). R. A. Adelseck, C. Bennhold, and L. E. Wright, [*Phys. Rev. C*]{} [**32**]{}, 1681 (1985). R. A. Adelseck and B. Saghai, [*Phys. Rev. C*]{} [**42**]{}, 108 (1990); J.-C. David, C. Fayard, G. H. Lamot, and B. Saghai, [*Phys. Rev. C*]{} [**53**]{}, 2613 (1996). T. Mart, C. Bennhold, and C. E. Hyde-Wright, [*Phys. Rev. C*]{} [**51**]{}, R1074 (1995); T. Mart and C. Bennhold, [*Nucl. Phys. A*]{} [**585**]{}, 369c (1995). Zhenping Li, Phys. Rev. C [**52**]{}, 1648 (1995). S. Steininger and U. G. Meissner, [*Phys. Lett.  B*]{} [**391**]{}, 446 (1997); B. Borasoy, P. C. Bruns, U. G. Meissner and R. Nissler, [*Eur. Phys. J.  A*]{} [**34**]{}, 161 (2007). N. Kaiser, T. Waas and W. Weise, [*Nucl. Phys.  A*]{} [**612**]{}, 297 (1997); T. Feuster and U. Mosel, [*Phys. Rev.  C*]{} [**59**]{}, 460 (1999). A. V. Anisovich [*et al.*]{}, [*Eur. Phys. J. A*]{} [**25**]{}, 427 (2005); B. Julia-Diaz, B. Saghai, T. S. Lee and F. Tabakin, [*Phys. Rev. C* ]{} [**73**]{}, 055204 (2006). M. Guidal, J. M. Laget and M. Vanderhaeghen, [*Nucl. Phys. A*]{} [**627**]{}, 645 (1997). T. Corthals, J. Ryckebusch and T. Van Cauteren, [*Phys. Rev.  C*]{} [**73**]{}, 045207 (2006), T. Mart and T. Wijaya, [*Acta Phys. Polon.  B*]{} [**34**]{}, 2651 (2003). T. Mart and C. Bennhold, [*Phys. Rev. C*]{} [**61**]{}, 012201 (1999). F. X. Lee, T. Mart, C. Bennhold and L. E. Wright, [*Nucl. Phys.  A*]{} [**695**]{}, 237 (2001); B. S. Han, M. K. Cheoun, K. S. Kim and I. T. Cheon, [*Nucl. Phys. A*]{} [**691**]{}, 713 (2001); A. de la Puente, O. V. Maxwell and B. A. Raue, arXiv:0809.3805 \[nucl-th\]. R.A. Adelseck and B. Saghai, [*Phys. Rev. C*]{} [**42**]{}, 108 (1990). M. Q. Tran [*et al*]{}., [*Phys. Lett. B*]{} [**445**]{}, 20 (1998). K. H. Glander [*et al.*]{}, [*Eur. Phys. J. A*]{} [**19**]{}, 251 (2004) R. Bradford [*et al.*]{}, [*Phys. Rev. C*]{} [**73**]{}, 035202 (2006) M. Sumihama [*et al.*]{}, [*Phys. Rev. C*]{} [**73**]{}, 035214 (2006) K. Hicks [*et al.*]{}, [*Phys. Rev. C*]{} [**76**]{}, 042201 (2007). A. Lleres [*et al.*]{}, [*Eur. Phys. J. A*]{} [**31**]{}, 79 (2007) R. Lawall [*et al.*]{}, [*Eur. Phys. J. A*]{} [**24**]{}, 275 (2005) H. Kohri [*et al.*]{}, [*Phys. Rev. Lett.*]{} [**97**]{}, 082003 (2006) R. Bradford [*et al.*]{}, [*Phys. Rev.  C*]{} [**75**]{}, 035205 (2007). A. Lleres [*et al.*]{}, Eur. Phys. J.  A [**39**]{}, 149 (2009) R. M. Mohring [*et al.*]{}, [*Phys. Rev. C*]{} [**67**]{}, 055205 (2003) D. S. Carman [*et al.*]{}, [*Phys. Rev. Lett.*]{} [**90**]{}, 131804 (2003) P. Ambrozewicz [*et al.*]{}, [*Phys. Rev. C*]{} [**75**]{}, 045203 (2007) T. Mart and A. Sulaksono, [*Phys. Rev. C*]{} [**74**]{}, 055203 (2006) P. Bydžovský and T. Mart, *Phys. Rev. C* [**76**]{}, 065202 (2007). T. Mart, [*Few Body Syst.*]{}  [**42**]{}, 125 (2008). S.B. Gerasimov, [*Sov. J. Nucl. Phys.*]{} [**2**]{}, 430 (1966); S.D. Drell and A.C. Hearn, [*Phys. Rev. Lett.*]{} [**16**]{}, 908 (1966). A. V. Anisovich [*et al.*]{}, [*Eur. Phys. J. A*]{} [**34**]{}, 243 (2007) M. Bockhorst et al., [*Z. Phys. C*]{} [**63**]{}, 37 (1994). S. Capstick and W. Roberts, [*Phys. Rev. D*]{} [**49**]{}, 4570 (1994); [**58**]{}, 074011 (1998). C. Caso [*et al*]{}., [*Eur. Phys. J. C*]{} [**3**]{}, 1 (1998). S. Janssen [*et al.*]{}, [*Eur. Phys. J.  A*]{} [**11**]{}, 105 (2001) A. Martinez Torres, K. P. Khemchandani, U. G. Meissner and E. Oset, arXiv:0902.3633 \[nucl-th\] (2009). R. Nelson and T. Mart, arXiv:0904.3598 \[nucl-th\], to be published. T. Mart and B. I. S. van der Ventel, [*Phys. Rev. C*]{} [**78**]{}, 014004 (2008); T. Mart, [*Nucl. Phys. A*]{} [**815**]{}, 18 (2009). K. Ohta, [*Phys. Rev. C*]{} [**40**]{}, 1335 (1989). H. Haberzettl, [*Phys. Rev. C*]{} [**56**]{}, 2041 (1997). H. Haberzettl, C. Bennhold, T. Mart, and T. Feuster, [*Phys. Rev. C*]{} [**58**]{}, R40 (1998). T. Mart, C. Bennhold, H. Haberzettl, and L. Tiator, KAON-MAID, available at http://www.kph.uni-mainz.de/MAID/kaon/kaonmaid.html. A. Salam [*et al.*]{}, [*Phys. Rev. C*]{} [**74**]{}, 044004 (2006). K. Tsukada [*et al.*]{}, [*Phys. Rev. C*]{} [**78**]{}, 014001 (2008). T. Mart, [*Mod. Phys. Lett.  A*]{} [**23**]{}, 3317 (2008) T. Mart and A. Sulaksono, arXiv:nucl-th/0701007 (2007). [^1]: see Section \[sec:data\] for the definition of old data
{ "pile_set_name": "ArXiv" }
--- abstract: 'We argue that in the two–loop approximation gauge coupling unification in the exceptional supersymmetric standard model (E$_6$SSM) can be achieved for any phenomenologically reasonable value of $\alpha_3(M_Z)$ consistent with the experimentally measured central value.' author: - 'P.  Athron ' - 'S.  F.  King ' - 'R.  Luo ' - 'D.  J.  Miller ' - 'S.  Moretti ' - 'R.  Nevzorov [^1]  [^2] ' title: 'Unification of Gauge Couplings in the E$_6$SSM' --- SHEP–09–21\ DFTT 61/2009\ [ address=[Institut für Kern- und Teilchenphysik, TU Dresden, D-01062, Germany]{}]{} [ address=[School of Physics and Astronomy, University of Southampton, Southampton, SO17 1BJ, UK]{} ]{} [ address=[Department of Physics and Astronomy, University of Glasgow, Glasgow G12 8QQ, UK]{} ]{} [ address=[Department of Physics and Astronomy, University of Glasgow, Glasgow G12 8QQ, UK]{} ]{} [ address=[School of Physics and Astronomy, University of Southampton, Southampton, SO17 1BJ, UK]{} ,altaddress=[Dipartimento di Fisica Teorica, Università di Torino, Via Pietro Giuria 1, 10125 Torino, Italy]{} ]{} [ address=[Department of Physics and Astronomy, University of Glasgow, Glasgow G12 8QQ, UK]{}]{} Introduction ============ The incorporation of weak and strong gauge interactions within Grand Unified Theories (GUTs) based on simple gauge groups such as $SU(5)$, $SO(10)$ or $E_6$ requires the unification of gauge couplings at some high energy scale $M_X$. On the other hand gauge coupling unification makes also possible partial unification of gauge interactions with gravity in the framework of superstring theories. At high energies the $E_6$ symmetry in the superstring inspired models can be broken to rank–5 subgroup $SU(3)_C\times SU(2)_W\times U(1)_Y\times U(1)'$ where $U(1)'=U(1)_{\chi} \cos\theta+U(1)_{\psi} \sin\theta$. Two anomaly-free $U(1)_{\psi}$ and $U(1)_{\chi}$ symmetries are defined by: $E_6\to SO(10)\times U(1)_{\psi},~ SO(10)\to SU(5)\times U(1)_{\chi}$. Here we concentrate on a particular $E_6$ inspired supersymmetric model with an extra $U(1)_{N}$ gauge symmetry that corresponds $\theta=\arctan\sqrt{15}$. Only in this exceptional supersymmetric standard model (E$_6$SSM) [@1]-[@2] the right–handed neutrinos do not participate in gauge interactions. The extra $U(1)_{N}$ gauge symmetry survives to low energies and forbids a bilinear term $\mu H_d H_u$ in the superpotential of the considered model but allows the interaction $\lambda S H_d H_u$. At the electroweak (EW) scale $S$ gets a non-zero vacuum expectation value (VEV), $\langle S \rangle=s/\sqrt{2}$, breaking $U(1)_N$ and an effective $\mu=\lambda s/\sqrt{2}$ term is automatically generated. The E$_6$SSM ============ The E$_6$SSM is based on the $SU(3)_C\times SU(2)_W\times U(1)_Y \times U(1)_N$ gauge group which is a subgroup of $E_6$. To ensure anomaly cancellation the particle content of the E$_6$SSM is extended to include three complete [**$27$**]{} representations of $E_6$ [@1]-[@2]. Each [**$27_i$**]{} multiplet contains a SM family of quarks and leptons, right–handed neutrino $N^c_i$, SM singlet field $S_i$ which carries a non–zero $U(1)_{N}$ charge, a pair of $SU(2)_W$–doublets $H_{1i}$ and $H_{2i}$ with the quantum numbers of Higgs doublets and a pair of colour triplets of exotic quarks $\overline{D}_i$ and $D_i$ which can be either diquarks (Model I) or leptoquarks (Model II) [@1]-[@2]. $H_{1i}$ and $H_{2i}$ form either Higgs or inert Higgs multiplets. We also require a further pair $L'$ and $\overline{L}'$ from incomplete extra $27'$ and $\overline{27'}$ representations to survive to low energies to ensure gauge coupling unification in the one–loop approximation. The presence of a $Z'$ boson and exotic quarks predicted by the E$_6$SSM provides spectacular new physics signals at the LHC which were discussed in [@1]-[@30]. In the two–loop approximation the lightest Higgs boson mass in the E$_6$SSM does not exceed $150-155\,\mbox{GeV}$ [@1]-[@2]. Recently the particle spectrum within the constrained version of the E$_6$SSM was studied [@4]-[@50]. Since right–handed neutrinos have zero charges in the considered model they are expected to gain Majorana masses at some intermediate scale. The heavy Majorana neutrinos may decay into final states with lepton number $L=\pm 1$, creating a lepton asymmetry in the early Universe. Due to the presence of exotic particles the substantial values of the CP asymmetries in the E$_6$SSM can be induced even for a relatively small mass of the lightest right–handed neutrino ($M_1 \sim 10^6\,\mbox{GeV}$) so that the successful thermal leptogenesis may be achieved without encountering gravitino problem [@6]. The superpotential in $E_6$ inspired models involves a lot of new Yukawa couplings that induce non–diagonal flavour transitions. To avoid a flavour changing neutral current (FCNC) problem an approximate $Z^{H}_2$ symmetry is postulated in the E$_6$SSM. Under this symmetry all superfields except $H_d\equiv H_{13}$, $H_u\equiv H_{23}$ and $S\equiv S_3$ are odd. The $Z^{H}_2$ symmetry reduces the structure of the Yukawa interactions to: $$\begin{array}{rcl} W_{\rm ESSM}&\simeq & \lambda_i S(H_{1i}H_{2i})+\kappa_i S(D_i\overline{D}_i)+f_{\alpha\beta}S_{\alpha}(H_d H_{2\beta})+ \tilde{f}_{\alpha\beta}S_{\alpha}(H_{1\beta}H_u)+\\ &&+\mu'(L'\overline{L}')+h^{E}_{4j} (H_d L') e^c_j +W_{\rm{MSSM}}(\mu=0), \end{array} \label{1}$$ where $\alpha,\beta=1,2$ and $i=1,2,3$. Here we assume that all right–handed neutrinos are heavy. The $SU(2)_W$ doublets $H_u$ and $H_d$ play the role of Higgs fields. Therefore it is natural to assume that only $S$, $H_u$ and $H_d$ acquire VEVs. The VEV of the field $S$ breaks $U(1)_N$ symmetry inducing effective $\mu$ term as well as the masses of exotic fermions and $Z'$ boson. To guarantee that only $H_u$, $H_d$ and $S$ develop VEVs in the E$_6$SSM a certain hierarchy between the Yukawa couplings is imposed, i.e. $\kappa_i\sim\lambda_i\gg f_{\alpha\beta}, \tilde{f}_{\alpha\beta}, h^{E}_{4j}$. RG flow of gauge couplings ========================== At high energies the two–loop renormalisation group (RG) flow of gauge couplings in the E$_6$SSM can be parametrised as $$\frac{1}{\alpha_i(t)}=\frac{1}{\alpha_i(M_Z)}-\frac{b_i}{2\pi} t-\frac{C_i}{12\pi}-\Theta_i(t) +\frac{b_i-b_i^{SM}}{2\pi}\ln\frac{T_i}{M_Z}\,, \qquad T_i=\prod_{k=1}^N\biggl(m_k\biggr)^{\frac{\Delta b^k_i}{b_i-b_i^{SM}}}\,, \label{2}$$ where index $i$ runs from $1$ to $3$ and corresponds to $U(1)_Y$, $SU(2)_W$ and $SU(3)_C$ interactions, $b_i$ and $b_i^{SM}$ are the coefficients of the one–loop beta functions in the E$_6$SSM and SM, $t=\ln\left(\mu/M_Z\right)$, $\mu$ is a renormalisation scale, $C_1=0$, $C_2=2$, $C_3=3$, $\Theta_i(t)$ are two–loop corrections and $T_i$ are effective threshold scales. In Eq. (\[2\]) $m_k$ and $\Delta b_i^k $ are masses of new particles and their contributions to the beta functions. In the E$_6$SSM the one–loop beta functions are $b_1=48/5$, $b_2=4$, $b_3=0$. The two–loop contributions to the beta functions of gauge couplings in the E$_6$SSM were calculated in [@8]. Using the approximate solution of the RG equations (RGEs) in Eq. (\[2\]) one can find the value of $\alpha_3(M_Z)$ for which exact gauge coupling unification can be achieved: $$\frac{1}{\alpha_3(M_Z)}=\frac{1}{b_1-b_2}\biggl[\frac{b_1-b_3}{\alpha_2(M_Z)}- \frac{b_2-b_3}{\alpha_1(M_Z)}\biggr]-\frac{1}{28\pi}+\Theta_s-\Delta_s\,,\quad \Delta_s=-\frac{19}{28\pi}\ln\frac{\tilde{M}_{S}}{M_Z}\,, \label{31}$$ $$\Theta_s=\biggl(\frac{b_2-b_3}{b_1-b_2}\Theta_1-\frac{b_1-b_3}{b_1-b_2}\Theta_2+\Theta_3\biggr)\,, \qquad \Theta_i=\Theta_i(M_X), \label{32}$$ where the effective threshold scale $\tilde{M}_S$ can be expressed in terms of the MSSM one $M_S$: $$\tilde{M}_S=\frac{T_2^{172/19}}{T_1^{55/19} T_3^{98/19}}= M_S\cdot\frac{m_{H_{\alpha}}^{12/19}\mu_{\tilde{H}_{\alpha}}^{24/19}\mu^{'18/19}} {m_{\tilde{D}_i}^{18/19}\mu_{D_i}^{36/19}}\simeq \mu'\cdot \biggl(\frac{M_{weak}}{M_{colour}}\biggr)^{4.5}\,. \label{5}$$ In Eq. (\[5\]) $\mu_{D_i}$ and $m_{\tilde{D}_i}$ are the masses of exotic quarks and their superpartners, $m_{H_{\alpha}}$ and $\mu_{\tilde{H}_{\alpha}}$ are the masses of inert Higgs and inert Higgsino fields, while the scalar and fermion components of $L'$ and $\overline{L}'$ are degenerate around $\mu'$. Because the unification of gauge couplings is determined by only one effective threshold scale $\tilde{M}_S$ one can simplify the analysis assuming that $T_1=T_2=T_3=\tilde{M}_S$. In our numerical study we parametrise the $\tilde{M}_S$ in terms of two scales. One of these scales is associated with the SUSY breaking scale $M_S$ while another one is set by the masses of $Z'$ and exotic fermions and bosons which we assume to be degenerate around $M_{Z'}$ for simplicity. Thus we use the SM beta functions to describe the running of $\alpha_i(t)$ between $M_Z$ and $M_S$, then we apply the two–loop RGEs of the MSSM to compute the flow of $\alpha_i(t)$ from $M_S$ to $M_{Z'}$ and the two–loop RGEs of the E$_6$SSM to calculate the evolution of $\alpha_i(t)$ between $M_{Z'}$ and $M_X$ which is equal to $3.5\cdot 10^{16}\,\mbox{GeV}$ in the case of the E$_6$SSM. Since $M_S\simeq \mu/6$ we choose $M_S\simeq 250\,\mbox{GeV}$. We also fix $M_{Z'}=1.5\,\mbox{TeV}$. --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Two–loop RG flow of gauge couplings from EW to GUT scale $M_X$ in the E$_6$SSM (left) and MSSM (right). Thick, dashed and solid lines correspond to the running of $SU(3)_C$, $SU(2)_W$ and $U(1)_Y$ couplings respectively. We used $\tan\beta=10$, $M_S=250\,\mbox{GeV}$, $M_{Z'}=1.5\,\mbox{TeV}$, $\alpha_s(M_Z)=0.118$, $\alpha(M_Z)=1/127.9$ and $\sin^2\theta_W=0.231$. The dotted lines represent the uncertainty in $\alpha_i(t)$ caused by the variation of the strong gauge coupling from 0.116 to 0.120 at the EW scale.](gc-unif16.eps "fig:"){height=".2\textheight"} ![Two–loop RG flow of gauge couplings from EW to GUT scale $M_X$ in the E$_6$SSM (left) and MSSM (right). Thick, dashed and solid lines correspond to the running of $SU(3)_C$, $SU(2)_W$ and $U(1)_Y$ couplings respectively. We used $\tan\beta=10$, $M_S=250\,\mbox{GeV}$, $M_{Z'}=1.5\,\mbox{TeV}$, $\alpha_s(M_Z)=0.118$, $\alpha(M_Z)=1/127.9$ and $\sin^2\theta_W=0.231$. The dotted lines represent the uncertainty in $\alpha_i(t)$ caused by the variation of the strong gauge coupling from 0.116 to 0.120 at the EW scale.](gc-unif14.eps "fig:"){height=".2\textheight"} --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- The results of our numerical studies of gauge coupling unification are summarised in Fig. 1. where the RG flow of gauge couplings in the E$_6$SSM and MSSM are shown. Due to the presence of exotic matter the E$_6$SSM gauge couplings are considerably larger at high energies than in the MSSM. Dotted lines in Fig. 1 represent the changes of the evolution of $\alpha_i(t)$ induced by the variations of $\alpha_3(M_Z)$ within $1\,\sigma$ around its average value. In the E$_6$SSM the two–loop effects lead to the mild growth of $\alpha_3(t)$ as renormalisation scale increases. As a consequence at high energies the uncertainty in $\alpha_3(t)$ caused by the variations of $\alpha_3(M_Z)$ is much bigger in the E$_6$SSM than in the MSSM. The relatively large uncertainty in $\alpha_3(M_X)$ allows one to achieve exact unification of $\alpha_i(t)$ even within $1\,\sigma $ deviation of $\alpha_3(M_Z)$ from its average value. It is worth noticing here that two–loop corrections to $\alpha_i(t)$ are large in the E$_6$SSM and could spoil the gauge coupling unification. Nevertheless the value of $\alpha_3(M_Z)$ that results in the exact gauge coupling unification in the E$_6$SSM is quite close to the experimentally measured central value. Without the inclusion of threshold effects Eq. (\[31\]) gives $\alpha_3(M_Z)\simeq 0.121$ which is very close to the one–loop prediction for $\alpha_3(M_Z)=0.118$. The small difference between one–loop and two–loop predictions for $\alpha_3(M_Z)$ is caused by the remarkable cancellation of different two–loop corrections in Eq. (\[32\]). As in the MSSM the inclusion of threshold effects lowers the prediction for the value of the strong gauge coupling at the EW scale. From Eq. (\[5\]) it is obvious the effective threshold scale in the E$_6$SSM is set by $\mu'$. The term $\mu'L'\overline{L}'$ in the superpotential is not involved in the process of EW symmetry breaking and is not suppressed by the $E_6$ symmetry. Therefore, although the effective threshold scale $\tilde{M}_S$ may be considerably less than $\mu'$, the corresponding mass parameter can be always chosen so that $\tilde{M}_S$ lies in a few hundred GeV range that allows to get the exact unification of gauge couplings for any value of $\alpha_3(M_Z)$ which is in agreement with current data. RN acknowledges support from the SHEFC grant HR03020 SUPA 36878. SM is financially supported in part by the scheme ‘Visiting Professor - Azione D - Atto Integrativo tra la Regione Piemonte e gli Atenei Piemontesi. [9]{} S. F. King, S. Moretti, and R. Nevzorov, *Phys. Rev. D* **73** 035009 (2006). S. F. King, S. Moretti, and R. Nevzorov, *Phys. Lett. B* **634** 278-284 (2006). S. F. King, S. Moretti, and R. Nevzorov, arXiv:hep-ph/0601269. S. F. King, S. Moretti, and R. Nevzorov, “E$_6$SSM,” in *Cairo International Conference on High Energy Physics (CICHEP II)*, edited by S. Khalil, AIP Conference Proceedings 881, American Institute of Physics, New York, 2007, pp 138-143. P. Athron, S. F. King, D. J. Miller, S. Moretti, and R. Nevzorov, *Phys. Rev. D* **80** 035009 (2009). P. Athron, S. F. King, D. J. Miller, S. Moretti, and R. Nevzorov, arXiv:0901.1192 \[hep-ph\]. P. Athron, S. F. King, D. J. Miller, S. Moretti, and R. Nevzorov, arXiv:0810.0617 \[hep-ph\]. S. F. King, R. Luo, D. J. Miller, and R. Nevzorov, *JHEP* **0812** 042 (2008). S. F. King, S. Moretti, and R. Nevzorov, *Phys. Lett. B* **650** 57-64 (2007). [^1]: Based on a talk presented by R. Nevzorov at the SUSY’09 Conference, Boston, USA, 5-10 June, 2009 [^2]: On leave of absence from the Theory Department, ITEP, Moscow, Russia
{ "pile_set_name": "ArXiv" }
--- abstract: 'We obtain new HI and $^{13}$CO images around Supernova Remnants (SNR) Kes 69 and G21.5-0.9. By comparing HI spectra with $^{13}$CO emission spectra, we significantly revise the kinematic distance for Kes 69 to $\sim$ 5.5 kpc, which was 11.2 kpc, and refine the kinematic distance for G21.5-0.9 to $\sim$ 4.8 kpc. For Kes 69, the highest velocity of absorption is $\sim$ 86 km s$^{-1}$ and a prominent HI emission feature at $\sim$ 112 km s$^{-1}$ has no respective absorption. These new results suggest that Kes 69 is associated with a newly detected extended 1720 MHz OH maser at velocity of $\sim$ 85 km s$^{-1}$ that originates from within the bright southern radio shell of Kes 69. For G21.5-0.9, the highest velocity of absorption is $\sim$ 67 km s$^{-1}$. The HI absorption spectra of the nearby bright source PMN J1832-1035 and of Kes 69 show a common absorption feature at velocity of $\sim$ 69 km s$^{-1}$, which is not seen for G21.5-0.9. The resulting velocity of $\sim$ 68 km s$^{-1}$ gives the best distance estimate of $\sim$ 4.8 kpc for G21.5-0.9 and associated young pulsar J1833-1034.' author: - | W. W. Tian [^1], D.A. Leahy\ Department of Physics $\&$ Astronomy, University of Calgary, Calgary, Alberta T2N 1N4, Canada date: 'Accepted Aug. 29 2008. Received 2008; in original form 2008 Aug. 1' title: 'The Distances of SNRs Kes 69 and G21.5-0.9 from HI and $^{13}$CO Spectra' --- \[firstpage\] supernova remnants - masers - pulsars:individual:J1833-1034 - cosmic rays - methods:data analysis Background ========== Determining distances to Galactic objects (HII regions, pulsars (PSR) and supernova remnants (SNR) etc.) may help understand the kinematics of the Milky Way. Also, distance measurement of SNRs is a key to obtain their basic parameters such as the luminosity, size and age. As an energetic class of objects, SNRs are associated with many highly active astrophysical phenomena, e.g. anomalous X-ray pulsars, soft $\gamma$-ray repeaters, Pulsar Wind Nebula (PWN), non-thermal X-rays and very-high-energy $\gamma$-rays (see Yang et al. 2008 for a review). The determination of distance of a SNR may test reality of a SNR/PSR/PWN or SNR/TeV $\gamma$-ray source association, and help constrain the mass range of the progenitor star and type of supernova responsible for the remnant. It can provide direct evidence whether a SNR is physically associated with molecular clouds, or if strong interaction between a SNR shock and surrounding clouds is possible, which is widely believed to be a source of TeV $\gamma$-ray and non-thermal X-ray emission in the Milky Way. The SNR/molecular cloud interaction may also potentially lead to new generations of star formation, so constitutes an important part of Galactic ecology. Comparing HI absorption spectra toward Galactic SNRs with respective HI and $^{13}$CO emission spectra along the line of sight, we previously revised distances to five SNRs and several overlapping HII regions (Tian et al. 2007; Tian & Leahy 2008; Leahy & Tian 2008a;). Some values differ greatly from previously published values, e.g. Kes 75 (Leahy & Tian 2008b), PWN G54.1+0.3 (Leahy, Tian & Wang 2008), because our methods implement improved background subtraction and spurious emission rejection, and resolve the near/far distance ambiguity in the inner Galaxy, which may have been previously done incorrectly. The distance changes can result in significant changes in interpretation of the SNR (and associated object) properties. As the newest paper of a series, we use the methods to directly re-measure distances of two more SNRs: Kes 69 which has 1720 MHz OH maser emission (Hewitt et al. 2008), and G21.5-0.9 which hosts a young pulsar (Camilo et al. 2006; Bietenholz & Bartel 2008) and also has TeV $\gamma$-ray emission (i.e. HESS J1833-105, Djannati-Atai et al. 2007). 1720 MHz masers are associated with warm, shocked molecular gas and are seen as signposts of SNR-molecular cloud interactions (Wardle & Yusef-Zadeh 2002), therefore they likely give reliable distance estimates to SNR/molecular cloud systems. An OH maser at velocity of 69.3$\pm$0.7 km s$^{-1}$ (hereafter, we use $\sim$ 70 km s$^{-1}$to describe the radial velocity of this maser) was detected toward Kes 69 (Green et al. 1997), leading to a distance of $\sim$ 11.2 kpc for the remnant. However, the masers’ site is outside of the northeastern radio and X-ray shells of Kes 69 (Yusef-Zadeh et al. 2003). This spurred us to re-measure its distance by analyzing HI+$^{13}$CO spectra in the direction of Kes 69 in order to test the reality of the claimed association. SNR G21.5-0.9 is an interesting SNR nearby Kes 69, so we also analyze HI and CO spectra in the direction of G21.5-0.9. In this paper, we significantly revise the distance to Kes 69, and refine the distance estimation to G21.5-0.9 by analyzing HI and CO spectra. We use the Galactic rotation curve model and recent measurements of the parameters for this (i.e. R$_{0} \sim$ 8 kpc, Eisenhauer et al. 2005; V$_{0}$$\sim$ 220 km s$^{-1}$, Feast & Whitelock 1997, Reid & Brunthaler 2004). The radio data come from 1420 MHz continuum plus HI-line observations of the VLA Galactic Plane Survey (VGPS, Stil et al. 2006) and the $^{13}$CO-line (J = 1-0) observations of the Galactic ring survey (Jackson et al. 2006) of the Five College Radio Astronomy Observatory 14 m telescope. The short-spacing information for the HI spectral line images is from additional observations with the 100 m Green Bank telescope (GBT) of the National Radio Astronomical Observatory (NRAO). Results ======= The Radio Continuum and HI Channel Images ----------------------------------------- The VLA 1420 MHz continuum image of Kes 69, G21.5-0.9 and background compact source PMN J1832-1035 (i.e. G21.35-0.63) is shown in Fig. 1 (upper left). We have searched the VGPS radial velocity range from -113 to 165 km s$^{-1}$ for features in the HI emission which might be related to the morphology of the two SNRs. Some negative HI features are found to be strongly correlated with the continuum intensity of the SNRs, indicating that they are caused by absorption in HI between the SNR and the earth (see the HI channel images shown in the second and third rows of Fig. 1). A CO channel image is also given in Fig. 1 (upper right). The continuum image clearly reveals that the site of the 1720 MH OH maser at $\sim$ 70 km s$^{-1}$ is outside of the detected radio emission of Kes 69 (marked by the plus in the first panel of Fig. 1). The CO channel image shows that a cloud at velocity of $\sim$ 86 km s$^{-1}$ overlaps Kes 69, which is likely responsible for producing the clear HI dip in the HI channel image at $\sim$ 86 km s$^{-1}$ at the location of the brightest continuum emission from Kes 69. This means that the CO cloud should be adjacent to or in front of Kes 69. The HI channel image at $\sim$ 112 km s$^{-1}$, which is close to the tangent point velocity of $\sim$ 120 km s$^{-1}$ in the direction towards Kes 69 (see Fig. 2), shows that an HI cloud overlaps Kes 69, but doesn’t produce any absorption from Kes 69, so the HI cloud is behind Kes 69. The HI channel image at $\sim$ 67 km s$^{-1}$ shows that HI covers Kes 69, G21.5-0.9 and PMN J1832-1035, and produces absorption for each source. So the HI cloud is in front of all three sources. Other HI at $\sim$ 69 km s$^{-1}$ (the lower left panel of Fig. 1) yields an upper limit distance to G21.5-0.9 because it covers all three sources, and produces absorption from Kes 69 and PMN J1832-1035 but not from G21.5-0.9. The HI and CO Spectra --------------------- We construct HI emission and absorption spectra of both SNRs and of PMN J1832-1035 in order to verify the above results in more detail. The extraction regions for these spectra are shown by the boxes in the first panel of Fig. 1. The $^{13}$CO emission spectra from boxes 1, 2, 3 and from the OH maser site are also analyzed. These spectra are shown in Fig. 2. We note that the CO (molecular cloud) emission features all have associated HI emission features indicating HI associated with H$_{2}$, likely due to the atomic envelope of the molecular cloud. Several HI emission features have no associated CO emission (i.e. H$_{2}$), as many atom clouds have no molecular component. The highest velocity of absorption to Kes 69 is $\sim$ 86 km s$^{-1}$ at which there is a CO cloud (see first and second rows of Fig. 2). The HI spectra of Kes 69 show a prominent HI emission feature at $\sim$ 112 km s$^{-1}$ with no respective absorption. These give a strict distance constraint to Kes 69, i.e. Kes 69 is in the range of distance of 5.5 kpc to 7.4 kpc. The $^{13}$CO spectrum of the OH maser site near Kes 69 shows little $^{13}$CO emission at $\sim$ 70 km s$^{-1}$ (above 0.2 K, see the lower right panel), showing there is no high column density cloud. However this cannot exclude there still possibly exists a small high density cloud at the OH maser site at 70 km s$^{-1}$. For G21.5-0.9, the highest velocity of HI absorption is $\sim$ 67 km s$^{-1}$ at which there is a CO cloud (third row of Fig. 2). Another CO cloud at $\sim$ 74 km s$^{-1}$ does not produce any respective HI absorption, so this cloud is behind G21.5-0.9. Further, the HI absorption spectra of both nearby PMN J1832-1035 and Kes 69 show a strong absorption feature at velocity of $\sim$ 70 km s$^{-1}$. These strongly support the above HI channel image analysis and constrain the distance, i.e. G21.5-0.9 is between clouds at velocities of $\sim$ 67 km s$^{-1}$ and $\sim$ 69 km s$^{-1}$, i.e. d $\sim$ 4.8 kpc for velocity of $\sim$ 68 km s$^{-1}$. Discussion ========== Distances to Kes 69 and G21.5-0.9 --------------------------------- The OH-maser determined distance should be reasonably consistent with the HI+CO determined distance to the same SNR. We note that distances are derived here from a circular rotation model, and contain errors due to uncertainties in the rotation model and non-circular motions. E.g. using the observed $l-V$ diagram, Weiner & Sellwood (1999) derived the radial velocity distribution in the inner galaxy. Applying to Kes 69, the distances are reduced by $\sim$ 0.4 kpc compared to the circular rotation model. Observed random motion of up to 7 km s$^{-1}$ (Shaveret al. 1982) yields a distance uncertainty of $\sim$ 0.3 kpc for this case. Using the circular rotation model, our analysis of the HI+CO spectra reveals that Kes 69 has a distance of 5.5 to 7.4 kpc, far smaller than 11.2 kpc determined by the OH maser at 70 km s$^{-1}$. This is strong evidence against the OH maser being associated with Kes 69. This is supported by further evidence: the site of the OH maser is outside the detected radio emission of Kes 69, and no CO emission at $\sim$ 70 km s$^{-1}$ is seen in the CO spectrum of the OH maser. A collisional pumping model shows that shock-excited OH maser emission at 1720 MHz appearing behind a SNR shock front results from the interaction of an SNR with an adjacent, warm, dense shocked molecular cloud (Wardle & Yusef-Zadeh 2002). Furthermore, an extended 1720 MHz OH maser at velocity of $\sim$ 85 km s$^{-1}$ has been detected from within the bright southern radio shell of Kes 69 by GBT and VLA observations (Hewitt et al. 2008). The new OH maser fits nicely with the highest velocity of HI absorption in the direction of Kes 69, therefore we conclude that Kes 69 is located at a distance of $\sim$ 5.5 kpc at the near side kinematic distance for the OH maser at $\sim$ 85 km s$^{-1}$ and consistent with the highest observed HI absorption velocity of $\sim$ 86 km s$^{-1}$. A bright $^{13}$CO cloud at $\sim$ 86 km s$^{-1}$ overlapping Kes 69 has been revealed here (Fig. 1 upper right). A research group from Nan Jing University newly finds $^{12}$CO, $^{13}$ CO and HCO$^{+}$ emissions near 85 km/s$^{-1}$ from the bright southern shell (Zhou et al. 2008), consistent with the previous formaldehyde molecular (H$_{2}$CO) and the 1665/7 MHz OH absorption line observations at velocities of 82 – 87 km s$^{-1}$ (Wilson 1972, Turner 1970). All these support our conclusion on the distance of 5.5 kpc to Kes 69. So the 1720 MHz maser at 70 km s$^{-1}$ is likely not associated with Kes 69. G21.5-0.9 is located between a cloud at velocity of $\sim$ 67 km s$^{-1}$ and that of $\sim$ 69 km s$^{-1}$ (i.e. d $\sim$ 4.8 kpc). G21.5-0.9 hosts a young pulsar PWN J1833-1034. New measurement refines the most recent velocity measurement to the SNR/PSR system by Camilo et al. 2006 who analyzed HI spectra data from the VLA and the Leiden/Dwingeloo 25 m telescope observations, and lower resolution $^{12}$CO data from the CfA 1.2 m telescope observations, to obtain lower and upper limits on the velocity of 65 km s$^{-1}$ and 76 km s$^{-1}$ (i.e. d $\sim$ 4.7$\pm$ 0.4 kpc). The Evolutionary State of Kes 69 -------------------------------- The 1420 MHz continuum map shows Kes 69 has a roughly elliptical outline 26$^\prime$ by 20$^\prime$ whereas the ROSAT PSPC image (Yusef-Zadeh et al. 2003), also elliptical and oriented in the same way, has smaller dimensions of 20$^\prime$ by 13$^\prime$. This is mainly due to absence of X-rays from the northwest radio filament (upper left side in Fig. 1 here) and from the northeast radio wing (lower left in Fig. 1), but also the X-rays do not extend out to the edge of the main radio filament in the south. We use 20$^\prime$ as the mean angular diameter of Kes 69, which is the mean of radio and X-ray values. At the distance of 5.5 kpc, the diameter is $\sim$ 32 pc. We apply a Sedov model (Cox 1972) to estimate the parameters of Kes 69, using the ROSAT PSPC temperature of 1.6 keV and X-ray luminosity of 8.4$\times 10^{34}$ erg/s (Yusef-Zadeh et al. 2003). The high temperature and moderate X-ray luminosity are indicative of explosion in a low density medium. Application of the Sedov model yields an age of 5000 years, an explosion energy of 0.8$\times 10^{51}$ erg, and a pre-explosion density of $\sim$0.1 cm$^{-3}$. Based on the estimated errors in the ROSAT PSPC spectral parameters, the errors in these values are $\sim$60% for age and explosion energy and $\sim$30% for density. These results are consistent with observation of OH maser emission from the main south filament of Kes 69 if the explosion occurred in a moderately low density cavity ($\sim$0.1 cm$^{-3}$) and has only recently ($<<$ 5000 years ago) run into dense molecular gas at the main southern filament. Supporting this are Spitzer infrared observations of line emissions from shocked molecular gas in Kes 69 (Reach et al. 2006). For the ROSAT temperature of 1.6 keV, the SNR shock velocity is 1250 km s$^{-1}$ in 0.1 cm$^{-3}$ gas, but is slowed down to $\sim$ 4 km s$^{-1}$ in cold molecular gas with density of $\sim$ $10^{4}$ cm$^{-3}$. The shock would still be supersonic, since the sound speed is 0.2 to 0.6 km s$^{-1}$ in 10 to 100 K molecular hydrogen, and would result in mild heating of the gas to $\sim$ 4000 K. This would subsequently result in good conditions for production of the observed OH maser emission (Wardle 1999). In summary, we significantly revise the distance to Kes 69 and obtain a better distance to G21.5-0.9, by an analysis using HI and $^{13}$CO spectra. The evolutionary state of Kes 69 is evaluated, and it is consistent with a moderate aged SNR, just recently encountering a dense molecular cloud. (100,100) (-180,360) (-10,590) (-235,385) (30,385) (-235, 210) (80, -80) (120,120) (-180,420) (60,560) (-180,275) (60,415) (-180,140) (60,280) (-180,0) (60,140) Acknowledgements {#acknowledgements .unnumbered} ================ WWT and DAL acknowledge support from the Natural Sciences and Engineering Research Council of Canada. WWT appreciates support from the Natural Science Foundation of China. This publication makes use of molecular line data from the Boston University-FCRAO Galactic Ring Survey (GRS). The GRS is a joint project of Boston University and Five College Radio Astronomy Observatory, funded by the National Science Foundation. The NRAO is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. Camilo, F., Ransom, S.M., Gaensler, B.M., Slane, P.O., Lorimer, D.R., Reynolds, J., Manchester, R.N., Murray, S.S. 2006, ApJ, 637, 456 Cox, D. 1972, ApJ, 178, 159 Djannati-Ataï, A. for the HESS Collaboration, 2007, in 30th ICRC, Merida, Mexico, arXiv:0710.2247 Bietenholz, M.F., Bartel, N. 2008, MNRAS, 386, 1411 Eisenhauer, F., Genzel, R., Alexander, T. et al. 2005, ApJ, 628, 246 Feast, M., Whitelock, P. 1997, MNRAS, 291, 683 Green, A. J., Frail, D. A., Goss, W. M., Otrupcek, R. 1997, AJ, 114, 2058 Hewitt, J.M., Yusef-Zadeh, F., Wardle. M. 2008, ApJ, 683, 189 Jackson, J.M., Rathborne, J.M., Shah, R.H. et al. 2006, ApJ Suppl., 163, 145 Leahy, D.A., Tian, W.W. 2008a, AJ, 135, 167 Leahy, D.A., Tian, W.W. 2008b, A&A, 480L, 25 Leahy, D.A., Tian, W.W., Wang, Q.D. 2008, AJ, 136, 1477 Reach, W.T., Rho, J., Tappe, A. et al. 2006, AJ, 131 1479 Reid, M.J., Brunthaler, A. 2004, ApJ, 616, 872 Shaver, P.A., Radhakrishnan, V., Anantharamaiah, K.R., et al. 1982, A&A, 106, 105 Stil, J.M., Taylor, A.R., Dickey, J.M. et al. 2006, AJ, 132, 1158 Tian, W.W., Leahy, D.A., Wang, Q.D. 2007, A&A, 474, 541 Tian, W.W., Leahy, D.A. 2008, ApJ, 677, 292 Turner, B.E. 1970, ApJ, 6, 99L Weiner, B.J., Sellwood, J.A. 1999, ApJ, 524, 112 Wardle, M., Yusef-Zadeh, F. 2002, Science, 28, 2350. Wardle, M. 1999, ApJ, 525, 101L Wilson, T.L. 1972, A&A, 19, 354 Yang, X.J., Lu, F., Tian, W.W. 2008, Progress in Natural Science, 9, 1057. arXiv:0805.3210 Yusef-Zadeh, F., Wardle, M., Rho, J., Sakano, M. 2003, ApJ, 585, 319 Zhou, X., Chen, Y., Su, Y., Yang, J. 2008, ApJ, accepted. arXiv:0809.4306. [^1]: E-mail: [email protected]
{ "pile_set_name": "ArXiv" }
--- abstract: 'Single quantum dots are solid-state emitters which mimic two-level atoms but with a highly enhanced spontaneous emission rate. A single quantum dot is the basis for a potentially excellent single photon source. One outstanding problem is that there is considerable noise in the emission frequency, making it very difficult to couple the quantum dot to another quantum system. We solve this problem here with a dynamic feedback technique that locks the quantum dot emission frequency to a reference. The incoherent scattering (resonance fluorescence) represents the single photon output whereas the coherent scattering (Rayleigh scattering) is used for the feedback control. The fluctuations in emission frequency are reduced to 20 MHz, just $\sim 5$% of the quantum dot optical linewidth, even over several hours. By eliminating the $1/f$-like noise, the relative fluctuations in resonance fluorescence intensity are reduced to $\sim 10^{-5}$ at low frequency. Under these conditions, the antibunching dip in the resonance fluorescence is described extremely well by the two-level atom result. The technique represents a way of removing charge noise from a quantum device.' author: - 'Jonathan H. Prechtel' - 'Andreas V. Kuhlmann' - Julien Houel - Lukas Greuter - Arne Ludwig - Dirk Reuter - 'Andreas D. Wieck' - 'Richard J. Warburton' title: 'A frequency-stabilized source of single photons from a solid-state qubit' --- \[sec:Introduction\] Single photons are ideal carriers of quantum information. A quantum state stored in one of the degrees of freedom of the photon’s wave packet (polarization, phase or time-bin) can be maintained over long distances. Single photons are therefore important in quantum communication, for coupling remote stationary qubits, the basis of a quantum repeater, or for coupling different elements in a quantum device. Furthermore, single photons are the seed for a variety of quantum optics experiments. A key challenge is to develop a single photon source [@Lounis2005]. Key parameters are fidelity of the antibunching, flux, wavelength and photon indistinguishability. Remarkably, solid-state emitters are presently better able to meet these demands than atomic systems (single atoms or parametric down conversion). In particular, spontaneous emission from individual quantum dots embedded in an inorganic semiconductor is a very promising source of highly antibunched, high flux, indistinguishable photons [@Michler2000; @Santori2002; @Shields2007]. The antibunching, particularly with resonant excitation, is very high [@Muller2007]. The radiative lifetime is very short, typically just less than 1 ns [@Dalgarno2008a]. The flux is usually limited by the poor collection efficiency: most of the light is internally reflected at the GaAs-vacuum interface. However, this problem can be solved by nano-structuring the photonic modes to create a micro-cavity [@Vahala2003] or a photonic nanowire [@Claudon2010]. In the latter case, collection efficiencies of $\sim 70$% have been achieved. The photon indistinguishability is very high for successive photons [@Santori2002]. Based on the optical linewidth, typically a factor of two above the transform limit when measured with resonant excitation [@Hogele2004; @Atature2006; @Houel2012; @andi], the indistinguishability is also reasonably high for photons emitted widely separated in time. Furthermore, a single quantum dot has also been developed as a spin qubit [@Warburton2013], facilitating an interface between stationary qubits and photons [@Yilmaz2010; @DeGreve2012; @Gao2012]. Unlike a real atom, the exact transition wavelength of a quantum dot is not locked to any particular wavelength and varies considerably from quantum dot to quantum dot. However, the host semiconductor can be designed so that considerable possibilities for tuning the emission wavelength exist. Electric field tuning [@Warburton2002; @Bennett2010] and strain tuning [@Seidl2006a; @Joens2011] allow the emission wavelength to be tuned over several nanometres. A major problem remains. The emission wavelength is not constant: it varies randomly over time, even in very controlled environments at low temperature. The culprit at low frequency is electrical noise in the semiconductor which shifts the emission wavelength via the Stark effect [@andi]. This noise has a $1/f$-like power spectrum resulting in, first, large and uncontrolled drifts at low frequencies and second, an undefined mean value. This noise, while poorly understood, is ubiquitous in semiconductors and makes it very difficult to couple an individual quantum dot to another quantum system, another quantum dot for instance, or an ensemble of cold atoms. We present here a new scheme which solves this problem: we create a stream of single photons with a wavelength which remains constant even over several hours. The output of our quantum device is a stream of single photons generated by resonance fluorescence (RF) from a single quantum dot. RF has considerable advantages over non-resonant excitation of photoluminescence: the linewidth is much lower[@Houel2012; @andi] and the antibunching is much better. We lock the wavelength of the quantum device to a stable reference. We generate an error signal, a signal with large slope at its zero-crossing, by measuring the differential transmission, $\Delta T/T$, simultaneously [@Alen2003; @Hogele2004; @Karrai2003]. The control variable is the voltage $V_g$ applied to a surface gate which influences the quantum dot frequency via the Stark effect. The performance of the feedback scheme is characterized by, first, measuring a series of snap-shots of the optical resonance to assess the residual frequency jitter; and second, by carrying out a full analysis of the noise in the RF. This scheme goes well beyond previous attempts at single emitter stabilization in the solid-state [@Acosta:2012; @Akopian:2013]. The absolute frequency of the quantum dot emission is locked with an uncertainty of just $20$ MHz. We observe a reduction in the noise power up to a frequency of $\sim 100$ Hz, high enough to eliminate the substantial drifts at low frequency. Arguably, these low frequency fluctuations have a classical nature, reflecting charge noise in our solid-state device on millisecond or second time-scales. On much shorter time scales, there are clear quantum effects: the intensity correlation coefficient exhibits a clear dip between 0 and $\pm 2$ ns; electron spins have decoherence times in the $\mu$s regime [@Warburton2013]. As such, this experiment represents a first step towards bridging these time scales, i.e. quantum control of a solid-state emitter. ![ (a) Schematic view of the experiment. The narrowband laser is stabilized to a fixed frequency by a wavemeter which in turn is stabilized to a HeNe laser. Laser light is guided through optical fibers (yellow) and microscope optics before it is focused onto the sample, driving the X$^{0}$ transition resonantly (BS = beam-splitter, PBS = polarizing BS, Pol. = linear polarizer). Two simultaneous measurements of X$^{0}$ scattering are performed: resonance fluorescence (RF) with a dark field technique, detected with an avalanche photo diode (APD) and absorption with a photodiode (PD) underneath the sample. The dynamic stabilization is realized with an active PID feedback loop which corrects for fluctuations in the transition energy using the gate voltage $V_{g}$ and the square wave modulation of a function generator (FG). (b) RF signal of the fine-structure splitted X$^{0}$ emission of a single quantum dot at wavelength 936.5 nm, a power corresponding to a Rabi energy $\Omega$ of 0.74 $\mu$eV and a temperature of . A detuning is achieved by sweeping the gate voltage. The solid red line is a Lorentzian fit to the data with linewidth $\Gamma$ = 1.28 $\mu$eV (309 MHz) and $\Gamma$ = 1.45 $\mu$eV (350 MHz) and with a fine structure splitting . (c) The differential transmission ($\Delta T/T$) signal on the same quantum dot with integration time 100 ms per point. The red curve is a fit to the derivative of the two Lorentzians. The signal around the zero crossing point ($\Delta T/T = 0$) is used to generate an error signal for the feedback scheme.[]{data-label="fig:1"}](figure1.pdf){width="\linewidth"} A sketch of the experimental concept is shown in . A linearly-polarized resonant laser is focused onto the sample surface and drives the optical transition. The resonance fluorescence of the quantum dot is collected with a polarization-based dark field technique [@Houel2012; @Matthiesen2012; @Yilmaz2010], described in detail elsewhere [@andiRSI]. Simultaneously, the optical resonance is detected in transmission by superimposing a sub-linewidth modulation to the gate. The transmission signal arises from an interference of quantum dot scattering with the driving laser [@Karrai2003]. The incoherent part, i.e. the resonance fluorescence, averages to zero in transmission; what is detected instead is the coherent scattering, i.e. the Rayleigh scattering. In this way, the experiment utilizes both incoherent and coherent parts of the scattered light, for the single photon output and control, respectively. With a small modulation, the transmission signal has a large slope with zero crossing at zero detuning and is therefore ideal for the generation of an error signal. $\Delta T/T$, the error signal, is recorded with a lock-in amplifier to reject noise and the lock-in output is fed into a classical feedback scheme. The feedback output is, like the modulation, applied to the gate electrode of the device. The set-point of the control loop is the zero crossing with the goal of locking the peak of the quantum dot RF spectrum to the laser. The laser itself is locked to a HeNe laser reference. \[sec:Experiment\] The self-assembled InGaAs quantum dots, grown by molecular beam epitaxy, are integrated into a semiconductor charge-tunable heterostructure [@Warburton2000]. The quantum dots are located 25 nm above a heavily n-doped GaAs back contact ($n=1.7 \times 10^{18}$ cm$^{-3}$). The intermediate layer, undoped GaAs (25 nm), acts as a tunneling barrier. A 150 nm GaAs layer caps the quantum dots and an AlAs/GaAs superlattice (68 periods of AlAs/GaAs 3 nm/1 nm) completes the heterostructure. A Ti/Au (5 nm/10 nm) Schottky gate is deposited on the sample surface; Ohmic contacts are prepared to the back contact. Bias $V_g$ is applied between the Schottky gate and the back contact. The sample is placed in a liquid helium bath cryostat at 4.2 K with a residual magnetic field of 10 mT. The single quantum dot spectroscopy is performed with a confocal microscope. The continuous wave laser has a short-term linewidth of 1 MHz. Long-term wavelength stability of $\sim 2$ MHz is achieved by locking the laser to a high resolution wavemeter, itself locked to a high quality HeNe laser. The size of the focal spot and the collection efficiency of the single quantum dot RF are both enhanced with a half-sphere zirconia solid immersion lens positioned on top of the Schottky gate. Fig. \[fig:1\](b) shows a RF signal from the neutral exciton transition, $\ket{0}\leftrightarrow \ket{X^{0}}$, where $\ket{X^{0}}$ represents an electron-hole complex and $\ket{0}$ the crystal ground state. The RF is detected with a silicon avalanche photodiode (APD) in single photon counting mode and the detuning of the quantum dot resonances relative to the constant frequency laser is achieved in this case with the Stark shift induced by the bias $V_{g}$. The X$^{0}$ exhibits a fine structure splitting of 11.8 $\mu$eV, the two lines having linewidths $\Gamma=1.45$, 1.28 $\mu$eV close to the transform limit of $\Gamma_0=\hbar/\tau_{\rm r} = 0.93$ $\mu$eV (220 MHz) where $\tau_r$ is the radiative lifetime of the exciton transition ($\tau_r=(0.71 \pm 0.01)$ ns here). A sub-linewidth square-wave modulation at 527 Hz is applied to the Schottky gate. This broadens both X$^{0}$ transitions slightly, here the “red" transition from $\Gamma=1.45$ to $\Gamma=2.58$ $\mu$eV. The transmitted light is detected with an in situ photodiode connected to a room temperature current-voltage preamplifier. Lock-in detection of the $\Delta T/T$ signal is shown in Fig. \[fig:1\](c). With the sub-linewidth modulation, the $\Delta T/T$ resonance is proportional to the derivative of the RF spectrum[@Alen2003]. There are two points which cross with high slope through zero, one for each X$^{0}$ transition. Both crossing points enable a feedback scheme: $\Delta T/T$ provides the error signal, $V_g$ the control parameter. For instance, if the transition energy increases due to electric fluctuations in the sample, $\Delta T/T$ moves away from zero. Once this is detected, a modified $V_g$ is applied to the gate to bring the resonance back to the set point. For the feedback circuit we use a PID loop. The proportional factor $P=0.1$ is chosen with respect to the slope of the error signal, while the integral $I=0.06$ and the derivative constant $D=6\times 10^{-5}$ were obtained by tuning methods. The signal:noise ratio in the $\Delta T/T$ circuit allows us to run the feedback scheme with a bandwidth up to $\sim 50$ Hz matching the frequency range of the device’s charge fluctuations. The fluctuations of the nuclear spins exceed the bandwidth of the feedback[@andi]. The “red" X$^{0}$ transition was used for the subsequent feedback experiments because it has a higher $\Delta T/T$ contrast than the “blue" X$^{0}$ transition. The performance of the single quantum dot frequency stabilization is put to the test in a stroboscopic experiment. The X$^0$ transition energy is mapped with a second laser (linewidth also ). The first laser stabilizes the transition with the feedback scheme at a power corresponding to a Rabi energy $\Omega$ of 0.74 $\mu$eV. A second laser of identical power is tuned with triangular function back and forth through the same transition with a rate of 8.0 $\mu$eV/s. The sum of the power of both lasers was selected to lie below the power at which power broadening becomes significant. A Lorentzian function is used to fit the data and the center position of the resonance is extracted. In this way, a “snap-shot" of the resonance position is recorded every 5 s with “exposure time" 100 ms for a total of 1,000 s. The distribution of the peak position can be seen in the histogram in Fig. \[fig:2\]. In Fig. \[fig:2\](a) and (b), the scanning laser results in an asymmetry: the resonance frequency is more likely to lie at positive detunings on sweeping from negative to positive detunings, and vice versa. This is probably related to the so-called “dragging" [@Latta2009] which is very pronounced on this quantum dot at high magnetic fields (above 0.1 T) [@andiRSI]: the nuclear spins polarize in such a way as to maintain the resonance with the laser over large detunings. In other words, it is likely that the asymmetries in Fig. \[fig:2\](a) and (b) are first hints of dragging. The histogram in (c) is a combination of the data sets of (a) and (b) which are influenced least by dragging (up-sweeps at negative detuning, down-sweeps at positive detuning). Without the stabilizing loop, the long term drift results in a broader distribution, Fig. \[fig:2\](d), the strong asymmetry reflecting the $1/f$-like noise, a drift to the red in this particular case. The fluctuations in resonance positions are quantified with the standard deviation $\sigma_{E}$ of the peak positions. Without stabilization Fig. \[fig:2\](d), $\sigma_{E}^{\rm OFF}=0.250$ $\mu$eV (61 MHz). With active stabilization, $\sigma=0.102$ $\mu$eV (25 MHz). This value is small enough to be influenced by shot noise in each data point which results in an energy uncertainty on fitting each spectrum to a Lorentzian. The shot noise results in an energy jitter of 0.049 $\mu$eV, giving $\sigma_{E}^{\rm ON}=0.089$ $\mu$eV (22 MHz), 36% of $\sigma_{E}^{\rm OFF}$. The measurement of $\sigma_{E}$ represents a measurement of the noise in a bandwidth from $\sim 1$ mHz to $\sim 3.1$ Hz. (Noise at higher frequencies is reflected in the linewidth $\Gamma$.) The ratio $\sigma_{E}^{\rm OFF}:\sigma_{E}^{\rm ON}$ would increase if lower frequencies were included on account of the $1/f$-like noise. ![image](figure2.pdf){width="\linewidth"} The ultimate operation capability of the stabilization system is limited by the random noise in the output of the PID electronics. In Fig. \[fig:1\](c) the noise in the $\Delta T/T$ signal is $\sigma_{\Delta T/T}=1.45 \times 10^{-4}$. In the ideal case, this determines the energy jitter of the quantum dot resonance position[@nag], $$\sigma_{E, {\rm min}}= \frac{d\delta}{d \Delta T} \sigma_{\Delta T/T} \simeq 0.013\; \mu{\rm eV} \; (3 \; {\rm MHz})$$ where $\delta$ is the detuning. This limit, $\sim 100$ times smaller than the linewidth, shows the power of this technique. We have not yet reached this limit in practice. Nevertheless, stabilization with a residual jitter down to just $\sigma_f=20$ MHz is achieved. ![(Color) (a) Time trace of the resonance fluorescence (RF) of a single quantum dot (the one from Fig. \[fig:1\]) with $\delta$ = 0 $\mu$eV recorded over several hours. The binning time was $t_{\rm bin}=100$ ms. The time trace is plotted with (red) and without (blue) the dynamic stabilization scheme. (b) Noise spectra of the normalized RF signal, $S(t)/\langle S(t)\rangle$, corresponding to the time traces of (a) after correction for external noise sources. []{data-label="fig:3"}](figure3.pdf){width="0.5\linewidth"} The frequency locking feedback scheme was also tested regarding its long term behaviour and bandwidth. The RF signal was recorded over several hours, Fig. \[fig:3\](a), without (blue) and with (red) the stabilizing loop. The measurements are accomplished by tuning the X$^{0}$ of the quantum dot via the Stark effect into resonance with the excitation laser ($\delta$ = 0 $\mu$eV) and then recording the arrival time of each single photon detected by the APD over the duration of the entire experiment $T$. Post-experiment, the data are analyzed by setting a binning time, $t_{\rm bin}=100$ ms in this case. For a fixed $V_{g}$, the RF counts show large fluctuations up to a factor of 2 (blue curve). The origin are slow electrical fluctuations in the sample which cause the transition to drift out of resonance with the laser. With the feedback on, these fluctuations disappear and the RF remains at a constant level (red curve). The fluctuations in the red curve arise almost entirely from shot noise in the detector. The average RF signal falls slightly with feedback because the applied modulation broadens slightly the resonance. Insight into the bandwidth of the stabilization mechanism is revealed by a fast Fourier transform (FFT) of the time trace. Although the shot noise dominates, the shot noise can be independently measured with a small amount of reflected laser light as a source, allowing us to determine the noise coming solely from the quantum dot. The FFT of the normalized RF signal $S(t)/\langle S(t)\rangle$ provides a noise spectrum[@andi]: $$N_{\rm RF}(f) = \vert {\rm FFT}[S(t)/\langle S(t)\rangle]\vert^2 (t_{\rm bin})^2/T.$$ For $N_{\rm RF}(f)$, $t_{\rm bin}=1$ $\mu$s and $T = 2$ hours. The noise spectrum of the quantum dot $N_{\rm QD}(f)$ is obtained by correcting the RF noise by the noise of the experiment $N_{\rm exp}(f)$ \[$N_{\rm QD}(f) = N_{\rm RF}(f) - N_{\rm exp}(f)$\]. $N_{\rm QD}(f)$ corresponding to the time traces of Fig. \[fig:3\](a) are shown in Fig. \[fig:3\](b). Without feedback, $N_{\rm QD}(f)$ has a $1/f$-like dependence on $f$ as a consequence of charge noise in the device. With feedback, $N_{\rm QD}(f)$ is reduced by up to a factor of 20 at the lowest frequencies, and is constant: the $1/f$-like noise is eliminated. The two curves meet at $f \simeq 130$ Hz once the bandwidth of the PID circuit has been exceeded. At higher frequency the noise spectrum is dominated by spin noise[@andi]. The quantum dot noise $N_{\rm QD}(f)$ under feedback can be linked to the jitter in the energy detuning, $\sigma_{E}$. The energy jitter is much less than the linewidth such that the change in the RF signal ($\Delta {\rm RF}$) is related quadratically to the detuning for fluctuations around $\delta=0$. The variance of the RF noise, $\sigma_{\rm RF}^{2}$, is related to an integral of the noise curve, $\sigma_{\rm RF}^{2}=\int N_{\rm QD}(f)df$[@Kogan]. Integrating up to frequency $\Delta f$ in the regime where $N_{\rm QD}(f)$ is approximately constant, $$\sigma_{E}^{\rm ON}=\frac{\Gamma}{2} \left(\frac{N_{\rm QD}(0) \Delta f}{3}\right)^{\frac{1}{4}}.$$ With $\Delta f=3.1$ Hz, $N_{\rm QD}(0)=1.0 \times 10^{-5}$, $\Gamma=2.58$ $\mu$eV this predicts $\sigma_{E}^{\rm ON}=0.073$ $\mu$eV, in excellent agreement with the measurement from the stroboscopic experiment (0.089 $\mu$eV). ![(Color) Second-order correlation $g^{2}(t)$ for the stabilized RF from the X$^{0}$ (black points). The red curve shows a convolution of the two-level atom result with a Gaussian distribution which describes the timing jitter of the detectors. The blue curve shows the two-level atom response alone. []{data-label="fig:4"}](figure4.pdf){width="0.5\linewidth"} An intensity correlation measurement $g^{(2)}(t)$ was performed with a Hanbury Brown-Twiss interferometer. Low noise $g^{(2)}(t)$ can only be determined at these count rates (50 kHz per APD) by integrating over several hours and the feedback is therefore important to ensure that the detuning of the quantum dot with respect to the laser remains constant. $g^{(2)}(t)$ is shown in Fig. \[fig:4\] from X$^{0}$ of the same quantum dot with zero detuning. $g^{(2)}(t)$ falls to 10% at $t=0$. This does not reflect $g^{(2)}(0)$ of the quantum dot but rather the timing jitter of the detectors which is comparable to the radiative lifetime. We attempt to describe $g^{(2)}(t)$ with a convolution of $g^{(2)}(t)$ for an ideal two-level atom, $g^{(2)}_{\rm atom}(t)$, and the response of the detectors $G(t)$: $$g^{(2)}(t) = g^{(2)}_{\rm atom}(t) \otimes G(t).$$ The detector response is a Gaussian function, $$G(t)=\frac{1}{\sqrt{2 \pi}\sigma_{D}} \exp\left(-\frac{t^2}{2\sigma_{D}^{2}}\right).$$ $g^{(2)}_{\rm atom}(t)$ of a 2-level system with resonant excitation is, $$g^{(2)}_{\rm atom}(t)= 1-\left[\cos(\lambda t)+ \frac{3}{4\tau_r} \lambda \sin (\lambda t)\right] \exp\left(-\frac{3t}{4\tau_r}\right)$$ with $\lambda = (\Omega^2-(1/4\tau_r)^2)^{1/2}$. The temporal jitter of the detector $\tau_D=0.40$ ns is measured independently. $\Omega$ and $\tau_r$ are known from other experiments to within $10-20$% and are allowed to vary in these windows by a fit routine. The convolution provides an excellent description of the measured $g^{(2)}(t)$ with $\Omega= (0.99\pm 0.1)$ $\mu$eV and $\tau_r= (0.78\pm 0.05)$ ns. In particular, with low systematic error we can set an upper bound to the quantum dot $g^{(2)}(0)$ of 1-2%. \[sec:Conclusion\] In conclusion, we have developed a dynamic method of locking the optical resonance of a single quantum dot to a stabilized laser in order to produce a stream of frequency-stabilized single photons via resonance fluorescence. Generally speaking, the scheme represents a way to reduce the local charge noise in a semiconductor. Now that the basic principle is established, there are options for improving the feedback scheme. First, the remaining jitter in the quantum dot resonance position can be reduced by reducing the noise in the transmission detection. Presently, we are far from the limit defined by the shot noise in the detector current. With lower noise, the feedback bandwidth can also be increased. The tantalizing prospect is to create transform-limited linewidths routinely with high bandwidth feedback. A bandwidth of about 50 kHz is required [@andi]. Secondly, the modulation required here to generate the error signal could be eliminated in a number of ways. For instance, a dispersive lineshape can arise naturally in reflectivity via weak coupling to a cavity [@Alen2006]; or the Faraday effect in a small magnetic field [@Atatuere2007] could be used. We acknowledge support from the Swiss National Science Foundation (SNF) and NCCR QSIT. A.L., D.R. and A.D.W. acknowledge gratefully support from DFG SPP1285 and BMBF QuaHLRep 01BQ1035. [33]{}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 [**](\doibase 10.1088/0034-4885/68/5/R04), [****,  ()]{} [**](\doibase 10.1126/science.290.5500.2282), [****,  ()]{} [**](\doibase 10.1038/nature01086), [****,  ()]{} [**](\doibase 10.1038/nphoton.2007.46), [****, ()]{} [**](\doibase 10.1103/PhysRevLett.99.187402), [****,  ()]{} [**](\doibase 10.1103/PhysRevB.77.245311), [****,  ()]{} @noop [**]{}, [****,  ()]{} [**](\doibase 10.1038/nphoton.2009.287), [****,  ()]{} [**](\doibase 10.1103/PhysRevLett.93.217401), [****,  ()]{} [**](\doibase 10.1126/science.1126074), [****,  ()]{} [**](\doibase 10.1103/PhysRevLett.108.107401), [****,  ()]{} @noop [**]{}, [ ()]{} [**](\doibase http://dx.doi.org/10.1038/nmat3585), [****,  ()]{} [**](\doibase 10.1103/PhysRevLett.105.033601), [****,  ()]{} [**](\doibase 10.1038/nature11577), [****,  ()]{} [**](\doibase 10.1038/nature11573), [****,  ()]{} [**](\doibase 10.1103/PhysRevB.65.113303), [****,  ()]{} [**](\doibase 10.1038/NPHYS1780), [****,  ()]{} [**](\doibase 10.1063/1.2204843), [****,  ()]{} [**](\doibase 10.1103/PhysRevLett.107.217402), [****,  ()]{} [**](\doibase 10.1063/1.1609243), [****,  ()]{} [**](\doibase 10.1016/j.spmi.2004.02.007), [****,  ()]{} [**](\doibase 10.1103/PhysRevLett.108.206401), [****,  ()]{} @noop [**]{}, [ ()]{} [**](\doibase 10.1103/PhysRevLett.108.093602), [****,  ()]{} @noop [**]{}, [ ()]{} [**](\doibase 10.1038/35016030), [****,  ()]{} [**](\doibase 10.1038/NPHYS1363), [****,  ()]{} @noop [**]{} (, ) @noop [**]{} (, ) [**](\doibase 10.1063/1.2354431), [****,  ()]{} [**](\doibase 10.1038/nphys521), [****,  ()]{}
{ "pile_set_name": "ArXiv" }
--- abstract: 'Virus binding to a surface results at least locally, at the contact area, in stress and potential structural perturbation of the virus cage. Here we address the question of the role of substrate-induced deformation in the overall virus mechanical response to the adsorption event. This question may be especially important for the broad category of viruses that have their shells stabilized by weak, non-covalent interactions. We utilize atomic force microscopy to measure the height change distributions of the brome mosaic virus upon adsorption from liquid on atomically flat substrates and present a continuum model which captures well the behavior. Height data fitting according the model provides, without recourse to indentation, estimates of virus elastic properties and of the interfacial energy.' author: - Cheng Zeng - 'Mercedes Hernando-Pérez' - Xiang Ma - Paul van der Schoot - Roya Zandi - Bogdan Dragnea bibliography: - 'IndentationlessAFM.bib' title: Contact Mechanics of a Small Icosahedral Virus --- The problem of how adhesion of a deformable object to a surface is driven by interfacial energy and opposed by elasticity is at the center of modern contact mechanics[@Maugis1999]. Cell membranes are naturally impermeable to virus particles. For viruses to cross plasma, endosomal, or nuclear membranes, the virus-cell interface has to change drastically after virus adsorption. This is often done in a system-specific manner. Nevertheless, before specific transformations to take place, virus particles must stick at the apical cell surface via generic interactions, e.g., hydrophobic or electrostatic[@Mercer2010]. Could this initial, random binding event already perturb the mechanochemistry of the virus particle in a way that would prime it for the next sequence in the entry process? Gao et al. have suggested a model for the clathrin-independent endocytosis mechanism by which interactions between ligands fixed on the particle surface and free receptors on the plasma membrane would result in bringing more of the membrane into contact with the particle, which in turn would lead to the particle being eventually engulfed by the plasma membrane[@Gao2005]. This receptor-mediated wrapping mechanism model was revisited by Yi et al. who allowed particles to deform under the influence of adhesion to the flexible membrane surface and pointed out the possibility of a strong effect of the elastic deformation of particles on their cellular uptake[@Yi2011]. Furthermore, more recent experimental studies provided indication that, at least in certain cases, virus stiffness may regulate entry[@Pang2013]. In contact mechanics of small soft-material particles, solid surface tension dominates elasticity[@Style2013]. Yet, this aspect has not been considered so far in approaches to mechanical measurements of viruses by atomic force microscopy (AFM) nanoindentation. Here we report on a case study aiming to determine how virus mechanics responds to virus adsorption to a surface. We find that a small icosahedral plant virus, the brome mosaic virus (BMV), will bind to atomically flat surfaces predominantly in one orientation, and that in order to achieve this preferred orientation it will deform, mainly at the contact interface. Moreover, in indentation experiments, the spring constant of the virus was independent of substrate-induced deformation. In other words, local stresses due to surface binding and distortion do not seem to propagate to the top, where the measurement is done. Interestingly, this would also be expected within the framework of thin shell theory[@buenemann2007]. Furthermore, with the aid of an elastic model, we show how the distribution of particle heights on the substrate informs on the magnitudes of elastic moduli and of the contact surface energy, without recourse to indentation experiments. Viruses are obligated biological systems much smaller in size than cells, but still composed of hundreds to tens of thousands of molecules working together. A complete understanding of their dynamic behavior requires a unifying framework including contributions from scale-dependent and scale-independent phenomena[@Phillips06]. In recent years, studies of virus mechanics under the influence of an external perturbation have begun to shed light on how energy flows between the different degrees of freedom describing these complex molecular assemblies. Osmotic pressure assays have provided important clues on how chemical energy is transformed into mechanical energy for phage genome injection[@Gelbart:2008wt] and single molecule pulling experiments with optical tweezers have helped elucidating the mechanisms of phage genome packaging[@Smith:2001cz; @Purohit2003; @Smith2011]. In the category of *in singulo* methods based on mechanical force application, AFM indentation[@Michel2006] has allowed the measurement of virus and protein cage deformation under uniaxial load[@Roos2010a; @Carrasco2011], and of the relationship between virus mechanics and chemistry, which includes contributions from environmental factors such as pH or hydration[@Wilts2015], and from the nucleic acid cargo[@Michel2006; @Carrasco2006a; @Vaughan2014; @Ahadi2013]. For sufficient imaging resolution and to perform reproducible indentation experiments, particles have to be immobilized strongly enough to resist lateral forces exerted by the AFM probe[@Baclayon2010]. Compared with contact modes, non-contact modes[@Hansma1994; @Moreno-Herrero2003; @Martinez-Martin2012] are generally considered as being the least intrusive. Even then, while mean forces during imaging are usually below 0.1 nN, peak force estimates in “tapping" mode can exceed 0.1 nN (albeit for only $\sim$1 ms per pixel)[@Xu2008]. Such forces require either virus immobilization in a crystalline lattice[@Kuznetsov1997] or, when single virus measurements are sought, strong adhesion forces between virus and substrate. This is why substrates are usually prepared by coating with molecules imparting a hydrophobic or charged character to the surface [@Muller1997; @Roos2011a]. Thus, when the virus binds to the surface through, say, hydrophobic interactions, an equilibrium is established between virus-substrate adhesive interactions and the cohesive interactions of the virus. Adhesion-related deformation was observed before[@Knez2004], but very little is actually known about this equilibrium. How does the balance between adhesion and mechanical stresses affect particle shape? How large is the adhesion area at equilibrium? What is the magnitude of surface energy? Does surface adhesion result in local structural perturbations that propagate through the virus lattice up to the top, at the indentation area? This study takes on the task of addressing these questions on one of the most-studied virus systems adsorbed on chemically well-defined, atomically flat substrates. BMV was the first virus to be imaged by AFM at molecular resolution[@Kuznetsov2001]. It is an established model[@Kao2000] for small (+) single–stranded RNA icosahedral viruses, the most plentiful viruses on this planet[@Flint2009]. BMV has a non-enveloped capsid formed from 180 copies of the same coat protein (CP), organized in a T=3 lattice with an average outer diameter of 284 Å[@Wang2014]. The outer surface of the BMV capsid is studded with hydrophobic patches surrounded by polar residues (Fig. SI-1) and thus BMV readily adsorbs on both hydrophobic surfaces and polar surfaces. In this work, we study the distributions of maximum particle heights measured by AC-mode AFM on two substrate materials that readily yield atomically flat surfaces: highly-oriented pyrolytic graphite (HOPG), and mica. The idea is that adhesion forces will tend to maximize the contact area by locally flattening the virus at the contact point. Assuming that, for small perturbations, the virus particle behaves approximately as an elastic shell[@Roos2010a], an increase in contact area can be accomplished at the energetic cost of bending the shell and of forming a rim, defined as the locus where the fluid, the substrate, and the shell outer surface meet. As a result, the maximum height of the virus over the surface support would change upon adsorption. Since measuring height is done relative to the substrate, it is beneficial to chemically homogeneous, atomically flat substrates for this work, as opposed to functionalized etched glass substrates customarily used in indentation experiments, which have higher local roughness[@Roos2011a]. Height measurements can be affected not only by substrate roughness, but also by virus shell anisotropy. For BMV, the root-mean-square deviation (rmsd) from a spherical surface is $\sim$ 20Å[@Lucas2002]. Since the measurement is made top-down, it is important to record the orientation of the virus particle relative to the substrate. Imaging at the experimental conditions reported here (see Supporting Information) leads to sufficient lateral resolution (Fig. SI-2) to distinguish not only broad icosahedral symmetry features, but individual capsomers on the virus surface, Fig. 1. In these conditions, we find a clear orientational bias, on both substrates. The most frequent orientation is with a three-fold axis normal to the substrate ($\sim70\%$ from a total of 38 particles which had enough resolution to be unambiguously analyzed). Note that, if particles were adsorbed with random orientation, one would expect the three-fold axis orientation to be observed significantly less often. Early work on cowpea chlorotic mottle virus done on KOH etched glass and silanized glass found random capsid orientations, in contrast with our findings[@Michel2006]. The difference is likely coming from the fact that etched glass is rough and chemically heterogeneous. As a consequence, particles may bind upon landing with an enhanced initial contact area, and hence with strong initial adhesion and without subsequent reorientation. The situation is likely different on atomically flat, chemically homogenous surfaces, where an initial small contact would require reorientation to avoid desorption. Orientational selection could come from the most exposed areas on the virus surface having a pronounced hydrophobic character and affinity for nonpolar surfaces such as HOPG, Fig. SI-1. Moreover, anionic residue patches bordering these areas may bind to divalent cations (such as Mg(II) present in buffer solution) and adsorbed on the mica surface[@Hansma1996]. Histograms of BMV maximum heights on HOPG and mica are presented in Fig. 2 a). A small (5 Å) correction to the apparent height values was made to account for compression under the imaging force (Supporting Information)[@Roos2011a]. There are two notable observationss: (1) On both substrates, height histograms are peaked at values 1.5-2.0 nm below the nominal 28.4 nm for BMV. (2) Histograms are asymmetric, with the longer tail extending towards lower heights, while the greater heights wing ends abruptly in the vicinity of the nominal BMV diameter. Height distribution peak position and peak width depend on the substrate, suggesting a chemical effect. Lower heights than the nominal diameter suggest particle deformation upon adhesion, Fig. 2 b). BMV particles have elastic constants of $\sim$ 0.2 N/m. The compression force that would have to act on the virus to obtain a deformation associated with the observed drop in height of $\sim$ 2 nm, is $\sim$ 400 pN. This value gives an order of magnitude estimate of substrate-induced interactions at work. An analytical model was setup that captures in a formal, albeit heuristic way, the interplay between elastic properties, capsid deformation, and adhesion. The model is inspired by the Helfrich treatment of the elastic properties of lipid bilayers[@Helfrich1973], but with significant differences as a viral shell is a very different object than a lipid vesicle. A specific assumption is made that, upon landing on a surface, adhesion can increase by local deformation and formation of a flat contact area (base) with circular symmetry, Fig. SI-3. In other words, there is a sharp boundary or fracture between the flat surface base and the spherical cap, in solution. We opted for this geometry instead of the one assuming continuous deformation of membrane vesicles adsorbed on a surface[@Seifert1990] because due to the discrete nature of shell subunits, line fracture rather than continuous deformation is a reasonable assumption. Moreover, a continuously-deformed particle should become pre-stressed by adsorption and presumably, show changes in apparent stiffness, a situation which, as we will see later, we do not observe. Stretching would imply deformation of the proteins and/or increase in capsomeric surface-to-surface distances. Both processes are expensive, the latter on account of the short-ranged nature of the interactions[@buenemann2007]. Moreover, attempts to fit the data including stretching/compression of the surface area either failed or showed that the contribution of stretching energy is negligible. Area conservation upon deformation is thus assumed, which leads to a relationship between height and the cap radius: $$a = \sqrt{\frac{4r_0^2-h^2}{2}}$$ where $r_0$ is the initial particle radius, and $h$ is the height on the surface after binding and deformation (Fig. SI-3). The spherical cap radius then obeys (Fig. SI-3): $$r = \frac{4r_0^2+h^2}{4h}$$ The total energy is partitioned into contributions from the bending and Gauss energies, as well as a surface energy associated with the contact area, and a line or rim energy associated with the contact perimeter. The total energy is (see Supporting Information): $$F=\frac{1}{2}\kappa \left ( \frac{2}{r}-\frac{2}{r_0} \right )^2 2 \pi r h + 2 \pi \kappa_G \frac{h}{r} -\gamma \pi a^2 + \tau 2 \pi a \label{eq:F}$$ where: $\kappa$ is the bending modulus, $\kappa_G$ is the Gauss modulus, $\gamma$ is the surface energy, and $\tau$ is the rim energy. We emphasize that we attempted to fit the data without the Gaussian term and in the presence of stretching energy, see below and the SI. The expression given in Eq. \[eq:F\] is the simplest equation with which we are able to fit the experimental data Within the thin shell approximation, the Gauss and bending moduli are related via Poisson’s ratio[@Landau1984]: $\kappa_G = \kappa(\nu-1)$. For small icosahedral ssRNA viruses, $\nu \approx 0.3 - 0.4$[@Ahadi2013; @Michel2006; @Gibbons:2008kv]. Here we take the value $\nu = 0.3$. The free energy change upon adsorption can be then written as a sole function of the reduced height, $H = h/ 2 r_0$. Parameters $\kappa$, $\kappa_G$, $\gamma$, and $\tau$ can be then in principle found from fitting experimental data with a Boltzmann distribution derived from the free energy as a function of $H$ (eq. \[eq:F\]). Note that the Gauss term would have a vanishing contribution on a continuous surface topologically equivalent to a sphere[@Seifert1990]. In our case the surface is not differentiable everywhere, and hence the Gauss term does contribute. Interestingly, if we remove the Gauss term, we obtain a substrate-dependent bending modulus, which should have been a property determined by the nature of the virus rather than the underlying surface. Furthermore, unreasonably high values for the bending modulus were also observed (see Supporting Information). If we keep the Gauss term, then our data could be fitted using same values for the bending modulus on different substrates. The fitting results and parameters for these conditions are summarized in Fig. 3 and Table 1. As discussed in the following, parameter values agree well with those previously reported by other methods. Since the bending of a shell involves compression of the inner surface and extension of the outer surface, the bending modulus, $\kappa$, is related to the stretching modulus, $\kappa_s$ through: $$\kappa=\kappa_s\cdot\frac{w^2}{\alpha} \label{eq:ks}$$ where: $w$ is the shell thickness, and $\alpha = $ 12, 24, or 48 depending on the shell model (12 for a uniform plate[@Landau1984], 24 for a polymer brush[@Rawicz2000], 48 for a two-leaflet structure[@Boal2002]). For virus capsids, $\alpha =$ 12 has been previously used[@gibbons2007nonlinear; @cuellar2010size; @roos2010b], which in our case leads to $\kappa_s\approx 43\ k_BT/nm^2$. In an examination of the low-frequency modes of a very similar virus to BMV, the chlorotic cowpea mosaic virus (CCMV), May *et al.* calculated, in the context of a spherical harmonic basis set, $\kappa_s$ values for the $l=0$ and $l=1$ modes at $81\ k_BT/nm^2$ and $60\ k_BT/nm^2$, respectively[@May:2011jx]. Note that while in AFM indentation experiments the $l=1$ is the dominating mode, both $l = 0$ and $l = 1$ modes are likely to be required in order to describe deformation in our case. Thus, estimates for the bending modulus from the particle height data lead to comparable values with those previously reported from similar systems. ------------- ---------- --------------- ------------- Parameters: $\kappa$ $\gamma$ $\tau$ $(k_BT)$ $(k_BT/nm^2)$ $(k_BT/nm)$ HOPG 32 0.16 0.09 Mica 32 0.10 -0.30 ------------- ---------- --------------- ------------- : \[Table 1\] Fit parameter values for data in Fig. 3 From the relation between the particle height, initial radius and radius of the flat part of the adsorbed virus, we can find the base area that corresponds to the most probable particle height. The base radius for HOPG is $\approx$ 9 nm and the corresponding base area is $\approx$ 250 nm$^2$. Creating the base lowers the particle energy by $\approx 40\ k_BT$ with a rim contribution of $\approx \ 5\ k_BT$. Note that different contributions dominate at different height ranges. For instance, the rim contribution dominates when the contact area is small. Populations at $ h \approx 2r_0$ are determined by the magnitude (and sign) of $\tau$. More specifically, a barrier to adsorption would occur if $\tau$ is positive (see Fig. SI-3). Such seems to be the case on HOPG, but not on mica (see Table 1). Because we would naively expect a positive line tension, we cannot explain this. As area conservation is assumed, it is not necessary to include a stretching term in our model. However, to verify how reliable this assumption is, we relaxed the constant area constraint (see Supporting Information for details). This necessitates the introduction of a stretching term in the free energy expression in order to account for the energy cost associated with any changes in the surface area. For simplicity, uniform stretching was assumed for the entire shell and the stretching modulus, $\kappa_s$, was related to bending modulus as we have seen above (eq. \[eq:ks\]). Fitting of the height histogram with the relaxed area constraint leads to a total surface area decrease for both HOPG and mica substrates. Still, the bending modulus showed negligible change. These results suggest that the contribution from stretching is minimal and that the assumption of constant surface area is valid. Together, our findings on the orientation bias and the estimates for the contact area suggest a possible mechanism for adsorption. We have seen that the most probable orientation on HOPG is with a 3-fold axis normal to the substrate. Considering the magnitude of the radius for the contact area, pentamers should be located on its circumference, i.e. touching the substrate. Keeping in mind that previous indentation experiments suggest compression to occur more readily along a 3-fold than along a 5-fold axis[@roos2010b] and that hexameric interfaces are thought to fail more readily than pentameric ones[@Zandi2005; @Krishnamani2016], we propose that the main displacement upon adsorption occurs along the three-fold axis, with the hexamer at the center radially shifting its position from the surface towards the particle center, and with the stiffer pentamers acting as a stabilizing tripod. As the interfacial area grows, a point is reached where the cost of continuing the flattening of the shell is greater than the energy drop due to adhesion, at which point the virus shell is stabilized. It is worth noting that, normal mode analysis of the mechanical properties of icosahedral virus capsids[@Tama2005] predicts pentamers to have greater propensity to move freely. However, continuum approaches based on elastic theory predict in certain cases the opposite, i.e., pentamers being stiffer than hexamers[@Buenemann2008]. The latter is valid for large ratios between elastic and bending energy contributions, for large viruses and when spontaneous curvature effects can be neglected[@Lidmar:2003gs]. It would be interesting to see how inclusion of substrate effects might affect these analyses. In any case, our experiments seem to support a scenario with stiffer pentamers, for a small virus. An issue of practical importance from a measurement perspective, is wether interactions at the substrate-virus interface affect readings of the virus stiffness in AFM indentation experiments. We have performed AFM indentation on BMV adsorbed on HOPG in SAMA buffer and plotted the elastic constants as a function of particle height. Within the framework of the proposed model, the smaller the height, the larger the virus-substrate interaction. Do particle height and elastic constants correlate? As expected, the joint histogram presented in Fig. 4 suggests that, within the experimental uncertainty, this is not the case. The particle height varied independently of the measured elastic constant $K_v$, which remained constant at $0.20\pm 0.06$ N/m. Note that, for the simple thin shell model, the elastic constant $K_v$ is proportional to the Young’s modulus, which in turn is directly proportional to the bending modulus $\kappa$. To avoid inhomogeneous broadening of $K_v$ in this experiment, and keep experimental uncertainty low, we produced a nearly homogeneous BMV virion population containing mainly a subset of the viral genome ($\sim$90% of RNA3/4) via an engineered *Agrobacterium* expression system[@Ni2014]. Moreover, natural variation in the average radius of the virus particle (from cryo-electron microscopy measurements) is $\sim$1 nm, much smaller than the deviations measured here. In conclusion, we have utilized AFM imaging in liquid on flat, chemically homogeneous substrates to show that orientation and height of viruses adsorbed on a substrate depend on the virus-substrate interaction. BMV was found to adsorb preferentially with a three-fold axis parallel to the surface normal. Local deformation, measurable as a change in virus height ensues as elastic and adhesive forces equilibrate. A simple model fitting experimental data suggests that interfacial energies of tens of $k_BT$ accompany the encounter of BMV with both charged and nonpolar model substrates. As we used the simplest possible free energy to obtain insights into the contribution of different elastic energies, our model is highly approximate, and it cannot reproduce the long tail in the distribution. Further investigations are required. However, local deformation at the contact area does not change the apparent elastic constant as measured by AFM indentation, which suggests that curvature elastic stress does not change upon adsorption. Since it appears that virus orientation and deformation at the surface stabilize interfacial interactions, an interesting question that might be raised is that of anisotropic deformability as yet another biologically beneficial facet of icosahedral symmetry in viruses. This work has been supported by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, under award DE-SC0010507 (to B.D., for work on atomic force microscopy) and by the Human Frontier Science Program, under award RGP0017/2012, and the National Science Foundation (DMR-1310687 to R.Z.), for modeling. The authors thank Dr. Irina Tsvetkova for her critical reading of the manuscript. Figure Captions =============== Figure 1. a) AFM images of BMV particles adsorbed on HOPG and mica. Several capsomers per capsid are observable at sufficient resolution to see broad morphological features of a single capsomer $\sim\ $6 nm in size. Scale bars: 10 nm. b) Orientations of the model icosahedron that correspond to virus orientations in a), respectively. c) Corresponding orientations of the BMV crystallographic model. d) and e) Distribution of surface normal directions, represented as colored lines from single measurements, with respect to an icosahedron attached to each particle and to the molecular model, on HOPG (N = 17 particles, green) and mica (N = 21 particles, blue). Figure 2. Particle height distributions for adsorbed BMV at equilibrium in SAMA buffer, at pH 4.5 (a), and cartoon representation of the mechanism by which particle heights may acquire a distribution when adsorbed on a surface (b). Figure 3. Model fit of the particle height distribution on HOPG (a) and mica (b). At close to nominal heights ($H=1$) contact area is minimal and likelihood of desorption increased, thus populations are low. At smaller heights, adhesion comes at the cost of structural perturbation, modeled here as elastic. Figure 4. Joint probability density histogram of heights and elastic constants for a homogeneous sample of BMV (N = 138) containing a genomic subset. The vertical streak aspect suggests negligible correlation between height variations due to adsorption and the magnitude of the elastic constant.
{ "pile_set_name": "ArXiv" }
--- author: - | M. J. White and F. Feroz\ Cavendish Laboratory. Madingley Road. Cambridge CB3 0HE, UK bibliography: - 'DM\_jhep.bib' title: MSSM dark matter measurements at the LHC without squarks and sleptons --- Introduction ============ The existence of a large amount of non-baryonic ‘dark’ matter (DM) in the universe is now relatively uncontroversial, and the precise nature of the matter is currently the source of much speculation. The WMAP observations of the cosmic microwave background have provided a precise measurement of the dark matter relic density [@Komatsu:2008hk], and also strongly support cold dark matter in the form of weakly interacting massive particles (WIMPs). Since the Standard Model (SM) of particle physics does not contain any particles that fit this description, searches for dark matter are intimately connected with searches for new physics. The Large Hadron Collider (LHC) at CERN, Geneva, is designed to probe the TeV scale for evidence of new fundamental theories, and can be expected to produce dark matter particles if a WIMP exists with an accessible mass. At the same time, direct search experiments continue to look for interactions between Earth-based targets and passing WIMPs, indirect search experiments look for evidence of WIMP annihilation in space, and the cosmic microwave background is being mapped with ever greater precision. If it is genuinely true that a single WIMP candidate explains dark matter, one should obtain a consistent set of observations from these experiments, but checking this consistency is hampered by large uncertainties in the relevant astrophysics. For example, if no annihilation signal is observed, how do we know if this is because the annihilation cross-section lies below the reach of our current apparatus or whether we simply looked in a region where there was too little dark matter to give a strong flux? If a direct search experiment fails to see a signal, how do we untangle the interaction cross-section from the uncertainties in the local density and velocity of the dark matter that affect the number of observed events? From a particle physics perspective, the only answer is to try and measure the microscopic properties of the WIMP as precisely as possible in accelerator experiments in order to make accurate predictions of the astrophysical data. These measurements can then either be used with halo models to predict annihilation fluxes and direct search cross-sections, or, in the exciting case of a *positive* result in one or more astrophysical experiments, to directly constrain the halo distributions themselves. Comparisons with the relic density could give important clues on the proportion of dark matter composed of the WIMP observed at the LHC or could help validate the standard picture of Big Bang cosmology. This paper examines dark matter arising from the Minimal Supersymmetric Standard Model (MSSM) in the popular case that the WIMP is the lightest neutralino of the model. Previous studies [@baltz-2006-74; @nojiri-2006-0603] have indicated that, under favourable conditions, one can measure the WIMP mass and couplings rather well at the LHC but, in the case that the scalars are too heavy to be produced, one might get the mass but certainly not the couplings. A classic example is the ‘focus point’ region of the mSUGRA parameter space [@Feng:2000gh], in which the scalar masses may be high while the model still obeys all experimental constraints, including the WMAP constraint on the relic density. A linear collider could resolve these issues but is still many years away and, while it would be possible to use astrophysical data in combination with the LHC data, this relies on assuming the model which we are trying to test. Hence there is a clear need to improve DM predictions at the LHC if we are to make proper use of the rich astrophysical data sets expected over the next ten years. We revisit the focus point SUSY case and demonstrate that the LHC can do significantly better than previously advertised, by using the shape of the dilepton invariant mass spectrum to constrain neutralino mixing. We also find that the improvements are still obtained even if one assumes that the sparticles producing kinematic endpoints at the LHC have not been correctly identified. In analysing our benchmark point we quote a study from members of the ATLAS collaboration [@:2008zzm], though our results would apply equally well to CMS [@:2008zzk]. We then use the anticipated ATLAS data in a Bayesian analysis of the 24 parameter MSSM space, and hence although we use an mSUGRA benchmark point, our analysis is not restricted to the mSUGRA parameter set. Moreover, it should be possible to apply our approach to other SUSY models with heavy scalars. The paper is structured as follows. Section \[Background\] briefly reviews the relevant background, defines the benchmark point used in this study and introduces the [MultiNest]{} sampling algorithm [@2009MNRAS.398.1601F] used to obtain our results. In Section \[Results\] we describe how to use the shape information to improve the results, and present the posteriors obtained using the ‘conventional’ and ‘new’ approaches. We also discuss potential pitfalls. We present conclusions in section \[Conclusions\]. Background {#Background} ========== Neutralino dark matter in the focus point region ------------------------------------------------ The Minimal Supersymmetric Standard Model (MSSM) describing broken SUSY permits 108 parameters and, although a large number of these must be small to suppress flavour changing processes, at least 24 parameters are normally considered necessary to investigate LHC phenomenology. Although SUSY provides several WIMP candidates (including the sneutrino and gravitino), the most popular is the lightest neutralino, which is an admixture of the superpartners of the standard model gauge bosons, with components set by the following mixing matrix: $$\begin{aligned} {\cal M}=\left(\begin{array}{cccc} M_1 & 0 & -m_Z \cos\beta s_W & m_Z \sin\beta s_W \\[2mm] 0 & M_2 & m_Z \cos\beta c_W & -m_Z \sin\beta c_W\\[2mm] -m_Z \cos\beta s_W & m_Z \cos\beta c_W & 0 & -\mu \\[2mm] m_Z \sin\beta s_W &-m_Z \sin\beta c_W & -\mu & 0 \end{array}\right)\ \label{eq:massmatrix}\end{aligned}$$ where $M_1$ and $M_2$ are the U(1) and SU(2) gaugino masses, $\mu$ is the Higgsino mass parameter, tan$\beta$ is the ratio of the vacuum expectation values of the two Higgs doublets and the other parameters are all from the standard model. Measuring the couplings of a neutralino WIMP essentially comes down to trying to constrain the components of this matrix which is not generically possible at the LHC. Many phenomenological studies have been performed using simpler frameworks such as the minimal supergravity model (mSUGRA) [@mSUGRA] where SUSY breaking, which is assumed to occur in a hidden sector, is communicated to the observable sector via gravitational interactions. The mSUGRA model unifies various GUT scale parameters, obtaining the following parameter set: the scalar mass $m_0$, the gaugino mass $m_{1/2}$, the trilinear coupling $A_0$, the ratio of Higgs expectation values tan$\beta$, and the sign of the SUSY Higgs mass parameter $\mu$. Assuming the lightest neutralino is the correct WIMP candidate, one finds that, in most of the mSUGRA parameter space, too many neutralinos would survive after the Big Bang, and hence the only allowed regions must have some kind of annihilation mechanism to bring the density down to within the limits set by astrophysical observation. There are four main mechanisms, each of which dominates in a particular region of the parameter space [@baltz-2006-74]: 1. Slepton exchange. This is suppressed unless the slepton masses are lighter than approximately 200 GeV and gives rise to the ‘bulk’ region at low values of the mass parameters. 2. Co-annihilation with light sleptons. This occurs when there are suitable mass degeneracies in the sparticle spectrum, giving rise to a suitably named ‘co-annihilation region’. 3. Annihilation to third-generation fermions. This is enhanced when the heavy Higgs boson $A$ is almost twice as massive as the LSP, giving rise to a ‘funnel region’ for models with reasonably large tan$\beta$. 4. Annihilation to vector bosons. This can occur if the neutralino LSP acquires a significant wino or higgsino component, which happens in the so-called ‘focus point’ region at high $m_0$. One must remember that the tight level of constraint provided by the WMAP data is only a feature of models with a small number of parameters. A recent review of mSUGRA results that investigates the effect of adding new parameters can be found in reference [@Baer:2008ih], while the results of a 24 parameter Bayesian analysis incorporating all current data can be found in [@AbdusSalam:2009qd]. The high value of $m_0$ in the focus point region puts the squark and slepton masses in the multi-TeV range and thus out of direct reach of the LHC, and this forms the basis of why the LHC is expected to struggle to measure the WIMP couplings in this region. The main method of constraining the parameters of the mixing matrix is to look for cascade decay processes featuring squarks and sleptons (see figure \[fig:decay\]) and use kinematic endpoints in the invariant mass distributions of the SM decay products to measure mass differences between the sparticles. If a large amount of the sparticle mass spectrum can be measured (for example in the bulk or coannihilation regions), one can tell that the LSP is almost purely bino and reconstruct the relic density to a precision of around 10% [@nojiri-2006-0603], dependant on assumptions about the LHC reach in the Higgs sector. This remains true in a 24 parameter MSSM analysis, and thus is not limited to an unreasonably restrictive mSUGRA analysis. Mass differences can still be measured in the focus point region using cascade decays headed by gluinos, but the lack of squarks and sleptons will clearly reduce the number of decays that we can observe, and it was previously argued that this leaves the dominant component of the neutralino ambiguous when one attempts to fit the SUSY parameters [@baltz-2006-74], leading to substantial (i.e. orders of magnitude) uncertainties in the gamma ray flux, direct detection cross-section and relic density results from the LHC. Nevertheless, the focus point is attractive for a number of reasons. From a DM perspective, it is interesting to note that the higgsino component of the neutralino WIMP implies large annihilation via $W^+W^-$ and $ZZ$ bosons, and a large number of gamma rays in the final state, thus improving the prospects for finding dark matter using the forthcoming generation of gamma ray telescopes. Indeed, a previous study comparing the search reach of the LHC and H.E.S.S. II revealed a region of overlap in the focus point region that would give clear signals in both experiments [@moulin-2008-77]. If a focus point type scenario were chosen by Nature, we therefore have the potential to learn much about dark matter. Furthermore, the focus point region is actually favoured by constrained MSSM fits to particle physics and cosmological data if one assumes ‘natural priors’ [@Allanach:2008iq]. Benchmark point --------------- Our study uses a particular benchmark point that gives a typical focus point phenomenology, and a discussion of the generality of our results is deferred to section \[Discussion\]. One of the problems of performing studies of the focus point region is that the results of spectrum calculations are highly sensitive to the top Yukawa coupling, and therefore to the physical top mass. Given that one of the leading aims of this paper is to show improvement over previously published results, we have chosen to use the same spectrum generator as that used in [@baltz-2006-74] (ISAJET 7.69 [@baer-2003]), with the same focus point region benchmark point (their ‘LCC2’) and the same value of the top mass. While it is true that a more recent and improved version of ISAJET is available, along with a revised measurement of the top mass, the results of this paper should be viewed as comparitive rather than definitive. We thus use the following values of the mSUGRA parameters: $$m_0=3280 GeV, m_{1/2}=300 GeV, \mbox{tan}\beta = 10, A_0=0, \mu > 0$$ with a top mass of 175 GeV. We ran ISASUGRA with these parameters, generating the equivalent 24 parameter MSSM input for ISASUSY. We then used this input to generate the expected values for the observables used in our analysis [^1]. The ISASUSY parameter set gives a slightly different mass spectrum and a relic density $\Omega_{\chi}h^2 = 0.089$ which is a little lower than the WMAP limit but this does not affect the validity of our conclusions. Important masses are given in Table \[tablemasses\], as generated from ISASUSY. Constraining the MSSM parameter space ------------------------------------- Given a set of LHC observables $\mathbf{O}$ and our set of hypothesised MSSM parameters $\mathbf{P_{MSSM}}$, we will shortly wish to evaluate $p(\mathbf{P_{MSSM}}|\textbf{O})$ and thus to find the region of parameter space consistent with the data. This posterior probability is related to the likelihood $\mathcal{L}(\mathbf{P_{MSSM}})=p(\mathbf{O}|\mathbf{P_{MSSM}})$ by Bayes theorem: $$p(\mathbf{P_{MSSM}}|\mathbf{O}) = \frac{p(\mathbf{O}|\mathbf{P_{MSSM}})p(\mathbf{P_{MSSM}})}{Z}$$ where $p(\mathbf{P_{MSSM}})$ represents our prior knowledge on the distribution of parameters, and the normalization constant $Z$ is the ‘Bayesian evidence’, given by the average of the likelihood over the prior: $$Z=\int \mathcal{L}(\mathbf{P_{MSSM}}) p(\mathbf{P_{MSSM}})d^{N_{par}}\mathbf{P_{MSSM}} \label{evidence}$$ In our study, we assign points the following likelihood: $$p(\mathbf{O}|\mathbf{P_{MSSM}})=\prod_i \mbox{exp} \left ( -\frac{(O_i(\mathbf{P_{MSSM}})-o_i)^2}{2\sigma_i^2}\right ) \label{likelihood}$$ where $O_i(\mathbf{P_{MSSM}})$ is the predicted value of the $i$th observable $o_i$, $\sigma_i$ is the error in $O_i$ and the product runs over all observables. In typical SUSY phenomenology problems such as that presented here, the posterior cannot be evaluated analytically and one can instead use a sampling approach. Most popular are Markov Chain Monte Carlo algorithms that draw samples in such a way that the distribution of samples tends towards the unnormalised posterior distribution. This allows one to make inferences on parameter constraints while bypassing the evaluation of the Bayesian evidence. Although this has proven successful in the past in MSSM case studies, the methods (such as the Metropolis algorithm) become far less efficient as the dimension of the parameter space increases, and additionally they struggle in cases with a multimodal posterior where the most likely islands of interest occupy a vanishingly small volume of the parameter space. Sadly, this is the quintessential MSSM posterior, and previous examples have either used many CPU years or have varied the step size as the sampling progresses from a large initial value tuned to find disparate islands to a small value designed to explore the region around each mode. One still has to choose sensible initial conditions for these algorithms, as well as finding sensible step sizes in each parameter if one chooses not to vary them. A more recent approach is that of ‘nested sampling’ which targets the Bayesian evidence first, subsequently obtaining samples from the posterior as a by product. We here give a brief description and refer the reader to [@2008MNRAS.384..449F; @2009MNRAS.398.1601F] for more details. Nested sampling [@2004AIPC..735..395S] is a Monte Carlo method that calculates the evidence by transforming the multi–dimensional evidence integral into a one–dimensional integral that is easy to evaluate numerically. This is accomplished by defining the prior volume $X$ as $dX = p(\mathbf{P_{MSSM}})d^{N_{par}}\mathbf{P_{MSSM}}$, so that $$X(\lambda) = \int_{\mathcal{L}\left( \mathbf{P_{MSSM}}\right) > \lambda} p(\mathbf{P_{MSSM}}) d^{N_{par}}\mathbf{P_{MSSM}}, \label{eq:Xdef}$$ where the integral extends over the region(s) of parameter space contained within the iso-likelihood contour $\mathcal{L}(\mathbf{P_{MSSM}}) = \lambda$. The evidence integral, Eq. (\[evidence\]), can then be written as $$\mathcal{Z}=\int_{0}^{1}{\mathcal{L}(X)}dX, \label{eq:nested}$$ where $\mathcal{L}(X)$, the inverse of Eq. (\[eq:Xdef\]), is a monotonically decreasing function of $X$. Thus, if one can evaluate the likelihoods $\mathcal{L}_{i}=\mathcal{L}(X_{i})$, where $X_{i}$ is a sequence of decreasing values, $$0<X_{M}<\cdots <X_{2}<X_{1}< X_{0}=1,$$ as shown schematically in Fig. \[fig:NS\], the evidence can be approximated numerically using standard quadrature methods as a weighted sum $$\mathcal{Z}={\textstyle {\displaystyle \sum_{i=1}^{M}}\mathcal{L}_{i}w_{i}}, \label{eq:NS_sum}$$ where the weights $w_{i}$ for the simple trapezium rule are given by $w_{i}=\frac{1}{2}(X_{i-1}-X_{i+1})$. An example of a posterior in two dimensions and its associated function $\mathcal{L}(X)$ is shown in Fig. \[fig:NS\]. The summation in Eq. (\[eq:NS\_sum\]) is performed as follows. The iteration counter is first set to $i=0$ and $N$ ‘active’ (or ‘live’) samples are drawn from the full prior $p(\mathbf{P_{MSSM}})$, so the initial prior volume is $X_{0} = 1$. The samples are then sorted in order of their likelihood and the smallest (with likelihood $\mathcal{L}_{0}$) is removed from the active set (hence becoming ‘inactive’) and replaced by a point drawn from the prior subject to the constraint that the point has a likelihood $\mathcal{L}>\mathcal{L}_{0}$. The corresponding prior volume contained within the iso-likelihood contour associated with the new live point will be a random variable given by $X_{1} = t_{1} X_{0}$, where $t_{1}$ follows the distribution $\Pr(t) = Nt^{N-1}$ (i.e., the probability distribution for the largest of $N$ samples drawn uniformly from the interval $[0,1]$). At each subsequent iteration $i$, the removal of the lowest likelihood point $\mathcal{L}_{i}$ in the active set, the drawing of a replacement with $\mathcal{L} > \mathcal{L}_{i}$ and the reduction of the corresponding prior volume $X_{i}=t_{i} X_{i-1}$ are repeated, until the entire prior volume has been traversed. The algorithm thus travels through nested shells of likelihood as the prior volume is reduced. The mean and standard deviation of $\log t$, which dominates the geometrical exploration, are: $$E[\log t] = -1/N, \quad \sigma[\log t] = 1/N.$$ Since each value of $\log t$ is independent, after $i$ iterations the prior volume will shrink down such that $\log X_{i} \approx -(i\pm\sqrt{i})/N$. Thus, one takes $X_{i} = \exp(-i/N)$. Once the evidence $\mathcal{Z}$ is found, posterior inferences can be easily generated using the final live points and the full sequence of discarded points from the nested sampling process, i.e., the points with the lowest likelihood value at each iteration $i$ of the algorithm. Each such point is simply assigned the probability weight $$p_{i}=\frac{\mathcal{L}_{i}w_{i}}{\mathcal{Z}}.\label{eq:12}$$ These samples can then be used to calculate inferences of posterior parameters such as means, standard deviations, covariances and so on, or to construct marginalised posterior distributions. The most challenging task in implementing nested sampling is to draw samples from the prior within the hard constraint $\mathcal{L}> \mathcal{L}_i$ at each iteration $i$. The [MultiNest]{} algorithm [@2008MNRAS.384..449F; @2009MNRAS.398.1601F] tackles this problem through an ellipsoidal rejection sampling scheme. The live point set is enclosed within a set of (possibly overlapping) ellipsoids and a new point is then drawn uniformly from the region enclosed by these ellipsoids. The ellipsoidal decomposition of the live point set is chosen to minimize the sum of volumes of the ellipsoids. The ellipsoidal decomposition is well suited to dealing with posteriors that have curving degeneracies, and allows mode identification in multi-modal posteriors. If there are subsets of the ellipsoid set that do not overlap with the remaining ellipsoids, these are identified as a distinct mode and subsequently evolved independently. The [MultiNest]{} algorithm has proven very useful for tackling inference problems in cosmology and particle physics(see e.g. [@2009MNRAS.400.1075B; @AbdusSalam:2009qd; @2009PhRvD..80c5017A; @2008JHEP...10..064F; @2008JHEP...12..024T; @2009arXiv0911.1986R; @2009CQGra..26u5003F]) typically showing two orders of magnitude improvement in efficiency over conventional techniques. We use this algorithm for exploring the posterior distributions in this paper. Focus point SUSY at the LHC {#Results} =========================== Conventional input measurements ------------------------------- Before explaining how to improve the LHC measurements in the focus point region, we will reproduce the existing approach to provide a comparison for our later results. Assuming that the LHC will not produce squarks or sleptons, mass measurements are restricted to cascade decays of the form: $$\tilde{g}\rightarrow \tilde{\chi}^0_{j} q \bar{q} \rightarrow \tilde{\chi}^0_i q \bar{q} l^+ l^-$$ where $j > i$. The neutralino decays give rise to kinematic endpoints in the dilepton invariant mass distribution that will be recognisable as arising from three body decays (by virtue of shape or possibly by using a wedgebox technique as in [@Kang:2009sk]). In fact, two endpoints are visible for the focus point model considered here, arising from decays of the $\tilde{\chi}^0_{2}$ and $\tilde{\chi}^0_{3}$, and their positions are simply given by the difference $m_{\tilde{\chi}^0_{i}}-m_{\tilde{\chi}^0_{1}}$. The most recent relevant study performed by members of the ATLAS collaboration [@desanctis-2007] gives an error of 0.5 GeV on $m_{\tilde{\chi}^0_{2}}-m_{\tilde{\chi}^0_{1}}$, and 1.5 GeV on $m_{\tilde{\chi}^0_{3}}-m_{\tilde{\chi}^0_{1}}$ if we assume, as these studies have, that we have correctly identified the neutralinos producing the bumps. The resulting distribution is shown in figure \[fig:dilep\] and would be very similar for the benchmark point used in this study. We therefore assume for our study that the neutralino mass differences have been determined to be the nominal values obtained from the ISASUSY mass spectrum for our MSSM benchmark point, with the same errors as those quoted in the ATLAS study. Reference [@baltz-2006-74] further assumes that the $\tilde{\chi}^0_{1}$ and $\tilde{g}$ masses can be fixed to within  10% by kinematic fits, and the lightest Higgs mass could be obtained with an accuracy of 0.25 GeV. Again, we assume these same errors, and use the nominal mass values from our benchmark point spectrum for the reconstructed masses. Finally, one can set limits on the squark and slepton masses based on their non-observation, which we impose approximately by restricting the prior range for the squark and slepton mass parameters. A summary of the observables used in our analysis is given in Table \[observables\], while the prior range used in our exploration of the MSSM parameter space is given in Table \[prior-range\]. We have performed a nested sampling of the MSSM parameter space, using a log prior on all parameters except the $A$ parameters for which we follow [@baltz-2006-74] in using the formula $A = M$sinh$(x)$, with $M=50$ GeV. A log prior is generally useful for parameters with a scale uncertainty, while $A = M$sinh$(x)$ is chosen since the $A$ parameters can take either sign. For simplicity, we have further restricted $M_1$ and $\mu$ to be positive as this will not affect the validity of our conclusions. For each point selected by [MultiNest]{} we run ISASUSY on the 24 MSSM parameters to generate the mass and decay spectrum. We then use the masses and neutralino mass differences to calculate the predicted values of the observables given in Table \[observables\] and evaluate the likelihood using equation \[likelihood\]. Figure \[m1mu-old\] shows the marginalised posterior in the ($M_1$,$\mu$) plane as evaluated by [MultiNest]{} for this set of input measurements, and one can clearly see three solutions. Each of these has a lightest neutralino of different dominant character, with the left-most island corresponding to the benchmark model[^2]. One can use the points in this posterior to calculate astrophysical observables, and we used DarkSUSY 5.05 [@gondolo-2004-0407; @frank-2007-047; @degrassi-2003-28; @heinemeyer-1998] with a custom interface to ISASUSY based on the Les Houches format [@skands-2004-0407]. The WIMP relic density and neutralino pair annihilation cross-section at threshold (used to calculate the flux expected in indirect search experiments) are shown in Figures \[relic-old\] and \[annihilation-old\], and we see that the effect of the extra islands is to give us false regions of high likelihood orders of magnitude away from the correct value. These would clearly destroy any reasonable prospect of accurately predicting the relic density or gamma ray flux based on LHC measurements. Direct search cross-sections divide into spin-dependent and spin-independent contributions, with the spin-independent term enhanced by a factor of the square of the target nucleus mass. The effect of false posterior regions at the LHC is to contribute an extra peak to each of these terms, as shown in Figure \[protonSI-old\] for the spin-independent neutralino-proton direct detection cross-section and in Figure \[neutronSD-old\] for the spin-dependent neutralino-neutron direct detection cross-section (the spin-dependent proton-neutralino and spin-independent neutralino-neutron cross-sections are very similar). We see a longer tail on the spin dependent distribution than in [@baltz-2006-74], but this is almost certainly due to [MultiNest]{} having explored the bino mode more thoroughly than their sampling method. New input measurements ---------------------- To remove the false regions of the posterior in Figure \[m1mu-old\], we need to find observables at the LHC that are sensitive to neutralino mixing. Here we suggest one example in the form of the shape of the dilepton invariant mass distribution, since the number of events contributing to each bump in the distribution is sensitive to the couplings of the relevant neutralinos producing the bumps. Tuning the WIMP components has the effect of raising and lowering the two bumps, and thus changing the shape. In general, tuning the neutralino couplings could also add or remove endpoints, or change their interpretation (i.e. making an endpoint appear from a $\tilde{\chi}^0_{4} \rightarrow \tilde{\chi}^0_1l^+l^-$ decay rather than from a $\tilde{\chi}^0_{3} \rightarrow \tilde{\chi}^0_1l^+l^-$, etc). Although the dependence of the shape on the neutralino components is non-trivial, we are helped in the case of heavy scalars by the fact that we know from the LHC data that there are no on shell sleptons contributing to the endpoints, and thus the shape is even more sensitive to the neutralino couplings than in the generic case where one would complicate matters by introducing greater dependence on the slepton masses. In principle, one could model the shape of the dilepton invariant mass precisely at a given point in parameter space by generating a large sample of Monte Carlo events at that point, and passing them through a detector simulation. For a thorough sampling of the 24 parameter MSSM involving millions of explored points, this is clearly unfeasible, however, and we must use some shortcuts to encode the shape information in a manner that is easy to calculate. Under the assumption that the three body neutralino decays proceed dominantly through an off-shell $Z$ boson [^3], the shape of the dilepton invariant mass distribution is a known function of the neutralino masses, whose fit to the measured distribution is shown in Figure \[fig:dilep\]. The integral can be used to calculate the ratio of the number of events arising from each neutralino decay contributing to the dilepton invariant mass, which is related to the branching ratios of the sparticles by: $$\frac{BR(\tilde{g} \rightarrow ... \rightarrow \tilde{\chi}^0_{2} + X) \times BR(\tilde{\chi}^0_{2} \rightarrow \tilde{\chi}^0_1 l^+ l^-)}{BR(\tilde{g} \rightarrow ... \rightarrow \tilde{\chi}^0_{3} + X) \times BR(\tilde{\chi}^0_{3} \rightarrow \tilde{\chi}^0_1 l^+ l^-)} \label{ratioequation}$$ where the first decay in the numerator and denominator can proceed through multiples steps provided they do not produce leptons. The ATLAS study in reference [@desanctis-2007] quotes a measured value of $1.4 \pm 0.3$ for this ratio, comparing favourably to their theoretical value of 1.19, even in the presence of background and a detailed simulation of the ATLAS detector resolution. Our benchmark point has a slightly different theoretical value of 1.46. The study in [@desanctis-2007] further assumed that the $\tilde{\chi}^0_{3}$ and $\tilde{\chi}^0_{2}$ had been correctly identified as the cause of the endpoints in the dilepton invariant mass spectrum, an assumption that we do not wish to work with here since it presupposes knowledge of neutralino mixing which we are trying to determine in a general manner. There are other more trivial details of the shape of the dilepton invariant mass distribution. It is clear that any parameter point capable of explaining the distribution in figure \[fig:dilep\] must have a taller endpoint on the left, and a shorter endpoint on the right (this is of course the same as using ratio information). It is also clear that *only* two endpoints were observed, and hence any model that would have given more or less than two endpoints is also not capable of explaining the observed distribution. If we now assume we have not identified the neutralinos producing the endpoints, we must examine the general case where the number of $\tilde{\chi}^0_{j} \rightarrow \tilde{\chi}^0_i l^+ l^-$ decays contributing to the dilepton invariant mass distribution is given by: $$N \propto BR(\tilde{g} \rightarrow ... \rightarrow \tilde{\chi}^0_{j} + X) \times BR(\tilde{\chi}^0_{j} \rightarrow \tilde{\chi}^0_i l^+ l^-) \label{BRequation}$$ where the exact number of events depends on the cross-section for gluino production and the efficiency of the kinematic cuts. Again, the first branching ratio shown could in fact be the product of various decay processes that end with a $\tilde{\chi}^0_{j}$ and anything else (excluding leptons), and could consist of multiple steps. Since we know (from measurement) that all decay chains contributing to our dilepton plot started with gluinos, however, we can confidently state that we would have seen two endpoints if the two largest combined branching ratios of the form of equation \[BRequation\] are close in value while the third highest branching ratio is much lower. This encodes the shape of the distribution in a way which is easy to evaluate quickly at many points in parameter space, making it possible to scan or sample the parameter space to reject points on the basis of shape. We therefore reran the [MultiNest]{} sampling with the following additions. At each point in the parameter space, we used the ISASUSY decay spectrum to work out the combined branching ratios for all processes of the form given in equation \[BRequation\]. The two highest of these branching ratios are assumed to have caused the observed bumps in the dilepton invariant mass distribution, and the ratio of the number of events contributing to each endpoint is obtained using the following generalisation of equation \[ratioequation\]: $$\frac{BR(\tilde{g} \rightarrow ... \rightarrow \tilde{\chi}^0_{p} + X) \times BR(\tilde{\chi}^0_{p} \rightarrow \tilde{\chi}^0_q l^+ l^-)}{BR(\tilde{g} \rightarrow ... \rightarrow \tilde{\chi}^0_{r} + X) \times BR(\tilde{\chi}^0_{r} \rightarrow \tilde{\chi}^0_s l^+ l^-)}$$ The gluino cross-section and reconstruction efficiency cancel in this division and thus do not need to be explicitly evaluated- we are indeed fortunate that squarks are not contributing otherwise this procedure would not have worked. Note that our two highest branching ratios can in principle come from any neutralino decay process, so in principle we have endpoints arising from the decays $\tilde{\chi}^0_{p} \rightarrow \tilde{\chi}^0_q l^+ l^-$ and $\tilde{\chi}^0_{r} \rightarrow \tilde{\chi}^0_s l^+ l^-$, where $p$,$q$,$r$ and $s$ can take any value as long as $p>q$ and $r>s$ (in the previous section, we had $p=2$, $r=3$ and $q=s=1$). We now assign a likelihood based on the observables listed in Table \[newobservables\], which simply comprise of generalisations of those in the previous section plus the addition of the ratio measurement. Furthermore, we assign zero likelihood to points that could not have produced the observed dilepton distribution because they fell into the following categories: 1. Points in which the ‘wrong’ endpoint is highest, i.e. the kinematic endpoint with the largest $m_{ll}$ value is taller than the endpoint with the lowest $m_{ll}$ value\ 2. Points in which the third highest combined branching ratio of the form of equation \[BRequation\] is more than 10% of the second highest branching ratio (this is considered a sensible boundary of visibility).\ Results of a [MultiNest]{} sampling in the same parameter space as the previous sub-section are shown in the ($M_1$,$\mu$) plane in figure \[m1mu-new\], and the extra regions have been completely removed. Further investigation reveals that the vetos listed above are not the dominant contribution to this improvement, since both leave many points in both of the incorrect regions. Rather, it is the ratio measurement that provides enough discrimination between them to improve the results, leading to a substantial decrease in likelihood in the false regions. Since this is the ratio between the two highest endpoints produced by the gluino cascade processes, this is really telling us that the “two bump” feature of the dilepton invariant mass plot is a feature of the focus point model. The large spread in this ratio between the different regions also indicates that even a reasonably imprecise measurement of this quantity would be enough to favour the correct region. We also note that the selected region has endpoints arising from $\tilde{\chi}^0_{2} \rightarrow \tilde{\chi}^0_1 l^+ l^-$ and $\tilde{\chi}^0_{3} \rightarrow \tilde{\chi}^0_1 l^+ l^-$ decays and hence, although we had not assumed this advance, we have successfully inferred it from the data. The relic density and annihilation cross-section are shown in figures \[relic-new\] and \[annihilation-new\], where it can immediately be seen that the ‘fake’ solution peaks are removed from the annihilation cross-section prediction, thus allowing an accurate comparison of LHC data with the forthcoming round of gamma ray results. The relic density prediction is improved but still suffers from a lack of constraint at low values which is caused by the effect of the $A$ pole; resonant annihilation through the $A$ boson leads to a large reduction of the relic density for any model in which $m_A \approx 2m_{\tilde{\chi}_1^0}$. The direct search cross-sections, meanwhile, enjoy a similar improvement to the annihilation cross-section, as seen in Figures \[protonSI-new\] and \[neutronSD-new\], though the spin-dependent cross-section has a tail to large values in which it is possible to get a larger cross-section by increasing the amount of bino-Higgsino mixing or by increasing tan$\beta$ [^4]. This would seem to be irreducible at the LHC. Discussion {#Discussion} ---------- Having demonstrated an improvement in the LHC capability at one benchmark point, we now consider the generality of our result, along with any potential pitfalls. Firstly, it is clear that the improvement in our results comes from using shape information, and one must therefore carefully consider whether there are issues that might distort this shape. The dilepton invariant mass is a particularly clean signature, since practically all SUSY backgrounds can be removed via ‘flavour-subtraction’ (i.e. one plots the combination $e^{+}e^{-} + \mu^{+}\mu^{-} - e^{+}\mu^{-} + \mu^{+} e^{-}$). This explains the precision of the quoted ATLAS result which survived a rigorous detector simulation. It is more likely that the method described here will fail if Nature presents a focus point scenario with larger mass differences between the neutralinos such that the endpoints are either obscured by the Z peak or, in the case where the mass difference exceeds the Z mass, are not produced at all since the decay $\tilde{\chi}^0_{j}\rightarrow Z \tilde{\chi}^0_{1}$ is open. One could also push the gluino mass to large values and reduce the overall production cross-section at the LHC without violating any current constraints. In such a case, the LHC is trivially bound to fail. On the plus side, it is possible that the precision of the measurement could be improved further by using inclusive signatures in addition to the endpoint information, using a method similar to that in reference [@lester-2006-0601]. The focus point is conceptually simpler than most other regions of the parameter space, as the fact that the LHC would only be producing gauginos means that any inclusive signature must be telling us something about the gaugino sector. Thus, if dark matter is a gaugino, the focus point could be one of the more promising cases to handle at the LHC rather than one of the hardest, and this is a promising avenue for future work. Finally, we consider the relic density prediction, and speculate as to whether the low values arising from the $A$ pole can be removed by any further LHC data. Although $m_A$ is well in excess of $2m_{\tilde{\chi_1^0}}$ in the focus point region, there is no way *a priori* to rule out low heavy Higgs masses. If one could measure a lower bound on $m_A$, it is worth noting that the low values are substantially reduced for $m_A> 300$ GeV (Figure \[relic-300\]). Previous studies have placed bounds on the Higgs mass by either examining the Higgs search reach for the LHC in the $m_{A}-$tan$\beta$ plane and determining which region is compatible with non-observation or by using the measured mass difference between sparticles to place bounds on the Higgs mass based on its non-appearance in cascade decays [@nojiri-2006-0603]. Neither method is particularly useful here since a rather large region of the $m_{A}-$tan$\beta$ search plane is compatible with the observation of only $m_h$, while $(m_{\tilde{\chi_4^0}}-m_{\tilde{\chi_1^0}})<2m_{\tilde{\chi_1^0}}$ for our benchmark point, meaning that even if the branching ratios were favourable enough to produce heavy Higgs bosons in cascade decays, the only lower limit we could place on the $A$ mass would still be insufficient to rule out the low values. We therefore consider it unlikely that the LHC could set a lower limit sufficient to reduce this uncertainty on the relic density calculation, though it is not inconceivable that new work on Higgs decays to SUSY particles might extend the search reach for the $A$ boson sufficiently to improve general limits. If evidence suggests that a focus point scenario is realised in Nature, there is a clear incentive to pursue this line of enquiry. Conclusions {#Conclusions} =========== We have revisited the case of focus point neutralino dark matter at the LHC, for which it was previously assumed that it was not possible to make accurate predictions of astrophysical observables due to poorly constrained neutralino mixing parameters. Using similar Bayesian sampling techniques as the previous literature, we have demonstrated a method by which the shape of the dilepton invariant mass distribution can be used to remove false solutions and substantially improve predictions of the annihilation cross-section at threshold and the direct search cross-sections. The relic density prediction is also improved and could be significantly enhanced if one could set a lower bound on the $A$ mass. If Nature has chosen a focus point scenario, we will thus be able to test the compatibility of LHC WIMP measurements with astrophysical data without having to wait for a linear collider. This is particularly important given the rich astrophysical data sets expected over the next few years, and given the high likelihood of obtaining a strong indirect detection signal in focus point models. Although there are still specific reasons why we might be unlucky at the LHC (we are still dependent on favourable details of the mass spectrum), we remain optimistic that the LHC will be able to make stronger statements on neutralino mixing than has previously been assumed if Nature has chosen a model with heavy scalars. Applying our technique to other such scenarios would be an interesting line of future work. Acknowledgements ================ MJW gratefully acknowledges the financial support of the Science and Technology Facilities Council, and wishes to thank Howie Baer, Michael Peskin, Edward Baltz, Are Raklev, Giacomo Polesello, Csaba Balazs and the members of the Cambridge Supersymmetry Working group for helpful discussions. FF is supported by a research fellowship from Trinity Hall, Cambridge. [^1]: Note that we do not set the top, bottom and tau masses, Higgs vacuum expectation value and strong coupling constant to the benchmark value as in [@baltz-2006-74], nor do we scale the Yukawa couplings to more closely match mSUGRA. [^2]: The results agree closely but not exactly with those shown in [@baltz-2006-74], but the differences can be explained either by our slightly different procedure of handling the MSSM weak scale spectrum or by the fact that we do not impose a Tevatron-inspired limit on the lightest chargino mass. This latter cut restricted the phase space for the false solutions in [@baltz-2006-74] and, without it, we essentially get extra points in the ‘wrong’ islands which contribute to raising the ‘wrong peaks’ in figures \[relic-old\] and \[annihilation-old\]. [^3]: Our decay through $Z$ boson assumption is in fact true in the focus point region since the sleptons are very heavy, though *a priori* one can only set lower bounds on the slepton masses and hence might encounter a systematic uncertainty due to this asssumption. In practise, however, we shall find the ratio of the number of events contributing to each bump in the invariant mass distribution is $\approx$50 times higher in one of the false solutions, and $\approx$5 times higher in the other which probably exceeds any systematic correction to the ratio measurement. [^4]: We thank Michael Peskin for this suggestion.
{ "pile_set_name": "ArXiv" }
--- author: - 'Karttikeya Mangalam$^{1}$ and Prof. K S Venkatesh$^{2}$[^1][^2][^3]' title: '**Bitwise Operations of Cellular Automaton on Gray-scale Images** ' --- Cellular Automata, Regression weights, Digital Image Processing, Noise Reduction. INTRODUCTION ============ \ CA MODEL FOR DIGITAL IMAGES =========================== CONCLUSIONS =========== ACKNOWLEDGMENT ============== The author would like to express his gratitude towards SURGE (Student Undergraduate Research Grant for Excellence) programme offered by IIT Kanpur for the grant and the opportunity to pursue the project. It would not have been possible without the kind support and help of many individuals and the staff of the dept. The author would like to extend his sincere thanks to all of them. [99]{} J. von Neumann, Theory of Self-Reproducing Automata (edited and completed by Arthur Burks), University of Illinois Press, 1966 Wolfram, Stephen. “Computation theory of cellular automata.” Communications in mathematical physics 96.1 (1984): 15-57. D. Nayak, P. Patra and A. Mahapatra. “A Survey on Two Dimensional Cellular Automata and Its Application in Image Processing.” IJCA Proceedings on International Conference on Emergent Trends in Computing and Communication (ETCC-2014) ETCC(1):78-87, Sept. 2014. P. J. Selvapeter and Wim Hordijk, “Cellular automata for image noise filtering,” 2009 World Congress on Nature & Biologically Inspired Computing (NaBIC), Coimbatore, 2009, pp. 193-197. Huang, T., G. J. T. G. Y. Yang, and G. Tang. “A fast two-dimensional median filtering algorithm.” IEEE Transactions on Acoustics, Speech, and Signal Processing 27.1 (1979): 13-18. I. Pitas, A.Venetsanopou. Nonlinear Digital Filters: Principles and Application, Norwell, MA: Kluwer,1990. . Popovici and D. Popovici, “Cellular automata in image processing,” in Proceedings of the 15th International Symposium on the Mathematical Theory of Networks and Systems, D. S. Gilliam and J. Rosenthal, Eds., 2002, electronic proceedings. Songtao Liu, Hongguan Chan, Shaoging Yang, “An effective filtering algorithm for salt-peper noises based on cellular automata”, IEEE congress on image and signal processing,2008 Rosin, Paul L. “Training cellular automata for image processing.” IEEE transactions on image processing 15.7 (2006): 2076-2087. Gardner, Martin (October 1970). “Mathematical Games – The fantastic combinations of John Conway’s new solitaire game ”life“”. Scientific American. 223: 120–123. ISBN 0-89454-001-7. Gray, L. “A Mathematician Looks at Wolfram’s New Kind of Science.” Not. Amer. Math. Soc. 50, 200-211, 2003. M. Sipper, “The evolution of parallel cellular machines toward evolware,” BioSystems, vol. 42, pp. 29–43, 1997. Nobuyuki Otsu (1979). “A threshold selection method from gray-level histograms”. IEEE Trans. Sys., Man., Cyber. 9 (1): 62–66. G. Adorni, F. Bergenti, and S. Cagnoni, A Cellular-Programming Approach to Pattern Classification, 1998, vol. 1391, Lecture Notes in Computer Sciences, pp. 142–150. Paul L. Rosin. 2010. Image processing using 3-state cellular automata. Comput. Vis. Image Underst. 114, 7 (July 2010), 790-802. Khan, A. R. 1998. Replacement of some Graphics Routines with the help of 2D Cellular Automata Algorithms for Faster Graphics Operations, PhD thesis, University of Kashmir. Huynh-Thu, Q.; Ghanbari, M. (2008). “Scope of validity of PSNR in image/video quality assessment”. Electronics Letters. 44 (13): 800. doi:10.1049/el:20080522 [^1]: \*This work was not supported by any organization [^2]: $^{1}$Karttikeya Mangalam is pursuing B.Tech. at Department of Electrical Engineering, Indian Institute of Technology, Kanpur, India [mangalam at iitk.ac.in]{} [^3]: $^{2}$Prof. K S Venkatesh is with the Department of Electrical Engineering, Indian Institute of Technology, Kanpur, India. [venkats at iitk.ac.in]{}
{ "pile_set_name": "ArXiv" }
--- address: 'Université Côte d’Azur, Observatoire de la Côte d’Azur, CNRS,Laboratoire Lagrange, Bd de l’Observatoire, CS 34229, 06304 NiceCedex 4, France' author: - 'F. MIGNARD' bibliography: - 'FM\_DS\_bibfile.bib' - 'BiblioICRF.bib' title: 'The Gaia MISSION and SIGNIFICANCE [^1]' --- The Gaia mission ================ Context ------- A major change in the precise measuring of position and displacement of celestial objects took place with the access to space. It is not exaggerated to refer to this period, starting with the selection in 1980 by ESA of the first space astrometric mission Hipparcos, as a new golden age for this discipline which rests upon a long history of at more than two millennia, at least the Mediterranean world. Although Milky Way stars are the main celestial sources concerned by modern astrometry, Gaia has more diverse targets with the positions and motions of $100,000$s of solar system bodies, stars within nearby external galaxies and the most remote quasars at the outskirts of the Universe. The extended review of Hipparcos results by Perryman [@2012aaa..book.....P] amply demonstrates that the mission products influenced many areas of astronomy, in particular the structure and evolution of stars, the kinematics of stars and stellar groups. Even with its limited sample in number and variety of stars and observed volume, Hipparcos also made significant advances in our knowledge of the structure and dynamics of the Milky Way. Hipparcos was a resounding and acclaimed international success allowing the Europeans to quickly submit several more ambitious proposals for space astrometry, at the same time as others were also proposed to NASA or to the Japanese space agency . Only one of these proposals survived the various examinations by selection committees and Gaia was eventually selected as an ESA cornerstone mission in April 2000 for a launch around 2011. The basic observing concept is directly drawn from Hipparcos, but with a much larger telescope (actually two telescopes), a mosaic of 106 CCD detectors replacing the outdated photoelectric detector of Hipparcos. Two other instruments were added to carry out spectrophotometry and spectroscopic measurements, the latter to measure the velocity along the line of sight. While Hipparcos catalogue was limited to $100,000$ pre-defined stars brighter than 13.2 mag, Gaia was designed to realise a sensitivity-limited survey to $20$ mag. Hipparcos could only take a star at a time while Gaia is able to record simultaneously several $10,000$s images mapped on its focal plane. About $1.5$ billion stars, amounting to $\approx 1$ percent of the Milky Way stellar content, are repeatedly observed during the nominal 5-year mission, leading to a final astrometric accuracy of $25$ [[as]{}]{} at $G = 15$ mag. (1 [[as]{}]{} = 0.001 mas = $10^{-6}$ arcsec). Data product or source type Number of sources ----------------------------------------------------------------------- ------------------------ Total (excluding Solar system) [$1\,692\,919\,135$]{} \[2pt\] Five-parameter astrometry (position, parallax, proper motion) [$1\,331\,909\,727$]{} Two-parameter astrometry (position only) [$361\,009\,408$]{} ICRF3 prototype sources (link to radio reference frame) [$2\,820$]{} Gaia-CRF2 extra-galactic sources (optical reference frame) [$556\,869$]{} $G$-band (330–1050 nm) [$1\,692\,919\,135$]{} $G_\mathrm{BP}$-band (330–680 nm) [$1\,381\,964\,755$]{} $G_\mathrm{RP}$-band (630–1050 nm) [$1\,383\,551\,713$]{} \[2pt\] Median radial velocity over 22 months [$7\,224\,631$]{} Classified as variable [$550\,737$]{} Variable type estimated [$363\,969$]{} Detailed characterisation of light curve [$390\,529$]{} \[2pt\] Effective temperature [$T_\mathrm{eff}$]{} [$161\,497\,595$]{} Extinction [$A_G$]{} [$87\,733\,672$]{} Colour excess [$E(G_\mathrm{BP}-G_\mathrm{RP})$]{} [$87\,733\,672$]{} Radius [$76\,956\,778$]{} Luminosity [$76\,956\,778$]{} \[2pt\] Solar system object epoch astrometry and photonetry [$14\,099$]{} : Overall content of the Gaia DR2. \[table:summary\] The mission main features ------------------------- Gaia’s main scientific goal is to clarify the origin and history of our Galaxy, from a quantitative census of the stellar populations and extremely accurate astrometric measurements to derive proper motions and parallaxes. See for the proposal and for a presentation of the actual mission, the spacecraft, the operations and the data acquisition strategy. The principle of the scanning satellite relies on a slowly spinning spacecraft to measure the crossing times of stellar images transiting on the focal plane. As for Hipparcos, there are two fields of view combined onto a single focal plane where astrometric measurements are done. The time relates at once the one-dimensional star position to the instrumental axes. The relation to the celestial frame is obtained with the satellite attitude, which is solved simultaneously with the star positions in a global solution as described technically by Lindegren et al. . The mission and early results ----------------------------- The Gaia satellite was launched on 19 December 2013 and the science data collection started after the in-flight qualification on 25 July 2014. A first batch of results was released on 15 September 2016 with only 14 months of data processed. This release comprised primarily a position catalogue (only two position parameters per source) for 1.14 billion stars, the largest such collection ever. A smaller catalogue combining Gaia and Hipparcos included parallaxes and proper motions for $\approx 2,000,000 $ stars with a sub-mas accuracy . The release contained also variable stars and a set of $2,200$ quasars common to Gaia and the radio ICRF used to align the Gaia and the radio frame. Therefore the Gaia reference frame and ICRF are nominally identical. However this first release (Gaia DR1) based on just above one year of data was far from touching the core objective of the mission, since the number of parallaxes available was very small, even though it was already about 20 times larger than with Hipparcos. The second release (Gaia DR2) came out on 25 April 2018. Gaia DR2 data is based on data collected between 25 July 2014 and 23 May 2016, spanning a period of 22 months of data collection, enough to disentangle the parallactic motion from the proper motion for most of the stars. The reference epoch for Gaia DR2 is J2015.5 and positions and proper motions are referred to the ICRS, to which the optical reference frame defined by Gaia DR2 is aligned. The main numbers describing the content are given in \[table:summary\] and everyone would agree that they are impressive, even for those not well aware of how many of these stellar parameters were known before. Actually we will see in \[sect:significance\] that the gap with existing data is really wide. An overall description of the content of this survey is given in a general paper from the Gaia collaboration from which I took the two main tables \[table:summary\]-\[table:accuracy\]. [lr]{} Data product or source type & Typical uncertainty\ Five-parameter astrometry (position & parallax) & $0.02$–$0.04$ mas at $G<15$\ & $0.1$ mas at $G=17$\ & $2$ mas at $G=21$\ Five-parameter astrometry (proper motion) & $0.07$ mas yr$^{-1}$ at $G<15$\ & $0.2$  mas yr$^{-1}$ at $G=17$\ & $3$ mas yr$^{-1}$ at $G=21$\ Systematic astrometric errors (sky averaged) & $<0.1$ mas\ Gaia-CRF2 alignment with ICRF & $0.02$ mas at $G=19$\ Gaia-CRF2 rotation with respect to ICRF & $<0.02$ mas yr$^{-1}$ at $G=19$\ Mean $G$-band photometry & $0.3$ mmag at $G<13$\ & $2$ mmag at $G=17$\ & $10$ mmag at $G=20$\ Mean $G_\mathrm{BP}$- and $G_\mathrm{RP}$-band photometry & $2$ mmag at $G<13$\ & $10$ mmag at $G=17$\ & $200$ mmag at $G=20$\ Median radial velocity over 22 months & $0.3$ km s$^{-1}$ at $G_{\rm RVS} < 8$\ & $1.8$ km s$^{-1}$ at $G_{\rm RVS} = 11.75$\ Systematic radial velocity errors & $<0.1$ km s$^{-1}$ at $G_{\rm RVS} < 9$\ & $0.5$ km s$^{-1}$ at $G_{\rm RVS} = 11.75$\ Effective temperature [$T_\mathrm{eff}$]{} & 324 K\ Extinction [$A_G$]{} & 0.46 mag\ Radius & 10%\ Luminosity & 15%\ Solar system object epoch astrometry & 1 mas (in scan direction)\ The DR2 place in history {#sect:significance} ======================== Even at the level of the Gaia DR2, the Gaia products are just amazing in terms of volume and quality. Although the numbers look impressive, but arguably this is true for many space missions, it is not straightforward to appreciate how much better or larger it stands compared to available astronomical data. I attempt to address this issue by showing that Gaia is producing much more than an incremental advance in the collection of basic astrophysical data, both in volume and accuracy, but that we are witnessing a true discontinuity of historical significance. For this purpose I compare the content of the Gaia survey to the astronomical knowledge at different times in the past. It is never easy to tell without ambiguity *how many parallaxes or proper motions* were known in 1850 or 1950. There is no clear boundary of what should be considered as known and published values may be wrong or so poor as not providing a real knowledge. Therefore the figures on the set of tables below are often rounded and in details could be disputed, but not by large margin. At the end the trend is clear and does not suffer discussion. The general trend ----------------- My main point regarding the jump or discontinuity created by the Gaia catalogue is supported by the Table \[table:history\]. It gives for the most relevant parameters covering astrometry, spectroscopy and photometry the size of the available material at three different epochs, namely 1900, 1980, 2018. Although the number of stars for which one knows the parallaxes or the radial velocity say nothing about how well this is known, the large difference in the figures speaks for itself, and usually the increase in size goes with an increase in accuracy. *Sky survey* refers to the number of stars that have been catalogued with a position good enough to identify the star at any further epoch, a census of the sky in short. Parameter ----------------------------- -------- -------------- -------- sky survey 800000 20 M$\,^a$ 1.7 B \[1pt\] fundamental sources 350 1500$\,^b$ 550000 \[1pt\] stellar parallaxes 100 10000 1.3 B \[1pt\] proper motions 1000 250000$\,^c$ 1.3 B \[1pt\] radial velocities 50 30000 7.2 M \[1pt\] variable stars 300 15000 500000 \[1pt\] stellar temperature 10000 300000 150 M \[1pt\] : General evolution in the size of general and specialised stellar catalogues. \[table:history\] The Survey ---------- In comparing size of surveys at different epochs, I looked at the largest cataloguing of the whole sky, without caring about the positional accuracy. The object of these censuses was just to list the place of the sources on the sky up to a certain magnitude, with a positional accuracy good enough to recover the star and tell later between a known star and a moving body or a new star (becoming bright enough to be seen, but not really new and often referred to as a *guest star*). At different epochs efforts were made to get a completeness, at least from one place, of all the stars brighter than a given limit and provide also an approximate magnitude to help in the identification. The numbers of stars growing exponentially with the limiting magnitude, these systematic surveys were quickly limited by the observational burden and the data handling, even when photographic plates became available. The compilation is given in Table \[table:stat\_surveys\] and provides an idea of how many stars appeared in the largest surveys in the past. By the end of XIXth century, all the stars to $V\approx 10$ were catalogued. The progresses were truly slow and most of the advances since 1950 were made possible with the advent of large computers with massive data storage. Gaia survey is today the largest, but smaller areas of sky have been investigated by optical or near IR telescopes at fainter magnitude than Gaia limit. In the coming years the LSST will outnumber Gaia by one order of magnitude. For Gaia this was not a science goal to make the biggest survey in the sense it was done 150 years ago. However the last column of the table show where the strength of Gaia really lies in terms of positional accuracy. mag $< $ ----------------------- --------------- -------- ---------- ------------------ Ptolemy II$^e$ siècle 1030 6 0.3 [$^\circ$]{} \[1pt\] Flamsteed 1700 2934 7 20 \[1pt\] Lacaille 1750 10000 7 5 \[1pt\] Lalande 1800 48000 9 3 \[3pt\] BD 1860 450000 9.5 20 \[1pt\] CD 1900 600000 10 20 \[1pt\] CPD 1900 450000 11 5 \[1pt\] Carte du Ciel 1950 2 M 12 0.5 \[1pt\] Tycho 2 2000 2.5 M 11.5 50 mas \[3pt\] 2MASS 2000 300 M 15 (IR) 500 mas \[1pt\] USNO B1 2000 1000 M 21 200 mas \[1pt\] **Gaia** 2015 1700 M 21 0.2 mas \[1pt\] LSST 2030 ? 15 B 24 5 mas : Historical sky surveys. \[table:stat\_surveys\] The reference system -------------------- For the general public a catalogue of stars is before all a list of sources one can locate in the sky thanks to the two angular coordinates given in each line. This is more or less the goal of the sky census just mentioned. But scientifically more important is the use of an accurate position catalogue to materialise a reference frame in space. As said by W. Fricke[@1985CeMec..36..207F], a recognised expert in this field, *the main purpose of a Fundamental Catalogue in astronomy consists in providing a coordinate system for describing the motion of the planetary systems $(\cdots$) and for the determination of the proper motions of stars*. In short one wishes to have consistent coordinates for a small set of point-sources on the sky so that this can be used to fix the three orthogonal directions $[x, y, z]$ of an inertial system, although the language should be slightly altered in the framework of General Relativity. Producing a fundamental catalogue requires absolute observations not depending on earlier positions of the same stars or of other stars used to derive relative positions. For the general principles and construction see the book by H.G. Walter and O.J. Sovers [@2000afce.conf.....W]. Building a fundamental catalogue is a very demanding task, extending over years and even decades, and a handful have ever existed containing few 100-1000s sources, stars until Hipparcos and quasars more recently. Table \[table:stat\_refcat\] provides an almost exhaustive list of fundamental catalogues since the concept has emerged in the mid-XVIIIth century. Comparing the 3rd column in this table and in Table \[table:stat\_surveys\] is enough to grasp the wide difference between the two kinds of catalogues, two different worlds indeed. Remarkably there is only one catalogue common to these two tables: Gaia, but not for the same sources. The Gaia CRF (Celestial Reference Frame) is built on a small subset of sources, namely the $550\,000$ QSOs meeting the overall principles of the ICRS (International Celestial Reference System), as explained in Mignard et al. . Gaia accuracy is on a par with the most recent version of the ICRF (the IAU approved fundamental reference frame) in the radio wavebands and resulting from VLBI astro-geodetic observations. However the density is 100 times higher and it is directly accessible in the optical domain. Both realisations are needed and used in different contexts in astronomy and geodesy. [lcrcr]{} & & & mag $< $ &\ Lacaille & 1760 & 397 & 7 & 10\ Maskelyne & 1774 & 36 & 5 & 5\ Piazzi & 1818 & 220 & 6 & 2\ Bessel & 1830 & 36 & 5 & 1\ Argelander & 1869 & 160 & 6 & 1\ Auwers & 1879 & 539 & 6 & 0.5\ FK4 & 1963 & 1535 & 7.5 & 0.2\ FK5 & 1988 & 1535 & 7.5 & 40 mas\ Hipparcos$^a$ & 1996 & 100000 & 11.5 & 1 mas\ ICRF1 (radio) & 1998 & 620 & - & 2 mas\ ICRF2 (radio) & 2009 & 3400 & - & 0.6 mas\ ICRF3 (radio) & 2018 & 4500 & - & 0.2 mas\ **Gaia** QSOs & 2018 & 550000 & 21 & 0.4 mas\ **Gaia** QSOs $G< 18$ & 2018 & 27000 & 18 & 0.12 mas\ \ The parallaxes -------------- Photometry, radial velocities, stellar parameters can in principle be obtained with ground based instruments, provided the necessary means and human resources are made available over several years to support this routine work. This is *easier* and more productive from space, but space is not technically required. This is not true for absolute parallaxes which need the finest astrometry not achievable through the atmosphere. The parallaxes of stars, or equivalently their distances, were searched as soon as the Heliocentric system became accepted by astronomers. Our displacement around the Sun should translate into an annual reflex motion of the stars whose extent is in direct relation to their distances. So the underlying geometric principles needed to ascertain the distances are extremely simple and were well understood well before any realistic attempt to detect the tiny parallactic effect was feasible. The first success cam with F. Bessel in 1838 and further progress were extremely slow. Parallaxes are small in regard of the achievable astrometric accuracy and their astrophysical value is low if not better than $10\%$ in fractional accuracy. Even in 1980, just before Hipparcos, many of the 8000 stars with published parallaxes did not reach this goal. In addition they were relative parallaxes and depended on additional calibration/assumptions for the reference stars used to record the parallactic displacement. ------------------- -------- ------------------------------- year \[2pt\] 1840 3 61 Cygni, Vega, $\alpha$ Cent \[2 pt\] 1850 20 C.A.F. Peters \[2 pt\] 1890 45 Ch. André (astron. stell.) \[2 pt\] 1910 300 incl. 52 with photography \[6 pt\] 1925 2000 photographic plates \[2 pt\] 1965 6000 Yale Catalogue \[2 pt\] 1990 8000 just before Hipparcos \[6 pt\] 1996 120000 Hipparcos \[2 pt\] **2018** 1300 M Gaia **2025** 1800 M Gaia ------------------- -------- ------------------------------- : Number of published stellar parallaxes.\[table:stat\_varpi\] Hipparcos astrometry was a truly new start for parallax survey. The total number of trigonometric parallaxes rose at once to more than $100,000$, with nearly $50,000$ better than 20% in fractional errors ($\sigma_\varpi/\varpi < 20\%$) and $20,000$ at the $10\%$ level. The new technique, allowing to get virtually absolute parallaxes, proved efficient and multiplied in only 3 years the number of measured parallaxes by a factor 15. This success paved the way for Gaia and the numbers given in Table \[table:stat\_varpi\] demonstrate without further comments the power of Gaia compared to the Ground-based methods, painstaking and with low yield. This is a true revolution in this field, something that any astronomer of the 1980s could not simply dream of. With Gaia DR2 one has parallax estimates for $1.3$ billion stars, but more important above 50 million with a relative accuracy better than $10\%$. With the coming releases this figure will soar above the 100 million. The Gaia parallax survey is by far the most comprehensive ever done and has no match in terms of size and accuracy, with the exception, regarding the accuracy, of a handful of radio masers observed with the VLBI technique. For the Galactic stars this is the crowning of nearly two centuries of parallax quest starting with F.W. Bessel in 1838. Until many years in the future there will be no such undertaking to get trigonometric parallaxes directly from astrometric observations and the Gaia survey is now actively being used to reconstruct the whole distance scale beyond the Galaxy, based on secondary indicators. The proper motions ------------------ The same historical comparison is done for the proper motions of stars with results given in Table \[table:stat\_pm\]. Detecting the displacement of stars is in principle not too difficult, since it required basically only patience and archives. Patience, since the changes in the relative positions of stars is very slow at the scale of human life, and then one must wait decades or centuries before it becomes manifest. But for that one needs also to keep track of where the stars were in the past, otherwise no comparison is feasible. It was not until the early years of the XVIIIth century that hints of a possible break in the doctrine of the *fixed stars* came to the fore and no indisputable quantitative measurement was available before J. Cassini confirmed beyond doubt that Arcturus has changed position over 40 years. Table \[table:stat\_pm\] is rather similar to that for the parallaxes and just illustrates the sluggish progress of astrometry until space astrometry became a reality. Accuracy is a key factor, since the computed positions degrade in proportion of the uncertainty of the proper motions. The impact of Gaia is obvious with an incredible increase in quantity and quality. And the icing on the cake with Gaia, angular displacement combined with parallaxes gives the true velocity in [[kms$^{-1}$]{}]{}. And for the bright end, Gaia spectrometer delivers also the third component on the line of sight. ------------------- -------- ------------------------------------------------------------ Year \[2pt\] 1738 1 J. Cassini \[2 pt\] 1760 15 T. Mayer \[2 pt\] 1790 36 N. Maskelyne \[2 pt\] 1835 390 Argelander, $\mu > 100 \text{ mas/a}$ \[2 pt\] 1856 3200 Mädler, (Bradley’s stars) \[6 pt\] 1939 33000 Gal. Cat., $ V < 7.5$ $\sigma \approx 10 \text{ mas/a}$ \[2 pt\] 1966 260000 SAO $\sigma \approx 15 \text{ mas/a}$ \[6 pt\] 1990 400000 PPM $\sigma \approx 4 \text{ mas/a}$ \[2 pt\] 1996 120000 Hipparcos $\sigma \approx 1 \text{ mas/a}$ \[2 pt\] 2002 2.5 M Tycho-2 $\sigma \approx 2.5 \text{ mas/a}$ \[6 pt\] **2018** 1.3 B Gaia $\sigma \approx 0.3 \text{ mas/a} \text{ @ } G =18$ **2025** 1.7 B Gaia $\sigma \approx 0.03 \text{ mas/a} \text{ @ } G =18$ ------------------- -------- ------------------------------------------------------------ : Number of known proper motions at different epochs.\[table:stat\_pm\] Acknowledgments {#acknowledgments .unnumbered} =============== The section summarising the Gaia mission takes heavily on the Gaia on-line documentation and summary papers prepared by the DPAC collaboration to accompany the data releases. The members of the DPAC are collectively thanked for providing this material and for their dedication to the success of Gaia. References {#references .unnumbered} ========== [^1]: To appear in Proceedings of Rencontres de Moriond 2019, Gravitation
{ "pile_set_name": "ArXiv" }
--- abstract: 'We apply an integral inequality to obtain a rigorous apriori estimate of the accuracy of the partial sum to the power series solution of the celebrated <span style="font-variant:small-caps;">Riccati</span>-<span style="font-variant:small-caps;">Bernoulli</span> differential equation.' author: - | Mark B. Villarino\ Depto. de Matemática, Universidad de Costa Rica,\ 2060 San José, Costa Rica date: 'December 16, 2007' title: 'An Integral Inequality and the Riccati-Bernoulli Differential Equation' --- Introduction ============ The deservedly celebrated <span style="font-variant:small-caps;">Riccati</span>-<span style="font-variant:small-caps;">Bernoulli</span> differential equation:$$\label{RB} \boxed{y'=x^2+y^2}$$has the general solution ([@EP]):$$\label{GS} \boxed{y(x)=x \frac{J_{\frac{3}{4}}\left(\frac{1}{2}x^2\right)-cJ_{\frac{-3}{4}}\left(\frac{1}{2}x^2\right)}{cJ_{\frac{1}{4}}\left(\frac{1}{2}x^2\right)+J_{\frac{-1}{4}}\left(\frac{1}{2}x^2\right)}}$$where $c$ is an arbitrary constant and where $$\label{ BS} \boxed{J_n(x):=\frac{x^n}{2^n\Gamma(n+1)}\left\{1-\frac{x^2}{2^2\cdot1!\cdot(n+1)}+\frac{x^4}{2^4\cdot2!\cdot(n+1)(n+2)}-\cdots\right\}}$$is ***the Bessel function of the first kind of order*** $n$, where $n$ is any real number, and $\Gamma(n+1)$ is the famous ***gamma function***. The very interesting history of is detailed in <span style="font-variant:small-caps;">Watson</span>’s standard treatise [@WA]. The equation is an example of a simple differential equation whose solutions form a family of transcendental functions which are essentially distinct from the elementary transcendents. Unfortunately, the general solution does not easily lend itself to a rigorous error analysis of its accuracy in a particular interval of the variable. We will show how a simple application of an ***integral inequality*** allows one to estimate the accuracy of the partial sum of the <span style="font-variant:small-caps;">Taylor</span> series expansion of the solution within the latter’s interval of convergence. Cauchy’s theorem ================ (In this section we follow [@IPM], Chapter IV, section 5). The general ***Cauchy Problem*** is to solve the ordinary differential equation (ODE) initial-value problem: $$\label{IVP} \boxed{y'=f(x,y), \ \ \ \ y(x_0):=y_0}$$Let $f(x,y)$ be expanded in the series $$\label{PS} f(x,y)=\sum_{i,j}A_{ij}(x-x_0)^i(y-y_0)^j$$convergent for $$\begin{aligned} \label{R} |x-x_0|<R_1& &|y-y_0|<R_2&&(R_1>0, \ \ R_2>0) \end{aligned}$$Then, according to <span style="font-variant:small-caps;">Cauchy</span>’s theorem in the theory of differential equations, the problem has a solution $y(x)$ represented by the series$$\label{y(x)} \boxed{y(x)=\sum_{k=0}^{\infty}\frac{y^{(k)}(x_0)}{k!}(x-x_0)^k}$$convergent in some neighborhood of the point $x_0$. <span style="font-variant:small-caps;">Cauchy</span>’s theorem allows one also to indicate the neighborhood of the point $x_0$, in which the series converges. Namely, let $M$ be a constant such that$$\label{M} |f(x,y)|{\leqslant}M$$with$$\begin{aligned} \label{r12} |x-x_0|{\leqslant}r_1<R_1& & |y-y_0|{\leqslant}r_2<R_2 \end{aligned}$$where $R_1$ and $R_2$ are numbers, defining the region of the convergence of the series , and $r_1$ and $r_2$ are some positive numbers. *Then the series converges for*$$\label{xx_0} |x-x_0|<r$$*where*$$\label{r} \boxed{r:=r_1\left\{1-e^{-\frac{1}{2M}\frac{r_2}{r_1}}\right\}}$$It should be noted that the *true* interval of convergence is usually much larger than , An integral inequality ====================== We use the notation of : Let $I$ be the interval, $x_0{\leqslant}x{\leqslant}x_1$, and suppose that for all $x\in I$, $f(x,y)>0$, and that the differential inequality $$\label{DE} y'{\leqslant}f\{x_1,y(x)\}$$also holds there. Then, the integral inequality $$\label{IE} \boxed{\int_{x_0}^{x}\left\{\frac{1}{f[x_1, y(t)]}\frac{dy}{dt}\right\}~dt{\leqslant}x-x_0}$$holds for all $x\in I$. This is a simple consequence of the elementary calculus sufficient condition that a function be *decreasing* in an interval. Define$$\label{ } g(x):=\int_{x_0}^{x}\left\{\frac{1}{f[x_1, y(t)]}\frac{dy}{dt}\right\}~dt- (x-x_0).$$Taking the derivative and using the fundamental theorem of calculus we obtain $$\label{ } g'(x)=\left\{\frac{1}{f[x_1, y(x)]}\frac{dy}{dx}\right\}-1$$But, the inequality shows that $$\label{ } g'(x){\leqslant}0$$for all $x\in I$. The Riccati-Bernoulli Initial Value Problem =========================================== (In this section we follow [@IPM], Chapter IV, section 5, but with some important refinements.) ***Problem:*** *It is required to find the first 11 terms of the power series expansion of the solution of the initial value problem*$$\label{IVP2} \boxed{y'=x^2+\frac{y^2}{4}, \ \ \ \ y(0):=-1}$$*and an interval of convergence for it.* If we specialize the general solution to this initial value problem we find the following formula for the exact solution:$$\label{GS1} \boxed{y(x)=\frac{x}{16} \cdot \left\{\frac{\Gamma\left(\frac{3}{4}\right)J_{\frac{3}{4}}\left(\frac{1}{2}x^2\right)-\sqrt{2}\Gamma\left(\frac{1}{4}\right)J_{\frac{-3}{4}}\left(\frac{1}{2}x^2\right)}{\sqrt{2}\Gamma\left(\frac{1}{4}\right)J_{\frac{1}{4}}\left(\frac{1}{2}x^2\right)+\Gamma\left(\frac{3}{4}\right)J_{\frac{-1}{4}}\left(\frac{1}{2}x^2\right)}\right\}}$$Unfortunately, the computation of the power series solution on the basis of the quotient , although theoretically possible, is computationally formidable. It is easier to use the equation to compute the derivatives of $y(x)$ at $x=0$ directly:$$\begin{aligned} y' = &x^2+\frac{y^2}{4}&=\frac{1}{4} \label{y1} \\ y'' =& 2x+\frac{1}{2}yy'&=-\frac{1}{8} \label{y2}\\ y'''= &2+\frac{1}{2}y'^2+\frac{1}{2}yy''&=\frac{67}{32} \label{y3}\\ y^{(4)}= &\frac{3}{2}y'y''+\frac{1}{2}yy'''&=-\frac{35}{32} \label{y4}\\ y^{(5)}= &\frac{3}{2}y''^2+2y'y'''+\frac{1}{2}yy^{(4)}&=\frac{207}{128} \label{y5}\\ y^{(6)}= &5y''y'''+\frac{5}{2}y'y^{(4)}+\frac{1}{2}yy^{(5)}&=-\frac{231}{64} \label{y6}\\ y^{(7)}= &5y'''^2+\frac{15}{2}y''y^{(4)}+3y'y^{(5)}+\frac{1}{2}yy^{(6)}&=\frac{26585}{1024} \label{y7}\\ y^{(8)}= &\frac{35}{2}y'''y^{(4)}+\frac{21}{2}y''y^{(5)}+\frac{7}{2}y'y^{(6)}+\frac{1}{2}yy^{(7)}&=-\frac{119475}{2048} \label{y8}\\ y^{(9)}= &\frac{35}{2}(y^{(4)})^2+28y'''y^{(5)}+14y''y^{(6)}+4y'y^{(7)}+\frac{1}{2}yy^{(8)}&=\frac{725769}{4096} \label{y9}\\ y^{(10)}=&63y^{(4)}y^{(5)}+42y'''y^{(6)}+18y''y^{(7)}+\frac{9}{2}y'y^{(8)}+\frac{1}{2}yy^{(9)}&=-\frac{10509885}{16384}\ \label{y10}\end{aligned}$$Therefore, by the formula , the first $11$ terms of the series solution are:$$\label{sol} \begin{split} y(x)&=-1+\frac{1}{4}x-\frac{1}{16}x^2+\frac{67}{192}x^3-\frac{35}{768}x^4+\frac{69}{5120}x^5-\frac{77} {15360}x^6+\frac{5317}{1032192}x^7\\ &\quad -\frac{2655}{1835008}x^8+\frac{80641}{165150720}x^9-\frac{77851}{62914560}x^{10}+\cdots \end{split}$$ To find an interval of convergence of this series we use the <span style="font-variant:small-caps;">Cauchy</span> theorem. It $x$ and $y$ satisfy the inequalities $$\begin{aligned} \label{} |x|{\leqslant}0.5 &&|y+1|{\leqslant}1 \end{aligned}$$then we may conclude that $$\begin{aligned} \label{} |f(x,y)|{\leqslant}&|0.25[(y+1)-1]^2+x^2| \\ {\leqslant}& 0.25(|y+1|+1)^2+|x|^2\\ {\leqslant}&1.25 \end{aligned}$$Therefore, in the formula , we may take $$\begin{aligned} \label{} r_1:=0.5&&r_2:=1&&M:=1.25 \end{aligned}$$and the value of $r$ we obtain is:$$\label{ } r=0.5\left(1-e^{-0.8}\right)=0.2753355\cdots$$Therefore, *the power series solution most certainly converges for $|x|{\leqslant}0.27$*. The Accuracy of a Partial Sum from the Integral Inequality ========================================================== We will consider the following concrete problem although the principles are of general applicability. ***Problem:*** *It is required to determine the accuracy of the partial sum of degree $9$ of the power series solution in the interval $0{\leqslant}x{\leqslant}0.2$*. Since the series is the <span style="font-variant:small-caps;">Maclaurin</span> expansion of $y(x)$, we must estimate the remainder term, $R_9(x)$, which we write in the <span style="font-variant:small-caps;">Lagrange</span> form:$$\label{Rem} \boxed{R_{9}(x):=\frac{y^{(10)}(\Theta_9)}{10!}x^{10}}$$where $0<\Theta_9<0.2$. We have to estimate $y^{(10)}(x)$, the formula for which is given in , for all values of $x$ in the interval $0{\leqslant}x{\leqslant}0.2$. The formulas through show us, finally, that *we must estimate $y(x)$, itself in $0{\leqslant}x{\leqslant}0.2$*. The estimate of $y(x)$ via *our integral inequality* constitutes the novelty in this paper. Maintaining the notation of we see that$$\label{ } x_0:=0 \ \ \ \ \ \ x_1:=0.2,$$that the right hand side $$\label{ } f(x,y):=x^2+\frac{y^2}{4}>0$$on $I$, and the differential inequality becomes: $$\label{ } \frac{dy}{dx}{\leqslant}0.04+\frac{y(x)^2}{4}$$Therefore, the integral inequality becomes $$\label{ } \int_{0}^{x}\left\{\frac{1}{0.04+\dfrac{y(t)^2}{4}}\frac{dy}{dt}\right\}~dt{\leqslant}x$$But, $$\label{ } \left\{\frac{1}{0.04+\dfrac{y(t)^2}{4}}\frac{dy}{dt}\right\}=\frac{d}{dt}\left\{10\arctan\frac{y(t)}{0.4}\right\}$$Therefore, $$\label{ } 10\arctan\frac{y(x)}{0.4}-10\arctan\frac{-1}{0.4}{\leqslant}x$$or $$\label{ } \arctan\frac{y(t)}{0.4}{\leqslant}\frac{x}{10}+\arctan\frac{-1}{0.4}$$and taking the tangent of both sides and reducing *we obtain the estimate* $$\label{y11} \boxed{y(x){\leqslant}\frac{\frac{2}{5}\tan\left(\frac{x}{10}\right)-1}{1+\frac{x}{4}}}$$*which holds for all* $x\in I$, The function on the right-hand side of is monotonically increasing in $I$, and$$\frac{\frac{2}{5}\tan\left(\frac{0.2}{10}\right)-1}{1+\frac{0.2}{4}}=-0.9447608\cdots<-0.94$$and *we have proved that the following inequality is true for all* $x\in I$: $$\label{Y1} \boxed{-1{\leqslant}y(x){\leqslant}-0.94.}$$(*Note:* the true value of $y(0.2)$ is $$y(0.2)=-0.9497771\cdots,$$ so the estimate , with an error of $-0.00501\cdots$, or about $0.53\%$, is quite good!). Now we must estimate $y'(x), \cdots, y^{(10)}(x)$ using the formulas through . The following inequalities are valid for all $x\in I$. $$\begin{aligned} -1 & {\leqslant}y(x) {\leqslant}& -0.94, \\ 0.22 & {\leqslant}y'(x) {\leqslant}& 0.29, \\ -0.15 & {\leqslant}y''(x) {\leqslant}& 0.3, \\ 1.87 & {\leqslant}y'''(x) {\leqslant}& 2.12 \\ -1.13 & {\leqslant}y^{(4)}(x) {\leqslant}& -0.74, \\ 1.17 & {\leqslant}y^{(5)}(x) {\leqslant}& 1.93, \\ -3.38 & {\leqslant}y^{(6)}(x) {\leqslant}& 2.23, \\ 14.59 & {\leqslant}y^{(7)}(x) {\leqslant}& 27.12, \\ -61.96 & {\leqslant}y^{(8)}(x) {\leqslant}& -22.73, \\ 92.03 & {\leqslant}y^{(9)}(x) {\leqslant}& 146.76, \\\end{aligned}$$and finally $$\label{ } \boxed{-665.9<y^{(10)}(x)<281.}$$ All of our estimates come from worst case values applied to each of the summands in the formulas. By $$\label{yy1} \boxed{y' = x^2+\frac{y^2}{4}.}$$Therefore, using and , we conclude that for all $x\in I$, $$y'< 0.2^2+\frac{(-1)^2}{4}=0.29,$$while $$y'> 0^2+\frac{(-.94)^2}{4}=.2209>.22.$$Therefore, we obtain the bounds $$\label{YY1} \boxed{.22<y'(x)<.29.}$$ By $$\label{yy2} \boxed{y'' = 2x+\frac{1}{2}yy'.}$$Therefore, using and , we conclude that for all $x\in I$, $$y''< 2(0.02)+\frac{1}{2}(-0.94)(.22)=0.2966<.3,$$while $$y''> 2(0)+\frac{1}{2}(-1)(.29)=-0.145>-0.15.$$Therefore, we obtain the bounds $$\label{YY2} \boxed{-0.15<y''(x)<0.3.}$$ By $$\label{yy3} \boxed{y'''=2+\frac{1}{2}y'^2+\frac{1}{2}yy''.}$$Therefore, using , and , we conclude that for all $x\in I$, $$y'''<2+\frac{1}{2}(0.29)^2+\frac{1}{2}(-0.15)(-1) =2.11705<2.12,$$while $$y'''>2+\frac{1}{2}(0.22)^2+\frac{1}{2}(0.3)(-1) =1.8742>1.87.$$Therefore, we obtain the bounds $$\label{YY3} \boxed{1.87<y'''(x)<2.12.}$$ By $$\label{yy4} \boxed{y^{(4)}=\frac{3}{2}y'y''+\frac{1}{2}yy'''.}$$Therefore, using , , , and , we conclude that for all $x\in I$, $$y^{(4)}<\frac{3}{2}(0.29)(0.3)+\frac{1}{2}(-0.94)(1.87)=-0.7484<-.074,$$while $$y^{(4)}>\frac{3}{2}(0.29)(-0.15)+\frac{1}{2}(-1)(2.12)=-1.12525>-1.13.$$Therefore, we obtain the bounds $$\label{YY4} \boxed{-1.13<y^{(4)}(x)<-0.74.}$$ By $$\label{yy5} \boxed{y^{(5)}=\frac{3}{2}y''^2+2y'y'''+\frac{1}{2}yy^{(4)}.}$$Therefore, using , , , and , we conclude that for all $x\in I$, $$y^{(5)}<\frac{3}{2}(0.3)^2+2(0.29)(2.12)+\frac{1}{2}(-1)(-1.13)=1.9296<1.93,$$while $$y^{(5)}>\frac{3}{2}(0)^2+2(0.22)(1.87)+\frac{1}{2}(-0.94)(-0.74)=1.1706>1.17.$$Therefore, we obtain the bounds $$\label{YY5} \boxed{1.17<y^{(5)}(x)<1.93.}$$ By $$\label{yy6} \boxed{y^{(6)}=5y''y'''+\frac{5}{2}y'y^{(4)}+\frac{1}{2}yy^{(5)}.}$$Therefore, using , , , , and , we conclude that for all $x\in I$, $$y^{(6)}<5(.3)(2.12)+\frac{5}{2}(.22)(-0.74)+\frac{1}{2}(-0.94)(1.17))=2.2231<2.23,$$while $$y^{(6)}>5(-0.15)(2.12)+\frac{5}{2}(.29)(-1.13)+\frac{1}{2}(-1)(1.93))=-3.37425>-3.38.$$Therefore, we obtain the bounds $$\label{YY6} \boxed{-3.38<y^{(6)}(x)<2.23.}$$ By $$\label{yy7} \boxed{y^{(7)}=5y'''^2+\frac{15}{2}y''y^{(4)}+3y'y^{(5)}+\frac{1}{2}yy^{(6)}.}$$Therefore, using , , , , , and , we conclude that for all $x\in I$, $$y^{(7)}<5(2.12)^2+\frac{15}{2}(-0.15)(-1.13)+3(.29)(1.93)+\frac{1}{2}(-1)(-3.38),$$which$$=27.11235<27.12,$$while $$y^{(7)}>5(1.87)^2+\frac{15}{2}(0.3)(-1.13)+3(.22)(1.17)+\frac{1}{2}(-1)(2.23),$$which$$=14.5992>14.59.$$Therefore, we obtain the bounds $$\label{YY7} \boxed{14.59<y^{(7)}(x)<27.12.}$$ By $$\label{yy8} \boxed{y^{(8)}=\frac{35}{2}y'''y^{(4)}+\frac{21}{2}y''y^{(5)}+\frac{7}{2}y'y^{(6)}+\frac{1}{2}yy^{(7)}.}$$Therefore, using , , , , , , and , we conclude that for all $x\in I$, $$y^{(8)}<\frac{35}{2}(1.87)(-0.74)+\frac{21}{2}(0.3)(1.93)+\frac{7}{2}(.29)(2.23)+\frac{1}{2}(-0.94)(14.59),$$which$$=-22.73085<-22.73,$$while $$y^{(8)}>\frac{35}{2}(2.12)(-1.13)+\frac{21}{2}(-0.15)(1.93)+\frac{7}{2}(.29)(-3.38)+\frac{1}{2}(-1)(27.12),$$which$$=-61.95345>-61.96.$$Therefore, we obtain the bounds $$\label{YY8} \boxed{-61.96<y^{(8)}(x)<-22.73.}$$ By $$\label{yy9} \boxed{y^{(9)}=\frac{35}{2}(y^{(4)})^2+28y'''y^{(5)}+14y''y^{(6)}+4y'y^{(7)}+\frac{1}{2}yy^{(8)}.}$$Therefore, using , , , , , , , and , we conclude that for all $x\in I$, $$y^{(9)}<\frac{35}{2}(-1.13)^2+28(2.12)(1.93)+14(0.3)(2.23)+4(.29)(27.12)+\frac{1}{2}(-1)(-61.96),$$which $$=146.75575<146.76,$$while $$y^{(9)}>\frac{35}{2}(-0.74)^2+28(1.87)(1.17)+14(-0.15)(2.23)+4(.22)(14.59)+\frac{1}{2}(-0.94)(-27.73),$$which$$=92.0335>92.03.$$Therefore, we obtain the bounds $$\label{YY9} \boxed{92.03<y^{(9)}(x)<146.76.}$$ By $$\label{yy10} \boxed{y^{(10)}=63y^{(4)}y^{(5)}+42y'''y^{(6)}+18y''y^{(7)}+\frac{9}{2}y'y^{(8)}+\frac{1}{2}yy^{(9)}.}$$Therefore, using , , , , , , , , and , we conclude that for all $x\in I$, $$y^{(10)}<63(-0.74)(1.17)+42(2.12)(2.23)+18(0.3)(27.12)+\frac{9}{2}(0.22)(-22.73)+\frac{1}{2}(-0.94)(92.03),$$which $$=280.9922<281,$$while $$y^{(10)}>63(-1.13)(1.93)+42(2.12)(-3.38)+18(-0.15)(27.12)+\frac{9}{2}(0.29)(-61.96)+\frac{1}{2}(-1)(146.76),$$which$$=-665.8137>-665.9.$$Therefore, we obtain the bounds $$\label{YY10} \boxed{-665.9<y^{(10)}(x)<281.}$$ *The inequality was the goal is this long and detailed computation!* Now we can state the accuracy of the partial sum: For all $x$ in $0{\leqslant}x{\leqslant}0.2$, the partial sum of degree $9$: $$\label{soll} \begin{split} y(x)&\approx-1+\frac{1}{4}x-\frac{1}{16}x^2+\frac{67}{192}x^3-\frac{35}{768}x^4+\frac{69}{5120}x^5-\frac{77} {15360}x^6+\frac{5317}{1032192}x^7\\ &\quad -\frac{2655}{1835008}x^8+\frac{80641}{165150720}x^9\end{split}$$approximates the true value, $y(x)$, of the solution series, , with an error that does not exceed $2$ units in the eleventh decimal place. The estimate $$\label{Remm} |R_{9}(x)|=\left|\frac{y^{(10)}(\Theta_9)}{10!}x^{10}\right|{\leqslant}\frac{665.9}{10!}(0.2)^{10}=1.878\cdots10^{-11}<2\cdot 10^{-11}$$completes the proof. Conclusions =========== Our integral inequality can be applied to wide classes of differential equations. For example, our method allows us to prove that in the interval $0{\leqslant}x{\leqslant}0.4$ the polynomial$$\label{ } \overline{y}(x):=1+\frac{x}{4}+\frac{3}{16}x^2+\frac{7}{192}x^3 +\frac{1}{96}x^4+\frac{1}{200}x^5$$approximates the true solution, $y(x)$, of the initial value problem$$\label{ } 4y'=x+y^2, \ \ \ \ \ y(0):=1$$with an error that does not exceed $2$ units in the fifth decimal place. We chose the <span style="font-variant:small-caps;">Riccati</span>-<span style="font-variant:small-caps;">Bernoulli</span> equation because it illustrates the process so perfectly and because a direct estimate of the accuracy of the partial sum of the series solution is troublesome. We did not investigate the accuracy for the *negative* half of the interval, i.e., for $-0.2{\leqslant}x{\leqslant}0$, which we leave as an exercise for the reader. The only change occurs in the estimate of $y''(x)$ since then $x$ can be equal to negative numbers. Finally we observe that we did not exploit the *sign* of the error. In fact, our estimate allows us to say that the error we commit is between an error in *defect* smaller than $2$ units in the eleventh decimal place and an error in *excess* smaller than $8$ units in the twelfth decimal place. Therefore we can *centralize the error* by adding the term $$\dfrac{281-665.9}{2\cdot 10!}x^{10}=\dfrac{-1283}{24192000}x^{10}$$ to our polynomial and obtain an approximating polynomial whose maximum error is between $\pm 1.4$ units in the 11th decimal place. ### Acknowledgment {#acknowledgment .unnumbered} Support from the Vicerrectoría de Investigación of the University of Costa Rica is acknowledged. C.H. Edwards, D.E. Penny, *Differential Equations and Boundary Value Probems*, Prentice Hall, Upper Saddle River, New Jersey, 2000. I.P. Mysovskih, *Lectures on Numerical Methods*, Wolters-Noordhoff, Gronigen, The Netherlands, 1969. G.N. Watson *A Treatise on the Theory of Bessel Functions*, Cambridge University Press, New York, New York, 1958.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We prove that a submaximal curve in ${\mathbb{P}}^2$ has sequence of multiplicities $(\mu,\nu, \dots, \nu)$, with $\mu < s \nu $ for every integer $s$ with $(s-1)^2(s+2)^2\ge 6.76(\, r-1)$.' author: - | Joaquim Roé\ \ bibliography: - 'Biblio.bib' title: On submaximal plane curves --- This note is a sequel to [@Roe01], where a specialization method was developed in order to bound the degree of singular plane curves. The problem under consideration is, given a system of multiplicities $(m)=(m_1,m_2,\dots,m_r) \in {\mathbb{Z}}^r$ and points $p_1, \dots, p_r \in {\mathbb{P}}^2$, which we shall always assume to be in general position, to determine the minimal degree $\alpha(m)$ of a curve with multiplicity $m_i$ at each point $p_i$. In [@Roe01], the focus was on homogeneous $(m)$, (i.e., $m_1=m_2=\dots=m_r$), but the method applies in general; here it is used to show that if one of the multiplicities is much bigger than the others, in a sense we make precise below (see theorem \[nohomo\]), then $$\label{eqnag} \alpha(m)> \frac{\sum_{i=1}^r m_i}{\sqrt{r}}.$$ In connection with his solution to the fourteenth problem of Hilbert, M. Nagata conjectured in 1959 that the inequality (\[eqnag\]) holds for all $(m)$ provided $r>9$, and proved it in the case when $r$ is a perfect square (see [@Nag59] or [@Nag60]). Since then, many partial results have been proved by several authors (see for instance [@Xu94], [@Eva98], [@Har01], [@HR??], [@Sze01], [@Tut??]), but as far as we know the conjecture remains open in general. One of the research lines in this area is the study of *submaximal curves* that arise in the context of Seshadri constants. A submaximal curve is an irreducible curve which causes the ($r$-point) Seshadri constant of a surface to be non-maximal; in the case of ${\mathbb{P}}^2$ it is just an irreducible curve which causes (\[eqnag\]) to fail, and Nagata’s conjecture can be equivalently stated by saying that there exist no submaximal curves for $r>9$. T. Szemberg proved in [@Sze01 4.6] that every submaximal curve on a surface with Picard number $\rho=1$ whose multiple points are in general position is quasi-homogeneous, i.e., has $m_2=\dots=m_r$ for a suitable ordering of the points. In the case of the projective plane, our result shows that moreover $m_1$ can not be much bigger than $m_2$, constraining further the range of possible counterexamples to Nagata’s conjecture. It is worth mentioning that quasi-homogeneous curves are relevant also for the method of C. Ciliberto and R. Miranda [@CM98a] to compute the dimension of (homogeneous) linear systems. The approach is based on the specialization introduced in [@Roe01]. Roughly speaking, one proves that if there exists a curve with given multiplicities at $r$ general points, then by semicontinuity there must also exist curves with the same degree and (virtual) multiplicities at $r$ points which satisfy some well chosen proximity relations. The proximity inequalities impose then that the effective multiplicity of the specialized curve must grow, and one uses this bigger multiplicity as a bound for the degree of the curve. In order to give a brief explicit description of the specialization, let us recall the notations of [@Roe01] (see [@Cas00] for generalities on clusters and unloading, and [@Roe?4] for a general approach to specializations parameterized by varieties of clusters). We work on ${\mathbb{P}}^2$, and consider both proper and infinitely near points (which are those lying on a smooth surface that dominates ${\mathbb{P}}^2$ birrationally. A *cluster* is a set $K$ of points of ${\mathbb{P}}^2$ such that if $p\in K$ and $p$ is infinitely near to $q$ (i.e., $p$ lies on the exceptional divisor of $q$ after blowing up a sequence of points) then $q \in K$. Write $\pi_K:S_K\longrightarrow {\mathbb{P}}^2$ for the blowing up of all points in $K$. For each $i=2, \dots, r$, we denote $U_{i}$ the set of clusters $K=\{p_1, \dots, p_r \}$ such that - $p_2, \dots, p_i$ are proximate to $p_1$, - $p_j$ is proximate to $p_{j-1}$ for all $j=2, \dots, r$, and - there are no other proximity relations. In other words, denoting by $E_j$ the (total) exceptional divisor of blowing up $p_j$ this can be expressed by saying that the divisors $\tilde E_1= E_1-E_2-\dots-E_i$ and $\tilde E_j=E_j-E_{j+1}$, $j=2, \dots r-1$ on the surface $S_K$ are effective and irreducible. The sets $U_{i}$ are nonempty and have a natural structure of smooth irreducible locally closed subvarieties in a projective variety (the iterated blowing-up $X_{r-1}$ of Kleiman [@Kle81a]), and they satisfy $$\overline{U_{2}} \supset \overline{U_{3}}\supset \dots \supset \overline{U_{r}}.$$ The specialization we use works stepwise. Begin with a general cluster of $r$ distinct points, $K=\{p_1,p_2,\dots,p_r\}$, and a curve $C$ with multiplicities $(m)=(m_1, m_2, \dots, m_r)$ at these points, assuming $m_1 \ge m_2 \ge \dots \ge m_r$. Then specialize $K$ to a cluster $K_3$ general in $U_{3}$. If $m_1<m_2+m_3$, then the specialized curve $C_3$ cuts negatively the irreducible divisor $\tilde E_1=E_1-E_2-E_3$, so $\tilde E_1$ is a component of $\pi_K^*(C_3)-m_1E_1-\dots-m_rE_r$, and the effective multiplicity of $C_3$ at $p_1$ is bigger than $m_1$. Call $(m^{(3)})$ the system of multiplicities obtained after unloading multiplicities (i.e., substracting the $\tilde E_j$ which are cut negatively); $C_3$ goes through the cluster $K_3$ with multiplicities $(m^{(3)})$. Then specialize $K_3$ to a general $K_4 \in U_4$, and successively to a $K_5 \in U_5$, …, to a $K_r\in U_r$, performing unloadings whenever necessary. The first multiplicity of the last system $(m^{(r)})$ is a lower bound for the degree of $C$, and therefore $\alpha(m)\ge m_1^{(r)}$. This multiplicity is not hard to compute in each particular case; in [@Roe01] a bound was given that holds in general and is asymptotically sharp but that in many particular cases can be improved, especially when the multiplicities are relatively small. Now we are interested in the case that $m_1$ is much bigger than the other multiplicities, in which one can show that the inequality (\[eqnag\]) holds: \[nohomo\] Let $s$ be such that $(s-1)^2(s+2)^2 \ge 6.76(r-1)$, $r >9$, and assume $m_1 \ge m_2 \ge \dots \ge m_r$. If moreover $$m_1 \ge \sum_{i=2}^{s+1} m_i$$ then $\alpha(m) > \sum_{i=1}^r m_i/\smash{\sqrt{r}}$. Using notations as above, the preceding discussion shows that it suffices to prove $m_1^{(r)} > \sum_{i=1}^r m_i/\smash{\sqrt{r}}$. The hypothesis implies that the system of multiplicities $(m)$ is consistent for all clusters in $U_3$, …, $U_{s+1}$ (no unloading is needed for these) so $(m)=(m^{(3)})=\dots=(m^{(s+1)})$. Then, apply [@Roe01 lemma 3.5] as in the proof of [@Roe01 theorem 4.1] to obtain $$m_1^{(r)} \ge \sum_{i=1}^r m_i \, \left(1-\frac{1}{r}\right) \prod_{k = s+1}^{r-1} \left(1 - {k\over {{k^2} + r-1}} \right).$$ We have to see that this is bigger than $\sum m_i / \smash{\sqrt{r}}$. Because of [@Roe01 proposition 5.1], it will be enough to prove $$\label{untros} \prod_{k = 2}^{s} \left(1 - {k\over {{k^2} + r-1}} \right)^{-1} > \frac{\sqrt{r}}{\sqrt{r-1}-\frac{\pi}{8}}.$$ Write $x^2=r-1$. As $r >9$, we have $x \ge 3$. The term on the left in (\[untros\]) is $$\prod_{k = 2}^{s} \left(1 + \frac {k}{k(k+1) + x^2} \right)> 1+\sum_{k = 2}^{s} \frac {k}{k(k+1) + x^2}.$$ Let $s_0$ be the minimum integer such that $(s_0-1)^2(s_0+2)^2 \ge 6.76x^2$; the hypothesis on $s$ says that $s \ge s_0$, and it is clear that $s_0(s_0+1) \le x^2$. Using this we get $$1+\sum_{k = 2}^{s} \frac {k}{k(k+\!1) + x^2}> 1+\sum_{k = 2}^{s_0} \frac {k}{2 x^2} = 1+ \frac{(s_0-\!1)(s_0+\!2)}{4 x^2} \ge 1+ \frac{\sqrt{6.76}x}{4 x^2}= 1 + \frac{.65}{x}.$$ On the other hand, the term on the right in \[untros\] can be written as $$\frac{x}{x-\pi/8} \sqrt{1+\frac{1}{x^2}}\le \left(1+\frac{\pi}{8x-\pi}\right)\left(1+\frac{1}{2x^2}\right)$$ which for $x \ge 3$ is less or equal to $1+.65/x$, and the proof is complete. \[quasi\] Let $C$ be a submaximal curve with respect to general points $p_1, p_2 \dots, p_r \in {\mathbb{P}}^2$. Then for some reordering of the points, the system $(m)$ of multiplicities of $C$ at $p_1, p_2 \dots, p_r$ is $(m)=(\mu,\nu, \dots,\nu)$ with $\mu < s \nu $ for every integer $s$ with $(s-1)^2(s+2)^2\ge 6.76(r-1)$. The system $(m)$ of multiplicities of $C$ at $p_1, p_2 \dots, p_r$ is $(m)=(\mu,\nu, \dots,\nu)$ because of [@Sze01 corollary 4.6], so it is enough to prove that $\mu \ge s \nu $ for some integer $s$ with $(s-1)^2(s+2)^2\ge 6.76(r-1)$ leads to contradiction. But theorem \[nohomo\] shows that there are no submaximal curves when $\mu \ge s \nu $ for some integer $s$ with $(s-1)^2(s+2)^2\ge 6.76(r-1)$, so we are done. We finish by an example, considering the smallest values $r$ for which Nagata’s conjecture is unknown. Corollary \[quasi\] says that a submaximal curve with respect to $r$ general points, $10 \le r \le 15$ has system of multiplicities $(m)=(\mu,\nu, \dots,\nu)$ with $\mu < 3 \nu $.
{ "pile_set_name": "ArXiv" }
--- abstract: | We discuss the consequences of recent experimental results from CDF, SLC, LEP and elsewhere for the Standard Model and for new physics. A global fit to all indirect precision data yields $m_t = 175 \pm 11^{+17}_{-19}$ GeV, $\sin^2 \hat\theta_{\overline {MS}} = 0.2317 (3) (2)$, and $\alpha_s = 0.127 (5) (2)$, where the central values are for $M_H = 300$ GeV and the second uncertainties are for $M_H \rightarrow 1000$ GeV (+) and 60 GeV ($-$). The $m_t$ value is in remarkable agreement with the value $m_t=174 \pm 16$ GeV suggested by the CDF candidate events. There is a slight preference for a light Higgs with $M_H < 730$ (880) GeV at 95% c.l. if the CDF $m_t$ value is (not) included. The sensitivity is, however, due almost entirely to the anomalously large observed values for the $Z \rightarrow b\bar{b}$ width and left-right asymmetry. The value of $\alpha_s$ (from the lineshape) is clean theoretically assuming the Standard Model, but is sensitive to the presence of new physics contributions to the $Z \rightarrow b\bar{b}$ vertex. Allowing a vertex correction $\delta_{b\bar{b}}^{\rm new}$ one obtains the significantly lower value $\alpha_s = 0.111 \pm 0.009$, in better agreement with low energy determinations, and $\delta_{b\bar{b}}^{\rm new} = 0.023 \pm 0.011$. There is now enough data to perform more general fits to parameters describing new physics effects and to separate these from $m_t$ and $M_H$. Allowing the parameter $\rho_0$, which describes sources of $SU(2)$ breaking beyond the Standard Model, to be free one finds $\rho_0=1.0012 \pm 0.0017 \pm 0.0017$, remarkably close to unity. One can also separate the new physics contributions to the oblique parameters $S_{\rm new}$, $T_{\rm new}$ and $U_{\rm new}$, which all take values consistent with zero. The effects of supersymmetry on the determination of the Standard Model parameters are discussed. address: ' Department of Physics, University of Pennsylvania, Philadelphia, PA 19104, U.S.A. ' author: - Jens Erler and Paul Langacker title: | \ Implications of High Precision Experiments and the CDF\ Top Quark Candidates --- 2[ m\^2 ]{} 22t[ \^22]{} INTRODUCTION: ============= In April 1994 the CDF collaboration [@CDF] reported evidence for top quark production at the Tevatron. The collected events are consistent with a top quark (pole) mass $m_t = 174 \pm 16$ GeV. Electroweak radiative corrections to Standard Model (SM) observables have large $m_t$ effects, so that its direct determination is of upmost importance. Significant indirect bounds on the Higgs mass $M_H$ can only be obtained after $m_t$ is known independently. Top quark effects impede the setting of bounds on or the discovery of new physics beyond the Standard Model from precision observables. In this paper we study the implications of the $m_t$ range suggested by the CDF candidate events for high precision measurements. For comparison, we carry out analyses of the Standard Model parameters with and without the CDF constraint. Only one month prior to the announcement from Tevatron, the SLD collaboration [@SLD2] at SLAC published a precise measurement of the left-right asymmetry $A_{LR}$. The polarization of the SLC electron beam was increased from 22% to 63% and the number of $Z$ events increased by a factor of 5 compared to the 1992 run. At about the same time the LEP groups [@LEP] presented a first analysis of their 1993 data. The integrated luminosity in 1993 amounted to 40 ${\rm pb}^{-1}$, a number only slightly smaller than the integrated luminosities from the previous years combined. Thus, the new (and preliminary) data contribute with a high statistical weight. Moreover, systematic uncertainties were significantly reduced, most notably in the total $Z$ width, increasing the significance of the 1993 run even further. The experimental results are summarized in table \[tblexp\], together with the SM expectations using the global best fit values $m_t = 175 \pm 11$ GeV (for $M_H = 300$ GeV) and $\alpha_s = 0.127 \pm 0.005$ (see below). The three errors in the SM predictions correspond respectively to (1) the uncertainties in $M_Z$ and $\alpha (M_Z)$, (2) the (correlated) uncertainties from $m_t$ and $M_H$ (which can vary from 60 to 1000 GeV, with a central value of 300 GeV), and (3) the uncertainty in $\alpha_s$. $\sigma_{\rm had}^0$ is the bare hadronic peak cross section, i.e., the cross section at $\sqrt{s} = M_Z$ after correcting for photonic contributions. Similarly, $A_{FB}^{0f} = {3\over 4} A_e^0 A_f^0$ is the bare forward-backward asymmetry for $e^+ e^- \rightarrow Z \rightarrow f\bar{f}$, $A_{FB}^{0l}$ is the asymmetry for charged leptons assuming family universality (after correcting for $m_\tau$), $\bar{s}^2_e (Q_{FB})$ is the effective weak angle determined from the hadronic charge asymmetry, and $A_{LR}^0 = A_e^0$ is the bare left-right polarization asymmetry. The quantity $A^0_f$ for flavor $f$ is defined by[^1] A\_f\^0 = . $N_\nu$ is the number of active neutrinos (with $m_\nu \leq M_Z/2$); $Q_W$ is the effective weak charge in atomic parity violation [@atompar]; $g_{V,A}^{\nu e}$ are effective 4-Fermi couplings for neutrino electron scattering [@nueel]; and $s^2_W = 1-M_W^2/M_Z^2$ is the on-shell weak mixing angle from deep inelastic neutrino nucleon scattering [@nuenuc]. Our fits include the older low energy data [@olddata] as well. The agreement between theory and experiment is generally excellent, with the exceptions[^2] of $A_{LR}^0$ and $R_b$. Of course, given the plethora of precision measurements one expects deviations at some level, and it would be premature to take the deviating quantities described below as serious problems for the SM. However, these quantities, especially $R_b$, have a significant effect on some of the conclusions, so it is worthwhile to comment on them. The left-right asymmetry, $A_{LR}$, as measured at SLC, is the most precise single determination of the effective weak angle $\bar{s}_e^2$ [@SLD2]. SLD quotes $A_{LR} (\sqrt{s} = 91.26\; {\rm GeV}) = 0.1628 \pm 0.0076$. Correcting for photon exchange, electroweak interference, and initial-state radiation yields A\_[LR]{}\^0 = A\_e\^0 = 0.1656 0.0076. This corresponds to \[sineff\] |[s]{}\_e\^2 = [14 |Q\_e|]{} (1 - [|[g]{}\_[V\_e]{}|[g]{}\_[A\_e]{}]{}) = 0.2292 0.0010. Inclusion of the 1992 result [@SLD1], $A_{LR}^0 = 0.100 \pm 0.044$, yields combined values $A_{LR}^0 = 0.1637 \pm 0.0075$ and $\bar{s}_e^2 = 0.2294 \pm 0.0010$, which is about 2.5 $\sigma$ deviation from global SM fits and at least 2 $\sigma$ from the values derived from LEP asymmetries. The 1992 + 1993 result (combined with the LEP value of $M_Z$) yields $m_t = 251^{+24}_{-26}$ GeV for $M_H = 300$ GeV. Relaxing the universality assumption or even allowing for the most general fermion couplings to the $Z$ do not considerably improve the goodness of the fits [@Teupitz]. Also, using results from reference [@LLM] it was argued in [@Teupitz] that no kind of new physics can account for the SLD result without simultaneously conflicting one or several other observables, most notably the $W$ mass. Thus assuming that these experiments are completely governed by $\gamma$ and $Z$ amplitudes we look at a direct experimental conflict. One possible loophole, namely the presence of new effective interactions which contribute significantly to the lineshape and asymmetries, is discussed in reference [@4Fermi]. The $Z \rightarrow b\bar{b}$ vertex has long been advertized [@Jegerlehner] as the “ideal top mass meter” since it is virtually independent of $M_H$, and the terms quadratic in $m_t$ enter in a different way than in $\hat\rho \equiv M_W^2/M_Z^2 \cos^2\hat{\theta}_{\overline {MS}}$ which governs other electroweak observables. At the same time it is sensitive to many kinds of physics beyond the SM. The LEP groups obtain [@LEP] R\_b = = 0.2202 0.0020, from a fit with $R_c = \Gamma (c\bar{c})/\Gamma ({\rm had})$ left free[^3]. This is 2.3 standard deviations from the SM prediction $R_b = 0.2156 \pm 0.0004$. $R_b$ drives the fits to smaller values of $m_t$, independent of $M_H$. Due to the correlation of top and Higgs effects in the $\hat\rho$ parameter, this in turn favors smaller values of $M_H$. With the possible exception of these measurements, experiments and the minimal Standard Model are in spectacular agreement with each other. RADIATIVE CORRECTIONS: ====================== One of the goals of the $Z$ factories at SLAC and CERN is to test electroweak theory at the quantum level. $m_t$, $M_H$ and $\alpha_s$ enter only through radiative corrections. They are obscured by pure QED corrections, which are large but calculable and under control, with the possible exception of small-angle ${\cal O} (\alpha^2)$ Bhabba scattering. Experimenters usually present data with all QED corrections other than final state radiation removed. The hadronic contribution to the vacuum polarization [@Jegerlehner] induces an uncertainty of 0.0003 in $\sin^2 \theta_W$. Omitting this error in the global fits can change the extracted value of $m_t$ by about 3 GeV. QCD corrections are calculated [@QCD3] and included up to ${\cal O} (\alpha_s^3)$. We did not include ${\cal O} (\alpha_s^4)$ corrections, which are estimated to contribute with a negative sign and with about 0.4 MeV to the hadronic $Z$ width [@QCD4], corresponding to an additional uncertainty of 0.001 in $\alpha_s$. As pointed out in [@QCDT2] the ${\cal O} (\alpha_s^2)$ corrections to the vector and axial vector parts of the partial $Z$ width into $b$ quarks exhibit different dependences on $m_t$. They are important and included along with the analogous results for the ${\cal O} (\alpha_s^3)$ corrections [@QCDT3]. Higher order QCD corrections proportional to $m_b^2/M_Z^2$ are incorporated as well [@QCDB]. As for the electroweak sector, full one loop corrections are taken into account. Due to the heavy top quark, two loop effects of ${\cal O} (\alpha^2 m_t^4)$ are included with their full $M_H$ dependence [@BBCCV], as well as ${\cal O} (\alpha \alpha_s m_t^2)$ corrections to the $\rho$ parameter [@QCDEWR] and to the $Z \rightarrow b\bar{b}$ vertex [@QCDEWB]. Threshold effects corresponding to ${\cal O} (\alpha \alpha_s^2 m_t^2)$ corrections are incorporated by making use of the detailed work of Fanchiotti, Kniehl and Sirlin [@FKS]. They can also be estimated [@SV] by employing $\alpha_s (0.15 m_t)$ rather than $\alpha_s (m_t)$. The numerical difference between the two approaches is negligible, and either way threshold effects increase the extracted top mass by about 3 GeV. In practice we used the routine ZFITTER [@ZFITTER] for the calculation of form factors. The improved Born formulae were then dressed with the aforementioned QED and QCD corrections for the $Z$ partial widths. The agreement with ZFITTER version 4.6 is excellent with differences being at the 0.1 MeV level in the total $Z$ width. For the most important form factor, $\hat{k}_e (M_Z^2)$ (see eq. (\[sinsin\]) below), we used the update by Gambino and Sirlin [@GS]. On the quantum level the exact definition of the weak mixing angle becomes ambiguous. Besides the conceptually most simple on-shell definition, $s^2_W \equiv 1 - M_W^2/M_Z^2 = 0.2243 \pm 0.0012$, there are two other definitions[^4] which are numerically very close to each other. One is based on the coupling constants, $\tan \hat{\theta}_W (M_Z) \equiv g^\prime/g$, which are radiatively corrected according to the ${\overline {\rm MS}}$ prescription. This makes the ${\overline {\rm MS}}$ quantity $\sin^2 \hat{\theta}_W \equiv \hat{s}^2_Z = 0.2317 \pm 0.0004$ particularly convenient for GUT predictions and insensitive to new physics. However, it is a quantity designed by theorists and is not related simply to any single observable. Rather, it is best determined by a global fit. Also, there are variant forms of $\hat{s}^2_Z$ which differ in the treatment of heavy top quark effects. A variant [@DFS] in which the heavy top quark is not decoupled is a few times $10^{-4}$ larger [@GS] than the one introduced in [@MR]. In the latter, which is used here, the ln $m_t$ effects in $\gamma$ – $Z$ mixing are decoupled, so that the $Z$-pole asymmetries are essentially independent of $m_t$. The other is the effective mixing angle defined in eq. (\[sineff\]), with analogous definitions for other flavors. It is defined through observables (the $Z$-pole asymmetries), which makes it conceptually simple, but for the exact relation to other quantities a computer code is needed due to the need to compute three point functions. That also makes it difficult to relate $\bar{s}^2_e$ to non $Z$-pole observables. The two definitions above share a smaller sensitivity to $m_t$ compared to the on-shell $s^2_W$. For the relation between $s^2_W$ (or $M_W$), $\hat{s}^2_Z$, and $M_Z$ we rely on reference [@FKS][^5]. $\hat{s}^2_Z$ and $\bar{s}^2_e$ are related by \[sinsin\] |[s]{}\^2\_e = \^2\_Z [Re]{} \_e (M\_Z\^2), with the form factor ${\rm Re}\; \hat{k}_e (M_Z^2)$ from reference [@GS]. Relation (\[sinsin\]) is a very good approximation due to the smallness of ${\rm Im}\; \hat{k}_e (M_Z^2)$. For $m_t$ in the relevant range, (\[sinsin\]) implies |[s]{}\_e\^2 \~\^2\_Z + 0.00028. FIT RESULTS: ============ We regard the deviations in some of the observables as consistent with statistical fluctuations and have therefore refrained from using scale factors to increase error bars, and instead simply combined the data[^6]. Table \[tblres\] summarizes the results of various fits to $\hat{s}^2_Z$, $\alpha_s (M_Z)$, and $m_t$ based on different data sets. The central values[^7] correspond to $M_H = 300$ GeV and the second errors indicate the results for $M_H =1000$ GeV (+) and $M_H = 60$ GeV $(-)$. The increase in $\chi^2$ when changing $M_H$ from 60 to 1000 GeV, $\Delta \chi^2_H \equiv \chi^2 (1000) - \chi^2 (60)$, is also indicated. The first row is the fit to all indirect precision data[^8]. The prediction $m_t = 175 \pm 11^{+17}_{-19}$ GeV is in remarkable agreement with the CDF value $174 \pm 16$ GeV. Not surprisingly, including the CDF value as an additional constraint (second row) has little impact on the global fit within the Standard Model. It will, however, be of great importance in the non-Standard Model fits. The third row is a fit in which the indirect data is combined with the additional constraint $\alpha_s (M_Z) = 0.116 \pm 0.005$ obtained from data other than the $Z$ lineshape [@PDG]. As expected, the extracted $\alpha_s$ (which can be regarded as a simultaneous fit to the lineshape and other $\alpha_s$ data) is somewhat lower than the value from the lineshape alone. The other rows are fits to subsets of the data, which show the sensitivity to the various inputs. From the fourth row (LEP + low energy) we see that the predicted $m_t$ is 7 GeV lower without $A_{LR}$ from SLD, while when averaging $A_e^0$ from $A_{LR}^0$ and from ${\cal P}_\tau$ with a scale factor of 2.2 (fifth row) it is lower by 5 GeV. The results from the $Z$-pole (LEP + SLD), LEP, and SLD + $M_Z$ are also shown. The large value of $m_t$ in the last case reflects the high value of $A_{LR}^0$. It is useful to compare these results with the fits performed by the LEP electroweak working group [@LEP]. Their fits for $Z$-pole, $M_W$, and recent neutrino data (which corresponds roughly to our first “All indirect” fit), as well as to the LEP data are displayed in the last two rows of Table \[tblres\]. The agreement between their results and ours is excellent, with the small (correlated) differences in $\alpha_s$ and $m_t$ a reflection of the completely independent implementation of radiative corrections. There is a slight preference for a light Higgs (as is predicted in the minimal supersymmetric extension of the SM) but it is weak statistically. Combining all indirect data we can set an upper limit on $M_H < 570$ (880) GeV at the 90% (95%) c.l. Adding the CDF result this limit is strengthened to $M_H < 510$ (730) GeV. Moreover, the preference is driven mainly by the anomalous values of $R_b$ and $A_{LR}$. Removing them from the data set leads to an almost flat $\chi^2$ distribution with respect to $M_H$, as is shown in Figure \[Higgs\]. Hence, caution is called for in drawing any conclusion on $M_H$ from the present data. In the context of the SM, $R = \Gamma ({\rm had}) / \Gamma (l^+ l^-)$ is a theoretically clean measurement of $\alpha_s (M_Z)$. In the presence of new physics which increases the hadronic or $b\bar{b}$ event sample, however, $R$ loses its sensitivity to the strong coupling constant. Similar remarks hold for $\Gamma_Z$, which is also sensitive to $\alpha_s$. In 1993 the LEP groups collected data at the $Z$ peak and at $\pm 1.8$ GeV away from the peak. That allowed for a precise measurement of the $Z$ lineshape. The extracted $\Gamma_Z$ is about 1 standard deviation higher than in 1992 and the error decreased by almost 50%. It should be noted, however, that a lineshape scan involving only 3 scan points cannot by itself be sensitive to any non-$Z$ pole contribution to the cross section. An overconstrained lineshape fit is only possible when the lower statistics 1990/91 scan is included. The extracted value of $\alpha_s = 0.127 \pm 0.005 \pm 0.002$ is consistent with the LEP jet event shape analysis, which yields $\alpha_s = 0.123 \pm 0.006$ and with the value $0.122 \pm 0.005$ from the hadronic $\tau$ decay fraction [@PDG; @Bethke; @BC]. It is also in perfect agreement with grand desert SUSY-GUT expectations, favoring $\alpha_s = 0.127 \pm 0.002 \pm 0.008$, where the first error is due to $m_t$ and $M_H$ and the second arises from the lack of knowledge of the sparticle and GUT particle spectra (thresholds) and from the unknown effects of possible nonrenormalizable operators [@Nir]. It is, however, significantly higher than the values [@PDG; @Bethke; @BC] obtained from deep inelastic neutrino and lepton scattering ($0.112 \pm 0.005$), from $J/\Psi$ and $\Upsilon$ decays ($0.113 \pm 0.006$), and from determinations relying on lattice calculations of the charmonium ($0.110 \pm 0.006$) [@ElKhadra] and bottomonium ($0.115 \pm 0.002$) [@Davies] spectra. (The lower energy determinations must of course be extrapolated to $M_Z$.) As will be discussed in the next section, if one allows for the possibility of new physics in the $Zb\bar{b}$ vertex to account for $R_b$, the extracted value of $\alpha_s$ decreases to a lower value ($0.111 \pm 0.009 \pm 0.001$) consistent with these latter values. New Physics: ============ In the context of the Minimal Supersymmetric Standard Model (MSSM) and all its extensions under discussion, the lightest Higgs eigenstate is known to be light, in the range 60 GeV $ < M_H < $ 150 GeV. Taking as a central value $M_H = M_Z$ the extracted top mass is lowered to $m_t = 160^{+11 \; +6}_{-12 \; -5}$ GeV because of the strong $m_t$ – $M_H$ correlation[^9]. In most parts of the MSSM parameter space, i.e., whenever the sparticles and second Higgs doublet are much heavier than $M_Z$, the decoupling theorem applies and the only signs of supersymmetry in the precision observables are a light Higgs and the [*absence*]{} of deviations from the SM. On the contrary, in Extended Technicolor (ETC) and compositeness models we expect a variety of effects, most notably the observation of large Flavor Changing Neutral Currents (FCNC). As an example, in models with composite fermions, the effective 4-Fermi operators formed by constituent interchange have to be strongly suppressed. If we call the compositeness scale $\Lambda$, so that a 4-Fermi operator takes the form L= |[f\_1]{}f\_2 |[f\_3]{}f\_4, from FCNC we must require $\Lambda \geq {\cal O}(100\; {\rm TeV})$ unless a fine tuning is invoked. Even then, atomic parity violation experiments set lower limits [@leptoquarks] $\Lambda \geq {\cal O}(10\; {\rm TeV})$. They are expected to be increased to ${\cal O}(40\; {\rm TeV})$ with upcoming experiments. In ETC models, again contrary to observations, $R_b$ is usually expected to be below the SM value. Other predictions, at least of the simplest versions based on scaled-up QCD dynamics, are $S_{\rm new} > 0$ and $T_{\rm new} \neq 0$, neither of which are in conformity with the data (see below). It has become customary to use three quantities, e.g. $S$, $T$, and $U$ [@STU], to parametrize the flavor independent oblique radiative corrections, or a single parameter, $\rho_0$, to characterize new sources of $SU(2)$ breaking. In addition, new physics may in particular affect the $Z \rightarrow b\bar{b}$ vertex. In the past it was difficult to disentangle possible new physics effects from the dominant top mass contributions. With the new CDF result for $m_t$, however, it is now possible to clearly determine the new physics effect from the data. We introduce new variables $S_{\rm new}$, $T_{\rm new}$ and $U_{\rm new}$ via S = S\_[new]{} + S\_[m\_t]{} + S\_[M\_H]{},\ T = T\_[new]{} + T\_[m\_t]{} + T\_[M\_H]{},\ U = U\_[new]{} + U\_[m\_t]{}, where $S_{m_t}$ and $S_{M_H}$ are respectively the $m_t$ and $M_H$ contributions to $S$, and similarly for $T$ and $U$. The effects of $S$, $T$ and $U$ on the SM expressions for observables are given in [@PDG27]. We parametrize new physics entering the $Z \rightarrow b\bar{b}$ width by $\delta_{b\bar{b}}^{\rm new}$, defined by [@ABC] \[zbb\] \_[b|[b]{}]{}=\_[b|[b]{}]{}\^[SM]{}(1+\_[b|[b]{}]{}\^[new]{}). $\delta_{b\bar{b}}^{\rm SUSY}$ for the MSSM was computed in references [@Zbbsusy] and found to be positive or negative depending on the part of parameter space considered. If the experimental deviation of $\Gamma_{b\bar{b}}$ from the SM is to be explained by supersymmetry, then there must be one sparticle light enough to be detected soon. In typical ETC models and in particular in the explicit model by Appelquist and Terning [@AT] in which the ETC gauge bosons are weak singlets and no fine tuning occurs, $\delta_{b\bar{b}}^{\rm ETC}$ is negative and proportional to $m_t$ [@ZbbETC1]. In models in which the ETC gauge bosons are weak doublets, the sign in the corresponding contribution to $\delta_{b\bar{b}}^{\rm ETC}$ is reversed. However, there is a competing effect from weak gauge boson mixing which tends to cancel the former. Hence, a model independent statement about the sign and the size of $\delta_{b\bar{b}}^{\rm ETC}$ in this class of models is not possible [@ZbbETC2]. $\delta_{b\bar{b}}^{\rm new}$ may also be used to set limits on the admixture of extra particles such as an additional (SU(2) singlet) $D_L$ quark, since $b_L - D_L$ mixing reduces $\Gamma_{b\bar{b}}$ [@LLM]. In the fits $\delta_{b\bar{b}}^{\rm new}$ affects and is determined by $R_b$, $R$, $\Gamma_Z$ and $\sigma^0_{\rm had}$. One can also study the quantity \_0 \_0\^[tree]{} + \_0\^[loop]{} \_0\^[tree]{} + T\_[new]{}, which describes non-standard sources of (vector) SU(2) breaking. If $\rho_0 \neq 1$ one has to replace M\_Z && [1 ]{} M\_Z\^[SM]{},\ \_Z && \_0 \_Z\^[SM]{},\ [L]{}\_[NC]{} && \_0 [L]{}\_[NC]{}\^[SM]{}, where ${\cal L}_{NC}$ is a neutral current amplitude (effective Lagrangian). $\rho_0^{\rm tree}$ differs from unity in the presence of Higgs triplets or higher Higgs representations, \_0\^[tree]{} = , where $t_i$ and ${t_3}_i$ are the weak isospin and its third component of the neutral Higgs field $\phi_i$. $\rho_0^{\rm loop}$ gets a positive definite[^10] contribution in the presence of additional non-degenerate scalar or fermion doublets, \_0\^[loop]{} = 1 + [3 G\_F 8 \^2]{} \_i [C\_i 3]{} F(m\_[1i]{},m\_[2i]{}), where $C_i$ is the color factor and $F$ a function of the internal particle masses. In typical (level 1) superstring models and in grand desert SUSY-GUT models $\rho_0$ is close to 1, while $\rho_0 \neq 1$ in most compositeness models. Allowing $\rho_0 \neq 1$ is a special case of the $S_{\rm new}$, $T_{\rm new}$, $U_{\rm new}$ parametrization, corresponding to $S_{\rm new} = U_{\rm new} =0$, and $\rho_0 = 1 + \alpha T_{\rm new}$. (Higher dimensional Higgs representations are technically not included in the standard definition of $T_{\rm new}$. In practice, however, they cannot be distinguished from oblique contributions from the precision observables alone, so we will include both in our definition of $T_{\rm new}$.) With the CDF result we can now simultaneously determine $\hat{s}^2_Z$, $\alpha_s$ and $m_t$ as well as a variety of parameters describing physics beyond the SM. Table \[NPfits\] shows the results of various fits allowing for different parameters left free. In these fits $m_t$ comes mainly from the direct CDF result and $\hat{s}^2_Z$ from the asymmetries, and since (given the value of $M_Z$) they are consistent with each other in the SM, they are largely insensitive to the presence of the new physics parameters. With them $T_{\rm new}$ can be extracted from $\Gamma_Z$, $S_{\rm new}$ is determined by $M_Z$ and $U_{\rm new}$ from $M_W$. When $\delta_{b\bar{b}}^{\rm new}$ is left free the large observed value of $R_b$ drives it significantly to positive values and contributes this way to the hadronic partial $Z$ width. Although this increase in $\Gamma ({\rm had})$ is only at the per mille level this has a large effect on the extracted value[^11] of $\alpha_s$ which is mainly determined by $R$: it is driven to lower values, though with a larger error. The reason is that this determination is only a loop effect. The other observables, however, are only weakly correlated. Note, that before the announcement of the CDF top quark candidates the oblique parameters $S$, $T$ and $U$ could only be discussed relative to some arbitrary reference value of $m_t$. In particular, it was difficult to separate the effects of a heavy top on $T$ (or equivalently on $\rho_0$) from those of new physics. Including the CDF top mass range such a separation became feasible and from table \[NPfits\] we see that $\rho_0$ is remarkably close to unity, leaving little room for any new physics which contributes to it. The allowed regions in $\rho_0$ vs. $\hat{s}^2_Z$ from various observables and the global fit are shown in Figure \[rho\]. Similarly, $S_{\rm new}$, $T_{\rm new}$ and $U_{\rm new}$ are well-constrained and consistent with zero[^12]. The global fit yields a negative central value for $S_{\rm new}$, but consistent with 0 at 1 $\sigma$. This is in contrast to $S < 0$, as was suggested by earlier data. ($A_{LR}^0$ by itself does favor $S < 0$.) The allowed region in $S_{\rm new}$ and $T_{\rm new}$ is shown in Figure \[ST\]. So far, we have allowed for new physics in the $Zb\bar{b}$ vertex only by an overall factor in (\[zbb\]). This implicitly assumes that the relative contributions to the vector and axial vertices are such that there is little effect on $A_{FB}^{0b}$. Indeed, $A_{FB}^{0b}$ can be seen in Table \[tblexp\] to be in good agreement with the SM expectation. However, one can do a more detailed analysis [@SZ] by allowing separate corrections to the left- and right handed couplings, i.e., the (lowest order) couplings are replaced by g\_[Lb]{} = [12]{} (g\_[Vb]{} + g\_[Ab]{}) - [12]{} + [13]{} \^2 \_W + \^b\_L,\ g\_[Rb]{} = [12]{} (g\_[Vb]{} - g\_[Ab]{}) \^2 \_W + \^b\_R. From a global fit we obtain \^b\_L = 0.0003 0.0047,\ \^b\_R = 0.026 0.018 (with a correlation of 0.86). That is, $A_{FB}^{0b}$, which is consistent with but slightly lower than the Standard Model prediction, forces the $R_b$ anomaly to be in the right-chiral coupling. SUMMARY: ======== The indirect determination of $m_t = 175 \pm 11^{+17}_{-19}$ GeV (for $M_H = 300^{+700}_{-240}$ GeV) is in spectacular agreement with the CDF range, $m_t = 174 \pm 16$ GeV, while the somewhat lower value $160^{+11\; +6}_{-12\; -5}$ expected in supersymmetry is still in reasonable agreement. Also most other observables are in excellent agreement with the predictions of the Standard Model. One exception is that the left-right asymmetry is in direct conflict with LEP asymmetries. $\alpha_s$ determinations from LEP (0.127(5)(2) from the lineshape, 0.123(6) from jets, and 0.122(5) from $R_\tau$) are significantly higher than the ones performed at lower energies. The $Z \rightarrow b\bar{b}$ partial width exceeds the Standard Model value by about 2.3 standard deviations. Interestingly, new physics which can account for $\delta_{b\bar{b}}^{\rm new}$ would simultaneously decrease the extracted $\alpha_s$ from $R$, bringing it in closer agreement with other measurements. It remains to be seen to what extend these deviations and the one in $A_{FB}^{0\tau}$ persist in the future. Inclusion of the CDF result does not alter the Standard Model fits significantly. It is, however, very useful in constraining new types of physics: one can now separate the effects of new physics from $m_t$. E.g, $\rho_0$, which describes sources of vector $SU(2)$ breaking beyond the SM, is now known to be very close to and consistent with the SM value (of unity), $\rho_0 = 1.0012 \pm 0.0017 \pm 0.0017$. The same is true for all the oblique parameters. High precision experiments continue to prefer non-positive values for $S_{\rm new}$ and a vanishing $T_{\rm new}$, but there is no longer a significant indication of $S_{\rm new} < 0$. This is in contrast to standard ETC/compositeness models, but is consistent with most of the parameter space of minimal supersymmetry. The new data shows a slight preference for a light Higgs mass close to the direct lower bound, but this is weak statistically. One finds $M_H < 510$ (730) GeV at 90% (95%) c.l. It should be kept in mind, however, that this limit depends almost entirely on $R_b$ and $A_{LR}^0$, both of which are high compared to the SM. Acknowledgements {#acknowledgements .unnumbered} ================ We are happy to thank Wolfgang Hollik, Bernd Kniehl and Alberto Sirlin for discussions. This work was supported by the Texas National Laboratory Research Commission and by the D.O.E. under contract DE-AC02-76-ERO-3071. One of us (J. E.) is supported in part by the Deutsche Forschungsgemeinschaft. The CDF Collaboration (F. Abe et al.), Phys. Rev. Lett. 73 (1994) 225. The SLD Collaboration (K. Abe et al.), Phys. Rev. Lett. 73 (1994) 25. For updated results, see D. Schaile, [*Precision Tests of the Electroweak Interaction*]{}, plenary talk (Pl-2) presented at the 27th International Conference on High Energy Physics, Glasgow, Scotland, July 1994. M. C. Noecker, B. P. Masterson and C. E. Wieman, Phys. Rev. Lett. 61 (1988) 310. The CHARM-II Collaboration (P. Vilain et al.), Phys. Lett. B281 (1992) 159. The CCFR Collaboration (C. Arroyo et al.), Phys. Rev. Lett. 72 (1994) 3452. Review of Particle Properties (M. Aguilar-Benitez et al.), Phys. Rev. D50 (1994) 1173;\ P. Langacker and M. Luo, Phys. Rev. D44 (1991) 817;\ U. Amaldi et al., Phys. Rev. D36 (1987) 1385. J. Erler, [*The SLD Asymmetry in View of the LEP Results*]{}, Univ. of Pennsylvania preprint UPR-0619T, to appear in the Proceedings of the Zeuthen Workshop on Elementary Particle Theory: Physics at LEP 200 and Beyond, Teupitz, Germany, April 1994. The SLD Collaboration (K. Abe et al.), Phys. Rev. Lett. 70 (1993) 2515. P. Langacker, M. Luo and A. K. Mann, Rev. Mod. Phys. 64 (1992) 87. J. Erler, [*Four-Fermi Operators in $e^+$ $e^-$ Annihilation Experiments and Uncertainties in $Z$ Boson Properties*]{}, Univ. of Pennsylvania preprint UPR-0633T (1994). F. Jegerlehner, [*Renormalizing the Standard Model*]{}, lectures given at the Theoretical Advanced Study Institute in Elementary Particle Physics, (TASI), Boulder, Colorado, April 1990, published in [*Testing the Standard Model*]{}, M. Cveti$\check{\rm c}$ and P. Langacker (eds.), World Scientific, Singapore, 1991. S. G. Gorishnii, A. L. Kataev and S. A. Larin, JETP Lett. 53 (1991) 127 and Pisma Zh. Eksp. Teor. Fiz. 53 (1991) 121. A. L. Kataev and V. V. Starshenko, [*Estimates of the ${\cal O} (\alpha_s^4)$ Corrections to $\sigma_{tot} (e^+e^- \rightarrow hadrons)$, $\Gamma (\tau \rightarrow \nu_\tau + hadrons)$ and Deep Inelastic Scattering Sum Rules*]{}, preprint CERN-TH.7198/94. B. A. Kniehl and J. H. Kühn, Phys. Lett. B224 (1989) 229 and Nucl. Phys. B329 (1990) 547. K. G. Chetyrkin and O. V. Tarasov, Phys. Lett. B327 (1994) 114. K. G. Chetyrkin and J. H. Kühn, Phys. Lett. B248 (1990) 359;\ K. G. Chetyrkin, J. H. Kühn and A. Kwiatkowski, Phys. Lett. B282 (1992) 221;\ K. G. Chetyrkin and A. Kwiatkowski, Phys. Lett. B305 (1993) 285. R. Barbieri, M. Beccaria, P. Ciafaloni, G. Curci and A. Vicere, Nucl. Phys. B409 (1993) 105. A. Djouadi and C. Verzegnassi, Phys. Lett. B195 (1987) 265;\ A. Djouadi, Nuovo Cim. 100A (1988) 357. J. Fleischer, F. Jegerlehner, P. Raczka and O. V. Tarasov, Phys. Lett. B293 (1992) 437;\ K. G. Chetyrkin, A. Kwiatkowski and M. Steinhauser, Mod. Phys. Lett. A8 (1993) 2785. S. Fanchiotti, B. Kniehl and A. Sirlin, Phys. Rev. D48 (1993) 307. B. H. Smith and M. B. Voloshin, [*On Normalization of QCD Effects in $O(m_t^2)$ Electroweak Corrections*]{}, Univ. of Minnesota preprint UMN-TH-1241/94. D. Bardin et al., [*ZFITTER — an Analytical Program for Fermion Pair Production in $e^+$ $e^-$ Annihilation*]{}, preprint CERN-TH.6443/92. P. Gambino and A. Sirlin, Phys. Rev. D49 (1994) 1160. V. A. Novikov, L. B. Okun and M. I. Vysotsky, Nucl. Phys. B397 (1993) 35. P. Langacker and J. Erler, in the [*Review of Particle Properties*]{}, Phys. Rev. D50 (1994) 1304;\ P. Langacker, [*Theoretical Study of the Electroweak Interaction: Present and Future*]{}, Univ. of Pennsylvania preprint UPR-0624T, talk presented at the 22nd INS International Symposium on Physics with High Energy Colliders, Tokyo, Japan, March 1994. G. Degrassi, S. Fanchiotti and A. Sirlin, Nucl. Phys. B351 (1991) 49. W. J. Marciano and J. L. Rosner, Phys. Rev. Lett. 65 (1990) 2963 and Erratum ibid. 68 (1992) 898. See the first of references [@olddata]. S. Bethke, [*Experimental Results on QCD and Jets at LEP and SLC*]{}, talk given at the Tennessee International Symposium on Radiative Corrections: Status and Outlook, Gatlinburg, Tennessee, June 1994. S. Bethke and S. Catani, [*A Summary of $\alpha_s$ Measurements*]{}, preprint CERN-TH.6484/92, summary presented at the Round Table Discussion of the XXVII$^{th}$ Rencontre de Moriond, “QCD and High Energy Hadronic Interactions”, Les Arcs, France, March 1992. P. Langacker and N. Polonsky, Phys. Rev. D47 (1993) 4028;\ N. Polonsky, [*Precision Data Parameters and Grand Unification Predictions*]{}, Univ. of Pennsylvania preprint UPR-0588T, talk presented at the 16th International Warsaw Meeting on Elementary Particle Physics: New Physics at Experiments, Kazimierz, Poland, May 1993. A. X. El-Khadra, G. Hockney, A. S. Kronfeld and P. B. Mackenzie, Phys. Rev. Lett. 69 (1992) 729;\ A. X. El-Khadra, [*The Strong Coupling From Quarkonia*]{}, Ohio State Univ. preprint OHSTPY-HEP-T-93-020, talk given at LATTICE 93: 11th International Symposium on Lattice Field Theory, Dallas, Texas, October 1993. C. T. H. Davies, K. Hornbostel, G. P. Lepage, A. Lidsey, J. Shigemitsu and J. Sloan, [*A Precise Determination of $\alpha_s$ from Lattice QCD*]{}, Ohio State Univ. preprint OHSTPY-HEP-T-94-013. P. Langacker, Phys. Lett. B256 (1991) 277;\ M. Leurer, Phys. Rev. D49 (1994) 333 and D50 (1994) 536. M. E. Peskin and T. Takeuchi, Phys. Rev. Lett. 65 (1990) 964 and Phys. Rev. D46 (1992) 381;\ M. Golden and L. Randall, Nucl. Phys. B361 (1991) 3. P. Langacker and J. Erler, in the [*Review of Particle Properties*]{}, Phys. Rev. D50 (1994) 1312, and references therein. G. Altarelli, R. Barbieri and F. Caravaglios, Nucl. Phys. B405 (1993) 3. A. Djouadi, G. Girardi, W. Hollik, F. M. Renard and C. Verzegnassi, Nucl. Phys. B349 (1991) 48;\ M. Boulware and D. Finnell, Phys. Rev. D44 (1991) 2054. T. Appelquist and J. Terning, [*An Extended Technicolor Model*]{}, Yale Univ. preprint YCTP-P21-93. R. S. Chivukula, B. Selipsky and E. H. Simmons, Phys. Rev. Lett. 69 (1992) 575;\ R. S. Chivukula, E. Gates, E. H. Simmons and J. Terning, Phys. Lett. B311 (1993) 157;\ N. Kitazawa, Phys. Lett. B313 (1993) 395. R. S. Chivukula, E. H. Simmons and J. Terning, Phys. Lett. B331 (1994) 383. D. Schaile and P. Zerwas, Phys. Rev. D45 (1992) 3262. [ccc]{} Quantity & Value & Standard Model\ $M_Z$ \[GeV\] & $91.1888 \pm 0.0044$ & input\ $\Gamma_Z$ \[GeV\] & $2.4974 \pm 0.0038$ & $2.497 \pm 0.001 \pm 0.003 \pm [0.002] $\ $R = \Gamma({\rm had})/\Gamma(l^+ l^-)$ & $20.795 \pm 0.040$ & $20.784 \pm 0.006 \pm 0.003 \pm [0.03]$\ $\sigma_{\rm had}^0 = \frac{12 \pi}{M_Z^2} \; \frac{\Gamma(e^+ e^-) \Gamma({\rm had})}{\Gamma_Z^2}$ \[nb\] & $41.49 \pm 0.12$ & $41.44 \pm 0.004 \pm 0.01 \pm [0.02]$\ $R_b = \Gamma(b \bar{b})/ \Gamma({\rm had})$ &$0.2202 \pm 0.0020$ & $0.2156 \pm 0 \pm 0.0004$\ $R_c = \Gamma(c\bar{c}) /\Gamma({\rm had})$ & $0.1583 \pm 0.0098$ & $0.171 \pm 0 \pm 0$\ $A_{FB}^{0l} = \frac{3}{4} \left( A_l^0 \right)^2$ & $0.0170 \pm 0.0016$ & $0.0151 \pm 0.0005 \pm 0.0006$\ $A_{\tau}^0 \left(P_\tau \right)$ & $0.143 \pm 0.010$ & $0.142 \pm 0.003 \pm 0.003$\ $A_e^0 \left( P_\tau\right)$ & $0.135 \pm 0.011$ & $0.142 \pm 0.003 \pm 0.003$\ $A_{FB}^{0b} = \frac{3}{4} A^0_e A^0_b$ & $0.0967 \pm 0.0038$ & $0.0994 \pm 0.002 \pm 0.002$\ $A_{FB}^{0c} = \frac{3}{4} A^0_e A^0_c$ & $0.0760 \pm 0.0091$ & $0.071 \pm 0.001 \pm 0.001$\ $\bar{s}_e^2$   ($Q_{FB}$) & $0.2320 \pm 0.0016$ & $0.2322 \pm 0.0003 \pm 0.0004$\ $A_e^0 = A^0_{LR}$   (SLD) & $0.1637 \pm 0.0075 \;\; (92 + 93)$ & $0.142 \pm 0.003 \pm 0.003$\ & $(0.1656 \pm 0.0076 \;\; (93))$ &  \ $N_\nu$ & $2.988 \pm 0.023$ & $3$\ \ \ \ \ Quantity & Value & Standard Model\ $M_W$ \[GeV\] & $80.17 \pm 0.18$ & $80.31 \pm 0.02 \pm 0.07$\ $M_W/M_Z$ (UA2) & $0.8813 \pm 0.0041$ & $0.8807 \pm 0.0002 \pm 0.0007$\ $Q_W$ (Cs) & $-71.04 \pm 1.58 \pm [0.88]$ & $-72.93 \pm 0.07 \pm 0.04$\ $g_A^{\nu e}$ (CHARM II) & $-0.503 \pm 0.017$ & $-0.506 \pm 0 \pm 0.001$\ $g_V^{\nu e}$ (CHARM II) & $-0.035 \pm 0.017$ & $-0.037 \pm 0.001 \pm 0$\ $s^2_W = 1 - \frac{M_W^2}{M_Z^2}$ & $0.2218 \pm 0.0059$ (CCFR) & $0.2245 \pm 0.0003 \pm 0.0013$\ & $0.2260 \pm 0.0048$ (All) &\ $M_H$ \[GeV\] & $> 61$ (LEP) & $< \left\{ \begin{array}{l} O (600) \; {\rm (theory)} \\ O (800) \; {\rm (indirect)} \end{array} \right.$\ $m_t$ \[GeV\] & $> 131$ (D0) & $175 \pm 11^{+17}_{-19}$ (indirect)\ & $174 \pm 16$ (CDF) &\ $\alpha_s (M_Z)$ & $0.123 \pm 0.006$ (event shapes) & $0.127 \pm 0.005 \pm 0.002 \pm [0.001]$\ & $0.116 \pm 0.005$ (event shapes + low energy) & ($Z$ lineshape)\ \[tblexp\] ---------------------------------------------------------------------------------------------------------------------------------- Set $\hat{s}^2_Z$ $\alpha_s (M_Z)$ $m_t$ \[GeV\] $\Delta \chi^2_H$ ------------------------------------------- --------------------- ------------------ --------------------------------- ----------- All indirect $0.2317(3)(^1_2)$ 0.127(5)(2) $175 \pm 11^{+17}_{-19}$ 4.4 Indirect + CDF ($174 \pm 16$) $0.2317(3) (^2_3)$ 0.127(5)(2) $175 \pm 9^{+12}_{-13}$ 4.4 Indirect + $\alpha_s$ ($0.116 \pm 0.005$) $0.2316(3)(^1_2)$ 0.122(3)(1) $178^{+10 \; +17}_{-11 \; -19}$ 6.0 LEP + low energy 0.2320(3)(2) 0.128(5)(2) $168^{+11 \; +17}_{-12 \; -19}$ 2.7 All indirect ($S=2.2$) $0.2319 (3) (^1_2)$ 0.128(5)(2) $170^{+11 \; +17}_{-12 \; -19}$ 3.3 $Z$-pole 0.2316(3)(1) 0.126(5)(2) $179^{+11 \; +17}_{-12 \; -19}$ 4.2 LEP $0.2320 (4) (^1_2)$ 0.128(5)(2) $170^{+12 \; +18}_{-13 \; -20}$ 2.6 SLD + $M_Z$ 0.2291(10)(0) — $251^{+24 \; +21}_{-26 \; -23}$ — $Z$-pole, $M_W$, recent $\nu$ (LEP-EWG) $0.2317(3) (^0_2)$ 0.125(5)(2) $178 \pm 11^{+18}_{-19}$ LEP (LEP-EWG) $0.2319(4) (^1_2)$ 0.126(5)(2) $173^{+12 \; +18}_{-13 \; -20}$ ---------------------------------------------------------------------------------------------------------------------------------- : Results for the electroweak parameters in the Standard Model from various sets of data. The central values assume $M_H = 300$ GeV, while the second errors are for $M_H \rightarrow 1000$ GeV (+) and 60 GeV ($-$). The last column is the increase in the overall $\chi^2$ to the fit as the Higgs mass increases from 60 to 1000 GeV. The last two rows are the results of fits performed by the LEP electroweak working group (LEP-EWG), with the appropriate translation of $\bar{s}^2_e$ into $\hat{s}^2_Z$. \[tblres\] --------------- ------------------ ------------------------- ---------------------------------- ---------------------------------- --------------- ------------------------------- $\hat{s}^2_Z$ $\alpha_s (M_Z)$ $m_t$ \[GeV\] $S_{\rm new}$ $T_{\rm new}$ $U_{\rm new}$ $\delta_{b\bar{b}}^{\rm new}$ ($\rho_0$) 0.2317(3)(3) 0.127(5)(2) $175 \pm 9^{+12}_{-13}$ — — — — 0.2316(3)(2) 0.111(9)(0) $177 \pm 9 \pm 13$ — — — $0.023 \pm 0.011 \pm 0.003$ 0.2316(3)(1) 0.125(6)(1) $166 \pm 15 \pm 0$ — $0.16 \pm 0.23 \pm 0.23$ — — (1.0012(17)(17)) 0.2316(3)(2) 0.111(9)(0) $174 \pm 16^{+1}_{-0}$ — $0.05 \pm 0.25 \pm 0.25$ — $0.022 \pm 0.011 \pm 0$ (1.0004(18)(18)) 0.2314(4)(1) 0.125(6)(0) $167 \pm 15 \pm 0$ $-0.21 \pm 0.24^{-0.08}_{+0.17}$ $0.03 \pm 0.30^{+0.17}_{-0.10}$ $-0.50 \pm — 0.61$ ($1.0002(22)(^{12}_{\; 7})$) 0.2313(4)(1) 0.112(9)(0) $175 \pm 16 \pm 0$ $-0.21 \pm 0.24^{-0.08}_{+0.17}$ $-0.09 \pm 0.32^{+0.16}_{-0.11}$ $-0.53 $0.022 \pm 0.011 \pm 0$ \pm 0.61$ ($0.9993(23)(^{12}_{\; 8})$) --------------- ------------------ ------------------------- ---------------------------------- ---------------------------------- --------------- ------------------------------- : Results for the electroweak parameters including additional fit parameters describing physics beyond the SM. All fits include the CDF constraint $m_t = 174 \pm 16$ GeV. The central values are for $M_H = 300$ GeV, the upper second errors for $M_H = 1000$ GeV and the lower ones for $M_H = 60$ GeV. For $T_{\rm new}$ we also list the equivalent $\rho_0 \equiv 1 + \alpha T_{\rm new}$. \[NPfits\] [^1]: $\bar{s}_f^2$ and $\bar{g}_{V,A_f}$ are flavor dependent effective mixing angles and vector (axial-vector) couplings. They include propagator and vertex corrections evaluated at $s=M_Z^2$. [^2]: Also, the individual forward-backward $\tau$ asymmetry, $A_{FB}^{0\tau} = 0.0228 \pm 0.0028$ is 2.8 $\sigma$ above SM expectations and renders the test of lepton universality only moderately successful. If the FB asymmetries into the 3 lepton species are used, one finds that universality is excluded at the 93% c.l. [@Teupitz]. There is also a direct discrepancy between $A_{FB}^{0\tau}$ and $A_\tau^0({\cal P}_{\tau})$ (which is consistent with the SM) of about 2.5 $\sigma$. [^3]: We will always use the experimental values of $R_b$ and $R_c$ with their correlation of $-0.4$ [@LEP]. Alternatively, one could use the value $R_b = 0.2192 \pm 0.0018$, obtained [@LEP] by fixing $R_c$ to its SM value of 0.171. We have checked that the two methods yield virtually identical results. [^4]: Yet another definition, $s_{M_Z}^2 = 0.2312 \pm 0.0003$, is obtained by removing the $m_t$ dependence from the expression for $M_Z$ [@NOV]. The $m_t$ uncertainty reenters when other observables are expressed in terms of $s_{M_Z}^2$. The various definitions are further discussed in [@PDG26]. [^5]: We are indebted to Bernd Kniehl, who made his computer code on which the numerical results of ref. [@FKS] are based available to us. [^6]: A scale factor of 2.2 for the uncertainties in $A_e$ from $A_{LR}$ and ${\cal P}_\tau$, as suggested by the Particle Data Group [@PDG], would decrease the value of $m_t$ predicted by the indirect data by 5 GeV. [^7]: The predictions in Table \[tblexp\], especially for $\bar{s}^2_e (Q_{FB})$, differ slightly from the values at the best fit point, because the former use the central value of $\alpha^{-1} (M_Z) = 127.9 \pm 0.1$ [@FKS], incorporating the $\pm 0.1$ in the first listed uncertainty, while the best fit occurs at $\alpha^{-1}(M_Z) = 128.0$. [^8]: The overall $\chi^2$ is 181 for 206 d.o.f. This is rather low (mainly due to the earlier neutral current data), but statistically acceptable: the probability of $\chi^2 \leq 181$ is 10%. The correlation coefficients are $\rho_{\hat{s}^2_Z \alpha_s} = 0.30$, $\rho_{\hat{s}^2_Z m_t} = -0.67$, $\rho_{\alpha_s m_t} = -0.20$. The correlations for the other data sets are similar. [^9]: The other parameters are $\hat{s}_Z^2 = 0.2316(3)(1)$ and $\alpha_s = 0.126(5)(1)$. [^10]: Non-degenerate multiplets involving Majorana fermions or scalars with non-zero vacuum expectation values can give contributions of either sign. [^11]: The same effect on $\alpha_s$ would be obtained if one used the measured value of $R_b$ rather than the Standard Model formula in the expression for $\Gamma ({\rm had})$. [^12]: Note, that the oblique parameters are defined with a factor $\alpha$ factored out so that they are expected to be of order unity if non-zero.
{ "pile_set_name": "ArXiv" }
--- abstract: '**Background:** Direct measurement of the electric dipole moment (EDM) of the neutron lies in the future; measurement of a nuclear EDM may well come first. The deuteron is one nucleus for which exact model calculations are feasible. **Purpose:** We explore the model dependence of deuteron EDM calculations. **Methods:** Using a separable potential formulation of the Hamiltonian, we examine the sensitivity of the deuteron EDM to variation in the nucleon-nucleon interaction. We write the EDM as the sum of two terms, the first depending on the target wave function with plane-wave intermediate states, and the second depending on intermediate multiple scattering in the $^3$P$_1$ channel, the latter being sensitive to the off-shell behavior of the $^3$P$_1$ amplitude. **Results:** We compare the full calculation with the plane-wave approximation result, examine the tensor force contribution to the model results, and explore the effect of short range repulsion found in realistic, contemporary potential models of the deuteron. **Conclusions:** Because one-pion exchange dominates the EDM calculation, separable potential model calculations will provide an adequate description of the $^2$H EDM until such time as a better than 10% measurement is obtained.' author: - 'Iraj R. Afnan' - 'Benjamin F. Gibson' title: Model Dependence of the $^2$H Electric Dipole Moment --- Introduction ============ With the discovery of parity ($P$) violation, which was suggested by Lee and Yang [@LY56], Landau [@La57] deduced that charge conjugation and parity ($CP$) invariance implies that the electric dipole moment (EDM) of particles, *e.g.* the neutron, should be zero. If the $CPT$ theorem is valid, which is the case for gauge theories, then any $CP$ violation would also imply a corresponding time reversal ($T$) invariance violation. Predating the discovery of parity violation in the weak interaction, Purcell and Ramsey [@PR50] had pointed out that there was lacking any experimental test of parity conservation in the strong interaction. With their student Smith [@SPR57] they set limits on the EDM of the neutron of the order of $d_n < 5 \times 10^{-20}\, e \, cm$. The Standard Model of fundamental interactions predicts values for EDMs (due to second order W boson exchange) which are significantly smaller than contemporary experiments can detect, of the order of $10^{-31}\, e \, cm$. Therefore, an unambiguous observation of a nonzero EDM at current capabilities would imply a yet to be discovered source of $CP$ violation [@He95; @KL97]. The new physics could arise in the strong interaction sector (*e.g.*, the $\theta$ term), or in the weak interaction sector \[*e.g.*, Super Symmetric models or Left/Right (boson mass) symmetry breaking\]. Current limits on the nucleon EDM are of the order of $10^{-26}\, e \, cm$. Even were one to establish a nonzero neutron and proton EDM, those two results would at best determine the isoscalar and isotensor components but would not isolate any isovector component. Thus, one would need a third measurement, such as the deuteron EDM, to fully elucidate the isospin nature of the EDM operator. Both $PT$ violating and $P$ conserving, $T$ violating potentials may give rise to an EDM [@He95], but one-pion exchange contributes only to the former. We concentrate here upon the effects due to $PT$ invariance violation in the nuclear potential. The deuteron is attractive as the focus of an EDM investigation, both theoretically and experimentally, because a method has been proposed to directly measure the EDM of charged ions in a storage ring [@Kh98; @Fa04; @Se04; @Or06]. A permanent EDM can arise because a $PT$ violating interaction can induce a small P-state admixture in the deuteron wave function, one which produces a non vanishing matrix element of the charge dipole operator $\tau^z_{-} e\vec{r}$. Although this two-body EDM contribution must be disentangled from the one-body contributions of the neutron and proton, the neutron and proton EDMs tend to cancel in the case of the isospin zero $^2$H. (If the nucleon EDM were a pure isoscalar as is the case in the $\theta$ model, then this cancellation would be exact.) Therefore, the $PT$ violating nucleon-nucleon (*NN*) interaction can contribute significantly to the deuteron EDM. Because the deuteron is reasonably understood and has been accurately modeled, reliable calculations are possible. Our purpose is to address the sensitivity of the deuteron EDM to the nuclear physics in the modeling of the nucleon-nucleon interaction. Beyond understanding the model dependence of the $^2$H EDM, our goal is to determine an appropriate model approximation with which one might reliably calculate the nuclear physics contribution to the $^3$He and $^3$H EDMs. Therefore, we examine the uncertainties in the deuteron EDM calculation arising from the short range repulsion in the ground state wave function, the dependence on the size of the deuteron D-state, and the properties of the $^3$P$_1$ continuum in intermediate states. For the purpose of completeness and to place our work in context, we note that Avishai [@Av85] first estimated the two-body deuteron EDM \[ see Eq. (\[eq:2\]) \] $d_D^{(2)}$ using a separable potential model due to Mongan [@Mo69]. He reported a value of $- 0.91\ A\,e \,fm$ when he utilized the physical pion mass for the exchanged meson. \[Note: To exclude the $PT$ violating and strong coupling constants in the one pion exchange nucleon-nucleon interaction for the quoted values of the EDM, we have introduced $A=\bar{g}^{(1)}_{\pi NN}\,g_{\pi NN}/(16\pi)$.\] However, there is an ambiguity in Avishai’s results, in that he states his final result in terms of $A/2$. Because the particular separable potentials used by Avishai were not specified, we were unable to fully confirm his reported numbers. Khriplovich and Korkin [@KK00] later estimated $d_D^{(2)}$ using a zero-range approximation in the chiral limit ($m_\pi \rightarrow 0$) and obtained a value of $- 0.92\ A\, e\, fm$. This result does not depend upon the $^3$P$_1$ interaction and should, therefore, be directly comparable to our ‘plane wave’ result. Finally, using the Argonne and Nijmegen contemporary realistic potential models A$v_{18}$, Reid93, and Nijm II [@SKTS94] Liu and Timmermans [@LT04] obtained for the polarization component of the two-body contribution to the deuteron EDM $d_D^{(2)}$ values of $- 0.72\ A\, e\, fm$, $- 0.73\ A\, e\, fm$, and $- 0.74\ A\, e\, fm$, respectively. These relatively model-independent results suggest that pion exchange is indeed the essential aspect of the model. The differing degree of softness of the three potentials at intermediate range correlates with the values for $d_D^{(2)}$, the Nijm II potential being the softest and producing the largest EDM. The important conclusion for our purpose is that all three models yield essentially the same result; within the range of uncertainty defined by the three models utilized, the value of the polarization component of $d_D^{(2)}$ can be said to be $\approx - 0.73\pm .01\ A\, e\, fm$. Moreover, Liu and Timmermans estimated that the meson exchange current contribution was substantially smaller, calculated to be less than 5% of the potential model contribution. In any case, our goal is to determine an appropriately simple model with which one can calculate reliably the $^2$H, $^3$He, and $^3$H EDMs, so that our numerical comparisons will be made with the $- 0.73\pm .01\ A\, e\, fm$ value. Nucleon contributions ===================== The total one-body contribution $d_D^{(1)}$ to the deuteron EDM due to the neutron and proton is the sum of the individual nucleon EDMs: $$d_D^{(1)} = d_n + d_p \; , \label{eq:1}$$ whereas the total deuteron EDM is the sum of this one-body contribution and the two-body contribution $d_D^{(2)}$, $$d_D = d_D^{(1)} + d_D^{(2)} = (d_n + d_p) + d_D^{(2)} \; .\label{eq:2}$$ As has been noted, the neutron and proton EDMs can arise from a variety of sources. Because we have nothing new to add to prior analyses of the nucleon EDM, we adopt the approach advanced by Liu and Timmermans [@LT04]: $$d_D^{(1)} \simeq 0.22\times10^{-2} \bar{G}_\pi^{(1)} + O(\bar{G}_\pi^{(0,2)},\bar{G}_{\rho,\omega,\eta}) \; ,\label{eq:3}$$ which is expressed in terms of $\bar{G}_X^{(i)}$, the product of the strong coupling constant $g_{XNN}$ and the associated $PT$ violating meson-nucleon coupling constant $\bar{g}_X^{(i)}$. (For example, $\bar{G}_\pi^{(1)} = \bar{g}_{\pi NN}^{(1)} \, g_{\pi NN}$.) As noted in Ref. [@LT04], the contributions from the neutron and proton EDMs have a sizable theoretical uncertainty, but the significant cancellation between $d_n$ and $d_p$ is clear. For the two-body contribution to $d_D^{(2)}$ the mean value obtained by Liu and Timmermans can be expressed as $$d_D^{(pol)} = 1.45\times10^{-2}\bar{G}_\pi^{(1)} \; ;\label{eq:4}$$ this corresponds to the EDM value of $- 0.73\ A\, e\, fm$. Hence, for the deuteron there can be little doubt that the nuclear physics contribution to $d_D^{(2)}$ dominates. Even an uncertainty of 50% in $d_D^{(1)}$ contributes only in a minor way. It is the nuclear model aspects of the $d_D^{(2)}$ dominant term in the $^2$H EDM that we investigate below in detail. Two-body Contributions ====================== The interaction Hamiltonian for the ground state of the system consists of two components: (i) The strong interaction component $v$ based on nucleon-nucleon potentials with parameters adjusted to fit the experimental phase shifts. (ii) The $PT$ violating component $V$ which we parametrize in terms of one pion exchange (OPE) with one strong interaction vertex $g_{\pi NN}$ and a $PT$ violating vertex $\bar{g}^{(1)}_{\pi NN}$. As a result our Hamiltonian takes the form $$H= H^S + H^{PT}\quad\mbox{where}\quad H^S= H_0 + v \quad\mbox{and}\quad H^{PT}=V\ .\label{eq:5}$$ Because $H^{PT}$ will mix different parity states, i.e., for the deuteron we get coupling between the $^3$S$_1$-$^3$D$_1$ large component and the $^3$P$_1$ small component, we can write the Schrödinger equation for the Hamiltonian in Eq. (\[eq:5\]) $$H\,|\Psi{\rangle}= E\,|\Psi{\rangle}\label{eq:6}$$ as a set of coupled equations of the form $$\begin{aligned} (E-H_0)\,|\Psi_L{\rangle}&=& v\,|\Psi_L{\rangle}+ V\,|\Psi_S{\rangle}\ \label{eq:7}\\ (E-H_0)\,|\Psi_S{\rangle}&=& v\,|\Psi_S{\rangle}+ V\,|\Psi_L{\rangle}\ ,\label{eq:8}\end{aligned}$$ where the total wave function is the sum of the large and small components: $|\Psi{\rangle}= |\Psi_L{\rangle}+ |\Psi_S{\rangle}$. Because $V\ll v$, we have that $V\,|\Psi_S{\rangle}\ll v\,|\Psi_L{\rangle}$, and we can, to a good approximation, write Eq. (\[eq:7\]) as $$(E-H_0)\,|\Psi_L{\rangle}= v\,|\Psi_L{\rangle}\ ,\label{eq:9}$$ which is the Scrödinger equation for the ground state of the system in the absence of the $PT$ violating interaction. On the other hand, the small component of the wave function $|\Psi_S{\rangle}$ is given by the solution of Eq. (\[eq:8\]) in terms of the amplitude $t(E)$ for the strong potential $v$ as $$|\Psi_S{\rangle}= G(E)\,V\,|\Psi_L{\rangle}\quad\mbox{with}\quad G(E) = G_0(E) + G_0(E)\,t(E)\,G_0(E)\ ,\label{eq:10}$$ where $G_0(E)=(E-H_0)^{-1}$ is the free Green’s function, and $t(E)$ is the amplitude in the partial wave of the small component of the wave function, e.g., for the deuteron $t(E)$ is the amplitude in the $^3$P$_1$ partial wave at the ground state energy. Because the dipole operator $$O_d = \frac{e}{2}\ \sum_{i} \vec{r}_i\ \tau_z(i) \label{eq:11}$$ is odd under parity, we can write the two-body deuteron EDM ($d_D^{(2)}$) in terms of the total ground state wave function $|\Psi{\rangle}= |\Psi_L{\rangle}+|\Psi_S{\rangle}$ as $$d_D^{(2)} = {\langle}\Psi|\,O_d\,|\Psi{\rangle}= {\langle}\Psi_L|\,O_d\,|\Psi_S{\rangle}+ {\langle}\Psi_S|\,O_d\,|\Psi_L{\rangle}\ ,\label{eq:12}$$ where the matrix element of the dipole operator between the small and large component of the wave function can be written in terms of the charge $e$ and the constant $A$ as $$\begin{aligned} {\langle}\Psi_L|\,O_d\,|\Psi_S{\rangle}&=& {\langle}\Psi_L|\,O_d\,G_0(E)\,V\,|\Psi_L{\rangle}+ {\langle}\Psi_L|\,O_d\,G_0(E)\,t(E)\,G_0(E)\,V\,|\Psi_L{\rangle}\ \label{eq:13}\\ &\equiv& \frac{e}{2}\,\left[d_{PW} + d_{MS}\right]\,A \quad\mbox{with}\quad A\equiv\frac{\bar{g}^{(1)}_{\pi NN}\,g_{\pi NN}}{16\pi}\ .\label{eq:14}\end{aligned}$$ In Eq.(\[eq:13\]) the first term on the right hand side (rhs) involves a complete set of intermediate plane wave states and is, up to a constant, the ‘plane wave’ contribution $d_{PW}$. The second term on the rhs of Eq. (\[eq:13\]) involves multiple scattering via the amplitude $t(E)$ and is the ‘multiple scattering’ contribution $d_{MS}$. One should note that $E<0$ is the ground state energy, and as a result we need the amplitude $t(E)$ at an unphysical point corresponding to the $^2$H bound state energy. Numerical results ================= The primary motivation for the present investigation is: (i) to determine the sensitivity of $d_D^{(2)}$ to properties of the deuteron, *e.g.* the $D$-state probability and the short range behavior of the deuteron wave function. (ii) to determine the relative importance of $d_{PW}$ and $d_{MS}$. This will suggest the significance of multiple scattering terms as one proceeds to heavier nuclei. (iii) The role of the $^3$P$_1$ interaction in determining the magnitude of $d_{MS}$ and therefore the appropriateness of the $d_{PW}$ approximation in heavier nuclei. Before we proceed to illustrate the sensitivity of the deuteron EDM to nuclear structure effects due to the nuclear interaction, we should detail our choice of nucleon-nucleon interactions and their fit to those aspects of the two-body data relevant to the determination of the EDM. Two-body potentials ------------------- The input two-body interactions consists of: (i) The $PT$ violating one pion exchange potential. (ii) The deuteron wave function in the absence of the $PT$ violating interaction. (iii) The $^3$P$_1$ interaction that couples to the deuteron $^3$S$_1$-$^3$D$_1$ potential as a result of the introduction of the $PT$ violating potential. The choice of these interactions is motivated by the questions raised regarding the sensitivity of the EDM to nuclear structure effects and the hope of extending the analysis to $^3$H and $^3$He using the $d_{PW}$ approximation. For the $PT$ violating interaction we have chosen the standard isovector one-pion exchange given by [@PH92] $$V = -A\left[(\vec{\sigma}^{(-)}\cdot\hat{r})\,\tau_z^{(+)} + (\vec{\sigma}^{(+)}\cdot\tau_z^{(-)}\right]\,f(r)\ ,\label{eq:15}$$ where the radial dependence is given by $$f(r) = - \frac{1}{m_\pi}\,\frac{d}{dr}\left(\frac{e^{-m_\pi r}}{r}\right)\ ,\label{eq:16}$$ with $m_\pi$ being the pion mass. Here we have combined the strength of the strong and $PT$ violating vertices in the constant $A$ given in Eq. (\[eq:14\]). This allows us to express the numerical value of the EDM in terms of $A\,e$ with $e$ the charge on the proton. Finally, the spin and isospin operators in Eq. (\[eq:15\]) are given by $\vec{\sigma}^{(\pm)} = (\vec{\sigma}^{(1)}\pm\vec{\sigma}^{(2)})$ and $\tau_z^{(\pm)} = (\tau_z^{(1)}\pm\tau_z^{(2)})$. The strong $^3$S$_1$-$^3$D$_1$ interaction basically defines the deuteron wave function. Here we resort to a separable representation of the interaction to simplify the computation when we proceed to the EDM for the three-nucleon system. As a result the partial wave expansion of the strong interaction in momentum space is written as $${\langle}\vec{k}|\,v\,|\vec{k}'{\rangle}= \sum_{Sjtm}\sum_{\ell\ell'}\ {\langle}\hat{k}|{\cal Y}^t_{(\ell S)jm}{\rangle}\ v^{Sjt}_{\ell\ell'}(k,k')\ {\langle}{\cal Y}^t_{(\ell'S)jm}|\hat{k}'{\rangle}\ , \label{eq:17}$$ with $|{\cal Y}^t_{(\ell S)jm}{\rangle}$ eigenstates of the orbital angular momentum $\ell$, spin $S$, total angular momentum $j$ and isospin $t$. The separability of the potential is defined by the requirement that $$v^\alpha_{\ell\ell'}(k,k') = g^\alpha_\ell(k)\ \lambda^\alpha_{\ell \ell'}\ g^\alpha_{\ell'}(k')\ ,\label{eq:18}$$ where $\alpha = (Sjt)$. Here we wish to examine the role of the $D$-state probability and short range nature of the nucleon-nucleon interaction. For that we consider two classes of interactions: (i) The Yamaguchi and Yamaguchi (YY) [@YY54] separable potential with 4% and 7% $D$-state probability. Each has a different $D$-state probability and no short range repulsion. (ii) The Unitary Pole Approximation (UPA) [@AR73; @AR75] to the original Reid soft core potential (Reid68) [@Re68] and the Nijmegen modified Reid potential (Reid93) [@SKTS94]. The UPA potential by definition generates the same deuteron wave function as the original potential [@AR75] that provided the optimum fit to the available data at the time the potentials were constructed and includes short range repulsion. In addition the models have different $D$-state probabilities for the deuteron. For the Yamaguchi and Yamaguchi potentials [@YY54] the form factor $g_\ell^\alpha(k)$ is given by $$g_\ell(k) = \frac{k^\ell}{(k^2+\beta_\ell^2)^{(\ell+2)/2}}\ , \label{eq:19}$$ where the parameters $\beta_\ell$ and $\lambda_{\ell\ell'}$ are detailed in Table \[table:1\]. Also included in this table are the binding energy $\epsilon_D$ and the quadrupole moment $Q_D$ for these two potentials.  $D$-state  $\beta_0$ $\beta_2$ $\lambda_{00}$ $\lambda_{02}$ $\lambda_{22}$ $\epsilon_D$ (MeV) $Q_D$ ------------- ----------- ----------- ---------------- ---------------- ---------------- -------------------- ------------   4%  1.3134    1.5283    -0.6419     1.0849     -1.8320   2.2234   0.2821     7% 1.2410 1.9480 -0.3776 1.6975 -7.6301 2.2265 0.2826 : Parameters for the Yamaguchi-Yamaguchi potentials [@YY54] with 4% and 7% $D$-state probability for the deuteron. Also included are the binding energy and quadrupole moments.[]{data-label="table:1"}  potential   $\lambda_{00}$ $\lambda_{02}$ $\lambda_{22}$ -------------- -------------------- --------------------- --------------------  Reid68    -5.2896725E-02      -2.4385786E+00      1.1850926E+00   Reid93 -4.7704789E-01 -1.8111764E+00 2.5825467E-01 : The strength $\lambda_{\ell \ell'}$ for the UPA approximation to the Reid68 [@Re68]. and Reid93 [@SKTS94] potentials[]{data-label="table:2"} In constructing the UPA to the Reid68 [@Re68] and Reid93 [@SKTS94] we have used the method of moments [@AR75] to solve the Schrödinger equation for the deuteron wave function in coordinate space using the original potentials. This was achieved by taking the form factors such that the resultant deuteron wave functions for the Reid68 and Reid93 are linear combinations of the Yamaguchi-Yamaguchi type wave functions with different range parameters $\beta_i$, and therefore of the form $$g_\ell(k) = \sum_{i=1}^{12}\ \frac{c_\ell^i\ k^\ell}{(k^2+\beta_i^2)^{(\ell+2)/2}}\ . \label{eq:20}$$ The strengths of the UPA potential ($\lambda_{\ell\ell'}$), adjusted to reproduce the matrix elements of the original Reid68 and Reid93 potentials, are given in Table \[table:2\], while the parameters of the UPA form factors $\beta_i$ and $c^i_\ell$ for $\ell=0$ and $2$ are given in Table \[table:3\]. Here, we have chosen the range parameters $\beta_i$ to be multiples of the pion mass with the hope of reproducing some of the analytic structure of the one pion tail in the original Reid potentials. ------- -------------------------- -------------------- ------------------- -------------------- --------------------- $i$   $\beta_i$ (fm$^{-1}$)  $c^i_0$ $c^i_2$ $c^i_0$ $c^i_2$   1   0.7   7.21186419E-03    -2.24457073E-03    6.30646724E-03     -3.08893140E-03   2 1.4 1.78826642E-01 -3.31063031E-01 2.12846533E-01 -3.01564884E-01 3 2.1 1.31260692E+00 -1.04745293E+00 6.05450638E+00 -1.78185516E+00 4 2.8 2.13430424E+00 -1.43628043E+00 -2.57777824E+01 7.87042755E-01 5 4.2 1.46578861E+02 -1.95695256E+01 3.20079733E+02 -2.53483826E+01 6 5.6 -8.10387728E+02 3.12782173E+00 -1.49174373E+03 4.67387261E+01 7 7.0 1.12934549E+03 1.51126963E+02 2.32746050E+03 3.37908596E+01 8 9.8 -5.87779728E+02 -4.26701986E+02 -2.57402658E+03 -2.10353562E+02 9 12.6 -2.27638508E+02 5.92398037E+02 2.53223423E+03 3.41412020E+02 10 15.4 5.33784864E+02 -3.73533199E+02 -1.31246553E+03 -2.42126156E+02 11 21.0 -2.53746105E+02 9.68400708E+01 2.66329930E+02 7.60609941E+01 12 26.6 6.63870056E+01 -2.09513706E+01 -4.84106437E+01 -1.89979113E+01 ------- -------------------------- -------------------- ------------------- -------------------- --------------------- : The form factor parameters of the UPA approximation to the Reid68 [@Re68] and Reid93 [@SKTS94] potentials .[]{data-label="table:3"} To establish the quality of the UPA deuteron wave function generated using the method of moments we present in Table \[table:4\] the deuteron properties for the original potential and the UPA for both Reid68 and Reid93. Also included are the effective range parameters to illustrate the domain of agreement in the scattering amplitude between the original and the UPA potential. It is clear from these results that the method of moments gives a very good representation of the original deuteron wave function and can reproduce the effective range parameters. --------------------- -------------- ------------- ------------ ------------ UPA Original UPA  Original  $\epsilon_D$ 2.2246 2.2246 2.2246 2.2246 $A_S$ 0.87893 0.87758 0.8863 0.8853   $\eta=A_D/A_S$      0.026556     0.026223    0.02565  0.0251 $Q_D$ 0.2800 0.27964 0.2709 0.2703 $P_D$ 6.4691 6.4696 5.699 5.699 $a_t$ 5.408 5.390 5.445 5.422 $r_t$ 1.752 1.720 1.799 1.755 --------------------- -------------- ------------- ------------ ------------ : Comparison of the deuteron properties for the original potential and the UPA potential for both Reid68 and Reid93. Tabulated are the binding energy $\epsilon_D$, the asymptotic $S$-wave normalization $A_S$, the ratio of the asymptotic $D$-wave to $S$-wave $\eta$, the quadrupole moment $Q_D$, and the $D$-state probability $P_D$. Also included are the scattering length $a_t$ and effective range $r_t$.[]{data-label="table:4"} Finally, to examine the importance of multiple scattering in determining the deuteron EDM, we need to introduce a $^3$P$_1$ interaction to calculate $d_{MS}$. Here we need to know how important is the fit to the data and the role of the off-shell amplitude in determining the magnitude of $d_{MS}$. To simplify the evaluation of $d_{MS}$, we have chosen to use separable potentials with different form factors. The Mongan [@Mo69] potentials used by Avishai [@Av85] come with different form factors, and therefore different off-shell properties. They are either rank one or rank two to optimize the fit to the data; *i.e.*, the potentials are of the form $$v_{^3P_1}(k,k') = \sum_{i=1}^n\ g_i(k)\,\lambda_i\,g_i(k')\ ,\label{eq:21}$$ where $n=1$ for rank-one potentials and $n=2$ for rank-two potentials. For the form factors $g_i(k)$ we will use the four different forms chosen by Mongan (see Table \[table5\]). Considering the fact that Mongan adjusted the parameters of his potentials to fit the Livermore data of the 1960’s, we need first compare the phase shifts predicted by the Mongan potentials and those that we constructed to fit the latest Nijmegen [@Nij93] $np$ data. In Fig. \[fig1\] we compare the $^3$P$_1$ phase shifts for rank-one and rank-two Case I form factors for Mongan’s potentials with those refitted to the Nijmegen data. Also included are the Nijmegen [@Nij93] $np$ phase shifts. It is clear from the the results in Fig. \[fig1\] that the original Mongan potentials give a poor fit to the current data, while the new fits reproduce the data to a much better degree. Since the $^3$P$_1$ amplitude required for the determination of $d_{MS}$ is evaluated at the deuteron binding energy, *i.e.*, *below* the elastic threshold, it is essential that we fit well the low energy phase shifts. Because these are small, we have chosen the criteria for a good fit $\chi^2$ defined as $$\chi^2 = \sum_{i=1}^{n}\ \frac{|\delta_i^{\rm th}-\delta_i^{\rm exp}|^2}{|\delta_i^{\rm exp}|^2}\ ,\label{eq:22}$$ where $n=11$ is the number of data points below 300 MeV. In Table \[table5\] we present new fits to the Nijmegen $np$ data for the different form factors used by Mongan [^1]. Included are rank-one and rank-two potentials and the $\chi^2$ for each potential. It is clear from the $\chi^2$ that the rank two potentials give a better fit. This is especially true for the Case I form factor. In the following discussion of the deuteron EDM we will consider these different $^3$P$_1$ potentials to establish the importance of fitting the data and the role of the off-shell behavior of the amplitude. ![Comparison of the $^3$P$_1$ phase shifts for the Mongan potentials (Old) with Case I form factor and rank one (R=1) and rank two (R=2) with the new fit (New) and the experimental ($Exp.$) Nijmegen [@Nij93] $np$ data. []{data-label="fig1"}](./figure1.pdf) Potential   form factor $g_i(k)$  Rank  $\beta_1$ $\lambda_1$ $\beta_2$ $\lambda_2$ $\chi^2$ ------------- ----------------------------------------------------------------------- -------- ----------- ------------- ----------- ------------- ---------- Case I $k/(k^2+\beta_i^2)$ 1  1.725  0.95 - - 0.62 2 0.90   0.059    3.58   -2.0   0.02  Case II $k/(k^2+\beta_i^2)^{3/2}$ 1 2.38 9.35 - - 0.81 Case III  $\left[\frac{1}{k^2\pi} Q_1(1+\frac{\beta_i^2}{2k^2})\right]^{1/2}~$ 1 1.68 60.0 - - 0.19 2 1.20 120.0 4.4 -2.3 0.12 Case IV $k/(k^2+\beta_i^2)^{2}$ 1 2.715 147.0 - - 0.78 : The parameters of the ‘New’ rank-one and rank-two potentials with the different Mongan form factors. The parameters are adjusted by minimizing the $\chi^2$ defined in Eq. (\[eq:22\]) taking the experimental phases from the latest Nijmegen [@Nij93] $np$ phase shift analysis. The form factor for Case III is written in terms of $Q_1(\xi)$ the Legendre function of the second kind. []{data-label="table5"} The deuteron EDM ---------------- We now turn to the study of the sensitivity of the deuteron EDM to the nuclear structure effects as defined by the strong nucleon-nucleon interactions detailed above. We first consider the sensitivity of the two-body deuteron EDM $d_D^{(2)}$ to the $D$-state probability ($P_D$). In Table \[table6\] we summarize the contributions to the deuteron EDM for the four different deuteron wave functions being considered. For the $^3$P$_1$ interaction we use a rank-two Mongan Case I potential (fitted to the latest Nijmegen phase shifts [@Nij93]). Also included are the results of Khriplovich and Korkin [@KK00]. We observe that in the plane wave approximation ($d_{PW}$) there is little variation with $P_D$, and the short range repulsion incorporated in the two Reid potential wave functions provides no more than a 10% reduction in $d_{PW}$. Moreover, the results are effectively consistent with the zero range (chiral limit) approximation of Khriplovich and Korkin. In particular, the plane wave results for the two YY models suggest that the dependence upon the deuteron $D$-state probability is such that an S-state deuteron result would approach that of Ref. [@KK00]. In contrast, the multiple scattering contribution ($d_{MS}$), which is of the opposite sign to the plane wave term, varies considerably depending upon the short range character of the deuteron wave function. In particular, the two Reid potentials with different $P_D$ values yield quite similar values of $d_{MS}$, but these are only half those generated by the YY potentials. The difference between the YY and Reid potential models can be understood in light of our knowledge that there is no explicit short range repulsion in the YY potentials. We will return to this difference when we address the role of the off-shell behavior of the $^3$P$_1$ amplitude in determining the magnitude of the multiple scattering contribution $d_{MS}$. From these results we may conclude that the strong repulsion at short distance in realistic nucleon-nucleon potentials reduces the effects of multiple scattering in the matrix element to such an extent that the multiple scattering contribution $d_{MS}$ is only about 20% of the plane wave contribution $d_{PW}$. Furthermore, as noted above, the final results are not particularly sensitive to $P_D$. $^3$S$_1$-$^3$D$_1$ $P_d$ $d_{PW} (A\,e\,fm)$ $d_{MS} (A\,e\,fm)$ $d_D^{(2)} (A\,e\,fm)$ ---------------------- ------- --------------------- --------------------- ------------------------ YY 4% 4% -1.035 0.4115 -0.6234 Reid93 5.7% -0.9715 0.2009 -0.7706 Reid68 6.5% -0.9620 0.1718 -0.7902 YY 7% 7% -1.083 0.4271 -0.6564 Khriplovich *et al.* -0.92 : The variation of the two-body EDM with $D$-state probability of the deuteron. For the $^3$P$_1$ interaction we use the ‘New’ fit Case I rank-two potential. Also included are the results of Khriplovich and Korkin [@KK00].[]{data-label="table6"} To establish the importance of the multiple scattering contribution ($d_{MS}$) to the total two-body deuteron EDM, we turn to the dependence of $d_D^{(2)}$ on the choice of the $^3$P$_1$ interaction. But first we need to examine the sensitivity of the multiple scattering contribution to the $^3$P$_1$ phase shifts. This can be achieved by comparing the results for the EDM using the ‘Old’ Mongan fit to the 1960’s Livermore phase shift analysis and the ‘New’ fit with the same separable potential form factors to the latest Nijmegen [@Nij93] $np$ data. We have in Table \[table7\] the EDM results for the rank-one separable potentials with Case I and III form factors. For the deuteron wave function we have used either the UPA to the Reid68 or the YY 4% potentials. It is clear from these results that the multiple scattering contribution ($d_{MS}$) is reduced as a result of the fit to the more recent phase shift analysis (compare rows four and five or rows six and seven in Table \[table7\]). This reduction in $d_{MS}$ is consistent with the observation that the ‘New’ $^3$P$_1$ potentials provide less repulsion (i.e. smaller phase shifts, see Fig. \[fig1\]) and, therefore, substantially smaller multiple scattering contributions than the old fits due to Mongan. This observation is encouraging for extending the above analysis based on $d_{PW}$ to the three-nucleon EDM, as the new $np$ data suggest a reduced contribution from the multiple scattering term. We now return to the role of the short range repulsion in the deuteron wave function on the magnitude of the multiple scattering term $d_{MS}$ as illustrated in Table \[table6\]. In comparing the results for the Reid68 and 4% YY deuterons (column three and five in Table \[table7\]) for the Case I and Case III $^3$P$_1$ potentials, we find that the multiple scattering term is suppressed for both $^3$P$_1$ potentials. This suggests that the effect tabulated in Table \[table6\] might be valid in general. which implies that the inclusion of multiple scattering will require a more realistic treatment of the deuteron wave function than is the case for the zero range approximation employed by Khriplovich and Korkin [@KK00]. In fact for some combination of deuteron wave function and $^3$P$_1$ interaction ( 4% YY and Case III Old) the multiple scattering contribution ($d_{MS}$) is about the same size as the plane-wave approximation ($d_{PW}$) and as a result the deuteron EDM $d_D^{(2)}$ is suppressed by an order of magnitude compared to the combination Reid68 and Case I New. ----------- ---------- ------------- ------------- ------------ ------------- Case $\chi^2$   $d_{MS}$  $d_D^{(2)}$  $d_{MS}$  $d_D^{(2)}$ I (New)   0.62  0.21 -0.75 0.57 -0.47 I (Old) 1.90 0.31 -0.66 0.78 -0.26 III (New) 0.19 0.25 -0.71 0.77 -0.27 III (Old) 6.67 0.42 -0.54 1.16 0.12 ----------- ---------- ------------- ------------- ------------ ------------- : Variation in the deuteron EDM with changes in the $np$ phase shifts for two rank-one separable potentials having different form factors as defined by Mongan [@Mo69]. Here ‘New’ refers to the fit to the latest Nijmegen [@Nij93] $np$ phase shifts while ‘Old’ refers to the original Mongan fit. []{data-label="table7"} ------ -------- ------------- ------------------------ ---------------------------- ------------------------ ---------------------------- Case  Rank    $\chi^2$  $~d_{MS} (A\,e\,fm)$     $d_D^{(2)} (A\,e\,fm)$   $~d_{MS} (A\,e\,fm)$     $d_D^{(2)} (A\,e\,fm)$   I 1   0.62  0.2583 -0.7132 0.5665 -0.4684 I 2 0.02 0.2009 -0.7706 0.4115 -0.6234 II 1 0.81 0.2229 -0.7486 0.3807 -0.6542 III 1 0.19 0.3075 -0.6640 0.7654 -0.2696 III 2 0.12 0.3805 -0.5910 1.108 0.0734 IV 1 0.78 0.2153 -0.7562 0.3277 -0.7072 ------ -------- ------------- ------------------------ ---------------------------- ------------------------ ---------------------------- : The dependence of $d_{MS}$ on the $^3$P$_1$ separable potential form factor as defined by Mongan [@Mo69] and that are fit to the latest $np$ phase shifts. The Reid93 or the 4% YY deuteron wave function is used in all cases as indicated.[]{data-label="table8"} We now turn to the role of the off-shell behavior of the $^3$P$_1$ amplitude in the deuteron EDM. Here again we make use of the different separable potentials with the different form factors used by Mongan after readjusting the parameters of the potential to fit the latest Nijmegen [@Nij93] $np$ phase shifts. The parameters of these ‘New’ potentials are given in Table \[table5\]. In Table \[table8\] we report the multiple scattering contribution $d_{MS}$ and the two-body EDM $d_D^{(2)}$ for these separable potentials. In each case we have made use of either the Reid93 or 4% YY deuteron wave functions in the calculations. Here we observe that for the Reid93 deuteron there is a smaller variation in $d_D^{(2)}$ than is the case for the 4% YY deuteron. This is due to the fact the the multiple scattering contributions, $d_{MS}$, for the 4% YY deuteron have a substantially larger variation for the different fits to the $np$ data. This is consistent with the results in Table \[table7\] and is due to the absence of short range repulsion in the YY potentials. Here we can raise a number of questions regarding the role of the $^3$P$_1$ amplitude in determining the magnitude of the multiple scattering contribution $d_{MS}$. These are: - Why is $d_{MS}$ almost a factor of two smaller for the Reid93 when compared to that for the 4% YY potential? - Why, for the Reid93 deuteron, is $d_{MS}$ about the same for all form factors with the possible exception of Case III which gives the largest contribution? - Why is it that for the 4% YY deuteron $d_{MS}$ has a much larger variation than is the case for Reid93? To address these questions and to try to correlate the results in Table \[table8\] with the off-shell behavior of the $^3$P$_1$ amplitude, we need to examine the analytic continuation of the $P$-wave scattering wave function to the deuteron pole. This is defined in momentum space in terms of the half off-shell $t$-matrix as $$\begin{aligned} \psi_\alpha(k) &=& G_0(-\epsilon_D,k)\ t_\alpha(k,i\kappa;-\epsilon_D) \nonumber \\ &=& G_0(-\epsilon_D,k)\ \mathbf{g}_\alpha(k)\,\bm{\tau}_\alpha(-\epsilon_D)\,\mathbf{g}_\alpha^\dag(i\kappa)\nonumber \\ &\equiv& \Psi_\alpha(k)\,\bm{\tau}_\alpha(-\epsilon_D)\,\mathbf{g}_\alpha^\dag(i\kappa)\ , \label{eq:23}\end{aligned}$$ where $\alpha$ labels the $^3$P$_1$ channel, $\epsilon_D =\frac{\kappa^2}{2\mu}$ is the binding energy of the deuteron, and $\mu$ is the $np$ reduced mass. Here the free Green’s function at the deuteron energy is given by $G_0(-\epsilon_D,k) = -(2\mu)(\kappa^2+k^2)^{-1}$, while the amplitude $t_\alpha(k,i\kappa;-\epsilon_D)$ is the half off-shell $^3$P$_1$ $t$-matrix evaluated at the deuteron pole. In the second line of Eq. (\[eq:23\]) we have written the off-shell $t$-matrix in its separable form with $$\bm{\tau}_\alpha(-\epsilon_D) = \left[\,\bm{\lambda}_\alpha +2\mu \int\limits_0^\infty dk\,k^2\ \frac{\mathbf{g}_\alpha^\dag(k)\,\mathbf{g}_\alpha(k)}{\kappa^2 + k^2}\,\right]^{-1}\ .\label{eq:24}$$ For rank-one potentials $\bm{\tau}_\alpha(-\epsilon_D)$ is positive definite since the potential is repulsive (*i.e.*, $\lambda_\alpha>0$). As a result, the scattering wave function can be written as $$\psi_\alpha(k) = \chi_\alpha(k)\ \sqrt{\tau_\alpha(-\epsilon_D)}\ g_\alpha(i\kappa)\ .\label{eq:25}$$ This definition of the function $\chi_\alpha(k)$ is motivated in the following discussion of the matrix elements of the dipole operator $O_d$ and the $PT$ violating one pion exchange potential $V$ that go into the evaluation of $d_{MS}$. ![Comparison of the $^3$P$_1$ ‘scattered function’ $\chi_\alpha(k)$ defined in Eq. (\[eq:25\]) for the rank-one separable potentials that fit the latest Nijmegen [@Nij93] $np$ phase shifts.[]{data-label="fig2"}](./figure2.pdf) In Fig. \[fig2\] we plot the function $\chi_\alpha(k)$ for all the rank-one potentials used in Table \[table8\]. A careful inspection of this figure reveals that: (i)  The ‘scattered function’ $\chi_\alpha(k)$ for the Case III form factor is substantially larger for $k<1.0$ fm$^{-1}$ than that of the other three form factors. (ii) For $k>3$ fm$^{-1}$ the Case III ‘scattered function’ has the longest range followed by Case I and then Case II and finally Case IV. This is clear from the choice of form factors as given in Table \[table5\]. To establish how this momentum dependence of the $^3$P$_1$ ‘scattered function’ effects the multiple scattering contribution $d_{MS}$ to the deuteron EDM, we recall from Eq. (\[eq:14\]) that $d_{MS}$ can be written as $$d_{MS} = -2\left[\,\mathbf{O}_{sp}+\mathbf{O}_{dp}\,\right]\,\bm{\tau}(-\epsilon_D)\, \left[\,\mathbf{V}_{ps}+\mathbf{V}_{pd}\,\right]\ ,\label{eq:26}$$ where $$\mathbf{O}_{D \alpha} = {\langle}\Psi_D|\,O_d\,|\Psi_\alpha{\rangle}\quad\mbox{and}\quad \mathbf{V}_{\alpha D} = {\langle}\Psi_\alpha|\,V\,|\psi_D{\rangle}\ ,\label{eq:27}$$ with $D=\ ^3$S$_1$ or $^3$D$_1$ and $\alpha=^3$P$_1$. For rank-one separable potentials we can absorb a factor of $\sqrt{\tau(-\epsilon_D)}$ into the matrix elements, i.e. ${\cal O}_{\alpha p}\equiv O_{\alpha\beta}\sqrt{\tau(-\epsilon_D)}$ and ${\cal V}_{p \alpha} \equiv \sqrt{\tau(-\epsilon_D)}\,V_{p \alpha}$ and therefore for rank-one potentials we have $$d_{MS} = -2\left[\,{\cal O}_{sp} + {\cal O}_{dp}\,\right]\ \left[\,{\cal V}_{ps} + {\cal V}_{pd}\,\right]\ .\label{eq:28}$$ The values of ${\cal O}_{\alpha p}$ and ${\cal V}_{p\alpha}$ for the four different form factors and with a deuteron wave function given by either the 4% YY or the Reid93 are presented in Table \[table9\]. It is clear from these results that the matrix elements of the dipole operator $O_d$, which is long range in coordinate space, are to a good approximation independent of the deuteron wave function and to within 20% independent of the $^3$P$_1$ potential. On the other hand the matrix elements of the $PT$ violating one pion exchange potential, which probes the short range behavior of both the $^3$P$_1$ and the deuteron wave function, are clearly model dependent. In particular, for the Reid93 deuteron with short range repulsion, the variation in ${\cal V}_{p\alpha}$ is small with the Case III form factor giving the largest contribution and Case IV yielding the smallest contribution followed by Case II and Case I. This is consistent with the observation made in the Fig. \[fig2\] insert regarding the asymptotic behavior of the function $\chi_\alpha(k)$. This is also consistent with the observation in Table \[table8\] for rank-one potentials. On the other hand, for the 4% YY deuteron, with no short range repulsion, the matrix elements are almost a factor of two larger with the Case III form factor giving the largest contribution and Case IV the smallest. From the results in Table \[table9\] we may conclude that it is the matrix element of the $PT$ violating one pion exchange potential that probes the short range behavior of the $^3$P$_1$ and deuteron wave functions and, as a result, determines the magnitude of $d_{MS}$. To that extent it is essential that one generate those two wave functions in a consistent frame work. On the other hand, when the deuteron includes the short range behavior dictated by modern nucleon-nucleon interactions, the contribution of the multiple scattering term $d_{MS}$ is suppressed ($\approx 20$%) in comparison to the plane wave contribution $d_{PW}$. This suggests that one may be able to evaluate the EDM for the three-nucleon system in the plane wave approximation in such a model with an error of the order of 20%.  deuteron     Case   ${\cal O}_{sp}$ ${\cal O}_{dp}$ ${\cal V}_{ps}$ ${\cal V}_{pd}$ ------------- ---------- ----------------- ----------------- ----------------- ----------------- 4% YY I   -0.4197     -0.05599     0.5533     0.04211   II -0.4039 -0.05422 0.3794 0.03618 III -0.4819 -0.06300 0.6578 0.04444 IV -0.4185 -0.05622 0.3124 0.03276 Reid93 I -0.4221 -0.06069 0.2169 0.09031 II -0.4068 -0.05906 0.1928 0.04641 III -0.4852 -0.06712 0.2269 0.05154 IV -0.4224 -0.06105 0.1793 0.04338 : The matrix elements of the dipole operator $O_d$ and the $PT$ violating one pion exchange potential $V$ for the four different form factors and two different deuteron wave functions.[]{data-label="table9"} Finally, the results in Table \[table9\] for ${\cal O}_{\alpha p}$ and ${\cal V}_{p\alpha}$ indicate that the contribution from the $D$-wave component of the deuteron wave function are an order of magnitude smaller than the $S$-wave component. This may suggest that one could neglect the $D$-wave component in the calculating $d_{MS}$ and a simplification of the calculation of the multiple scattering term in heavier nuclei. This observation is consistent with the results in Table \[table6\] where the changes in the multiple scattering contribution has a variation of about 10% with $D$-state probability. Conclusions =========== From our analysis we offer the following conclusions: (i) In the absence of multiple scattering ($d_{MS}=0$) the variation in $d^{(2)}_D$ due to differences in the deuteron wave functions is less than 5%, and the value of $d_{PW}$ is consistent with the zero range (chiral limit) results of Khriplovich and Korkin [@KK00]. (ii) The contribution from multiple scattering $d_{MS}$ is sensitive to the short range behavior of the deuteron wave function, and the $d_{MS}$ contribution is about 20% for realistic parametrizations of the deuteron such as those represented by the Reid93 potential model. This suggests that we can extend the analysis to heavier nuclei in the plane wave approximation with an estimated error of $\approx 20$%. (iii) As suggested by Liu and Timmermans, one pion exchange dominates the deuteron EDM calculation. (iv) The contribution from the $^3$P$_1$ interaction via $d_{MS}$ depends on the phase shifts in this channel as well as the off-shell behavior of the amplitude. (v) A comparison of our Reid93 results with those of Liu and Timmermans [@LT04] indicates that one can use a separable potential approximation in heavier nuclei, *e.g.*, $^3$He and $^3$H, with minimal loss in accuracy. Moreover, until deuteron EDM experiments attain an uncertainty of less than 10%, simple separable potential model calculations should provide an adequate description. Acknowledgement =============== The work of BFG was performed under the auspices of the National Nuclear Security Administration of the U.S. Department of Energy at Los Alamos National Laboratory under Contract No DE-AC52-06NA25396. [11]{} T. D. Lee and C. N. Yang, Phys. Rev. **104**, 254 (1956). L. Landau, Nucl. Phys. **3**. 127 (1957). E. M. Purcell and N. F. Ramsey, Phys. Rev. **78**, 807 (1950). J. H. Smith, E. M. Purcell, and N. F. Ramsey, Phys. Rev.  **108**, 120 (1957). P. Herczeg, *Symmetries and Fundamental Interactions in Nuclei* edited by W. C. Haxton and E. M. Henley (World Scientific, Singapore), 1995, pp. 89-125. B. Khriplovich and S. K. Lamoreaux, *CP Violation without Strangeness: Electric Dipole Moments of Partllces, Atoms, and Molecules* (Springer-Verlag, Berlin, 1997). I. B. Khriplovich, Phys. Lett. B **444**, 98 (1998). F. J. M. Farley, K. Jungmann, J. P. Miller, W. M. Morse, Y. F. Orlov, B. L. Roberts, Y. K. Semertzidis, A. Silenko, and E. J.  Stephenson, Phys. Rev. Lett. **93**, 052001 (2004). Y. K. Semertzidis, M. Aoki, M. Auzinsh, V. Balankin, A. Bazhan, G. W. Bennett, R. M. Carey, P. Cushman, P. T. Debevec, A. Dudnikov, F. J. M. Farley, D. W. Hertzog, *et al.*, in *Intersections of Particle and Nuclear Physics: 8th Conference, CIPANP2003*, edited by Z. Parsa, AIP Conf. Proc. **698**, 200 (2004). Y. F. Orlov, W. M. Morse, and Y. K. Semertzidis, Phys.  Rev. Lett. **96**, 214802 (2006). Y. Avishai, Phys. Rev. D **32**, 314 (1985). T. R. Mongan, Phys. Rev. **175**, 1260 (1968); Phys.  Rev. **178**, 1597 (1969). I. B. Khriplovich and R. V. Korkin, Nucl. Phys. A **665**, 365 (2000). V. G. J. Stoks, R. A. M. Klomp, C. P. F. Terheggen, and J. J. de Swart, Phys. Rev. C **49**, 2950 (1994). C.-P. Liu and R. G. E. Timmermans, Phys. Rev. C **70**, 055501 (2004). P. Herczeg, *Hyperfine Interaction*, **75**, 127 (1992). Y. Yamaguchi and Y. Yamaguchi, Phys. Rev. **95**, 1635 (1954). I. R. Afnan and J. M. Read, Australian J. Phys. **26**, 725 (1973); I. R. Afnan and J. M. Read, Phys. Rev **12**, 293 (1975). R. V. Reid, Ann. Phys. (N.Y.) **50**, 411 (1968). V. G. J. Stoks, R. A. M. Klomp, M. C. M. Rentmeester, and J. J. de Swart, Phys. Rev. C **48**, 792 (1993). [^1]: The Case III form factor was motivated by the observation that the on-shell Born amplitude for a rank-one separable potential is identical to the on-shell Born amplitude resulting from meson exchange potential with a meson mass $\beta_1$.
{ "pile_set_name": "ArXiv" }
--- abstract: | By introducing a phase field and solving the eigen-functional equation of particles, we obtain the exact expressions of the ground state energy as a functional of the particle density for interacting electron/boson systems, and a two-dimensional electron gas under an external magnetic field, respectively. With the eigen-functionals of the particles, we can construct the ground state wave-function of the systems. Moreover, with the expressions of the ground state energy, we can exactly determine the ground state energy and the ground state particle density of the systems by taking $% \delta E_g[\rho ]/\delta \rho (x)=0$. address: | Center for Advanced Study, Tsinghua University, Beijing 100084,\ People’s Republic of China author: - 'Yu-Liang Liu' title: 'Exact expression of the ground state energy of quantum many-particle systems as a functional of the particle density' --- Quantum many-particle systems are the main topics of the condensed matter physics, in which strongly correlated electron systems are the most interesting and hard problems, such as heavy fermion systems, high Tc cuprate superconductors, fractional quantum Hall effects, and some one-dimensional interacting fermion systems. In general, the quantum many-particle (fermion) systems can be divided into two categories, one is represented by the Landau Fermi liquid theory[@1; @2], and may be called as weakly correlated systems, and another one is represented by non-Fermi liquid theory, such as the Tomonaga-Luttinger liquid theory[@3; @4; @5] and marginal Fermi liquid theory[@6; @7], and may be called as strongly correlated fermion systems. Therefore, there exists a key parameter hidden in the quantum many-particle systems, which represents the fermion/boson correlation strength. In Ref.[@8], we found this parameter we called the phase field, and with it we can unifiably represent the weakly and strongly correlated fermion systems, thus the Landau Fermi liquid theory and the non-Fermi liquid theory can be unified under our eigen-functional bosonization theory. According to the Hohenberg-Kohn theorem[@9], the ground state energy of the quantum many-particle systems is uniquely determined by their ground state particle density. This theorem does not tell us how to construct the ground state energy by the particle density, but it clearly tells us that the ground state energy can be exactly represented by the particle density. However, in the Kohn-Sham scheme[@10; @11], we can only approximately obtain the expression of the ground state energy as a functional of the ground state particle density. Thus it is very important to write out an exact expression of the ground state energy of the quantum many-particle systems as a functional of the particle density. In fact, in usual bosonization representation of one-dimensional interacting fermion systems[@12; @13], we learn how to construct the Hamiltonian by the fermion density operators. With the eigen-functional bosonization theory, we can easily treat any quantum many-particle system by using the fermion/boson density field to represent the kinetic energy of the systems. In Ref.[@8], we demonstrated how to unifiably represent the weakly and strongly correlated fermion systems, and how to calculate their correlation functions. In this paper, we give exact expressions of the ground state energy as the particle density field for different quantum many-particle systems. These expressions are universal, they are valid not only for weak fermion/boson interactions, but also for strong fermion/boson interactions. We first give the exact expression of the ground state energy as a functional of the electron density field for interacting electron systems; secondly we give the exact expression of the ground state energy for interacting boson systems, and their ground state wave-functions that are very similar to the correlated basis functions used in the study of the liquid $^{4}He$; finally we give the exact expression of the ground state energy of two-dimensional electron gas under an external magnetic field, and its ground state wave-functions that have a little similarity with Laughlin’s trial wave-functions[@14; @15]. This expression of the ground state energy has more advantages comparing with the Laughlin’s trial wave-functions, because it directly derives from the microscopic theory. In general, we consider the interacting electron system described by the Hamiltonian, $$H=\psi ^{\dagger }(x)\left( \frac{\hat{p}^2}{2m}-\mu \right) \psi (x)+\frac 1% 2\int d^Dyv(x-y)\rho (y)\rho (x) \label{1}$$ where $\hat{p}=-i\hbar {\bf \nabla }$, $\rho (x)=\psi ^{\dagger }(x)\psi (x)$ is the electron density operator, and $D$ the dimensions of the system. In order to calculate the ground state energy, we introduce a Lagrangian multiplier (boson) field $\phi (x)$ which takes $\rho (x)=\psi ^{\dagger }(x)\psi (x)$ as a constraint condition. The Hamiltonian (\[1\]) can be re-written as, $$\begin{aligned} H[\phi ,\rho ] &=&\displaystyle{\ \psi ^{\dagger }(x)\left( \frac{\hat{p}^2}{% 2m}-\mu +\phi (x)\right) \psi } \nonumber \\ &-&\displaystyle{\ \phi (x)\rho (x)+\frac 12\int d^Dyv(x-y)\rho (y)\rho (x)} \label{2}\end{aligned}$$ where $\phi (x)$ and $\rho (x)$ are independent boson fields. Analogy to the eigen-functional bosonization theory[@8], here we solve the following eigen-functional equation, $$\left( \frac{\hat{p}^2}{2m}-\mu +\phi (x)\right) \Psi _k(x,[\phi ])=E_k[\phi ]\Psi _k(x,[\phi ]) \label{3}$$ Using the Helmann-Feynman theorem, we have the expression of the eigen-values, $$\begin{aligned} E_k[\phi ] &=&\epsilon _k+\Sigma _k[\phi ] \nonumber \\ \Sigma _k[\phi ] &=&\displaystyle{\ \int_0^1d\xi \int d^Dx\phi (x)|\Psi _k(x,[\xi \phi ])|^2} \label{4}\end{aligned}$$ where $\epsilon _k=(\hbar k)^2/(2m)-\mu $. The eigen-functionals can be written as, $$\Psi _k(x,[\xi \phi ])=\frac{A_k}{L^{D/2}}e^{i{\bf k}\cdot {\bf x}% }e^{Q_k(x,\xi )} \label{5}$$ where $A_k$ is the normalization constant, and the phase (boson) field $% Q_k(x,\xi )$ satisfies usual Eikonal equation, $$\left( \frac{\hat{p}^2}{2m}+\frac \hbar m{\bf k}\cdot \hat{p}\right) Q_k(x,\xi )+\frac{[\hat{p}Q_k(x,\xi )]^2}{2m}+\xi \phi (x)=0 \label{6}$$ It is noted that the eigen-functionals are composed of two parts, one represents the free electron, and another one represents the contributions from other electrons by the interaction potential $v(x-y)$, therefore the eigen-functionals are the eigen-wave-functions of the interacting electrons corresponding to the definite boson field $\phi \left( x\right) $. With them, we can construct the ground state wave-function, and calculate the correlation functions of the system by taking a functional average over the boson field $\phi \left( x\right) $. The electron operators $\psi ^{\dagger }(x)$ and $\psi (x)$ can be represented as, $$\begin{aligned} \psi ^{\dagger }(x) &=&\displaystyle{\sum_k\Psi _k^{*}(x,[\phi ])\hat{c}% _k^{\dagger }} \nonumber \\ \psi (x) &=&\displaystyle{\sum_k\Psi _k(x,[\phi ])\hat{c}_k} \label{7}\end{aligned}$$ where $\hat{c}_k^{\dagger }$ ($\hat{c}_k$) is the electron’s creation ( annihilation) operator with momentum $\hbar k$. With equations. (\[3\]) and (\[7\]), we can obtain the exact expression of ground state energy as a functional of the boson field $\phi (x)$ and the electron density $\rho (x) $, $$\begin{aligned} E[\phi ,\rho ] &=&\displaystyle{\int d^DxH[\phi ]} \nonumber \\ &=&\displaystyle{\sum_k\theta (-E_k[\phi ])E_k[\phi ]-\int d^Dx\phi (x)\rho (x)+\frac 12\int d^Dxd^Dyv(x-y)\rho (x)\rho (y)} \label{8}\end{aligned}$$ where the boson field $\phi (x)$ is determined by the condition $\delta E[\phi ,\rho ]/\delta \phi (x)=0$, which induces the equation, $$\sum_k\theta (-E_k[\phi ])\left( G_k(x)+\int d^Dy\phi (y)\frac{\delta G_k(y)% }{\delta \phi (x)}\right) =\rho (x) \label{9}$$ where $G_k(x)=\int_0^1d\xi e^{2Q_k^R(x,\xi )}/\int d^Dxe^{2Q_k^R(x,\xi )}$, and $Q_k(x,\xi )=Q_k^R(x,\xi )+iQ_k^I(x,\xi )$. This equation shows that the boson field $\phi (x)$ is the functional of the electron density field $\rho (x)$. The chemical potential is determined by the constraint equation, $$\sum_k\theta (-E_k[\phi ])=N \label{10}$$ where $N$ is the total electron number. With equation (\[9\]), the ground state energy can be represented as another form, which is only the functional of the electron density field ($\phi (x)$ is the functional of $% \rho (x)$ determined by (\[9\])), $$E_g[\rho ]=E_0+\frac 12\int d^Dxd^Dy\left( v(x-y)\rho (x)\rho (y)-2\phi (x)\phi (y)\sum_k\theta (-E_k[\phi ])\frac{\delta G_k(y)}{\delta \phi (x)}% \right) \label{11}$$ where $E_0=\sum_k\theta (-E_k[\phi ])\epsilon _k$, and the self-energy can be written as a simple form, $\Sigma _k[\phi ]=\int d^Dx\phi (x)G_k(x)$. The equations (\[6\]), (\[9\]), (\[10\]) and (\[11\]) give the exact expression of the ground state energy of the system as a functional of the electron density, we can easily numerically calculate it and the ground state electron density by taking $\delta E_g[\rho ]/\delta \rho (x)=0$. These equations have more advantages than that in the Kohn-Sham scheme, because that, 1). the expression of the ground state energy as the functional of the electron density is exact, with it we can exactly determine the ground state energy and the ground state electron density; 2). with these equations, we can easily estimate the contributions from high order terms, and obtain enough accurate results we hoped for some special considerations; 3). Using these equations to calculate the ground state energy and the ground state electron density, we may need much less computer time than that in the Kohn-Sham scheme. The ground state wave-function of the system can be obtained by the eigen-functionals, however, we cannot write it as a simple form, because the boson field $Q_k(x,\xi =1)$ is the function of the momentum $k$. The expression of the ground state energy (\[11\]) is valid not only for Landau Fermi liquid (weak correlation systems), but also for non-Fermi liquid (strongly correlated systems), because it is universal for weak and strong electron interactions. Thus the Landau Fermi liquid and non-Fermi liquid can be unifiably represented by our eigen-functional bosonization theory, because the equation (\[11\]) gives the physical properties of the ground state, and in Ref.[@8] we show how to study the physical properties of the excitation states, and how to calculate the correlation functions of the systems. The phase field $Q_{k}(x,\xi)$ is a key parameter for unifiably representing the Landau Fermi liquid and non-Fermi liquid. Its imaginary part represents the electron correlation, and its real part ($% D\geq 2$) only contributes to the ground state energy and the action of the systems, which can be clearly seen in the expression of the ground state energy (\[11\]). For a boson system, such as the liquid $^4He$, at zero temperature it has Bose-Einstein condensation, and the bosons only occupy the state of the momentum $k=0$, thus the equation (\[10\]) is trivial, and the ground state energy can be written as, $$E_g[\rho ]=\frac 12\int d^Dxd^Dy[v(x-y)\rho (x)\rho (y)-2\phi (x)\rho (x)] \label{b1}$$ where we have taken $\mu =\Sigma _0[\phi ]$. The equation (\[9\]) reduces, $$G_0(x)+\int d^Dy\phi (y)\frac{\delta G_0(y)}{\delta \phi (x)}=\frac 1N\rho (x) \label{b2}$$ It is noted that the last term in (\[b1\]) is the contributions of kinetic energy of the bosons, and in general it may be non-zero for interacting boson systems. Due to the condensation of the bosons, we can easily obtain the ground state wave-function of the boson systems, $$\begin{aligned} \Psi (x_1,x_2,...,x_N) &=&\displaystyle{\ \left( \frac{A_0}{L^{D/2}}\right) ^N<\Psi (x_1,x_2,...,x_N,[\phi ])>_\phi } \nonumber \\ \Psi (x_1,x_2,...,x_N,[\phi ]) &=&\displaystyle{\ e^{\sum_{i=1}^NQ_0(x_i,\xi =1)}} \label{b3}\end{aligned}$$ where $<...>_\phi $ means the functional average over the boson field $\phi (x)$. Using the method in Ref.[@8], we can easily obtain the ground state wave-function which is uniquely determined by single effective potenial function, and has the expression very similar to usual correlated basis functions[@16] that are the type of wave-function most often employed in the study of the ground state properties of $^4He$. While the functional $\Psi (x_1,x_2,...,x_N,[\phi ])$ is very similar to the generalized London wave-function[@17], where $f(x_i)=\exp \{Q_0^R(x_i,\xi =1)\}$ and $S(x_i)=Q_0^I(x_i,\xi =1)$. We now consider the electron motion under an external magnetic field ${\bf B}% =(0,0,B)$, where the response of the system to the magnetic field cannot be written as a simple density form. It is well-known that for enough strong megnetic field, in low temperature limit a two-dimensional electron gas shows the fractional quantum Hall effects due to the Coulomb interaction of the electrons. Here we only give the exact expression of the ground state energy of this system, and do not compare it with that obtained by Laughlin’s trial wave-functions[@14], because that needs more detail numerical calculations. Under the magnetic field, the Hamiltonian (\[2\]) becomes, $$\begin{aligned} H[\phi ,\rho ] &=&\displaystyle{\psi ^{\dagger }(x)\left( \frac 1{2m}(\hat{p}% +\frac ec{\bf A})^2-\mu +\phi (x)\right) \psi (x)} \nonumber \\ &-&\displaystyle{\phi (x)\rho (x)+\frac 12\int d^2yv(x-y)\rho (y)\rho (x)} \label{12}\end{aligned}$$ where the gauge field ${\bf A}=(-yB/2,xB/2,0)$, and for simplicity we take $% D=2$. The eigen-functional equation corresponding to equation (\[3\]) reads, $$\left( \frac 1{2m}(\hat{p}+\frac ec{\bf A})^2-\mu +\phi (x)\right) \Psi _{nl}(x,[\phi ])=E_{nl}[\phi ]\Psi _{nl}(x,[\phi ]) \label{13}$$ and the eigen-values are $$\begin{aligned} E_{nl}[\phi ] &=&\epsilon _n+\Sigma _{nl}[\phi ] \nonumber \\ \Sigma _{nl}[\phi ] &=&\displaystyle\int_0^1{d\xi }\int {d^2x\phi (x)|\Psi _{nl}(x,[\xi \phi ])|^2} \label{14}\end{aligned}$$ where $\epsilon _n=\hbar \omega _0(n+1/2)-\mu $, $n=0,1,2,...$, and $\omega _0=eB/(mc)$ is the cyclotron frequency. It is well-known that at $\phi (x)=0$ the eigen-equation (\[13\]) has exact solutions, $\Psi _{nl}(x,[0])=\psi _{nl}(x)=a_{nl}(z-\partial _{z^{*}})^l(z^{*}-\partial _z)^n\exp \{-zz^{*}\}$, $l=0,1,2,..,L_{max}=\Phi /\Phi _0$, where $\Phi =BS$ is the total flux, $% \Phi _0=2\pi \hbar c/e$ the flux quantum, $z=(x-iy)/(2l_B)$, $% z^{*}=(x+iy)/(2l_B)$, $l_B=(\hbar c/(eB))^{1/2}$ is the magnetic length, and $a_{nl}$ a normalization constant. Therefore, the eigen-functionals have the following form, $$\Psi _{nl}(x,[\xi \phi ])=A_{nl}\psi _{nl}(x)e^{Q_{nl}(x,\xi )} \label{15}$$ where $A_{nl}$ is a normalization constant, and the phase field $% Q_{nl}(x,\xi )$ satisfies Eikonal-type equation, $$\left( \frac{\hat{p}^2}{2m}+(\frac e{mc}{\bf A}+{\bf a}_{nl})\cdot \hat{p}% \right) Q_{nl}(x,\xi )+\frac{[\hat{p}Q_{nl}(x,\xi )]^2}{2m}+\xi \phi (x)=0 \label{16}$$ where ${\bf a}_{nl}=(1/m)\hat{p}\ln \psi _{nl}(x)$. Following the above same procedures, we have the exact expression of the ground state energy as the electron density field $\rho(x)$, $$E_{g}[\rho]=E_{0}+\frac{1}{2}\int d^{2}x d^{2}y \left( v(x-y)\rho(x)\rho(y)- 2\phi(x)\phi(y)\sum_{nl}\theta(-E_{nl}[\phi])\frac{\delta G_{nl}(y)}{\delta \phi(x)}\right) \label{17}$$ where $E_{0}=\sum_{nl}\theta(-E_{nl}[\phi])\epsilon_{n}$, and the boson field $\phi(x)$ is the function of the electron density $\rho(x)$, and is determined by the equation, $$\sum_{nl}\theta(-E_{nl}[\phi])\left( G_{nl}(x)+\int d^{2}y\phi(y) \frac{% \delta G_{nl}(y)}{\delta\phi(x)}\right)=\rho(x) \label{18}$$ where $G_{nl}(x)=\int^{1}_{0} d\xi |\psi_{nl}(x)|^{2}e^{2Q^{R}_{nl}(x,\xi)} /\int d^{2}x |\psi_{nl}(x)|^{2}e^{2Q^{R}_{nl}(x,\xi)}$, and the self-energy can written as a simple form, $\Sigma_{nl}[\phi]=\int d^{2}x \phi(x)G_{nl}(x) $. The chemical potential is determined by the constraint condition, $$\sum_{nl}\theta(-E_{nl}[\phi])=N \label{19}$$ where $N$ is the total electron number. The equations (\[16\]), (\[17\]), (\[18\]) and (\[19\]) can be used to exactly determine the ground state energy and the ground state electron density of the system by taking $% \delta E_{g}[\rho]/\delta\rho(x)=0$. In the low temperature limit, for enough strong magnetic field, the system shows the fractional quantum Hall effects for some special filling factors $\nu=N\Phi_{0}/\Phi$. The expression of the ground state energy (\[17\]) has a great advantages comparing with the Laughlin’s trial wave-functions: 1). it is a microscopic theory expression. 2). it shows that the odd- and even-denominator’s fractional quantum Hall states have the same expression of the ground state energy. 3). it can exactly determine the ground state electron density of the system. 4). it is very simple to study the fractional quantum Hall effects in higher Landau levels ($n\geq 1$). However, the ground state wave-function of the system cannot be written as a simple form even for the lowest Landau level ($n=0$), because the phase field $Q_{nl}(x,\xi=1)$ depends on the quantum numbers $n$ and $l$. In the lowest Landau level, if we approximately take $Q_{0l}(x,\xi=1)\sim Q_{00}(x)$, we can obtain the following expression of the ground state wave-function, $$\Psi(x_{1},x_{2},...,x_{N})=A\prod^{N}_{i=1}z_{i}\prod_{i>j}(z_{i}-z_{j}) e^{\sum^{N}_{i=1}|z_{i}|^{2}}<e^{ \sum^{N}_{i=1}Q_{00}(x_{i})}>_{\phi} \label{20}$$ where $A$ is a normalization constant, and the factor $\prod_{i>j}(z_{i}- z_{j})$ guarantees the anti-commutation of the electrons. The last factor $% <\exp\{\sum^{N}_{i=1}Q_{00}(x_{i})\}>_{\phi}$ is very similar to the ground state wave-function of the boson systems (\[b3\]), and it is very clear that this factor is the contribution of the electron interactions. This ground state wave-function is different from the Laughlin’s trial wave-functions of the odd-dinominator fractional quantum Hall states. We believe that with the equations of the ground state energy and the ground state wave-function (\[17\]) and (\[20\]), respectively, we can obtain more important informations of the fractional quantum Hall effects than that by the Laughlin’s trial wave-functions, because they are directly derived from the microscopic theory. In summary, by introducing the Lagrangian multiplier (boson field) which makes the Hamiltonian only have the quadratic form of the electron/boson operators, and solving the eigen-functional equation of the electrons/bosons, we have obtained the exact expressions of the ground state energy as the functional of the electron/boson density for interacting electron systems, interacting boson systems, and the two-dimensional electron gas under an external magnetic field, respectively. By taking $% \delta E_g[\rho ]/\delta \rho (x)=0$, we can exactly determine the ground state energy and the ground state electron/boson density of the quantum many-particle systems. With the present method and the eigen-functional bosonization theory[@8] which can exactly calculate the action and the correlation functions of the quantum many-particle systems, we can establish an unified theory of the quantum many-particle systems, which is valid not only for weakly correlated fermion/boson systems, but also for strongly correlated fermion/boson systems. The key points of this unified theory are that: 1). we introduce the Lagrangian multiplier (boson field) $\phi (x)$ which takes the particle density $\rho (x)=\psi ^{\dagger }(x)\psi (x)$ as a constraint condition, so that the Hamiltonian can only have the quadratic form of the particle (fermion/boson) operators. 2). by introducing phase field $Q_k(x,\xi )$ which is a functional of the boson field $\phi (x)$, we solve the eigen-functional equation of the particles, so that we can use the phase field to completely represent the kinetic energy of the systems. 3). the phase field $Q_k(x,\xi )$ is a key parameter hidden in the quantum many-particle systems, its imaginary part represents the particle correlation strength, and its real part only contributes to the ground state energy and action of the systems. 4). we are able to use the particle density field $\rho (x)$ to exactly represent the ground state energy and action of the quantum many-particle systems. L. D. Landau, Sov. Phys. JEPT, [**3**]{}, 920(1956); [**5**]{}, 101(1957); [**8**]{}, 70(1959). P. Nozières, [*Theory of Interacting Fermi System*]{}, (Addison-Wesley, 1974). S. Tomonaga, Prog. Theor. Phys. [**5**]{}, 544(1950). J. M. Luttinger, J. Math. Phys. [**4**]{}, 1154(1963). D. C. Mattis and E. H. Lieb, J. Math. Phys. [**6**]{}, 304(1965). C. M. Varma [*et. al*]{}., Phys. Rev. Lett. [**63**]{}, 1986(1989). P. W. Anderson, [*The Theory of Superconductivity in the High-Tc Cuprates*]{}, (Princeton Univ. Press, NJ, 1997). Y. L. Liu, [*“Unified theory of strongly correlated electron systems”*]{}, preprint, cond-mat/0011254. P. Hohenberg and W. Kohn, Phys. Rev. [**136**]{}, 864(1964). W. Kohn and L. J. Sham, Phys. Rev. [**140**]{}, 1133(1965). R. M. Dreizler and E. K. U. Gross, [*Density Functional Theory*]{}, (Springer-Verlag, 1990). A. Luther and V. J. Emery, Phys. Rev. Lett. [**33**]{}, 389(1974). F. D. M. Haldane, J. Phys. C[**14**]{}, 2585(1981). R. B. Laughlin, Phys. Rev. Lett. [**50**]{}, 1395(1983). R. F. Prange and S. M. Girvin (Eds.), [*The Quantum Hall Effect*]{}, (Springer-Verlag, 1987). G. D. Mahan, [*Many-Particle Physics*]{}, (Plenum, NY, 1981). P. Nozières and D. Pines, [*The Theory of Quantum Liquids, Volume II: Superfluid Bose Liquids*]{}, (Addison-Wesley, 1990).
{ "pile_set_name": "ArXiv" }
--- abstract: 'In this article we obtain new results for the task of converting a *single* $N$-qubit W-class state (of the form $\sqrt{x_0}{|00...0\rangle}+\sqrt{x_1}{|10...0\rangle}+...+\sqrt{x_N}{|00...1\rangle}$) into maximum entanglement shared between two random parties. Previous studies in random distillation have not considered how the particular choice of target pairs affects the transformation, and here we develop a strategy for distilling into *general* configurations of target pairs. We completely solve the problem of determining the optimal distillation probability for all three qubit configurations and most four qubit configurations when $x_0=0$. Our proof involves deriving new entanglement monotones defined on the set of four qubit W-class states. As an additional application of our results, we present new upper bounds for converting a generic W-class state into the standard W state ${|W_N\rangle}=\sqrt{\frac{1}{N}}\left({|10...0\rangle}+...+{|00...1\rangle}\right)$.' author: - 'W. Cui' - 'E. Chitambar' - 'H.K. Lo' bibliography: - 'EricQuantumBib.bib' title: 'Randomly distilling W-class states into general configurations of two-party entanglement' --- [^1] Introduction ============ In quantum information processing, the two-qubit EPR state ${|\Phi\rangle}=\sqrt{\frac{1}{2}}\left({|00\rangle}+{|11\rangle}\right)$ provides a key resource for performing non-classical tasks such as teleportation [@Bennett-1993a] and super-dense coding [@Bennett-1992b]. Thus, for a multi-partite state ${|\varphi\rangle}_{1...N}$, it is important to know the optimal ways in which EPR entanglement can be obtained between two parties without having to introduce any more entanglement into the system. This latter constraint is known as the LOCC constraint because it requires each party to perform only local quantum operations (LO) while coordinating their operations through classical communication (CC). In general, the optimal conversion of ${|\varphi\rangle}_{1...N}$ into bipartite entanglement depends on which two final parties are left sharing the entanglement. One scenario to consider is when two specific parties are designated as the target pair, and the transformation is considered a success if and only if these two parties end up sharing the state ${|\Phi\rangle}$. A transformation of this sort is known as a **specified-pair distillation**. In this setting, an important problem is to determine the greatest success probability $p_{ij}$ for which the conversion $${|\varphi\rangle}_{1...N}\to{|\Phi^{(ij)}\rangle}$$ is possible by LOCC. Here, ${|\Phi^{(ij)}\rangle}$ denotes an EPR pair between parties $i$ and $j$, and it is assumed that all other parties are in some product state. While no full solution to this problem is known, some partial results exist [@Gour-2005a; @Chitambar-2010a]. ![\[RDABACtutorial\] A specified-pair versus random-pair distillation. For random distillations, it is convenient to combine all the desired outcomes into one configuration graph $\mathcal{G}=(V,E)$ whose edge set encodes the target pairs. Here, the target pairs are AB and AC. The “$\equiv$” indicates equivalent representations.](RDABACtutorial.png) A more general question can be posed by allowing the two EPR-entangled parties to vary across the different outcomes in the transformation (see Fig. \[RDABACtutorial\]). Any transformation of this form is known as a **random-pair distillation** (or just simply *random distillation*) because the final two entangled parties are *a priori* unspecified. Additional constraints to the problem can be added by demanding that the possible target pairs be limited only to some particular subset of all possible pairs. For example, in the random distillation of Fig. \[RDABACtutorial\], the transformation is considered a success only if AB or AC obtain an EPR pair, and not if BC become EPR entangled. For an $N$-party system, a random distillation can be written as $$\label{Eq:trans} {|\varphi\rangle}_{1...N}\to\{p_{ij},{|\Phi^{(ij)}\rangle}\}_{(i,j)\in E}$$ where $p_{ij}$ is the probability of obtaining ${|\Phi^{(ij)}\rangle}$ and $E\subset [N]\times [N]$ is some designated set of target bipartite pairs. The transformation is considered a success if an EPR state is obtained by any pair in $E$. A convenient way to represent random distillations is through a **configuration graph** $\mathcal{G}=(V,E)$. Each party $k$ is identified with a node $v_k\in V$, and an edge $e_{jk}\in E$ connects $v_j$ and $v_k$ if parties $j$ and $k$ form a desired target pair in the distillation (see Fig. \[RDABACtutorial\]). It should be emphasized that we are strictly dealing with a single copy of ${|\varphi\rangle}_{1...N}$, and each edge corresponds to one possible outcome. Variations to this question in the asymptotic regime have been studied elsewhere [@Smolin-2005a; @Yang-2009a]. Given some graph $\mathcal{G}$ and initial state ${|\varphi\rangle}_{1...N}$, the greatest success probability is given by: $$P(\varphi,\mathcal{G}):=\sup\sum_{(i,j)\in E} p_{ij}$$ where the supremum is taken across all LOCC protocols. ![\[RDcomplete\] An $N=8$ example of the “complete-type” distillations considered by Fortescue and Lo in Ref. [@Fortescue-2007a]. Such a transformation is a success if *any* two parties become EPR entangled, and this can be achieved with a probability arbitrarily close to one. Previous research has not considered more general types of configuration graphs than this.](RDcomplete) The subject of single-copy random distillation was first introduced and subsequently studied by Fortescue and Lo [@Fortescue-2007a; @Fortescue-2008a; @Fortescue-2009a]. One prominent finding of their work is that random distillations are, in general, strictly more powerful than specified-pair distillations. Perhaps the most dramatic example of this is the $N$-qubit state ${|W_N\rangle}=\sqrt{\frac{1}{N}}\left({|10...0\rangle}+{|01...0\rangle}+...+{|00...1\rangle}\right)$ and its random distillation into EPR pairs shared between any two parties (see Fig. \[RDcomplete\]). In terms of the terminology introduced above, the initial state is ${|W_N\rangle}$, and the configuration graph is complete (each node connected to every other) such that the conversion is a success if any two parties become EPR entangled. Fortescue and Lo were able to show that this transformation can be completed with probability arbitrarily close to one [@Fortescue-2007a]. On the other hand, for any two parties, the optimal specified-pair distillation probability is $2/N$. In this article, we turn to one largely unexplored question in Fortescue and Lo’s work which is the random distillation to *general* configuration graphs, and not just complete graphs. Specifically, we consider how the target configuration graph affects the random distillation in terms of overall success probability as well as the actual LOCC protocol the parties implement during the transformation. For example, one particular problem we are able to solve is the four qubit random distillation depicted in Fig. \[RDABCDtutorial\] which was left as an open problem in Ref. [@Fortescue-2009a; @Oppenheim-2011a]. ![\[RDABCDtutorial\] In Section \[Sect:fourqubits\] we show that the optimal LOCC probability of achieving this transformation is $2/3$, thus resolving an open problem in Refs. [@Fortescue-2009a; @Oppenheim-2011a]. The initial state is ${|W_4\rangle}=1/2({|1000\rangle}+{|0100\rangle}+{|0010\rangle}+{|0001\rangle})$.](RDABCDtutorial) Our focus is on the single-copy random distillation of $N$-party W-class of states, which is the collection of all states reversibly obtainable from ${|W_N\rangle}$ with a nonzero probability by LOCC. The choice to limit investigation to this class of states is motivated by multiple factors. First, from an experimental perspective, W-type entanglement seems relatively easier to generate than other forms of multipartite entanglement, with the state ${|W_4\rangle}$ already being realized in the laboratory [@Wieczorek-2009a]. In $N>4$ qubit systems, setups have also been proposed for the production of W-class states [@Bastin-2009a]. And for the particular task of random distillation, Fortescue has devised an experimental implementation of W-type random distillation using currently available technology, e.g. ion trap quantum computers [@Fortescue-2009a]. Second, as we will see in the next section, W-class states have a very simple structure which allows us to carefully analyze their behavior under LOCC evolution. Finally, a large amount of previous research conducted by Fortescue and Lo on random distillations involved W-class states. Thus, there is an established point of comparison for new results on the subject. We summarize our results and outline the structure of this article. - In Section \[Sect:FLNotation\], we begin by reviewing the results of the Fortescue-Lo Protocol and a described generalization, as well as some related work by Kintaş and Turgut on entanglement transformations within the W-class [@Kintas-2010a]. - In Section \[Sect:LPO\] we construct the “Least Party Out” Protocol which distills an arbitrary $N$-qubit W-class state given some target configuration $\mathcal{G}$. The protocol is similar in nature to the Fortescue-Lo Protocol but we show it to be strictly stronger. - In Section \[Sect:fourqubits\], we apply our protocol to three and four qubit systems to obtain the main results of the article. Every possible three and four qubit configuration graph is considered, and we introduce new four qubit entanglement monotones to show that our protocol is optimal in most cases when $x_0=0$. - In Section \[Sect:standardW\], we further apply our results to study the transformation ${|\varphi\rangle}_{1...N}\to{|W_N\rangle}$ where ${|\varphi\rangle}$ is a generic W-class state. New upper bounds on the optimal conversion probability are obtained. - In the Conclusion, we return to the question of LOCC versus separable operations investigated more heavily in our companion paper [@Chitambar-2011b]. Throughout this article, we will also recall a few other results from that paper. Previous Results and Notation {#Sect:FLNotation} ============================= The Generalized Fortescue-Lo Protocol {#the-generalized-fortescue-lo-protocol .unnumbered} ------------------------------------- In Ref. [@Fortescue-2007a], Fortescue and Lo developed a protocol which randomly distills the state ${|W_N\rangle}=\sqrt{\frac{1}{N}}\left({|10...0\rangle}+{|01...0\rangle}+...+{|00...1\rangle}\right)$ according to a complete configuration graph with success probability arbitrarily close to one (see Fig. \[RDcomplete\]). We briefly review the case when $N=3$. For some $\epsilon>0$, the parties locally perform the measurement given by $M_1=diag[\sqrt{1-\epsilon},1]$ and $M_2=diag[\sqrt{\epsilon},0]$. If all parties obtain outcome “1”, the final state is the original state ${|W_3\rangle}$. The parties then repeat the same measurement again. On the other hand if only two parties obtain outcome “1”, then this pair is left EPR entangled. But, if one or fewer parties obtain outcome “1”, all entanglement is destroyed and transformation is a failure. With the possible recursive step, this protocol can continue for an indefinite number of measurement rounds. In the end, the total probability of obtaining some EPR pair is $1-O(\epsilon)$. For $N>3$, the protocol generalizes and likewise the probability of success is $1-O(\epsilon)$. Here, the probabilities are distributed equally among all possible pairs; that is, with probability $\binom{N}{2}^{-1}-O(\epsilon)$, any two parties $i$ and $j$ obtain an EPR pair. In Ref. [@Fortescue-2009a], Fortescue briefly considered the problem of applying their protocol to more general configuration graphs, but only a limited discussion is given. Nevertheless, for a general outcome configuration graph $\mathcal{G}=(V,E)$, we can here describe an obvious way to apply the Fortescue-Lo Protocol. Starting with the state ${|W_N\rangle}$, it is converted with equal probability into the $\binom{N}{N-1}$ different ${|W_{N-1}\rangle}$ states. Here the difference between these states lies in which of the $N$ parties are entangled. If all the entangled parties in a particular ${|W_{N-1}\rangle}$ state are connected according to the graph $\mathcal{G}$, then the state is broken into EPR pairs with probability arbitrarily close to one. Otherwise, it is converted into the $\binom{N-1}{N-2}$ different ${|W_{N-2}\rangle}$ states. This process continues until ${|W_3\rangle}$ states are obtained. Either all these parties sharing the ${|W_3\rangle}$ state are connected in $\mathcal{G}$ or at most two are. In the former case, EPR pairs are obtained with probability $\approx 1$ whereas in the former, the distillation can be completed with probability $2/3$. We will let $P_{FL}(W_N,\mathcal{G})$ denote the distillation success probability of this Generalized Fortescue-Lo Protocol for some configuration $\mathcal{G}$. Obviously $P(W_N,\mathcal{G})\geq P_{FL}(W_N,\mathcal{G})$. The “Least Party Out” protocol described in the next section will be able to obtain a greater success probability than $P_{FL}(W_N,\mathcal{G})$ in general, and thus tighten the lower bound on $P(W_N,\mathcal{G})$. Additional notation and the K-T Monotones {#additional-notation-and-the-k-t-monotones .unnumbered} ----------------------------------------- In an $N$-partite system, if a “standard” W state ${|W_M\rangle}$ is shared among parties $S\subset [N]:=\{1,2,...,N\}$ with $|S|=M$, we will often explicate this by writing ${|W^{(S)}_{|S|}\rangle}$. Equivalently we can write this state as ${|W_{|S|}^{(\overline{T})}\rangle}$ where $T=[N]\setminus S$. Also, we define $${|W_2^{(ij)}\rangle}:=\sqrt{\tfrac{1}{2}}\left({|01\rangle}_{ij}+{|10\rangle}_{ij}\right)$$ which is local unitarily (LU) equivalent to ${|\Phi^{(ij)}\rangle}$. We will often represent a generic W-class state $\sqrt{x_0}{|00...0\rangle}+\sqrt{x_1}{|01...0\rangle}+...+\sqrt{x_n}{|00...1\rangle}$ by an $N$-component vector: $$\begin{aligned} \vec{x}=(x_1,&x_2,...,x_N)\notag\\ &\updownarrow\notag\\ \sqrt{x_0}{|00...0\rangle}+\sqrt{x_1}&{|10...0\rangle}+...+\sqrt{x_n}{|00...1\rangle},\end{aligned}$$ and $x_0=1-\sum_{i=1}^Nx_i$. More importantly, even after a basis change - ${|0\rangle}\to{|0'\rangle}$ and ${|1\rangle}\to{|1'\rangle}$ - the component values $\sqrt{x_i}$ always remain unchanged for $N\geq 3$ [@Kintas-2010a]. When $N=2$, uniqueness can be ensured by demanding that $x_0=0$ and $x_1\geq x_2$. Therefore, for any number of parties, the vector $\vec{x}$ uniquely specifies the state up to an LU transformation. For the state $\vec{x}$, we denote $$x_{n_1}=\max_{1\leq k\leq N} x_k.$$ By disregarding LU transformations and decomposing a general measurement into a sequence of binary outcome POVMS [@Anderson-2008a], we can assume that a local measurement by party $k$ consists of two upper triangular matrices $\{M^{(k)}_1, M^{(k)}_2\}$ whose entries are $$\begin{aligned} \label{Eq:constraints} M^{(k)}_1&=\begin{pmatrix}\sqrt{a_1} & b_1 \\0 & \sqrt{c_1}\end{pmatrix} & M^{(k)}_2&=\begin{pmatrix}\sqrt{a_2} & b_2 \\0 & \sqrt{c_2}\end{pmatrix}\end{aligned}$$ with $a_1+a_2=1$ and $c_1+c_2\leq 1$, where equality is achieved by the latter if and only if $M^{(k)}_1$ and $M^{(k)}_2$ are both diagonal. It is easy to see that this measurement by party $k$ on state $\sqrt{x_0}{|00...0\rangle}+\sqrt{x_1}{|10...0\rangle}+...+\sqrt{x_N}{|00...1\rangle}$ will transform the components as: $$\begin{aligned} \label{Eq:components} x_k&\to\frac{c_\lambda}{p_\lambda}x_k, &x_j&\to\frac{a_\lambda}{p_\lambda}x_j\;\;\;1\leq j\not=k\leq N,\end{aligned}$$ with $p_\lambda$ being the probability that outcome $\lambda$ occurs. From this it is easy to see the following, $$\begin{aligned} \label{Eq:KTmono} x_0&\leq\sum_\lambda p_\lambda x_{\lambda,0}& x_i&\geq\sum_\lambda p_\lambda x_{\lambda,i}\end{aligned}$$ for all $1\leq i\leq N$. We will refer to these as the K-T monotones after Kintaş and Turgut who first proved the inequalities [@Kintas-2010a]. The “Least Party Out” Protocol {#Sect:LPO} ============================== Here we describe our W-class random distillation protocol for a given configuration graph $\mathcal{G}$. It’s called the “Least Party Out” (LPO) protocol because it involves systematically removing parties from the $N$-party entanglement with a probability that decreases according to the number of edges connected to the party’s node in $\mathcal{G}$. For some group of parties $S$, we let $\mathcal{G}\setminus S$ denote the subgraph of $\mathcal{G}$ obtained by removing the nodes corresponding to the parties in $S$. Our protocol can be divided into three phases. Phase I takes a generic W-class state $\vec{x}$ and converts it into a state $\vec{x'}$ such that $x'_0=0$. Phase II converts an $x_0=0$ W-class state into standard W states ${|W_{|S|}^{(S)}\rangle}$ for $2\leq |S|\leq N$ using an “equal or vanish” (e/v) measuring scheme. Phase III then converts the standard W states into EPR pairs given by the configuration graph $\mathcal{G}$. Phase III is largely inspired by the Fortescue-Lo Protocol in that it involves an indefinite round measurement procedure: each party performs a measurement which, with some probability, leaves the state invariant and thus subject to a repeated round of identical measurement, which again leaves the state invariant with some probability, etc. **Phase I: Remove $x_0$ component:** Input $(\vec{x},\mathcal{G})$ where $\vec{x}$ is an $N$-partite W-class state and $\mathcal{G}$ is some configuration graph with $N$ nodes. If $x_0=0$, proceed to Phase II. Otherwise, choose some party $n_1$ with the largest component value to perform the measurement with $a_1=c_1=\lambda$, $b_1=-\sqrt{\lambda\frac{x_0}{x_1}}$, and $c_2=0$. The values for $a_2$, $b_2$ and $\lambda$ are fixed by the measurement being complete [@Cui-2010a]. Outcome “1” occurs with probability $$\label{Eq:x0prob} \frac{2x_{n_1}(1-x_0)}{x_0+2x_{n_1}+\sqrt{x_0^2+4x_{n_1}x_0}}.$$ and the resultant state has no zeroth component. For outcome “2”, the state is either a product state, in which case the protocol halts as a failure, or the state is entangled with party $n_1$’s component being zero and the state still having a zeroth component. In both cases, redefine $\vec{x}$ as the post-measurement state, but set $\mathcal{G}$ as $\mathcal{G}\setminus n_1$ only after outcome “2”. Repeat Phase I with input $(\vec{x},\mathcal{G})$. **Phase II: “Equal or Vanish” (e/v) Subroutine:** Input $(\vec{x},\mathcal{G})$ where $x_0=0$ and $\vec{x}$ is shared between $2\leq|S|\leq N$ parties. - If there does not exist an isolated node $v_k$ in $\mathcal{G}$ (one without any outgoing edges), proceed to the next step (2). Otherwise, when $|S|=2$ the protocol halts as a failure, and when $|S|>2$, the “isolated” party “k” performs the dis-entangling measurement $M^{(k)}_1=diag[1,0]$ and $M^{(k)}_2=diag[0,1]$. If outcome “1” occurs, redefine $\vec{x}$ as the post-measurement state and set $\mathcal{G}$ as $\mathcal{G}\setminus k$; repeat the e/v subroutine on input $(\vec{x},\mathcal{G})$. If outcome “2” happens, the protocol terminates as a failure. - If every component in $\vec{x}$ is maximal, then $\vec{x}$ is a standard W state ${|W_{|S'|}^{(S')}\rangle}$ and proceed to Phase III. Otherwise, choose some party $k$ such that (i) $x_k$ is non-maximal, and (ii) party $k$ is connected to some party whose component is maximal. If no party satisfies both these conditions, then choose some party $k$ which just satisfies condition (i). He/she then performs a two-outcome measurement with operators $M_1^{(k)}=diag[\sqrt{\tfrac{x_{k}}{x_{n_1}}},1]$ and $M_2^{(k)}=diag[\sqrt{1-\tfrac{x_{k}}{x_{n_1}}},0]$. Party $k$’s component value equals the maximum upon outcome “1” and vanishes upon outcome “2”. In both cases, redefine $\vec{x}$ as the post-measurement state, but set $\mathcal{G}$ as $\mathcal{G}\setminus k$ only after outcome “2”. Repeat the e/v subroutine on the new input $(\vec{x},\mathcal{G})$ (see Fig. \[evfigtree\]). For Phase II input $(\vec{x},\mathcal{G})$, the final success states of the e/v subroutine are ${|W^{(S')}_{|S'|}\rangle}$ for $2\leq |S'|\leq |S|$ where for each party $*$ such that $x_*=x_{n_1}$, either $*\in S'$ or no party in $S'$ is connected to $*$ in $\mathcal{G}$. The latter case occurs when there is only one party with a maximum component and all parties connected to $*$ measure a “vanish” outcome; consequently, $*$ becomes an isolated party and removes itself from the system via step (1) above. Let $$\begin{aligned} \lambda_{\vec{x},\mathcal{G}}(W_{|S'|}^{(S')}):=&\text{ the probability or obtaining ${|W^{(S')}_{|S'|}\rangle}$}\notag\\ &\text{ via the e/v subroutine for input $(\vec{x},\mathcal{G})$.}\end{aligned}$$ Note that $\lambda_{\vec{x},\mathcal{G}}$ is a smooth function of the component values $x_i$ and can be explicitly computed from the measurement operators given above. For example, $\lambda_{\vec{x},\mathcal{G}}(W_N)=N\frac{\prod_{k\not=n_1}x_k}{x_{n_1}^{N-2}}$. Also, $\lambda_{W_N,\mathcal{G}}(W_M)=\delta_{MN}$. ![\[evfigtree\] “Equal or Vanish” Subroutine (Phase II) for the normalized state $\frac{1}{1+3\alpha}(\alpha,\alpha,\alpha,1)$ and the configuration graph with edges $\{AB,AC,AD, BC\}$. 1. David’s component is largest and Alice is a connected to him with a lesser component value. She performs an e/v measurement. 2. For outcome “vanish” (right branch) she is separated from the system, and since David is not connected to either Bob or Charlie, he immediately removes himself from the system leaving ${|\Phi^{(BC)}\rangle}$ with some probability. For outcome ”equal” (left branch) the components of all other parties receive a factor of $\alpha$, and Alice’s component is now maximum equaling David’s. Bob is a connected party to Alice with a lesser component value and he performs an e/v measurement. 3. Again, either Bob vanishes (right branch) or all other components except his receive a factor of $\alpha$. In both cases, Charlie is then a connected party to Bob with a lesser component value and he performs an e/v measurement. 4. The final outcome states along these branches are ${|W_4\rangle}$, ${|W_3^{(ABD)}\rangle}$, ${|W_3^{(ACD)}\rangle}$, and ${|\Phi^{(AD)}\rangle}$.](evfigtree) **Phase III: Obtaining EPR Pairs:** Input $(W_{|S|}^{(S)},\mathcal{G})$ with $\mathcal{G}$ having $|S|$ nodes and at least one edge. If $|S|=2$, the state is an EPR pair and protocol halts as a success. Otherwise, Phase III of the protocol is defined recursively such that for $|S|>2$, the procedure depends on a pre-defined random distillation protocol for ${|W_{|S'|}^{(S')}\rangle}$ with $|S'|<|S|$. Let $$\begin{aligned} P_{III}(W_{|S|}^{(S)},\mathcal{G}):=&\text{ the probability of distilling ${|W_{|S|}^{(S)}\rangle}$}\notag\\ &\text{ into $\mathcal{G}$ via the Phase III procedure.}\notag\end{aligned}$$ If $|S|=2$, set $P_{III}(W_2,\mathcal{G})=1$ by definition. For $|S|>2$, identify the party $k$ whose node in $\mathcal{G}$ has a least number of connected edges. He/she performs the measurement with operators given by $M_1^{(k)}=diag[\sqrt{\alpha},1]$ and $M_2^{(k)}=diag[\sqrt{1-\alpha},0]$ where $\alpha$ is determined according to the discussion below Eq. . Outcome “2” occurs with probability $(1-\alpha)\frac{|S|-1}{|S|}$ and the resultant state is ${|W_{|S|-1}^{(\overline{k})}\rangle}$. Phase III is then repeated on this state and the reduced graph $\mathcal{G}\setminus k$. ![\[Phase3fig\] Phase III receives an input state $W_{|S|}^{(S)}$ and a configuration graph $\mathcal{G}$. Party $k$ performs an e/v measurement. One outcome is a standard $W$ state with party $k$ removed, and the other is the state $\tfrac{1}{|S|p_\alpha}(\alpha,...,\alpha,1,\alpha,...,\alpha)$. Phase II is applied on this state outputting either W states or a product (failure) state. Phase III will next be initiated on each of the W states, and for any W state $W_{|S'|}^{(S')}$ with $|S'|<|S|$, the transformation success probability from this point onward is given by $P_{III}(W_{|S'|}^{(S')},\mathcal{G}\setminus \overline{S'})$; this value is already known by recursion. However, for the state $W_{|S|}^{(S)}$, performing Phase III again will generate an indefinite loop, but one whose overall success probability converges to $\frac{f(\alpha)}{1-\alpha^{|S|-1}}$ (see Eqs. and ).](Phase3fig) Outcome “1” happens with probability $p_\alpha=\frac{1+\alpha(|S|-1)}{|S|}$ and the post-measurement state is $\vec{y}_\alpha$, which (up to a permutation between party 1 and $k$) takes the form: $\tfrac{1}{|S|p_\alpha}(1,\alpha,\alpha,...,\alpha)$. Party $k$ then has the largest component value in $\vec{y}_\alpha$, and the e/v subroutine (Phase II) is performed on the input $(\vec{y}_\alpha,\mathcal{G})$. The e/v subroutine will either output the states ${|W^{(S')}_{|S'|}\rangle}$ where $|S'|<|S|$ with respective probabilities $\lambda_{\vec{y}_\alpha,\mathcal{G}}(W^{(S')}_{|S'|})$, or the original state ${|W_{|S|}^{(S)}\rangle}$ with probability $\alpha^{|S|-1}$. In the former case, the Phase III procedure is performed on the input $(W^{(S')}_{|S'|},\mathcal{G}\setminus \overline{S'})$. Accounting for all states with $|S'|<|S|$, their total distillation success probability is $$\begin{aligned} \label{Eq:falpha} &f(\alpha) =(1-\alpha)\tfrac{|S|-1}{|S|}\cdot P_{III}(W_{|S|-1}^{(\overline{k})},\mathcal{G}\setminus k)\notag\\ &+\tfrac{1+\alpha(|S|-1)}{|S|}\sum_{2\leq |S'|<|S|}\lambda_{\vec{y}_\alpha,\mathcal{G}}(W^{(S')}_{|S'|}) P_{III}(W^{(S')}_{|S'|},\mathcal{G}\setminus \overline{S'}),\end{aligned}$$ where the sum is taken over all subsets $S'$ such that $2\leq |S'|<|S|$ and either $k\in S'$ or no party in $S'$ is connected to $k$ in $\mathcal{G}$. If the e/v subroutine outputs the original ${|W_{|S|}^{(S)}\rangle}$, repeat Phase III again on the same input $(W_{|S|}^{(S)},\mathcal{G})$. This will generate an indefinite loop in which for each cycle, the probability of distillation success is $f(\alpha)$, and the probability of continuing on for another cycle is $\alpha^{|S|-1}$. Therefore, the total success probability across all cycles is given by the geometric sum $\sum_{r=0}^\infty[\alpha^{|S|-1}]^rf(\alpha)$. To maximize this value, we set $$\begin{aligned} \label{Eq:probG} P_{III}(W_{|S|}^{(S)},\mathcal{G})&=\sup_{0\leq \alpha< 1}\frac{f(\alpha)}{1-\alpha^{|S|-1}}.\end{aligned}$$ This determines the original value of $\alpha$ in the Phase III measurement operators: if obtains its supremum in the interval $[0,1)$, then $\alpha$ is chosen to be any of these critical points; if the supremum is obtained in the limit $\alpha\to 1$, then $\alpha=1-\epsilon$ for any desired $\epsilon>0$. The smaller the value of $\epsilon$, the closer the success probability approaches $P_{III}(W_{|S|}^{(S)},\mathcal{G})$. Observe that when the supremum is obtained at $\alpha=0$, the first measurement by party $k$ will deterministically dis-entangle the party from the rest of the system. It is also important to note that the optimization of can always be efficiently performed. By the recursive construction, the values for $P_{III}(W_{|S|-1}^{(\overline{k})},\mathcal{G}\setminus k)$ and $P_{III}(W^{(S')}_{|S'|},\mathcal{G}\setminus \overline{S'})$ are just real numbers known *a priori*. Furthermore, the functions $\lambda_{\vec{y}_\alpha,\mathcal{G}}(W^{(S')}_{|S'|})$ are smooth functions of $\alpha$, and thus, Eq. represents a single-variable smooth function whose supremum value can be easily computed. In total then, for a state $N$-partite state $\vec{x}$ with $x_0=0$ and configuration graph $\mathcal{G}$, the overall success probability of the LPO protocol is given by $$\label{Eq:LPOprob} P_{LPO}(\vec{x},\mathcal{G}):=\sum_{2\leq |S'|\leq N}\lambda_{\vec{x},\mathcal{G}}(W^{(S')}_{|S'|})\cdot P_{III}(W^{(S')}_{|S'|},\mathcal{G}\setminus \overline{S'})$$ where the sum is taken over all subsets $S'$ such that $2\leq |S'|\leq N$ and for each party $*$ such that $x_*=x_{n_1}$, either $*\in S'$ or no party in $S'$ is connected to $*$ in $\mathcal{G}$. Main results: the LPO Protocol on three and four qubits {#Sect:fourqubits} ======================================================= [**Summary of Results:**]{} Before working through the LPO Protocol in detail on three and four qubit systems, we summarize the overall results. For three qubits, the possible configuration graphs are depicted in Fig. \[RDABAC\], and upper bounds on the transformation success probabilities are given by Eqs. and respectively. In both cases, when $x_0=0$ these bounds can be approached arbitrarily close. For four qubits, there are six different families of configurations depicted in Figs. \[RDABACAD\] – \[RDAll\]. For states with $x_0=0$, we have completely solved the random distillation problem for all configurations except VI. Upper bounds on configurations I – V are given by Eqs. – respectively. Three qubits {#three-qubits .unnumbered} ------------ ![\[RDABAC\] (Left) Configuration $\mathcal{G}_{\wedge}$. (Right) Configuration $\mathcal{G}_{\triangle}$. An upper bound on the success probability is given by Eqs. and respectively which is effectively tight when $x_0=0$. For ${|W_3\rangle}$, these probabilties are $2/3$ and $1$ respectively.](RDABAC) As a first example of the LPO protocol, consider the state ${|W_3\rangle}$ and the configuration graph given by $\mathcal{G}_{\wedge}$ in Fig. \[RDABAC\]. In Phase III of the protocol, we can choose the “least party” to be either Bob or Charlie (say it’s Charlie). He performs a measurement as described above, and either ${|\Phi^{(AB)}\rangle}$ is obtained or the post-measurement state is $\vec{y}_\alpha=\tfrac{1}{2\alpha+1}(\alpha,\alpha,1)$. For the latter, the e/v subroutine obtains ${|\Phi^{(AC)}\rangle}$ with probability $\lambda_{\vec{y}_\alpha,\mathcal{G}_\wedge}(W_2^{(AC)})=\frac{2\alpha(1-\alpha)}{2\alpha+1}$. Thus, we have $f(\alpha)=\frac{2}{3}(1-\alpha)+\frac{2}{3}(1-\alpha)\alpha$ and therefore $\frac{f(\alpha)}{1-\alpha^{2}}$ takes a constant value of $\frac{2}{3}$. Hence, $\alpha$ can be chosen as $0$ in Charlie’s measurement and $$\label{Eq:W3wedge} P_{III}(W_3,\mathcal{G}_\wedge)=\frac{2}{3}.$$ For a more general state $\vec{x}=(x_A,x_B,x_C)$ with $x_A\geq x_B\geq x_C$ and $x_0=0$, we have $\lambda_{\vec{x},\mathcal{G}_\wedge}(W^{(AB)}_{2})=2x_B(1-\frac{x_C}{x_A})$, $\lambda_{\vec{x},\mathcal{G}_\wedge}(W^{(AC)}_{2})=2x_C(1-\frac{x_B}{x_A})$, and $\lambda_{\vec{x},\mathcal{G}_\wedge}(W_{3})=3\frac{x_Bx_C}{x_A}$. Therefore, by Eq. \[Eq:LPOprob\], the distillation probability is given by $$\begin{aligned} \label{Eq:RDABAC} P_{LPO}(\vec{x},\mathcal{G}_\wedge)&=2x_B(1-\frac{x_C}{x_A})+2x_C(1-\frac{x_B}{x_A})+2\frac{x_Bx_C}{x_A}\notag\\ &=2x_B+2x_C-2\frac{x_Bx_C}{x_A}.\end{aligned}$$ One might wonder if this probability is optimal. It turns out that the answer is yes. See Eq. below and the discussion there. On the other hand, consider configuration $\mathcal{G}_\triangle$ given in Fig. \[RDABAC\]. We can still choose Charlie as the “least” party, and this time, the possible EPR success states of the e/v subroutine are ${|\Phi^{(AC)}\rangle}$ and ${|\Phi^{(BC)}\rangle}$. We have $\lambda_{\vec{y}_\alpha,\mathcal{G}_\triangle}(W_2^{(AC)})=\lambda_{\vec{y}_\alpha,\mathcal{G}_\triangle}(W_2^{(AC)})=\frac{2\alpha(1-\alpha)}{2\alpha+1}$ and so $f(\alpha)=\frac{2}{3}(1-\alpha)+\frac{4}{3}(1-\alpha)\alpha$. Thus, $$P_{LPO}(W_3,\mathcal{G}_\triangle)=\sup_{0\leq \alpha< 1}\frac{2}{3}\frac{(1-\alpha)(1+2\alpha)}{1-\alpha^{2}}=1.$$ This value is obtained in the limit $\alpha\to 1$ which means the LPO protocol calls for infinitesimal measurements with $\alpha=1-\epsilon$. Hence, for three qubits, the LPO protocol reduces to the Fortescue-Lo Protocol for distilling the state ${|W_3\rangle}$. Since $P_{III}(W_3,\mathcal{G}_\triangle)=1$ for the three-edge configuration, when considering the state $\vec{x}=(x_A,x_B,x_C)$ with $x_A\geq x_B\geq x_C$ and $x_0=0$, we obtain the distillation probability $$\begin{aligned} \label{Eq:ABACBC} P_{LPO}(\vec{x},\mathcal{G}_\triangle)=&2x_B(1-\frac{x_C}{x_A})+2x_C(1-\frac{x_B}{x_A})+3\frac{x_Bx_C}{x_A}\notag\\ &=2x_B+2x_C-\frac{x_Bx_C}{x_A}.\end{aligned}$$ Just as with the configuration graph $\mathcal{G}_\wedge$, this probability is optimal as we will see from Eq. below. We now turn to four qubits where, unlike the two cases just examined, there exists configurations for which $f(\alpha)$ obtains a maximum in the interval $(0,1)$. Four qubits {#four-qubits .unnumbered} ----------- Next, we apply the LPO protocol to four qubit W-class states. We will only consider a subset of possible configuration graphs, but any other can be obtained by a permutation of parties. **Configurations I (Fig. \[RDABACAD\]):** ![\[RDABACAD\] Let $\mathcal{G}_I$, $\mathcal{G}_I'$ and $\mathcal{G}_I''$ be the first, second and third of the above configurations respectively. An upper bound on the success probability is given by Eq. which is effectively tight when $x_0=0$. For ${|W_4\rangle}$, this probability is $1/4$ for each configuration.](RDABACAD) For a generic W-class state $\vec{x}=(x_A, x_B, x_C, x_D)$, whenever $x_A<x_j$ for some $j\in \{B,C,D\}$, an upper bound on distilling to any of these configurations is $2 x_{A}$ by the K-T monotones. However, when $x_A$ is the largest component value, we have $$\begin{aligned} \label{Eq:RDABACAD} P(\vec{x},\mathcal{G}_I)&\leq 2 x_{B}\notag\\ P(\vec{x},\mathcal{G}'_I)&\leq 2x_A-2\tfrac{(x_A-x_B)(x_A-x_C)}{x_A}\notag\\ P(\vec{x},\mathcal{G}''_I)&\leq 2x_A-2\tfrac{(x_A-x_B)(x_A-x_C)(x_A-x_D)}{x_A^2}\end{aligned}$$ as proven in Ref. [@Chitambar-2011b]. When $x_0=0$, these are precisely the rates obtained by the LPO Protocol, and so our protocol is optimal for such states. Note that setting $x_D=0$ proves to be optimal. **Configuration II (Fig. \[RDABBDAD\]):** ![\[RDABBDAD\] Let $\mathcal{G}_{II}$ be the above configuration. An upper bound on the success probability is given by Eq. which is effectively tight when $x_0=0$. For ${|W_4\rangle}$, this probability is $3/4$](RDABBDAD) For a generic W-class state $\vec{x}=(x_A, x_B, x_C, x_D)$, if we assume without loss of generality that $x_A\geq x_B\geq x_C$, then we have $$\label{Eq:RDABBDAD} P(\vec{x},\mathcal{G}_{II})\leq 1-x_0-x_D-\frac{(x_A-x_B)(x_A-x_C)}{x_A},$$ as also proven in Ref. [@Chitambar-2011b]. When $x_0=0$, the LPO protocol can approach this upper bound arbitrarily close. Note that this also proves Eq. to be optimal. **Configurations III (Fig. \[RDABBCCDAD\]):** ![\[RDABBCCDAD\] Let $\mathcal{G}_{III}$ be any of the above configurations. In each of these, $(A,C)$ and $(B,D)$ are unconnected pairs. An upper bound on the success probability is given by Eq. which is effectively tight when $x_0=0$. For ${|W_4\rangle}$, this probability is $2/3$ for each of these configurations.](RDABBCCDAD) A common feature to all of these configurations is that for each party, there is at least one other party to whom he/she is not connected. We will refer to such a pair as unconnected. For example $(A,C)$ and $(B,D)$ form unconnected pairs in each of the above configurations. We introduce the following entanglement monotones to put an upper bound on the probability for transformations of Configurations III. For a generic W-class state $\vec{x}=(x_A,x_B,x_C,x_D)$, let $n_1$ be some party whose component is maximum, $n_1'$ the party unconnected to $n_1$ with largest component value, and $p$ and $p'$ the other two parties. For definitiveness, if party $n_1$ has two unconnected parties (which is possible in the first two of Configurations III), take $p'$ to be the other one besides $n_1'$. Define the function $$\begin{aligned} \tau(\vec{x})&=2x_{p}+2x_{p'}-2\frac{x_{p}x_{p'}}{x_{n_1}}+\frac{2}{3}\frac{x_{p}x_{p'}x_{n_1'}}{x_{n_1}^2}.\notag\end{aligned}$$ Note that $\tau(W_4)=2/3$. \[Thm:tau\] The function $\tau$ is an entanglement monotone. See Appendix \[Apx:tau\] As a result of this theorem, we have that for a state $\vec{x}$, $$\label{Eq:tau} P(\vec{x},\mathcal{G}_{III})\leq \tau(\vec{x}).$$ For the LPO Protocol, first consider the initial state ${|W_4\rangle}$. In each of the configurations, either party A or D can be chosen as the “least” party. Regardless of the choice, we have $$\sum_{(i,j)\in E}\lambda_{\vec{y}_\alpha,\mathcal{G}_{III}}(W_2^{(ij)})=(4\alpha+2\alpha^2)(1-\alpha).$$ Here $E$ denotes the edge set of whatever configuration is considered. By Eq. , we also know that $P_{III}(W_3^{(BCD)},\mathcal{G}\setminus A)=2/3$. Thus, $$P_{III}(W_4,\mathcal{G}_{III})=\sup_{0\leq \alpha< 1}\frac{\tfrac{1}{2}+\alpha+\tfrac{1}{2}\alpha^2}{1+\alpha+\alpha^2}=\frac{2}{3}$$ which obtains this value as $\alpha\to 1$. Thus, for any $\epsilon>0$, $$2/3-\epsilon<P(W_4,\mathcal{G}_{III})\leq 2/3.$$ When $x=0$ and a state $\vec{x}=(x_A,x_B,x_C,x_D)$ is considered, it is straight forward to compute the probabilities for $\lambda_{\vec{x},\mathcal{G}_{III}}(W_{|S|}^{(S)})$. Doing so and using Eq. shows that the probability $\tau(\vec{x})$ can be approached arbitrarily close using the LPO protocol. **Configuration IV (Fig. \[RDABBDADACBC\]):** ![\[RDABBDADACBC\] Let $\mathcal{G}_{IV}$ be the above configuration. We say two parties are edge complementary if their nodes have a different number of connected edges. For example, $A$ is edge complementary to both $C$ and $D$. An upper bound on the success probability is given by Eq. which is effectively tight when $x_0=0$. For ${|W_4\rangle}$, this probability is $5/6$.](RDABBDADACBC) We first consider the transformation of the standard W state ${|W_4\rangle}$. The Phase II probabilities are $$\begin{aligned} \lambda_{\vec{y}_\alpha,\mathcal{G}_{IV}}(W_2^{(AD)})&=\frac{1}{1+3\alpha}\cdot2\alpha(1-\alpha)^2,\notag\\ \lambda_{\vec{y}_\alpha,\mathcal{G}_{IV}}(W_2^{(BD)})&=\frac{1}{1+3\alpha}\cdot2\alpha(1-\alpha)^2,\notag\\ \lambda_{\vec{y}_\alpha,\mathcal{G}_{IV}}(W_3^{(ABD)})&=\frac{1}{1+3\alpha}\cdot3\alpha^2(1-\alpha),\notag\\ \lambda_{\vec{y}_\alpha,\mathcal{G}_{IV}}(W_3^{(ACD)})&=\frac{1}{1+3\alpha}\cdot3\alpha^2(1-\alpha),\notag\\ \lambda_{\vec{y}_\alpha,\mathcal{G}_{IV}}(W_3^{(BCD)})&=\frac{1}{1+3\alpha}\cdot3\alpha^2(1-\alpha).\notag\end{aligned}$$ This gives $$P_{III}(W_4,\mathcal{G}_{IV})=\sup_{0\leq \alpha< 1}\frac{f(\alpha)}{1-\alpha^3}=\frac{\tfrac{3}{4}+\alpha+\tfrac{3\alpha^2}{4}}{1+\alpha+\alpha^2}=5/6$$ for which the value is obtained as $\alpha\to 1$. Consider a generic W-class state $\vec{x}=(x_A,x_B,x_C,x_D)$. We say that a party is edge complementary to a party if there corresponding nodes in $\mathcal{G}_{IV}$ have a different number of connected edges. For the particular configuration $\mathcal{G}_{IV}$, let $n_1$ denote some party with the largest component, $n_1'$ the party edge complementary to $n_1$ with the largest component, and $e_2$ and $e_3$ the other two parties having 2 and 3 edges respectively. Define the function: $$\begin{aligned} \Gamma(\vec{x})=\begin{cases} 2x_{e_3}+(x_{e_2}+x_{n_1'})\left(2-\frac{x_{e_3}}{x_{n_1}}\right)-2\frac{x_{n_1'}x_{e_2}}{x_{n_1}}\\\hspace{.5cm}+\frac{4}{3}\frac{x_{e_2}x_{e_3}x_{n_1'}}{x_{n_1}^2}\;\;\;\text{if $n_1$ has 3 connected edges},\\ 2x_{n_1'}+2x_{e_3}-\frac{x_{n_1'}x_{e_3}}{x_{n_1}}+\frac{x_{e_2}x_{e_3}x_{n_1'}}{3x_{n_1}^2}\;\;\;\text{if $n_1$ has 2}\\ \hspace{.5cm}\text{connected edges}. \end{cases}\end{aligned}$$ Note that $\Gamma(W_4)=5/6$. \[Thm:gamma\] The function $\Gamma$ is an entanglement monotone. The proof is nearly identical in structure to the one given for $\tau$ in Appendix \[Apx:tau\]. As a result, it immediately follows that $$\label{Eq:Gamma} P(\vec{x},\mathcal{G}_{IV})\leq \Gamma(\vec{x}).$$ And just as in the case of Configurations III, the LPO protocol can approach this upper bound arbitrarily close whenever $x_0=0$. Highlighting the standard W state, we have that for any $\epsilon>0$, $$5/6-\epsilon<P(W_4,\mathcal{G}_{IV})\leq 5/6.$$ It should be noted that 5/6 is the also the optimal transformation probability if one considers transformations within the more general class of separable operations [@Chitambar-2011b]. **Configuration V (Fig. \[RDAll\]):** ![\[RDAll\] Let $\mathcal{G}_{V}$ be the above configuration. An upper bound on the success probability is given by Eq. which is effectively tight when $x_0=0$. For ${|W_4\rangle}$, this probability is $1$.](RDAll) For the state ${|W_4\rangle}$, the Fortescue-Lo Protocol achieves this distillation configuration with probability arbitrarily close to one. For more general states, we recall the results from Ref. [@Chitambar-2011b]: $$\label{Eq:kappa} P(\vec{x},\mathcal{G}_{V})\leq 1-x_0-\frac{(x_A-x_B)(x_A-x_C)(x_A-x_D)}{x_A^2}$$ where we have assumed without loss of generality that $x_A$ is the largest component value. The LPO protocol can achieve this probability as close as desired whenever $x_0=0$. **Configuration VI (Fig. \[RDABBDADAC\]):** ![\[RDABBDADAC\] Let $\mathcal{G}_{VI}$ be the above configuration. For ${|W_4\rangle}$, the LPO Protocol gives a success probability of $\frac{1}{6}(3+\sqrt{3})$. We conjecture this to be optimal.](RDABBDADAC) For this configuration, we only work out the Phase III calculation for the standard W state ${|W_4\rangle}$. In this case, David is the “least” party and he measures first. Outcome “2” is the state ${|W^{(ABC)}_3\rangle}$ obtained with probability $3/4(1-\alpha)$; from here, we have $P_{III}(W_3^{(ABC)},\mathcal{G}_{VI}\setminus D)=1$. Outcome “1” is the state $\frac{1}{1+3\alpha}(\alpha,\alpha,\alpha,1)$. The ensuing e/v subroutine is described in Figure \[evfigtree\]. We have $$\begin{aligned} \lambda_{\vec{y}_\alpha,\mathcal{G}_{VI}}(W_2^{(BC)})&=\frac{1}{1+3\alpha}\cdot2\alpha(1-\alpha),\notag\\ \lambda_{\vec{y}_\alpha,\mathcal{G}_{VI}}(W_2^{(AD)})&=\frac{1}{1+3\alpha}\cdot2\alpha(1-\alpha)^2,\notag\\ \lambda_{\vec{y}_\alpha,\mathcal{G}_{VI}}(W_3^{(ABD)})&=\frac{1}{1+3\alpha}\cdot3\alpha^2(1-\alpha),\notag\\ \lambda_{\vec{y}_\alpha,\mathcal{G}_{VI}}(W_3^{(ACD)})&=\frac{1}{1+3\alpha}\cdot3\alpha^2(1-\alpha).\notag\end{aligned}$$ This gives $$\begin{aligned} P_{III}(W_4,\mathcal{G}_{VI})=\sup_{0\leq \alpha< 1}\frac{f(\alpha)}{1-\alpha^3}&=\frac{\tfrac{3}{4}+\alpha+\tfrac{\alpha^2}{2}}{1+\alpha+\alpha^2}\notag\\ &=\frac{1}{6}(3+\sqrt{3})\end{aligned}$$ which obtains this maximum when $\alpha=\frac{1}{2}(\sqrt{3}-1)$. The generalized Fortescue-Lo Protocol gives a rate of 3/4 so we see an improvement in our protocol. For an upper bound, it is known that this transformation cannot be accomplished with any probability greater than 5/6 by the more general class of separable operations [@Chitambar-2011b]. Thus, we summarize our result by $$\begin{aligned} P_{FL}(W_4,\mathcal{G}_{VI})<P_{LPO}(W_4,\mathcal{G}_{VI})&\approx\tfrac{3+\sqrt{3}}{6}\notag\\ &\leq P(W_4,\mathcal{G}_{VI})\leq \tfrac{5}{6}.\end{aligned}$$ We use the “$\approx$” symbol for the $P_{LPO}$ value since it can be approached arbitrarily close. While we conjecture that this protocol is optimal for the state ${|W_4\rangle}$ and the configuration graph $\mathcal{G}_{VI}$, unfortunately it does not appear optimal for more general four qubits states. Indeed, suppose that we begin with state $\vec{x}=(x_A,x_B,x_C,x_D)$ with $x_0=0$ and $x_A\geq x_B\geq x_C\geq x_D$. The LPO Protocol says that we should first perform the e/v subroutine with respect to party 1, and then implement Phase III on the state ${|W_4\rangle}$. The total probability is then given by Eq. . Explicitly computing it yields: $$\begin{aligned} P_{LPO}(\vec{x},\mathcal{G}_{VI})&=2(x_B+x_C+x_D)-\frac{x_Bx_D}{x_A}-2\frac{x_Bx_C}{x_A}\notag\\ &-2\frac{x_Cx_D}{x_A}+\frac{3+2\sqrt{3}}{3}\frac{x_Bx_Cx_D}{x_A^2}.\end{aligned}$$ Now suppose we have $x_A=1-3t$ and $x_B=x_C=x_D=t$ with $t<1/4$. Since Alice’s component is strictly greater than all other components, she can make a weak measurement such that her component value is still the largest in both post-measurement states. Specifically, when she performs the measurement given by Eq. with $(a_1,c_1,a_2,c_2)$ in some neighborhood of $(1/2,1/2,1/2,1/2)$, the average change in $P_{LPO}$ is $$\overline{\Delta P_{LPO}}=-(a-c)^2\cdot\frac{t^2}{1-3t}\left(20-t\frac{12-8\sqrt{3}}{1-3t}\right)$$ which can be positive for $t$ close to $1/4$. Therefore, $P_{LPO}$ cannot be the optimal probability for the initial state $\vec{x}$ since a weak measurement by Alice increases the overall transformation probability. It should be emphasized that for the transformation of ${|W_4\rangle}$ according to the LPO protocol, we never encounter a state like $\vec{x}$. The only time Alice’s component is larger than David’s is after David performs an e/v measurement and his component value is zero. Consequently, we still believe the protocol to optimal for ${|W_4\rangle}$. Application to the transformation varphi {#Sect:standardW} ======================================== For a generic W-class state ${|\varphi\rangle}_{1...N}$, there has been promising progress on the SLOCC transformation of ${|\varphi\rangle}_{1...N}\to{|W_N\rangle}$ since the discovery of the unique form possessed by multipartite W-class states [@Kintas-2010a]. However, the upper bound on the transformation success probability determined by the K-T monotones is not tight when the $x_0$ component of the initial state is not zero. A canonical example of this is the transformation of W-class state $\vec{x}=(tx_1, tx_2, \cdots, tx_n)$ into $\vec{y}=(x_1, x_2, ..., x_n)$ for $0<t<1$, which cannot be accomplished with probability $t$, and thus does not saturate the K-T monotones [@Kintas-2010a]. In the following, we improve on the general upper bound of $Nt$ set by the K-T monotones for the transformation ${|\vec{t}\rangle}\to{|W_N\rangle}$, where $\vec{t}=(t,t,\cdots, t)$. We do this by first considering the random distillation of $\vec{t}$ into EPR pairs between party 1 and any other party. \[lemma: x0nonzero\] The optimal LOCC success probability for randomly distilling the $N$-partite W-class state $\vec{t}=(t,t, \cdots, t), 0\leq t\leq \frac{1}{N}$ into EPR pairs between party 1 and any other party is upper bounded by $$p\leq 1-\sqrt{1-4(N-1)t^2}.$$ The proof is straightforward. We can “merge” together all parties other then party 1 so that we have a state unitarily equivalent to ${|\psi\rangle}=\sqrt{1-Nt}{|00\rangle}+\sqrt{t}{|10\rangle}+\sqrt{(N-1)t}{|01\rangle}$, whose smallest Schmidt component is $\sqrt{\tfrac{1-\sqrt{1-4(N-1)t^2}}{2}}$. Therefore, an upper bound on the probability for distilling EPRs across the bipartition $1:23\cdots N$ is $1-\sqrt{1-4(N-1)t^2}$. The following theorem then shows the desired result. The optimal LOCC transformation probability from $N$-partite W class state $\vec{t}=(t,t, \cdots, t), 0\leq t\leq \frac{1}{N}$ into the standard W state ${|W_N\rangle}=(\frac{1}{N}, \frac{1}{N}, \cdots, \frac{1}{N})$ is upper bounded by $$P(\vec{t}\rightarrow W_N)\leq \tfrac{N}{2}(1-\sqrt{1-4(N-1)t^2})< Nt.$$ We know that the optimal random-pair transformation of ${|W_N\rangle}$ into an EPR state shared between party 1 and some other party has probability $\frac{2}{N}$. If the transformation probability from $\vec{t}$ into ${|W_N\rangle}$ is higher than $\tfrac{N}{2}(1-\sqrt{1-4(N-1)t^2})$, then we can firstly transform $\vec{t}$ into ${|W_N\rangle}$, and then distill EPR pairs between party 1 and the other parties with an overall successful probability larger than $1-\sqrt{1-4(N-1)t^2}$, contradicting with lemma 2. Then to finish proving the theorem, we must show that $\tfrac{N}{2}(1-\sqrt{1-4(N-1)t^2})< Nt$ when $0<t<\frac{1}{N}$. It is an elementary optimization exercise to see that $ 1- 2t - \sqrt{1-4(N-1)t^2}<0$ whenever $0<t<\frac{1}{N}$. This “grouping” argument given for state $\vec{t}$ can be generalized to any ${|\varphi\rangle}_{1...N}$ having $x_0\not=0$ in order to obtain an upper bound on the transformation probability of ${|\varphi\rangle}_{1...N}\to{|W_N\rangle}$. While our upper bound is an improvement over the K-T monotones, it is not tight in general. Proving optimal transformation probability when $x_0\not=0$ remains an open problem. Conclusion ========== To conclude this article, let us first summarize the overall idea of the “Least Party Out” Protocol. Given a generic W-class state, we first remove the $x_0$ component with some probability. We then proceed to symmetrize by converting to standard W states ${|W_{|S|}^{(S)}\rangle}$. This is what the “Equal or Vanish” subroutine accomplishes, and it does so in such a way that the symmetry exists only between parties connected by $\mathcal{G}$ or any subgraph of $\mathcal{G}$. Finally, given a standard W state, the desired EPR pairings are obtained by removing parties from the entanglement in order of their connectivity in $\mathcal{G}$, the “least” parties being removed first. For three qubit random distillations, our protocol is optimal, and for four qubits, it is proven optimal when $x_0=0$ for all but Configuration VI, although it still may be optimal for the standard W state. In proving optimality for Configurations III and IV, the strategy was to compute the general expression for the LPO probability when $x_0=0$, and then show that this expression is an entanglement monotone. We have applied the same approach to study random distillations in systems with a greater number of parties. Unfortunately, the general expression for the LPO probability becomes quite complicated. This can be explicitly seen from Eq. in which the number of terms in the sum scales as $O(2^N)$ for a general configuration graph $\mathcal{G}$. Open Questions and Concluding Remarks {#open-questions-and-concluding-remarks .unnumbered} ------------------------------------- **I.** An obvious unresolved problem is to complete the four qubit picture by solving the random distillation of Configuration VI. We know the LPO Protocol is not optimal for non-standard W-class states, but it is not clear why this is the case. One possibility is the existence of $k$-cliques (a set of $k$ nodes all connected to one another) and the fact that all but one party belongs to a 3-clique. While Configurations IV and V also have 3-cliques, each party belongs to at least one. This may be the reason why the protocol behaves optimally in these two cases. Understanding precisely the limitations of our protocol for Configuration VI may also prove helpful when considering the same configuration of random distillations for more general states beyond the W Class. **II.** Another open problem is to generalize some of our results to a larger number of parties, especially the random distillations whose configuration graphs have relatively few edges. For example, consider the first graph in Configurations III for which we know the LPO Protocol reduces to the Fortescue-Lo Protocol, and it is optimal. For a six qubit system, this configuration generalizes to three disjoint pairings of the parties: $(1,2)$, $(3,4)$, and $(5,6)$. If, for this configuration, we perform the LPO Protocol on the state $(x_1,x_2,x_3,x_4,x_5,x_6)$ with $x_1\geq x_2\geq...\geq x_6$, the resultant probability function is $$\begin{aligned} \tau_6&= 2(x_2+x_4+x_6-\frac{x_2x_4}{x_1}-\frac{x_2x_6}{x_1}-\frac{x_4x_6}{x_3})\notag\\ &+2\frac{x_2x_4x_6}{x_1x_3}+2\frac{x_4x_5x_6}{x_3^2}+\frac{2x_2x_3x_4}{3x_1^2}+\frac{2x_2x_5x_6}{3x_1^2}\notag\\ &-2\frac{x_2x_4x_5x_6}{x_1x_3^2}-\frac{14x_2x_3x_4x_5x_6}{15x_1^4}.\end{aligned}$$ We strongly suspect that this probability is optimal, but we have no proof at this point. As in the four qubit case, the LPO Protocol reduces to the Generalized Fortescue-Lo Protocol. Note that for the state ${|W_6\rangle}$, the success probability is $2/5$. ![\[Fig:SEPvsLOCC3\] The relative difference between the optimal separable operation and the LPO Protocol. The configuration graph consists of $N$ disjoint pairs. Separable operations performs as $P_{SEP}=\sqrt{\frac{1}{N}}$ where as the LPO Protocol obtains the rate of $P_{LPO}=\frac{2}{2N-1}$. We conjecture that the LPO protocol is LOCC optimal for this configuration graph, as its known to be when $N=4$.](SepvsLOCC3.png) The generalization of this configuration to $2N$ qubits consists of a graph $\mathcal{G}_{2N}$ with $N$ disjoint pairings. Intuitively, the LPO protocol will again reduce to the Generalized Fortescue-Lo Protocol since there exists no particular “least” party. That is, the procedure will be for each party to perform weak measurements to randomly obtain three qubit W states ${|W_3\rangle}$ from which an EPR state can be obtained by a specified pair with probability $2/3$. One particular trio will obtain a W-state with probability $1/\tbinom{2N}{3}$, and there are a total of $2N-2$ trios in which this particular duo can belong. And finally, there are $N$ possible pairs. Thus, the total probability of some specified pair $(i,j)$ obtaining an EPR state is: $$P_{LPO}(W_{2N},\mathcal{G}_{2N})=\frac{1}{\tbinom{2N}{3}}\cdot(2N-2)\cdot\frac{2}{3}\cdot N=\frac{2}{2N-1}.$$ What is particularly interesting is when this transformation is compared to the optimal distillation probability by separable operations. As shown in Ref. [@Chitambar-2011b], this probability is given by $$P_{SEP}(W_{2N},\mathcal{G}_{2N})=\sqrt{\frac{1}{N}}.$$ Thus, if the LPO procedure is optimal for this particular transformation, which we strongly believe it is, then we see that the performance gap between separable operations and LOCC grows arbitrarily large. We depict this relative difference in Fig. \[Fig:SEPvsLOCC3\]. For example, in the four qubit case where the LPO procedure is optimal, we have $P_{LOCC}=\frac{2}{3}<P_{SEP}=\sqrt{\frac{1}{2}}$. **III.** Beyond the W-class of states, very little is known about single copy random distillations. Partial extensions of the Fortescue-Lo Protocol to symmetric Dicke states have been made, it has been shown that even within the three qubit GHZ class, distilling to randomly chosen pairs outperforms distilling to a specified pair [@Fortescue-2008a]. Nevertheless, how the topology of the outcome configuration graph $\mathcal{G}$ affects these transformations has yet to be studied in general. We hope the results of this article shed light on this question and provide a new insight into the structure of multipartite entanglement. We thank Jonathan Oppenheim, Ben Fortescue and Sandu Popescu for providing helpful discussions during the development of this work. We also thank the financial support from funding agencies including NSERC, QuantumWorks, the CRC program and CIFAR. Proof of Theorem \[Thm:tau\] {#Apx:tau} ============================ We consider case-by-case measurements in which each party acts according to . The function $\tau$ transforms as $\tau\to\tau_\lambda$ for $\lambda\in\{1,2\}$, and we are interested in the average change: $\overline{\tau_\lambda}=p_1\tau_1+p_2\tau_2$. By the universality of weak measurements [@Bennett-1999a; @Oreshkov-2005a], it is sufficient to prove $\tau$ monotonic in the weak measurement setting, i.e. with $(a_1,c_1,a_2,c_2)$ in some neighborhood of $(1/2,1/2,1/2,1/2)$. We consider three cases. **Case I, $x_{n_1}>x_{n_1'}$:** First consider when party $n_1$ performs a measurement. We can assume the measurement is weak enough such that $n_1$, $n_1'$, $p$ and $p'$ are the same for both pre-measurement and post-measurement states. Consider the measurement outcome $\lambda\in\{0,1\}$ with $a_\lambda > c_\lambda$. Then $$p_\lambda\tau_\lambda=2a_\lambda x_{p}+2a_\lambda x_{p'}-2\frac{a_\lambda^2}{c_\lambda}\left(1-\frac{1}{3}\frac{a_\lambda}{c_\lambda}\frac{x_{n_1'}}{x_{n_1}}\right)\frac{x_{p}x_{p'}}{x_{n_1}}.$$ We have $\frac{\partial\overline{\tau_\lambda}}{\partial c_\lambda}|_{a_\lambda=1/2,c_\lambda=1/2}\geq 0$ which implies that $\tau-\overline{\tau_\lambda}$ will be minimized by the choice $c_1+c_2=1$. Then writing $c\equiv c_1<a\equiv a_1$, we have that $$\begin{aligned} \tau-\overline{\tau_\lambda}=2\left(-1+\frac{a^2}{c}+\frac{(1-a)^2}{(1-c)}\right)\frac{x_{p}x_{p'}}{x_{n_1}}+\frac{2}{3}\left(1-\frac{a^3}{c^2}-\frac{(1-a)^3}{(1-c)^2}\right)\frac{x_{p}x_{p'}x_{n_1'}}{x_{n_1}^2}.\end{aligned}$$ Expanding this expression about the point $(1/2,1/2)$ to second order gives $$\begin{aligned} \tau-\overline{\tau_\lambda}\approx&\;\; 8\left((a-\tfrac{1}{2})^2+(c-\tfrac{1}{2})^2-2(a-\tfrac{1}{2})(c-\tfrac{1}{2})\right)\left(1-\frac{x_{n_1'}}{x_{n_1}}\right)\frac{x_{p}x_{p'}}{x_{n_1}}\notag\\ =&\;\;8\left(a-c\right)^2\left(1-\frac{x_{n_1'}}{x_{n_1}}\right)\frac{x_{p}x_{p'}}{x_{n_1}}\geq 0.\end{aligned}$$ Now consider when the other parties measure. Since the coefficient of $x_p$ is non-negative, the monotonicity of $\tau$ when party $p$ measures follows from the K-T monotones. For $n_1'$ and $p'$, there are two possibilities. **Subcase, $x_{n_1'}>x_{p'}$:** Here we can assume the measurements are weak enough such that their ordering does not change. Then since the coefficients of $x_{p'}$ and $x_{n_1'}$ are non-negative, the K-T monotones imply the monotonicity of $\tau$. **Subcase, $x_{n_1'}=x_{p'}$:** We have $\tau=2x_p+2x_{p'}-2\frac{x_px_{p'}}{x_{n_1}}+\frac{2}{3}\frac{x_px_{p'}^2}{x_{n_1}^2}$. It is easy to see that again $\tau-\overline{\tau_\lambda}$ is minimized when $c_1+c_2=1$. So parameterizing the measurement by $a$ and $c$ with $a>c$, we have that the average change in $x_{n_1'}$ is $(a+1-c)x_{p'}$ while the average change in $x_{p'}$ is $(1-a+c)x_{p'}$. It follows that $$\tau-\overline{\tau_\lambda}=2x_{p'}(a-c)-2(a-c)\frac{x_{p}x_{p'}}{x_{n_1}}\geq 0.$$ **Case II, $x_{n_1}=x_p$:** Here, $\tau=2x_p+\frac{2}{3}\frac{x_{p'}x_{n_1'}}{x_{n_1}}$. When either party $n_1$ or $p$ measures with $a_\lambda>c_\lambda$, the new components are $x_{\lambda,n_1}=\frac{a_\lambda}{p_\lambda} x_p$, $x_{\lambda,p}=\frac{c_\lambda}{p_\lambda}x_p$, $x_{\lambda,n_1'}=\frac{a_\lambda}{p_\lambda}x_{p'}$ and $x_{\lambda,p'}=\frac{a_\lambda}{p_\lambda}x_{n_1'}$. Thus, $$p_\lambda\tau_\lambda=2(a_\lambda-c_\lambda)x_{n_1'}+2c_\lambda x_p+\frac{2c_\lambda}{3}\frac{x_{p'}x_{n_1'}}{x_p}.$$ Since $x_p\geq x_{n_1'}$, we have $\frac{\partial\overline{\tau_\lambda}}{\partial c_\lambda}|_{a_\lambda=1/2,c_\lambda=1/2}\geq 0$. Again, this means that $\tau-\overline{\tau_\lambda}$ will be minimized by the choice $c_1+c_2=1$. Taking $a>c$, we have that $$\begin{aligned} \tau-\overline{\tau_\lambda}&=2(1-c-(1-a))x_{p}-2\left(1-c-\frac{(1-a)^2}{1-c}\right)x_{n_1'}+\frac{2}{3}\left(1-c-\frac{(1-a)^3}{(1-c)^2}\right)\frac{x_{p}x_{p'}}{x_{n_1'}}\end{aligned}$$ which to first order about the point $(1/2,1/2)$ takes the form $$2(a-c)(x_{p}-2x_{n_1'}+\frac{x_{p}x_{p'}}{x_{n_1'}}\geq 2(a-c)\frac{(x_{n_p}-x_{n_1'})^2}{x_{n_p}}\geq 0.$$ If either $x_{p'}$ or $x_{n_1'}$ measures, then the monotonicity of $\tau$ follows from the K-T monotones. **Case III, $x_{n_1}=x_{n_1'}$:** We have $\tau=2x_{p}+2x_{p'}-\frac{4}{3}\frac{x_{p}x_{p'}}{x_{n_1}}$. When either party $n_1$ or $n_1'$ measures, parties $p$ and $p'$ remain the same. With $a_\lambda>c_\lambda$, we have $$p_\lambda\tau=2a_\lambda x_p+2a_\lambda x_{p'}-2a_\lambda\frac{x_px_{p'}}{x_{n_1}}+\frac{2}{3}c_\lambda\frac{x_px_{p'}}{x_{n_1}}$$ which has $\frac{\partial\overline{\tau_\lambda}}{\partial c_\lambda}|_{a_\lambda=1/2,c_\lambda=1/2}\geq 0$. So again we assume $c_1=1-c_2\equiv c<a$ and we find that $$\begin{aligned} \tau-\overline{\tau_\lambda}&=-2\left(\frac{2}{3}-a-\frac{(1-a)^2}{1-c}\right)\frac{x_{p}x_{p'}}{x_{n_1}}-\frac{2}{3}\left(c+\frac{(1-a)^3}{(1-c)^2}\right)\frac{x_{p}x_{p'}}{x_{n_1}}\end{aligned}$$ which to third order about the point $(1/2,1/2)$ takes the form $$\frac{8}{3}\left((a-\tfrac{1}{2})^3-(c-\tfrac{1}{2})^3\right)-8\left((c-\tfrac{1}{2})(a-\tfrac{1}{2})^2-(a-\tfrac{1}{2})(c-\tfrac{1}{2})^2\right)\frac{x_{p}x_{p'}}{x_{n_1}}=\frac{8}{3}(a-c)^3\frac{x_{n_p}x_{n_p'}}{x_{n_1}}\geq 0.$$ Finally, since the coefficients of $x_p$ and $x_{p'}$ are positive in $\tau$, by the K-T monotones, $\tau$ is monotonic when either of these parties measures. [^1]: \ $^*{}^\dagger$ Authors contributed equally to this project.
{ "pile_set_name": "ArXiv" }
--- author: - 'Lihui Xue$^{1}$, Bin Wang$^{1}$ and Ru-Keng Su$^{2,1}$' title: 'Numerical simulation of the massive scalar field evolution in the Reissner-Nordström black hole background' --- $^{1.}$ Department of Physics, Fudan University, Shanghai 200433,\ People’s Republic of China\ $^{2.}$ China Center of Advanced Science and Technology (World Lab),\ P. O. Box. 8730, Beijing 100080, People’s Republic of China\ Introduction ============ The study of the evolution of various fields outside black holes plays an important role in black hole physics. In virtue of previous works, we now have the schematic picture regarding the dynamics of waves outside a spherical collapsing body. One intriguing subject is that a static observer outside the black hole can indicate a quasinormal ringing, which has frequencies directly relating to the parameters of the black hole. This quasinormal ringing is believed to be a unique fingerprint which would lead to the direct identification of the black hole existence and expected to be detected through gravitational wave observation in the near future [@01]. In order to extract as much information as possible from gravitational wave signal, it is of interest to understand quasinormal modes of various fields outside different black hole spacetimes.\ The massless neutral external perturbations were first studied by Price [@02]. Studying the behavior of a massless scalar field propagating on a fixed Schwarzschild background, he showed that for an observer on a fixed position the field dies off with a power-law tail. The most complete picture to date on this problem was presented by Gundlach, Price, and Pullin [@03]. They found that the result given in [@02] also holds at the event horizon and the future null infinity. Similar results for a massless scalar field propagating on Reissner-Nordström (RN) background have been obtained in [@04][@05][@06]. A close dependence of the quasinormal frequencies on the mass and charge of the RN black hole has been disclosed. The evolutions of charged massless scarlar field around a RN black hole and dilaton black hole have been investigated in detail in [@Hod] and [@M]. A number of authors have studied radiative dynamics in black hole spacetimes that are not asymptotically flat. Brady, et al [@07] considered the evolution of a massless scalar field in Schwarzschild de Sitter and RN de Sitter spacetimes. They found that at late times the field decays exponentially, not as an inverse power-law. Motivated by the recent discovery of the AdS/CFT correspondence, the investigation of the quasinormal modes of AdS black holes becomes appealing. The analysis of the quasinormal modes for massless scalar field in four, five, and seven dimensional Schwarzschild Ads black holes was performed in [@08]. It was argued that the quasinormal frequencies have different properties concerning different size of black holes. This argument was confirmed by studying the small Schwarzschild AdS black holes [@09]. Extending the study of the evolution of the massless scalar field to RN AdS and topological black hole backgrounds [@10][@11], richer dependence of the quasinormal frequencies on the black hole parameters has been exhibited.\ All these previous works are concentrated on massless scalar field perturbations. The evolution of massive scalar fields in black hole backgrounds is also of interest. Studying analytically the massive scalar field on a RN spacetime [@12] Hod and Piran argued that the intermediate asymptotic behavior of the field depends only on the field’s parameter and does not have relation to the spacetime properties. Their result was obtained by expanding the wave equations for the massive scalar field in the black hole background as a power series in $M/r$, $Q/r$ and neglecting terms of order $O \left[ (Mm/r)^2 \right]$ and higher. Similar behavior was also claimed by Koyama et al in Schwarzschild black hole [@13] and Moderski et al in dilaton black hole cases [@14]. These results imply that one cannot get any information about the structure of the background spacetime from the intermediate asymptotic tails of massive scalar fields. It’s quite different from the properties of the quasinormal modes of massless scalar field. Recent exact expressions of the quasinormal frequencies of massive scalar field for three-dimensional AdS [@15] and dS [@16] spacetimes got by either solving the wave equation in the bulk or studying the perturbation of conformal field theory on the boundary showed a contrary property to that of four-dimensional cases [@12][@13][@14]. The dependence of black hole parameters has been exhibited. Since the exact solution of quasinormal modes in four-dimensional spacetimes cannot be obtained analytically, it is of interest to carry out the numerical simulation of the collapse of a massive scalar field and compare with the asymptotic results got before. This is the motivation of the present paper.\ The outline of this paper is as follows. In Sec. II we describe the physical system and formulate the evolution equations. We also discuss in detail the intermediate range and the approximations involved. In Sec. III we give the numerical results. Sec. IV will be a brief summary.\ Approximate Solution ==================== We consider the evolution of massive scalar fields in the Reissner-Nordström black hole background described by the metric $$ds^2=-\left( 1-\frac{2M}{r}+\frac{Q^2}{r^2} \right) dt^2+\left( 1-\frac{2M}{r}+\frac{Q^2}{r^2} \right)^{-1}dr^2+r^2d \Omega^2, \label{metric}$$ where $M$ and $Q$ correspond to the mass and charge of the black hole. The tortoise radial coordinate $y$ is defined by $$dy=\frac{dr}{\lambda^2},$$ where $$\lambda^2=1-\frac{2M}{r}+\frac{Q^2}{r^2},$$ Using the tortoise coordinate, Eq.(\[metric\]) becomes $$ds^2=\lambda^2(-dt^2+dy^2)+r^2d \Omega^2.$$ The scalar field with mass $m$ satisfies the wave equation $$(\Box-m^2)\phi=0.$$ Resolving the field into spherical harmonics $\phi=\sum_{l,m}\psi_m^l(t,r)Y_l^m(\theta,\varphi)/r$, one obtains a wave equation for each multipole moment $$\psi_{,tt}-\psi_{,yy}+V\psi=0, \label{field_eq}$$ where $$V=\left( 1-\frac{2M}{r}+\frac{Q^2}{r^2} \right)\left[ \frac{l(l+1)}{r^2}+\frac{2M}{r^3}-\frac{2Q^2}{r^4}+m^2 \right]. \label{potential}$$ The background spacetime parameters $M$ and $Q$ are obviously included in Eq.(\[potential\]).\ In order to study the time evolution of a massive scalar field, the retarded Green’s function $G(y,y';t)$ is defined as [@12] $$\left[ \frac{\partial^2}{\partial t^2}-\frac{\partial^2}{\partial y^2}+V(r) \right]G(y,y';t)=\delta(t)\delta(y-y'),$$ and Eq.(\[field\_eq\]) can be written as $$\psi(y,t)=\int \left[ G(y,y';t)\psi_t(y',0)+G_t(y,y';t)\psi(y',0) \right]dy'.$$ The initial condition reads $G(y,y';t)=0$ for $t \le 0$. $G(y,y';t)$ can be found by using the Fourier transformation $$\tilde{G}(y,y';\omega)=\int_{0^-}^{\infty}G(y,y';t)e^{i \omega t}dt.$$ The Fourier component of the Green’s function $\tilde{G}(y,y';\omega)$ can be expressed in terms of two linearly independent solutions $\tilde{\psi}_1(y,\omega)$ and $\tilde{\psi}_2(y,\omega)$ to the homogeneous equation $$\left( \frac{d^2}{dy^2}+\omega^2-V \right)\tilde{\psi}_i(y,\omega)=0; i=1,2.$$ Introduce an auxiliary variable $\xi$ in such a way that $\xi=\lambda \tilde{\psi}$ and remember $dy=dr/\lambda^2$ one can rewrite the equation of $\tilde{\psi}_i$ as $$\frac{d^2\xi}{dr^2}+\left( \frac{\omega^2}{\lambda^4}-\frac{\frac{l(l+1)}{r^2}+\frac{2M}{r^3}-\frac{2Q^2}{r^4}+m^2}{\lambda^2}-\frac{1}{\lambda} \frac{d^2\lambda}{dr^2} \right)\xi=0. \label{eq_of_xi}$$ In order to calculate $\xi$, some approximations have been made in [@12] to simplify the equation. Assuming that the observer is situated far away from the black hole $(M \ll r)$, Eq.(\[eq\_of\_xi\]) can be expanded in power series of $M/r$ and $Q/r$ to the order $O[M^2/r^2]$ and we arrive at $$\begin{aligned} &&\xi ''+\left[ \omega^2-m^2+\frac{-2Mm^2+4M\omega^2}{r}+\frac{-l(l+1)}{r^2} \right.\nonumber \\ &&\left.+\frac{M^2-Q^2}{r^4}+O[M^2/r^4] \right]\xi=0. \label{xi_power}\end{aligned}$$ Eq.(\[xi\_power\]) can be further approximated to the form $$\xi ''+\left( \omega^2-m^2-\frac{l(l+1)}{r^2} \right)\xi=0, \label{xi_inter}$$ if $\frac{Mm^2}{r} \ll m^2$, $\frac{Mm^2}{r} \ll \frac{l(l+1)}{r^2}$, $\frac{M^2}{r^4} \ll m^2$, and $\frac{M^2}{r^4} \ll \frac{l(l+1)}{r^2}$. One can read from these four inequalities the intermediate range $\sqrt{\frac{M}{m}} \ll r \ll \frac{1}{Mm^2}$ in the limit $Mm \ll 1$. Eq.(\[xi\_inter\]) shows that the backscattering of the field from asymptotically far region has been neglected. From this equation, it was argued that the intermediate behavior of massive scalar field depends only on the field’s parameter and it does not depend on the spacetime parameters [@12].\ With the increase of $Mm$, the intermediate range will shrink. When $Mm \ge 1$, the intermediate range will no longer exist.\ Numerical Simulation ==================== In this section we do not introduce any approximation and directly solve Eq.(\[field\_eq\]) numerically. It should be noted that Eq.(\[field\_eq\]) is invariant under the rescaling $$r \rightarrow ar; t \rightarrow at; M \rightarrow aM; Q \rightarrow aQ; m \rightarrow m/a,$$ where $a$ is a positive constant. In such a rescaling $Mm$ is invariant. Using the null coordinates $u=t-y$ and $v=t+y$, Eq.(\[field\_eq\]) can be recast as $$4\psi_{,uv}+V(r)\psi=0. \label{eq_of_uv}$$ Eq.(\[eq\_of\_uv\]) can be integrated numerically by the finite difference method suggested in [@03]. The late time evolution of a massive scalar field should not depend on the form of the initial data. In our numerical calculations, we use a Gaussian pulse of the form $$\psi(u=0,v)=A \exp \{-\left[ \left( v-v_0 \right)/\sigma \right]^2 \}.$$ After the integration is completed, the value $\psi(u_{max},v)$ is extracted, where $u_{max}$ is the maximum value of $u$ on the numerical grid. Taking sufficiently large $u_{max}$, $\psi(u_{max},v)$ represents a good approximation for the wave function at the event horizon. Since it has been shown that wave behavior is the same near or far away from the event horizon [@03], we will study the dependence of wave behavior on background parameters in different ranges of $Mm$.\ To test our numerical program, we first take $m=0$ in our simulation and investigate the dependence of the decay rate of the massless scalar field on the black hole charge. The results are shown in Fig. 1. There is a maximum of the slope at a critical value of the black hole charge $Q_{C}$. When $Q>Q_{C}$, the massless scalar field settles down slower as the increase of $Q$ until its extreme value $Q=M$. Considering the relation between the damping timescale $\tau$ and the imaginary part of the quasinormal frequencies $\omega_{I}$, $\tau=1/\left| \omega_{I} \right|$, our numerical result exhibits that $\left| \omega_{I} \right|$ has a maximum value for a critical value of the black hole charge. This wiggle behavior of $\left| \omega_{I} \right|$ is in accordance with the lowest normal mode frequencies studied in [@04][@05].\ ![image](fig1.ps){width="4in" height="3in"}\ *Fig. 1. The decay rate for different black hole charges $Q=0$, $20$, $40$, $60$, $80$, $90$, $95$, $97$ when $l=0$. The black hole mass is taken to be $M=100$, and the mass of the scalar field is $m=0$.\ * We now report our numerical simulation results of evolving massive scalar field on a RN black hole background. In the first series of numerical calculations, we choose appropriate values of the black hole mass $M$ and scalar field mass $m$ to satisfy $Mm \ll 1$. For different value of $M$, the results are shown in Fig. 2 with the multipoles $l=0$. In order to display the property of the relaxation of the perturbation clearly, we here just connected highest points of each wave crest and neglected the oscillations. It is obvious from this figure that in the region $Mm \ll 1$ the decay rate keeps approximately the same for different values of the black hole mass. The small difference in the slope to the order $10^{-2}$ was just caused by selecting points in getting the slope of the damping. This result also holds by taking different values of the charge of the black hole, which is different from that of the massless scalar field as shown in Fig. 1. The numerical pictures obtained here support the approximate solution in [@12]. The intermediate asymptotic behavior of massive scalar perturbation does not depend on the background black hole parameters.\ ![image](fig2a.ps){width="4in" height="3in"}\ *Fig. 2a. Evolution of the massive field $|\psi|$ on the Reissner-Nordström background, for $m=0.01$ and $l=0$. The field at a fixed radius ($r=50$) is shown as a function of time. The decay rate of the perturbation (the slope of the curve) $B$ is $-1.470$ for $M=0.5, Q=0.45$ (bottom curve) and $-1.432$ for $M=1.0, Q=0.45$ (top curve).\ * ![image](fig2b.ps){width="4in" height="3in"}\ *Fig. 2b. The amplitude of the field along the event horizon for $m=0.01$ and $l=0$. The field is shown as a function of $v$. The black hole parameter is $M=0.5, Q=0.45$ for the bottom curve and $M=1.0, Q=0.45$ for the top curve. For $Mm \ll 1$, the decay rate for the field along the event horizon is almost identical to that at a fixed radius.\ * The dependence of the decay rate of the perturbation on the black hole parameters appears for big value of $Mm$. For $Mm \gg 1$, we have found a difference of decay rate caused by the background parameters. Results displayed in Fig. 3 tell us that the bigger black hole mass leads to the faster decay of the massive scalar field perturbation. The difference of the relaxation process caused by the black hole charge $Q$ has also be exhibited in this region in Fig. 4 with the black hole mass fixed to be $M=200$. We see that the slope (the decay rate) of the scalar perturbation, first decreases with the increase of $Q$ and then over a critical value of $Q$, the slope increases with the increase of $Q$ until $Q$ reaches the extreme value. This behavior also holds for other values of black hole mass and for $M$ is bigger, the dependence is clearer. This means that before some critical value of $Q$, the larger black hole charge is, the slower the outside perturbation dies out, corresponding to the decrease of the imaginary quasinormal frequency; while over the critical value of $Q$, we have the opposite decay behavior with the increase of the imaginary frequency. This wiggle behavior of the imaginary quasinormal frequency caused by black hole charge arise in the massive scalar field perturbation is qualitatively different from the case of the massless scalar field. ![image](fig3.ps){width="4in" height="3in"}\ *Fig. 3. Evolution of the massive field on the RN background for $Mm \gg 1$. The field along the event horizon is shown as a function of $v$. From top to bottom, the black hole mass and charge are $M=50, Q=45$; $M=200, Q=45$, and $M=100, Q=45$, respectively. The field mass is $m=1.0$ and the multipole index is $l=0$.\ * ![image](fig4.ps){width="4in" height="3in"}\ *Fig. 4. The decay rate for different black hole charge $Q=0$, $40$, $80$, $120$, $160$, $180$, $190$, $195$ and $200$. The black hole mass is taken to be $M=200$, and $m=1.0$, $l=0$.\ * Conclusions =========== We have studied the massive scalar wave propagation in the background of RN black hole by using numerical simulations instead of solving the wave equation approximately. The field evolution behavior we found is different from that of massless scalar field. We have learnt that the value $Mm$ plays an important role in determining the properties of the relaxation of the perturbation. For $Mm \ll 1$, our numerical result confirmed the approximate argument given in [@12], that the intermediate behavior of massive scalar perturbations on a black hole background approximately depends only on the field’s parameter and it does not depend on the spacetime parameters. For $Mm \ge 1$, the intermediate approximation condition breaks down. Our numerical results show that for $Mm>1$, the dependence of the relaxation of the perturbation on the background black hole mass appears. This kind of dependence becomes much clearer for $Mm \gg 1$. In the region $Mm \gg 1$, the influence given by the black hole charge $Q$ on the massive scalar field’s relaxation has also been observed. For small values of $Q$, the settle down of massive scalar field perturbation is slower for larger charge. However, after the black hole charge reaches a critical value, we got an opposite behavior: the field decays faster with the increase of the black hole charge. The picture of the massive scalar field relaxation caused by the black hole charge is different from that of the massless scalar field. New phenomenon found here are quite interesting and the inclusion of the scalar field mass have enriched the spectrum of the wave dynamics outside the black hole.\ \ Acknowledgements: This work was partially supported by NNSF of China. We would like to acknowledge the helpful discussions with Weigang Qiu.\ [ll]{} K. D. Kokkotas, B. G. Schmidt, Living Rev. Rel. 2 (1999) and references therein. R. H. Price, Phys. Rev. D 5, 2419 (1972) C. Gundlach, R. H. Price and J. Pullin, Phys. Rev. D 49, 883 (1994) K. D. Kokkotas and B. F. Schutz, Phys. Rev. D 37, 3378 (1988) N. Andersson, Proc. R. Soc. Lond. A 442, 427 (1993) E. W. Leaver, Phys. Rev. D 41, 2986 (1990) S. Hod and T. Piran, Phys. Rev. D 58, 024017; 024018; 024019 (1998) R. Moderski and M. Rogatko, Phys. Rev. D 63, 084014 (2001) P. R. Brady, C. M. Chambers, W. Krivan and P. Laguna, Phys. Rev. D 55, 7538 (1997)\ P. R. Brady, C. M. Chambers, W. G. Laarakkers and E. Poisson, Phys. Rev. D 60, 064003 (1999) G. T. Horowitz, V. E. Hubeny, Phys. Rev. D 62, 024027 (2000) J. M. Zhu, B. Wang and E. Abdalla, Phys. Rev. D 63, 124004 (2001) B. Wang, C. Y. Lin and E. Abdalla, Phys. Lett. B 451, 79 (2000)\ B. Wang, C. M. Mendes and E. Abdalla, Phys. Rev. D 63, 084001 (2001)\ B. Wang, E. Abdalla and R. B. Mann, Phys. Rev. D 65, 084006 (2002) S. Hod, T. Piran, Phys. Rev. D 58, 044018 (1998) H. Koyama and A. Tomimatsu, Phys. Rev. D 63, 064032 (2001)\ H. Koyama and A. Tomimatsu, Phys. Rev. D 64, 044014 (2001) R. Moderski and M. Rogatko, Phys. Rev. D 64, 044024 (2001) D. Birmingham, I. Sachs and S. N. Solodukhin, Phys. Rev. Lett. 88, 151301 (2002) E. Abdalla, B. Wang, A. Lima-Santos and W. G. Qiu, hep-th/0204030
{ "pile_set_name": "ArXiv" }
--- abstract: 'Based on the 2PI quantum effective action of the linear sigma model with constituent quarks, we develop a transport approach to study systems out of equilibrium. In particular, we focus on the chiral phase transition as well as the critical point, where nonequilibrium effects near the phase transition give rise to critical behavior such as the fluctuation of the baryon number density. Predictions for long-range correlations and fluctuations of observables in our model could be used to study fundamental properties of the QCD phase transition. In the previous version of our transport model the chiral fields are implemented as mean fields, whereas quarks are treated as on-shell particles in the Vlasov equation with a dynamical force term. The current update includes also the distribution functions of sigma mesons and pions in a self-consistent way. On this basis a dissipation kernel between the mean fields and particle modes can be implemented.' address: - '$^1$ Institut für Theoretische Physik, Goethe-Universität Frankfurt, Max-von-Laue-Stra[ß]{}e 1, D-60438 Frankfurt, Germany' - '$^2$ Frankfurt Institute for Advanced Studies, Ruth-Moufang-Stra[ß]{}e 1, D-60438 Frankfurt, Germany' author: - 'A Meistrenko$^1$, C Wesp$^1$, H van Hees$^{1,2}$ and C Greiner$^1$' title: 'Nonequilibrium dynamics and transport near the chiral phase transition of a quark-meson model' --- Introduction {#sec:a1} ============ We concentrate our study on nonequilibrium effects and the chiral phase transition within the linear sigma model with constituent quarks, which is an effective theory of QCD in the low-energy limit. Our motivation is to investigate critical phenomena at the phase transition and the critical point, where time-dependent long-range correlations can arise. Such effects can be studied in a (3+1)-dimensional numerical approach. We present our model in the following section \[sec:a2\] and discuss an improved set of transport equations in section \[sec:a3\]. Finally, we summarize the progress in section \[sec:a4\] and give an outlook to further improvements. Classical transport equations within a linear sigma framework {#sec:a2} ============================================================= The linear sigma model with constituent quarks is a well known $O\left(4\right)$-model [@Scavenius:2000qd], which is suited for studying the chiral phase transition. Because of the spontaneously and explicitly broken chiral symmetry the mesonic part of this theory consists of a massive scalar sigma field and three isoscalar pion fields, which form the chiral field $\Phi=\left(\sigma, \vec\pi\right)$. The sigma field represents the order parameter for the chiral phase transition and mimics the properties of the quark condensate in QCD, since both transform equally under chiral transformation. Without explicit symmetry breaking a $\mathrm{SU_L}\left(2\right)\times \mathrm{SU_R}\left(2\right)$ symmetry transformation would let the Lagrangian invariant. It has the following form: L=&\_i|\_i \_i +(\_\^ +\_\^)\ &-(\^2+\^2-\^2)\^2+f\_m\_\^2+U\_0, \[eqn:a1\] where the field shift term and the zero potential constant are given by $\nu^2=f_\pi^2-m_\pi^2/\lambda$, $U_0=m^4_\pi/\left(4\lambda\right)-f_\pi^2m_\pi^2$. The sum runs over included quark flavor $\psi_i$ and the parameters of the model are adjusted to match the vacuum values of the pion decay constant $f_\pi=93\,\mbox{MeV}$, the pion mass $m_\pi=138\,\mbox{MeV}$ as well as an estimated sigma mass $m_\sigma\approx604\,\mbox{MeV}$. The inhomogeneous Klein-Gordon equations of motion for the mean fields follow directly from the functional derivative of the classical action with respect to the chiral field components and treating quarks at one-loop level \_\^+(\^2+\^2-\^2)-f\_m\_\^2+g&=0,\ \_\^+(\^2+\^2-\^2)+g&=0. \[eqn:a2\] The scalar and pseudoscalar densities are given by (t,x)&=gd\_q(t,x)d\^3p,\ (t,x)&=gd\_q(t,x)d\^3p, \[eqn:a3a\] where $f_q$ and $f_{\bar q}$ denote the phase-space distribution functions of quarks and antiquarks with their degeneracy factor $d_q$. Numerically the quarks are treated as test particles, which propagate according to the Vlasov equation f\_[q||q]{}(t,x,p)=0. \[eqn:a3b\] Thereby the force term is dynamically generated through the effective mass of quarks, depending on the mean-field values E(t,x,p)&=,\ M\^2(t,x)&=g\^2. \[eqn:a3c\] The resulting equilibrium properties of such a classical system are shown in Fig. \[fig:a1\] for the order parameter and in Fig. \[fig:a2\] for the effective sigma mass as a function of the temperature. Depending on the Yukawa coupling constant $g$ different order of a phase transition can be generated. More results, also with a constant binary cross section for quarks, can be found in [@H]. ![\[fig:a2\]Same as Fig. \[fig:a1\] but for the sigma mass spectrum.](sigma_values){width="18pc"} ![\[fig:a2\]Same as Fig. \[fig:a1\] but for the sigma mass spectrum.](sigma_mass){width="18pc"} Transport equations from the 2PI effective action {#sec:a3} ================================================= Since we are interested in nonequilibrium dynamics at the phase transition, a consistent real-time description of the interaction between the mean-field (soft) and the particle (hard) modes of the chiral field as well as quarks is needed. The model should also include a dissipation term, which would drive the system to thermal equilibrium. In the following we use the 2PI quantum effective action to derive a set of coupled equations of motion. Thereby 2PI stands for two-particle irreducible, which means that a 2PI diagram does not become disconnected by cutting $2$ inner lines (propagators). This effective action for the mean fields and propagators preserves global symmetries of the original theory, guarantees thermodynamic consistency [@Baym:1962] and can be renormalized with vacuum counter terms [@vanHees:2001ik; @VanHees:2001pf]. At the same time such a self-consistent approximation is suitable to derive off-shell transport equations, which also respect conservation laws and the correct equilibrium limit. Furthermore, a systematic inclusion of collisional memory effects is possible [@Ivanov:1998nv; @Ivanov:1999tj; @Knoll:2001jx; @Ivanov:2003wa]. A more phenomenological study, motivated by the Langevin equation, is discussed in [@H], where a new statistical approach for the scattering effects between quarks as quasi particles and mean fields is introduced. Within the linear sigma model with constituent quarks the 2PI effective action reads = & S\_[cl]{}\[,\]+G\^[-1]{}+G\_0\^[-1]{}G\ & -iD\^[-1]{}-iD\_0\^[-1]{}D+\_2\[,,G,D\], \[eqn:a4\] where $G_0,\,D_0$ denote the free and $G,\,D$ the fully dressed propagators for bosons and fermions, which fulfills the Schwinger-Dyson equation and are formally given by G\^[-1]{}(x,y)&=G\_0\^[-1]{}(x,y)-(x,y),(x,y)=2i,\ D\^[-1]{}(x,y)&=D\_0\^[-1]{}(x,y)-(x,y),(x,y)=-i. \[eqn:a5\] Here, $\Pi$ and $\Sigma$ denote the bosonic and fermionic self-energies. These are derived from the 2PI part of the effective action, where we include only $1$- and $2$-point diagrams as shown in Fig. \[fig:a3\], since a first order gradient expansion of Kadanoff-Baym and Schwinger-Dyson equations in Wigner space then reduces to a Markov-like collisional dynamics without memory effects [@Ivanov:1998nv; @Ivanov:1999tj; @Knoll:2001jx; @Ivanov:2003wa]. Unfortunately, finite truncations of the 2PI effective action violate Ward-Takahashi-identities of global and local symmetries in the first neglected order of the expansion parameter, that is a direct consequence of the resummation in the $2$-point function. It follows, that also the Goldstone theorem is violated [@Baym:1977; @vanHees:2002bv], and the pions acquire a non-vanishing (temperature dependent) mass in the broken phase of the linear sigma model, even without explicitly broken chiral symmetry. Possible modifications for a symmetry-improved 2PI effective action are discussed in [@Pilaftsis:2013xna]. ![Approximation for the 2PI part of the effective action: the upper line shows bosonic diagrams (Hartree, sunset and basketball), where dashed lines stand for boson propagators and external lines with a cross represent mean fields. The lower line shows interactions between bosons and fermions, here a solid line stands for a fermion propagator.[]{data-label="fig:a3"}](Gamma2){width="28pc"} In a first attempt we extend our classical set of equations , by including also the Vlasov equation for the phase-space distribution functions of $\sigma$ and $\pi$ mesons in a Hartree-like approximation, where only the local part of $\Gamma_2$ is considered. Such a procedure requires the computation of self-consistently derived effective mass terms, which follow from the gap equation of the propagator. Starting from the general expression for the bosonic propagator and inserting the self-energy expressions from the Hartree diagrams, leads to the following self-consistent equations M\_\^2(x)&=(3\^2+3\^2-\^2)+3G\_+3G\_,\ M\_\^2(x)&=(\^2+5\^2-\^2)+G\_+5G\_, \[eqn:a6\] where the bosonic loop integrals are given by G\_(t,x)={,}. \[eqn:a7\] The Vlasov equations for the phase-space distributions of $\sigma$ and $\pi$ mesons are then given by f\_(t,x,p)=0, \[eqn:a8\] where the force term depends on the self-consistent mass expressions from as gradient of E\_(t,x,p)=. \[eqn:a9\] The Vlasov equations of the chiral field components are solved as differential equations without using a test particle ansatz. Including also the tadpole contribution of $\Gamma$, we end up with the following mean-field equations \_\^+(\^2+\^2-\^2+3G\_+3G\_)-f\_m\_\^2+g&=0,\ \_\^+(\^2+\^2-\^2+G\_+5G\_)+g&=0. \[eqn:a10\] The system of equations , and serves as a basis to study nonequilibrium effects. Nevertheless, it does not account for dissipation as well as $q,\bar q$ creation and annihilation processes. In an upcoming work we will implement a dissipation kernel [@Greiner:1996dx; @Rischke:1998qy], which arises from the sunset diagram (see Fig. \[fig:a3\]) by deriving $\Gamma_2$ with respect to the mean fields. Conclusion and outlook {#sec:a4} ====================== In this proceeding we presented an improved set of equations of motion to study systems out of equilibrium, which is included in a numerical simulation. In comparison to the phenomenological work [@H], here we use the 2PI effective action of the linear sigma model with constituent quarks. In the current version the set consists of mean-field equations for the components of the chiral field, the Vlasov equation for the phase-space distribution functions of $\sigma$ and $\pi$ mesons as well as the Vlasov equation for the quarks, which are treated as test particles. In a further study we plan to implement a dissipation kernel for the mean-field equations, which results from the sunset diagram. For a consistent treatment it requires also the inclusion of a collision term between soft and hard modes for the Vlasov equation of chiral partners. Furthermore, an interaction between $\sigma$, $\pi$ and quarks on the particle level has to be implemented to account for chemical equilibration in a dynamical and consistent way. Because of numerical challenges such a direct Boltzmann-like transport approach will be primarily suited for the study of time-dependent fluctuations in homogeneous systems. High order cumulants of conserved quantities like baryon number density offer a possibility to observe such fluctuations. Nevertheless, it is an open question whether fluctuations can also arise on the time scale of a heavy ion collision and survive the hadronization, that needs further investigation. This work has been supported by the German Federal Ministry of Education and Research (BMBF Förderkennzeichen 05P12RFFTS). A. M. and C. W. acknowledge financial support from the Helmholtz Research School for Quark Matter Studies (H-QM) and HIC for FAIR. References {#references .unnumbered} ========== [9]{} Scavenius O, Mocsy A, Mishustin I N and Rischke D H 2001 [*Phys. Rev.*]{} C [**64**]{}, 045202 van Hees H, Wesp C, Meistrenko A and Greiner C 2013 Dynamics of the chiral phase transition [*Preprint to be published in the proceedings of the XXXI Max Born Symposium and HIC for FAIR Workshop, 14-16 June 2013, Wroclaw, Poland*]{} Baym G 1962 [*Phys. Rev.*]{} [**127**]{}, 1391 van Hees H and Knoll J 2002 [*Phys. Rev.*]{} D [**65**]{}, 025010 van Hees H and Knoll J 2002 [*Phys. Rev.*]{} D [**65**]{}, 105005 Ivanov Y B, Knoll J and Voskresensky D N 1999 [*Nucl. Phys.*]{} A [**657**]{}, 413 Ivanov Y B, Knoll J and Voskresensky D N 2000 [*Nucl. Phys.*]{} A [**672**]{}, 313 Knoll J, Ivanov Y B and Voskresensky D N 2001 [*Annals Phys.*]{} [**293**]{}, 126 Ivanov Y B, Knoll J and Voskresensky D N 2003 [*Phys. Atom. Nucl.*]{} [**66**]{}, 1902 Arrizabalaga A and Smit J 2002 [*Phys. Rev.*]{} D [**66**]{}, 065014 Baym G and Grinstein G 1977 [*Phys. Rev.*]{} D [**15**]{}, 2897–2912 van Hees H and Knoll J 2002 [*Phys. Rev.*]{} D [**66**]{}, 025028 Pilaftsis A and Teresi D 2013 [*Nucl. Phys.*]{} B [**874**]{}, 594 Greiner C and Muller B 1997 [*Phys. Rev.*]{} D [**55**]{}, 1026 Rischke D H 1998 [*Phys. Rev.*]{} C [**58**]{}, 2331
{ "pile_set_name": "ArXiv" }
--- abstract: 'Mild Cognitive Impairment (MCI) is a mental disorder difficult to diagnose. Linguistic features, mainly from parsers, have been used to detect MCI, but this is not suitable for large-scale assessments. MCI disfluencies produce non-grammatical speech that requires manual or high precision automatic correction of transcripts. In this paper, we modeled transcripts into complex networks and enriched them with word embedding (CNE) to better represent short texts produced in neuropsychological assessments. The network measurements were applied with well-known classifiers to automatically identify MCI in transcripts, in a binary classification task. A comparison was made with the performance of traditional approaches using Bag of Words (BoW) and linguistic features for three datasets: DementiaBank in English, and Cinderella and Arizona-Battery in Portuguese. Overall, CNE provided higher accuracy than using only complex networks, while Support Vector Machine was superior to other classifiers. CNE provided the highest accuracies for DementiaBank and Cinderella, but BoW was more efficient for the Arizona-Battery dataset probably owing to its short narratives. The approach using linguistic features yielded higher accuracy if the transcriptions of the Cinderella dataset were manually revised. Taken together, the results indicate that complex networks enriched with embedding is promising for detecting MCI in large-scale assessments.' author: - | Leandro B. dos Santos[^1^]{}, Edilson A. Corr[ê]{}a Jr[^1^]{}, Osvaldo N. Oliveira Jr[^2^]{}, Diego R. Amancio[^1^]{},\ **Letícia L. Mansur[^3^]{}**, **Sandra M. Aluísio[^1^]{}**\ [^1^]{} Institute of Mathematics and Computer Science, University of São Paulo, São Carlos, São Paulo, Brazil\ [^2^]{} São Carlos Institute of Physics, University of São Paulo, São Carlos, São Paulo, Brazil\ [^3^]{} Department of Physiotherapy, Speech Pathology and Occupational Therapy,\ University of São Paulo, São Paulo, São Paulo, Brazil\ [{leandrobs,edilsonacjr,lamansur}@usp.br]{}, [[email protected]]{}\ [{diego,sandra}@icmc.usp.br]{} bibliography: - 'acl2017.bib' title: Enriching Complex Networks with Word Embeddings for Detecting Mild Cognitive Impairment from Speech Transcripts --- Introduction\[sec:intro\] ========================= Mild Cognitive Impairment (MCI) can affect one or multiple cognitive domains (e.g. memory, language, visuospatial skills and executive functions), and may represent a pre-clinical stage of Alzheimer’s disease (AD). The impairment that affects memory, referred to as amnestic MCI, is the most frequent, with the highest conversion rate for AD, at 15% per year versus 1 to 2% for the general population. Since dementias are chronic and progressive diseases, their early diagnosis ensures a greater chance of success to engage patients in non-pharmacological treatment strategies such as cognitive training, physical activity and socialization [@Art:Teixeira:2012:Non:pharmacological]. Language is one of the most efficient information sources to assess cognitive functions. Changes in language usage are frequent in patients with dementia and are normally first recognized by the patients themselves or their family members. Therefore, the automatic analysis of discourse production is promising in diagnosing MCI at early stages, which may address potentially reversible factors [@Art:Muangpaisan:2012:Prevalence]. Proposals to detect language-related impairment in dementias include machine learning [@Inc:Jarrold:2010:Language; @Art:Roark:2011:Spoken; @Art:Fraser:2014:Automated; @Art:Fraser:2015:linguistic], magnetic resonance imaging [@Artc:Dyrba:2015:Predicting], and data screening tests added to demographic information [@Art:Weakley:2015:Neuropsychological]. Discourse production (mainly narratives) is attractive because it allows the analysis of linguistic microstructures, including phonetic-phonological, morphosyntactic and semantic-lexical components, as well as semantic-pragmatic macrostructures. Automated discourse analysis based on Natural Language Processing (NLP) resources and tools to diagnose dementias via machine learning methods has been used for English language [@Inp:Lehr:2012:Fully; @Inp:Jarrold:2014:Aided; @Inp:Orimaye:2014:Learning; @Art:Fraser:2015:linguistic; @Inp:Davy:2016:Towards] and for Brazilian Portuguese [@Inp:Aluisio:2016:Evaluating]. A variety of features are required for this analysis, including Part-of-Speech (PoS), syntactic complexity, lexical diversity and acoustic features. Producing robust tools to extract these features is extremely difficult because speech transcripts used in neuropsychological evaluations contain disfluencies (repetitions, revisions, paraphasias) and patient’s comments about the task being evaluated. Another problem in using linguistic knowledge is the high dependence on manually created resources, such as hand-crafted linguistic rules and/or annotated corpora. Even when traditional statistical techniques (Bag of Words or ngrams) are applied, problems still appear in dealing with disfluencies, because mispronounced words will not be counted together. Indeed, other types of disfluencies (repetition, amendments, patient’s comments about the task) will be counted, thus increasing the vocabulary. An approach applied successfully to several areas of NLP [@Boo:Mihalcea:2011:Graph:NLP], which may suffer less from the problems mentioned above, relies on the use of complex networks and graph theory. The word adjacency network model [@Art:Cancho:2001:Small; @Art:Roxas:2010:prose:poetry; @Art:Amancio:2012:Extractive; @Art:Amancio:2015:Complex] has provided good results in text classification [@Art:Arruda:2016:Classification:Texts] and related tasks, namely author detection [@Art:Amancio:2015:Authorship], identification of literary movements [@Art:Amancio:2012:Literary:movements], authenticity verification [@10.1371/journal.pone.0067310] and word sense discrimination [@0295-5075-98-1-18002]. In this paper, we show that speech transcripts (narratives or descriptions) can be modeled into complex networks that are enriched with word embedding in order to better represent short texts produced in these assessments. When applied to a machine learning classifier, the complex network features were able to distinguish between control participants and mild cognitive impairment participants. Discrimination of the two classes could be improved by combining complex networks with linguistic and traditional statistical features. With regard to the task of detecting MCI from transcripts, this paper is, to the best of our knowledge, the first to: a) show that classifiers using features extracted from transcripts modeled into complex networks enriched with word embedding present higher accuracy than using only complex networks for 3 datasets; and b) show that for languages that do not have competitive dependency and constituency parsers to exploit syntactic features, e.g. Brazilian Portuguese, complex networks enriched with word embedding constitute a source to extract new, language independent features from transcripts. Related Work\[sec:related\] =========================== Detection of memory impairment has been based on linguistic, acoustic, and demographic features, in addition to scores of neuropsychological tests. Linguistic and acoustic features were used to automatically detect aphasia [@Art:Fraser:2014:Automated]; and AD [@Art:Fraser:2015:linguistic] or dementia [@Inp:Orimaye:2014:Learning] in the public corpora of DementiaBank[^1]. Other studies distinguished different types of dementia [@Art:Garrard:2014:ML:WAB; @Inp:Jarrold:2014:Aided], in which speech samples were elicited using the Picnic picture of the Western Aphasia Battery [@Book:Kertesz:1982:Western]. @Inp:Davy:2016:Towards also used the Picnic scene to detect MCI, where the subjects were asked to write (by hand) a detailed description of the scene. As for automatic detection of MCI in narrative speech, @Art:Roark:2011:Spoken extracted speech features and linguistic complexity measures of speech samples obtained with the Wechsler Logical Memory (WLM) subtest [@Book:Wechsler:1997:WLM], and @Inp:Lehr:2012:Fully fully automatized the WLM subtest. In this test, the examiner tells a short narrative to a subject, who then retells the story to the examiner, immediately and after a 30-minute delay. WLM scores are obtained by counting the number of story elements recalled. @Inp:Toth:2015:Automatic and @Inp:Vincze:2016:Detecting used short animated films to evaluate immediate and delayed recalls in MCI patients who were asked to talk about the first film shown, then about their previous day, and finally about another film shown last. @Inp:Toth:2015:Automatic adopted automatic speech recognition (ASR) to extract a phonetic level segmentation, which was used to calculate acoustic features. @Inp:Vincze:2016:Detecting used speech, morphological, semantic, and demographic features collected from their speech transcripts to automatically identify patients suffering from MCI. For the Portuguese language, machine learning algorithms were used to identify subjects with AD and MCI. @Inp:Aluisio:2016:Evaluating used a variety of linguistic metrics, such as syntactic complexity, idea density [@Inp:Cunha:2015:Automatic], and text cohesion through latent semantics. NLP tools with high precision are needed to compute these metrics, which is a problem for Portuguese since no robust dependency or constituency parsers exist. Therefore, the transcriptions had to be manually revised; they were segmented into sentences, following a semantic-structural criterion and capitalization was applied. The authors also removed disfluencies and inserted omitted subjects when they were hidden, in order to reduce parsing errors. This process is obviously expensive, which has motivated us to use complex networks in the present study to model transcriptions and avoid a manual preprocessing step. Modeling and Characterizing Texts as Complex Networks ===================================================== The theory and concepts of complex networks have been used in several NLP tasks [@Boo:Mihalcea:2011:Graph:NLP; @Cong2014598], such as text classification [@Art:Arruda:2016:Classification:Texts], summarization [@Antiqueira2009584; @Art:Amancio:2012:Extractive] and word sense disambiguation [@0295-5075-98-5-58001]. In this study, we used the word co-occurrence model (also called word adjacency model) because most of the syntactical relations occur among neighboring words [@Art:Cancho:2004:Patterns]. Each distinct word becomes a node and words that are adjacent in the text are connected by an edge. Mathematically, a network is defined as an undirected graph $G = \{V, E\}$, formed by a set $V = \{v_1, v_2, ..., v_n\}$ of nodes (words) and a set $E = \{e_1, e_2, ..., e_m\}$ of edges (co-occurrence) that are represented by an adjacency matrix $A$, whose elements $A_{ij}$ are equal to $1$ whenever there is an edge connecting nodes (words) $i$ and $j$, and equal to $0$ otherwise. Before modeling texts into complex networks, it is often necessary to do some preprocessing in the raw text. Preprocessing starts with tokenization where each document/text is divided into tokens (meaningful elements, e.g., words and punctuation marks) and then *stopwords* and punctuation marks are removed, since they have little semantic meaning. One last step we decided to eliminate from the preprocessing pipeline is lemmatization, which transforms each word into its canonical form. This decision was made based on two factors. First, a recent work has shown that lemmatization has little or no influence when network modeling is adopted in related tasks [@Art:Machicao:2016:Lemma:Influence]. Second, the lemmatization process requires part-of-speech (POS) tagging that may introduce undesirable noises/errors in the text, since the transcriptions in our work contain disfluencies. ![Example of co-occurrence network enriched with semantic information for the following transcription: “*The water’s running on the floor. Boy’s taking cookies out of cookie out of the cookie jar. The stool is falling over. The girl was asking for a cookie.*”. The solid edges of the network represent co-occurrence edges and the dotted edges represent connections between words that had similarity higher than $0.5$. \[fig:complex:network\]](rede_ex_leandro.pdf) Another problem with transcriptions in our work is their size. As demonstrated by @Art:Amancio:2015:Short:Texts, classification of small texts using networks can be impaired, since short texts have almost linear networks, and the topological measures of these networks have little or no information relevant to classification. To solve this problem, we adapted the approach of inducing language networks from word embeddings, proposed by @Col:Perozzi:2014:Inducing to enrich the networks with semantic information. In their work, language networks were generated from continuous word representations, in which each word is represented by a dense, real-valued vector obtained by training neural networks in the language model task (or variations, such as context prediction) [@Art:Bengio:2003:Neural; @Art:Collobert:2011:NLP; @Art:Mikolov:2013:Exploiting; @Inp:Mikolov:2013:Distributed]. This structure is known to capture syntactic and semantic information. @Col:Perozzi:2014:Inducing, in particular, take advantage of word embeddings to build networks where each word is a vertex and edges are defined by similarity between words established by the proximity of the word vectors. Following this methodology, in our model we added new edges to the co-occurrence networks considering similarities between words, that is, for all pairs of words in the text that were not connected, an edge was created if their vectors (from word embedding) had a cosine similarity higher than a given threshold. Figure \[fig:complex:network\] shows an example of a co-occurrence network enriched by similarity links (the dotted edges). The gain in information by enriching a co-occurrence network with semantic information is readily apparent in Figure  \[fig:complex:network:transcriotion\]. [0.22]{} ![Example of (a) co-occurrence network created for a transcript of the Cookie Theft dataset (see Supplementary Information, Section \[sec:supplemental\]) and (b) the same co-occurrence network enriched with semantic information. Note that (b) is a more informative network than (a), since (a) is practically a linear network. \[fig:complex:network:transcriotion\]](rede_en_coo "fig:"){width="\textwidth"} [0.22]{} ![Example of (a) co-occurrence network created for a transcript of the Cookie Theft dataset (see Supplementary Information, Section \[sec:supplemental\]) and (b) the same co-occurrence network enriched with semantic information. Note that (b) is a more informative network than (a), since (a) is practically a linear network. \[fig:complex:network:transcriotion\]](rede_en_emb "fig:"){width="\textwidth"} Datasets, Features and Methods\[sec:methods\] ============================================= Datasets -------- The datasets[^2] used in our study consisted of: (i) manually segmented and transcribed samples from the DementiaBank and Cinderella story and (ii) transcribed samples of Arizona Battery for Communication Disorders of Dementia (ABCD) automatically segmented into sentences, since we are working towards a fully automated system to detect MCI in transcripts and would like to evaluate a dataset which was automatically processed. The DementiaBank dataset is composed of short English descriptions, while the Cinderella dataset contains longer Brazilian Portuguese narratives. ABCD dataset is composed of very short narratives, also in Portuguese. Below, we describe in further detail the datasets, participants, and the task in which they were used. ### The Cookie Theft Picture Description Dataset The clinical dataset used for the English language was created during a longitudinal study conducted by the University of Pittsburgh School of Medicine on Alzheimer’s and related dementia, funded by the National Institute of Aging. To be eligible for inclusion in the study, all participants were required to be above 44 years of age, have at least 7 years of education, no history of nervous system disorders nor be taking neuroleptic medication, have an initial Mini-Mental State Exam (MMSE) score of 10 or greater, and be able to give informed consent. The dataset contains transcripts of verbal interviews with AD and related Dementia patients, including those with MCI (for further details see [@Art:Becker:1994:Natural]). We used 43 transcriptions with MCI in addition to another 43 transcriptions sampled from 242 healthy elderly people to be used as the control group. Table \[tab:demographic:talkbank\] shows the demographic information for the two diagnostic groups. Demographic Control MCI -------------------- ------------ ------------ Avg. Age (SD) 64.1 (7.2) 69.3 (8.2) No. of Male/Female 23/20 27/16 : Demographic information of participants in the Cookie Theft dataset.[]{data-label="tab:demographic:talkbank"} For this dataset, interviews were conducted in English and narrative speech was elicited using the Cookie Theft picture [@Book:Goodglass:2001:Assessment] (Figure \[fig:book:Cookie:Theft\] from @Book:Goodglass:2001:Assessment in Section \[sec:supplemental:example\]). During the interview, patients were given the picture and were told to discuss everything they could see happening in the picture. The patients’ verbal utterances were recorded and then transcribed into the CHAT (Codes for the Human Analysis of Transcripts) transcription format [@Book:Macwhinney:2000:Childes]. We extracted the word-level transcript patient sentences from the CHAT files and discarded the annotations, as our goal was to create a fully automated system that does not require the input of a human annotator. We automatically removed filled pauses such as *uh*, *um* , *er* , and *ah* (e.g. *uh it seems to be summer out*), short false starts (e.g. *just t the ones* ), and repetition (e.g. *mother’s finished certain of the the dishes* ), as in [@Art:Fraser:2015:linguistic]. The control group had an average of 9.58 sentences per narrative, with each sentence having an average of 9.18 words; while the MCI group had an average of 10.97 sentences per narrative, with 10.33 words per sentence in average. ### The Cinderella Narrative Dataset The dataset examined in this study included 20 subjects with MCI and 20 normal elderly control subjects, as diagnosed at the Medical School of the University of São Paulo (FMUSP). Table \[tab:demographic:data\] shows the demographic information of the two diagnostic groups, which were also used in @Inp:Aluisio:2016:Evaluating. Demographic Control MCI -------------------- ------------- ------------ Avg. Age (SD) 74.8 (11.3) 73.3 (5.9) Avg. Years of Education (SD) No. of Male/Female 27/16 29/14 : Demographic information of participants in the Cinderella dataset.[]{data-label="tab:demographic:data"} The criteria used to diagnose MCI came from @Art:Petersen:2004:MCI. Diagnostics were carried out by a multidisciplinary team consisting of psychiatrists, geriatricians, neurologists, neuropsychologists, speech pathologists, and occupational therapists, by a criterion of consensus. Inclusion criteria for the control group were elderlies with no cognitive deficits and preservation of functional capacity in everyday life. The exclusion criteria for the normal group were: poorly controlled clinical diseases, sensitive deficits that were not being compensated for and interfered with the performance in tests, and other neurological or psychiatric diagnoses associated with dementia or cognitive deficits and use of medications in doses that affected cognition. Speech narrative samples were elicited by having participants tell the Cinderella story; participants were given as much time as they needed to examine a picture book illustrating the story (Figure \[fig:book:cinderella\] in Section \[sec:supplemental\]). When each participant had finished looking at the pictures, the examiner asked the subject to tell the story in their own words, as in @Art:Saffran:1989:Quantitative. The time was recorded, but there was no limit imposed to the narrative length. If the participant had difficulty initiating or continuing speech, or took a long pause, an evaluator would use the stimulus question “What happens next ?”, seeking to encourage the participant to continue his/her narrative. When the subject was unable to proceed with the narrative, the examiner asked if he/she had finished the story and had something to add. Each speech sample was recorded and then manually transcribed at the word level following the NURC/SP N. 338 EF and 331 D2 transcription norms[^3]. Other tests were applied after the narrative, in the following sequence: phonemic verbal fluency test, action verbal fluency, Camel and Cactus test [@Art:Bozeat:2000:Non:Verbal], and Boston Naming test [@Book:Kaplan:2005:Boston], in order to diagnose the groups. Since our ultimate goal is to create a fully automated system that does not require the input of a human annotator, we manually segmented sentences to simulate a high-quality ASR transcript with sentence segmentation, and we automatically removed the disfluencies following the same guidelines of TalkBank project. However, other disfluencies (revisions, elaboration, paraphasias and comments about the task) were kept. The control group had an average of 30.80 sentences per narrative, and each sentence averaged 12.17 words. As for the MCI group, it had an average of 29.90 sentences per narrative, and each sentence averaged 13.03 words. We also evaluated a different version of the dataset used in @Inp:Aluisio:2016:Evaluating, where narratives were manually annotated and revised to improve parsing results. The revision process was the following: (i) in the original transcript, segments with hesitations or repetitions of more than one word or segment of a single word were annotated to become a feature and then removed from the narrative to allow the extraction of features from parsing; (ii) empty emissions, which were comments unrelated to the topic of narration or confirmations, such as “*né*” (alright), were also annotated and removed; (iii) prolongations of vowels, short pauses and long pauses were also annotated and removed; and (iv) omitted subjects in sentences were inserted. In this revised dataset, the control group had an average of 45.10 sentences per narrative, and each sentence averaged 8.17 words. The MCI group had an average of 31.40 sentences per narrative, with each sentence averaging 10.91 words. ### The ABCD Dataset The subtest of immediate/delayed recall of narratives of the ABCD battery was administered to 23 participants with a diagnosis of MCI and 20 normal elderly control participants, as diagnosed at the Medical School of the University of São Paulo (FMUSP). MCI subjects produced 46 narratives while the control group produced 39 ones. In order to carry out experiments with a balanced corpus, as with the previous two datasets, we excluded seven transcriptions from the MCI group. We used the automatic sentence segmentation method referred to as DeepBond [@Inp:Treviso:2017:DeepBond] in the transcripts. Table \[tab:demographic:data:abcd\] shows the demographic information. The control group had an average of 5.23 sentences per narrative, with 11 words per sentence on average, and the MCI group had an average of 4.95 sentences per narrative, with an average of 12.04 words per sentence. Interviews were conducted in Portuguese and the subject listened to the examiner read a short narrative. The subject then retold the narrative to the examiner twice: once immediately upon hearing it and again after a 30-minute delay [@Book:Bayles:1991:ABCD]. Each speech sample was recorded and then manually transcribed at the word level following the NURC/SP N. 338 EF and 331 D2 transcription norms. Demographic Control MCI -------------------- ---------- ------------ Avg. Age (SD) 61 (7.5) 72,0 (7.4) Avg. Years of Education (SD) No. of Male/Female 6/14 16/7 : Demographic information of participants in the ABCD dataset.[]{data-label="tab:demographic:data:abcd"} Features -------- Features of three distinct natures were used to classify the transcribed texts: topological metrics of co-occurrence networks, linguistic features and bag of words representations. ### Topological Characterization of Networks Each transcription was mapped into a co-occurrence network, and then enriched via word embeddings using the cosine similarity of words. Since the occurrence of out-of-vocabulary words is common in texts of neuropsychological assessments, we used the method proposed by @Art:Bojanowski:2016:Enriching to generate word embeddings. This method extends the skip-gram model to use character-level information, with each word being represented as a bag of character $n$-grams. It provides some improvement in comparison with the traditional skip-gram model in terms of syntactic evaluation [@Inp:Mikolov:2013:Distributed] but not for semantic evaluation. Once the network has been enriched, we characterize its topology using the following ten measurements: 1. **PageRank:** is a centrality measurement that reflects the relevance of a node based on its connections to other relevant nodes [@Inp:Brin:2012:PageRank]; 2. **Betweenness:** is a centrality measurement that considers a node as relevant if it is highly accessed via shortest paths. The betweenness of a node $v$ is defined as the fraction of shortest paths going through node $v$; 3. **Eccentricity:** of a node is calculated by measuring the shortest distance from the node to all other vertices in the graph and taking the maximum; 4. **Eigenvector centrality:** is a measurement that defines the importance of a node based on its connectivity to high-rank nodes; 5. **Average Degree of the Neighbors of a Node:** is the average of the degrees of all its direct neighbors; 6. **Average Shortest Path Length of a Node:** is the average distance between this node and all other nodes of the network; 7. **Degree:** is the number of edges connected to the node; 8. **Assortativity Degree:** or degree correlation measures the tendency of nodes to connect to other nodes that have similar degree; 9. **Diameter:** is defined as the maximum shortest path; 10. **Clustering Coefficient:** measures the probability that two neighbors of a node are connected. Most of the measurements described above are local measurements, i.e. each node $i$ possesses a value $X_i$, so we calculated the average $\mu(X)$, standard deviation $\sigma(X)$ and skewness $\gamma(X)$ for each measurement [@Art:Amancio:2015:Complex]. ### Linguistic Features Linguistic features for classification of neuropsychological assessments have been used in several studies [@Art:Roark:2011:Spoken; @Inp:Jarrold:2014:Aided; @Art:Fraser:2014:Automated; @Inp:Orimaye:2014:Learning; @Art:Fraser:2015:linguistic; @Inp:Vincze:2016:Detecting; @Inp:Davy:2016:Towards]. We used the Coh-Metrix[^4][@Art:Graesser:2004:Coh:Metrix] tool to extract features from English transcripts, resulting in 106 features. The metrics are divided into eleven categories: Descriptive, Text Easability Principal Component, Referential Cohesion, Latent Semantic Analysis (LSA), Lexical Diversity, Connectives, Situation Model, Syntactic Complexity, Syntactic Pattern Density, Word Information, and Readability (Flesch Reading Ease, Flesch-Kincaid Grade Level, Coh-Metrix L2 Readability). For Portuguese, Coh-Metrix-Dementia [@Inp:Aluisio:2016:Evaluating] was used. The metrics affected by constituency and dependency parsing were not used because they are not robust with disfluencies. Metrics based on manual annotation (such as proportion short pauses, mean pause duration, mean number of empty words, and others) were also discarded. The metrics of Coh-Metrix-Dementia are divided into twelve categories: Ambiguity, Anaphoras, Basic Counts, Connectives, Co-reference Measures, Content Word Frequencies, Hypernyms, Logic Operators, Latent Semantic Analysis, Semantic Density, Syntactical Complexity, and Tokens. The metrics used are shown in detail in Section \[sec:supplemental:metrics\]. In total, 58 metrics were used, from the 73 available on the website[^5]. ### Bag of Words The representation of text collections under the BoW assumption (i.e., with no information relating to word order) has been a robust solution for text classification. In this methodology, transcripts are represented by a table in which the columns represent the terms (or existing words) in the transcripts and the values represent frequency of a term in a document. Classification Algorithms ------------------------- In order to quantify the ability of the topological characterization of networks, linguistic metrics and BoW features were used to distinguish subjects with MCI from healthy controls. We employed four machine learning algorithms to induce classifiers from a training set. These techniques were the Gaussian Naive Bayes (G-NB), $k$-Nearest Neighbor ($k$-NN), Support Vector Machine (SVM), linear and radial bases functions (RBF), and Random Forest (RF). We also combined these classifiers through ensemble and multi-view learning. In ensemble learning, multiple models/classifiers are generated and combined using a majority vote or the average of class probabilities to produce a single result [@Book:Zhou:2012:Ensemble]. In multi-view learning, multiple classifiers are trained in different feature spaces and thus combined to produce a single result. This approach is an elegant solution in comparison to combining all features in the same vector or space, for two main reasons. First, combination is not a straightforward step and may lead to noise insertion since the data have different natures. Second, using different classifiers for each feature space allows for different weights to be given for each type of feature, and these weights can be learned by a regression method to improve the model. In this work, we used majority voting to combine different feature spaces. Experiments and Results\[sec:experiments\] ========================================== All experiments were conducted using the Scikit-learn[^6] [@Art:Pedregosa:scikit-learn:2011], with classifiers evaluated on the basis of classification accuracy i.e. the total proportion of narratives which were correctly classified. The evaluation was performed using 5-fold cross-validation instead of the well-accepted 10-fold cross-validation because the datasets in our study were small and the test set would have shrunk, leading to less precise measurements of accuracy. The threshold parameter was optimized with the best values being $0.7$ in the Cookie Theft dataset and $0.4$ in both the Cinderella and ABCD datasets. We used the model proposed by @Art:Bojanowski:2016:Enriching with default parameters (100 dimensional embeddings, context window equal to 5 and 5 epochs) to generate word embedding. We trained the models in Portuguese and English Wikipedia dumps from October and November 2016 respectively. The accuracy in classification is given in Tables \[results:en\] through \[results:pt:abcd\]. CN, CNE, LM, and BoW denote, respectively, complex networks, complex network enriched with embedding, linguistic metrics and Bag of Words, and CNE-LM, CNE-BoW, LM-BoW and CNE-LM-BoW refer to combinations of the feature spaces (multiview learning), using the majority vote. Cells with the “–” sign mean that it was not possible to apply majority voting because there were two classifiers. The last line represents the use of an ensemble of machine learning algorithms, in which the combination used was the majority voting in both ensemble and multiview learning. Classifier CN CNE LM BoW CNE-LM CNE-BoW LM-BoW CNE-LM-BoW ------------ -------- -------- -------- -------- -------- --------- -------- ------------ SVM-Linear 52 55 56 59 – – – 60 SVM-RBF 56 **62** **58** **60** – – – **65** $k$-NN **59** 61 46 57 – – – 59 RF 52 47 45 48 – – – 50 G-NB 51 48 56 55 – – – 50 Ensemble 56 60 54 58 57 60 63 **65** Classifier CN CNE LM BoW CNE-LM CNE-BoW LM-BoW CNE-LM-BoW ------------ -------- -------- -------- -------- -------- --------- -------- ------------ SVM-Linear 52 60 **52** 50 – – – **52** SVM-RBF **57** **65** 47 37 – – – 50 $k$-NN 47 50 47 37 – – – 37 RF 55 57 47 45 – – – **52** G-NB 47 52 47 **55** – – – **52** Ensemble 52 60 50 37 57 52 50 47 Classifier CN CNE LM BoW CNE-LM CNE-BoW LM-BoW CNE-LM-BoW ------------ -------- -------- -------- -------- -------- --------- -------- ------------ -- -- SVM-Linear 56 **69** 51 **75** – – – **74** SVM-RBF 54 57 66 67 – – – 71 $k$-NN 56 56 69 63 – – – 71 RF 54 62 **70** 64 – – – 69 G-NB **61** 55 55 65 – – – 65 Ensemble 55 61 62 72 69 68 75 73 In general, CNE outperforms the approach using only complex networks (CN), while SVM (Linear or RBF kernel) provides higher accuracy than other machine learning algorithms. The results for the three datasets show that characterizing transcriptions into complex networks is competitive with other traditional methods, such as the use of linguistic metrics. In fact, among the three types of features, using enriched networks (CNE) provided the highest accuracies in two datasets (Cookie Theft and original Cinderella). For the ABCD dataset, which contains short narratives, the small length of the transcriptions may have had an effect, since BoW features led to the highest accuracy. In the case of the revised Cinderella dataset, segmented into sentences and capitalized as reported in @Inp:Aluisio:2016:Evaluating, Table \[results:pt:andre\] shows that the manual revision was an important factor, since the highest accuracies were obtained with the approach based on linguistic metrics (LM). However, this process of manually removing disfluencies demands time; therefore it is not practical for large-scale assessments. Ensemble and multi-view learning were helpful for the Cookie Theft dataset, in which multi-view learning achieved the highest accuracy (65% of accuracy for narrative texts, a 3% of improvement compared to the best individual classifier). However, neither multi-view or ensemble learning enhanced accuracy in the Cinderella dataset, where SVM-RBF with CNE space achieved the highest accuracy (65%). For the ABCD dataset, multi-view CNE-LM-BoW with SVM-RBF and KNN classifiers improved the accuracy to 4% and 2%, respectively. Somewhat surprising were the results of SVM with linear kernel in BoW feature space (75% of accuracy). Classifier CN CNE LM BoW ------------ -------- -------- -------- -------- -- -- -- SVM-Linear 50 65 65 52 SVM-RBF **57** **67** **72** **55** KNN 42 47 55 50 RF 52 47 70 45 G-NB 52 65 62 45 Ensemble 52 60 **72** 45 : Classification accuracy achieved on Cinderella dataset manually processed to revise non-grammatical sentences.[]{data-label="results:pt:andre"} Conclusions and Future Work \[sec:conclusion\] ============================================== In this study, we employed metrics of topological properties of CN in a machine learning classification approach to distinguish between healthy patients and patients with MCI. To the best of our knowledge, these metrics have never been used to detect MCI in speech transcripts; CN were enriched with word embeddings to better represent short texts produced in neuropsychological assessments. The topological properties of CN outperform traditional linguistic metrics in individual classifiers’ results. Linguistic features depend on grammatical texts to present good results, as can be seen in the results of the manually processed Cinderella dataset (Table \[results:pt:andre\]). Furthermore, we found that combining machine and multi-view learning can improve accuracy. The accuracies found here are comparable to the values reported by other authors, ranging from 60% to 85% [@Inp:Prud'hommeaux:2011:Alignment; @Inp:Lehr:2012:Fully; @Inp:Toth:2015:Automatic; @Inp:Vincze:2016:Detecting], which means that it is not easy to distinguish between healthy subjects and those with cognitive impairments. The comparison with our results is not straightforward, though, because the databases used in the studies are different. There is a clear need for publicly available datasets to compare different methods, which would optimize the detection of MCI in elderly people. In future work, we intend to explore other methods to enrich CN, such as the Recurrent Language Model, and use other metrics to characterize an adjacency network. The pursuit of these strategies is relevant because language is one of the most efficient information sources to evaluate cognitive functions, commonly used in neuropsychological assessments. As this work is ongoing, we will keep collecting new transcriptions of the ABCD retelling subtest to increase the corpus size and obtain more reliable results in our studies. Our final goal is to apply neuropsychological assessment batteries, such as the ABCD retelling subtest, to mobile devices, specifically tablets. This adaptation will enable large-scale applications in hospitals and facilitate the maintenance of application history in longitudinal studies, by storing the results in databases immediately after the test application. Acknowledgments {#acknowledgments .unnumbered} =============== This work was supported by CAPES, CNPq, FAPESP, and Google Research Awards in Latin America. We would like to thank NVIDIA for their donation of GPU. Supplementary Material {#sec:supplemental} ====================== Figure \[fig:book:Cookie:Theft\] is Cookie Theft picture, which was used in DementiaBank project. Figure \[fig:book:cinderella\] is a sequence of pictures from the Cinderella story, which were used to elicit speech narratives. ![The Cookie Theft Picture, taken from the Boston Diagnostic Aphasia Examination [@Book:Goodglass:2001:Assessment].[]{data-label="fig:book:Cookie:Theft"}](cookie_theft) ![Sequence of Pictures of the of Cinderella story.[]{data-label="fig:book:cinderella"}](Cinderela) Examples of transcriptions\[sec:supplemental:example\] ------------------------------------------------------ Below follows an example of a transcript of the Cookie Theft dataset. You just want me to start talking ? Well the little girl is asking her brother we ’ll say for a cookie . Now he ’s getting the cookie one for him and one for her . He unbalances the step the little stool and he ’s about to fall . And the lid ’s off the cookie jar . And the mother is drying the dishes abstractly so she ’s left the water running in the sink and it is spilling onto the floor . And there are two there ’s look like two cups and a plate on the sink and board . And that boy ’s wearing shorts and the little girl is in a short skirt . And the mother has an apron on . And she ’s standing at the window . The window ’s opened . It must be summer or spring . And the curtains are pulled back . And they have a nice walk around their house . And there ’s this nice shrubbery it appears and grass . And there ’s a big picture window in the background that has the drapes pulled off . There ’s a not pulled off but pulled aside . And there ’s a tree in the background . And the house with the kitchen has a lot of cupboard space under the sink board and under the cabinet from which the cookie you know cookies are being removed . Below follows an excerpt of a transcript of the Cinderella dataset. **Original transcript in Portuguese:** ela morava com a madrasta as irmã né e ela era diferenciada das três era maltratada ela tinha que fazer limpeza na casa toda no castelo alias e as irmãs não faziam nada até que um dia chegou um convite do rei ele ia fazer um baile e a madrasta então é colocou que todas as filhas elas iam menos a cinderela bom como ela não tinha o vestido sapato as coisas tudo então ela mesmo teve que fazer a roupa dela começou a fazer ... **Translation of the transcript in English:** she lived with the stepmother the sister right and she was differentiated from the three was mistreated she had to do the cleaning in the entire house actually in the castle and the sisters didn’t do anything until one day the king’s invitation arrived he would invite everyone to a ball and then the stepmother is said that all the daughters they would go except for cinderella well since she didn’t have a dress shoes all the things she had to make her own clothes she started to make them ... Coh-Metrix-Dementia metrics\[sec:supplemental:metrics\] ------------------------------------------------------- 1. **Ambiguity**: verb ambiguity, noun ambiguity, adjective ambiguity, adverb ambiguity; 2. **Anaphoras**: adjacent anaphoric references, anaphoric references; 3. **Basic Counts**: Flesch index, number of word, number of sentences, number of paragraphs, words per sentence, sentences per paragraph, syllables per content word, verb incidence, noun incidence, adjective incidence, adverb incidence, pronoun incidence, content word incidence, function word incidence; 4. **Connectives**: connectives incidence, additive positive connectives incidence, additive negative connectives incidence, temporal positive connectives incidence, temporal negative connectives incidence, casual positive connectives incidence, casual negative connectives incidence, logical positive connectives incidence, logical negative connectives incidence; 5. **Co-reference Measures**: adjacent argument overlap, argument overlap, adjacent stem overlap, stem overlap, adjacent content word overlap; 6. **Content Word Frequencies**: Content words frequency, minimum among content words frequency; 7. **Hypernyms**: Mean hypernyms per verb; 8. **Logic Operators**: Logic operators incidence, and incidence, or incidence, if incidence, negation incidence; 9. **Latent Semantic Analysis (LSA)**: Average and standard deviation similarity between pairs of adjacent sentences in the text, Average and standard deviation similarity between all sentence pairs in the text, Average and standard deviation similarity between pairs of adjacent paragraphs in the text, Givenness average and standard deviation of each sentence in the text; 10. **Semantic Density**: content density; 11. **Syntactical Complexity**: only cross entropy; 12. **Tokens**: personal pronouns incidence, type-token ratio, Brunet index, Honoré Statistics. [^1]: [talkbank.org/DementiaBank/](talkbank.org/DementiaBank/) [^2]: All datasets are made available in the same representations used in this work, upon request to the authors. [^3]: [albertofedel.blogspot.com.br/2010\_11\_01\_archive.html](albertofedel.blogspot.com.br/2010_11_01_archive.html) [^4]: [cohmetrix.com](cohmetrix.com) [^5]: <http://143.107.183.175:22380> [^6]: <http://scikit-learn.org>
{ "pile_set_name": "ArXiv" }
--- abstract: 'We present a detailed multi-epoch analysis of 31 potential southern hemisphere radio calibrators that were originally observed as part of a program to maintain the International Celestial Reference Frame (ICRF). At radio wavelengths, the primary calibrators are Active Galactic Nuclei (AGN), powerful radio emitters which exist at the centre of most galaxies. These are known to vary at all wavelengths at which they have been observed. By determining the amount of radio source structure and variability of these AGN, we determine their suitability as phase calibrators for long baseline radio interferometry at 2.3 GHz. For this purpose, we have used a set of complementary metrics to classify these 31 southern sources into five categories pertaining to their suitability as VLBI calibrators. We find that all of the sources in our sample would be good interferometric calibrators and almost ninety per cent would be very good calibrators.' author: - | F. Hungwe$^{1,2}$[^1], R. Ojha$^{3,4}$, R.S Booth$^{1,2}$, M.F Bietenholz$^{1,5}$, A. Collioud$^{6,7}$, P. Charlot$^{6,7}$ D. Boboltz$^{4}$, A.L Fey$^{4}$\ $^{1}$ Dept. of Physics and Electronics, Rhodes University, P.O Box 94, Grahamstown, 6140, South Africa\ $^{2}$ Hartebeesthoek Radio Astronomy Observatory, P.O Box 443, Krugersdorp, 1740, South Africa\ $^{3}$ NVI, Inc., 7257D Hanover Parkway, Greenbelt, MD 20770, USA\ $^{4}$ United States Naval Observatory, 3450 Massachusetts Ave., NW, Washington D.C 20392, USA\ $^{5}$ Dept. of Physics and Astronomy, York University, Toronto, M3J 1P3, Ontario, Canada\ $^{6}$ Université de Bordeaux, Observatoire Aquitain des Sciences de l’Univers, BP 89, 33271 Floirac Cedex, France\ $^{7}$ CNRS, Laboratoire d’Astrophysique de Bordeaux-UMR 5804, BP 89, 33271 Floirac Cedex, France date: 'Accepted 2011 June 11. Received 2011 May 24; in original form 2010 October 29' title: 'Characterisation of Long Baseline Calibrators at 2.3 GHz' --- \[firstpage\] quasar: general - galaxies: jets - radio continuum: galaxies - techniques: interferometric Introduction ============ High angular resolution observations of weak radio sources (where self calibration is not possible) require calibrator sources for correction of systematic effects and effects of the atmosphere on the measured visibilities. Atmospheric fluctuations cause perturbations in visibility phase which, if not corrected, seriously limit both the sensitivity and image quality of an interferometric array. Phase calibrators are also required for astrometric observations. An ideal calibrator would look the same on all observing baselines. It should be bright, unresolved, or at least compact and should not vary. A calibrator source should also be separated from the target sources by as small an angle as possible in order to look along the same line of sight through the atmosphere. Therefore, it is desirable to have calibrator sources evenly distributed across the whole sky. In practice, at radio wavelengths, calibrators are mostly Active Galactic Nuclei (AGN), whose fundamental source of power is believed to be the accretion of matter onto a super-massive black hole [@Rees:97]. They are known to vary at every wavelength at which they have been studied. AGN are very compact and isotropically distributed around the sky. They are also very distant objects and therefore, generally, have no discernible proper motions on the sky. It is these qualities that make them suitable as calibrators at radio wavelengths. Due to the limited number of radio telescopes, few surveys for calibrators have been carried out in the southern hemisphere [@Ojha:04a; @Ojha:05; @Fey:04a; @Fey:04b; @Fey:06]. Hence, there are fewer known calibrators in the south. A major expansion of radio astronomy observing capability is underway in the southern hemisphere. Two SKA (Square Kilometre Array) precursors, the South African MeerKAT (Karoo Array Telescope, @Booth:09) and ASKAP (Australian SKA Pathfinder, @Johnston:08) are presently under construction, leading to the SKA itself. It is clear that interferometry and Very Long Baseline Interferometry (VLBI) in the southern hemisphere needs a dense network of calibration sources at high resolutions and a range of frequencies. MeerKAT is the South African SKA demonstrator telescope. When completed, it will be the largest radio telescope array in the southern hemisphere. It will operate in two frequency ranges, 0.58-2.5 GHz and 8-14.5 GHZ [@Booth:09]. MeerKAT will participate in VLBI observations with the European VLBI Network and other VLBI arrays. The SKA will operate in the frequency range 0.7 to 25 GHz [@Carilli:04; @Schilizzi:07] and is expected to have much longer baselines for which a southern VLBI calibrator list will be essential. The United States Naval Observatory, in collaboration with NASA, the Laboratoire d’Astrophysique de Bordeaux (LAB) and the National Radio Astronomy Observatory (NRAO[^2]) has, since 1994 been observing AGN, once every two months, through the Research and Development VLBI (RDV) experiments. This has led to a wealth of data of sources observed simultaneously at 2.3 GHz and 8.4 GHz at regular epochs. Images of these AGN form the USNO’s Radio Reference Frame Image Database (RRFID) and LAB’s Bordeaux VLBI Image Database (BVID)[^3] [@Fey:96; @Fey:97; @Fey:00; @Colli:09]. The RRFID is a database of about 6700 images of over 700 AGN sources compiled from geodetic and astrometric VLBI experiments. The BVID contains over 1800 images of 824 radio sources. It is the goal of the ongoing RDV program to image the radio reference frame sources on a regular basis, to monitor them for variability and structural change. @Piner:07 discuss jet kinematics in a subset of the RRFID sources, concentrating on the 8.4 GHz observations. Their survey is made up of RRFID sources that have been observed at 3 or more epochs from July 1994 to December 1998. In this paper, we focus on 2.3 GHz observations of a sample of 31 potential southern VLBI calibrators (with declinations between 0 and $-60^{\circ}$) from the kinematic survey of @Piner:07. The choice of 2.3 GHz was made because both MeerKAT and ASKAP will be operating at frequencies close to this. We seek to characterise this selection of southern radio sources, and determine their suitability as calibrators for southern VLBI experiments, especially those using MeerKAT and the SKA when completed. In the following sections, we will describe the observations and data reduction process. We will then describe the imaging and modelfitting process and go on to analyse the data and discuss the results. ------------- --------- --------- ------------- ------------------- ------------------- ------------- ------------ ---------- -- B1950 Other Optical $z$ Latest Source Name Name ID RA (hh mm ss) DEC (deg mm ss) $S_{total}$ $S_{core}$ Epoch 0003-066 BL Lac 0.35$^{1}$ 00 06 13.89288849 -06 23 35.3353162 2.29 1.79 Jan 2008 0104-408 BL Lac 0.58$^{2}$ 01 06 45.10796851 -40 34 19.9602291 1.51 1.52 Jan 2008 0238-084 NGC1052 Galaxy 0.005$^{3}$ 02 41 04.79850256 -08 15 20.7517956 0.53 0.26 Dec 2007 0336-019 CTA26 Quasar 0.85$^{4}$ 03 39 30.93778751 -01 46 35.8041062 2.53 2.16 Jan 2008 0402-362 Quasar 1.42$^{5}$ 04 03 53.74989835 -36 05 01.9131085 0.91 0.99 Mar 2007 0454-234 Quasar 1.00$^{1}$ 04 57 03.17922863 -23 24 52.0201418 3.41 3.01 Dec 2007 0458-020 Quasar 2.29$^{6}$ 05 01 12.80988366 -01 59 14.2562534 0.82 0.65 Dec 2007 0727-115 Quasar 1.59$^{7}$ 07 30 19.11247420 -11 41 12.6005110 3.77 3.47 Jan 2008 0919-260 Quasar 2.30$^{8}$ 09 21 29.35385535 -26 18 43.3861684 1.58 1.50 Sep 2006 0920-397 Quasar 0.59$^{9}$ 09 22 46.41826064 -39 59 35.0683561 1.56 1.49 Jan 2008 1034-293 Quasar 0.31$^{1}$ 10 37 16.07973476 -29 34 02.8133345 1.60 1.51 Jan 2008 1124-186 Quasar 1.05$^{10}$ 11 27 04.39244958 -18 57 17.4416582 1.16 1.16 Jan 2008 1144-379 Quasar 1.05$^{1}$ 11 47 01.37070177 -38 12 11.0234199 1.09 1.09 Jan 2007 1145-071 Quasar 1.34$^{11}$ 11 47 51.55402876 -07 24 41.1410887 0.95 0.81 Jan 2007 1253-055 3C279 Quasar 0.54$^{12}$ 12 56 11.16656541 -05 47 21.5247030 7.60 7.30 Dec 1998 1255-316 Quasar 1.92$^{13}$ 12 57 59.06081737 -31 55 16.8516980 1.51 1.38 Jan 2008 1313-333 Quasar 1.21$^{14}$ 13 16 07.98593995 -33 38 59.1725057 0.76 0.53 Feb 2004 1334-127 Quasar 0.54$^{15}$ 13 37 39.78277768 -12 57 24.6932620 2.61 2.57 Jan 2008 1351-018 Quasar 3.71$^{16}$ 13 54 06.89532213 -02 06 03.1904447 1.00 0.98 Jan 2008 1424-418 Quasar 1.52$^{2}$ 14 27 56.29756536 -42 06 19.4375991 2.04 1.66 Jan 2008 1451-375 Quasar 0.31$^{17}$ 14 54 27.40975442 -37 47 33.1448724 0.56 0.51 Jul 2006 1514-241 BL Lac 0.05$^{18}$ 15 17 41.81313221 -24 22 19.4760251 2.94 2.24 Jan 2007 1622-253 Quasar 0.79$^{19}$ 16 25 46.89164010 -25 27 38.3267989 1.01 0.99 Jan 2008 1741-038 Quasar 1.05$^{2}$ 17 43 58.85613396 -03 50 04.6166450 4.78 4.73 Jan 2008 1908-201 Quasar 1.12$^{20}$ 19 11 09.65289198 -20 06 55.1089891 2.10 1.78 Dec 2007 1921-293 Quasar 0.35$^{9}$ 19 24 51.05595514 -29 14 30.1210524 6.18 5.43 Dec 2007 1954-388 Quasar 0.63$^{21}$ 19 57 59.81927470 -38 45 06.3557585 2.75 2.69 Jan 2008 1958-179 Quasar 0.65$^{21}$ 20 00 57.09044485 -17 48 57.6725440 1.53 1.56 Dec 2007 2052-474 Quasar 1.49$^{13}$ 20 56 16.35981874 -47 14 47.6276461 1.37 1.43 Dec 2007 2243-123 Quasar 0.63$^{21}$ 22 46 18.23197613 -12 06 51.2774796 2.57 1.33 Jan 2008 2255-282 Quasar 0.93$^{5}$ 22 58 05.96288481 -27 58 21.2567425 0.95 0.87 Jan 2008 ------------- --------- --------- ------------- ------------------- ------------------- ------------- ------------ ---------- -- Observations and Data Reduction =============================== ![Sky distribution of the southern sources plotted on an Aitoff equal-area projection of the celestial sphere. The dotted line represents the Galactic plane while the dashed line is the ecliptic.](Figure1.eps){height="6cm"} \[skycoverage\] ![image](Figure2a.ps){height="7.6cm"} ![image](Figure2b.ps){height="7.6cm"} ![image](Figure2c.ps){height="7.6cm"} ![image](Figure2d.ps){height="7.6cm"} ![image](Figure2e.ps){height="7.6cm"} ![image](Figure2f.ps){height="7.6cm"} Research and Development VLBI (RDV) experiments are carried out using the 10 antennas of the NRAO’s VLBA and up to 10 other antennas across the globe, including Hartebeesthoek (South Africa) when available. The use of the global array greatly improves the  coverage. Hartebeesthoek greatly improves the  coverage for sources in the south, which make up the sample in this paper. The 31 southern sources included in our sample are listed in Table \[tsouthernsample\] along with their optical properties. The source sky distribution is shown in Figure \[skycoverage\]. The RDV observing epochs are from July 1994 to January 2008, giving a total of 32 epochs and an average of 20 epochs per source. About 100 sources were observed in each 24hr observing run, with an average on-source time of 15 minutes. The on-source time is not continuous, but divided into scans of between 1 and several minutes long, spaced in time to give optimal  coverage. Observations were made in a dual frequency bandwidth synthesis mode to facilitate delay measurements for astrometry. Observations in this mode also allow imaging at both frequency bands. Eight individual frequency bands (IFs) were recorded simultaneously, each 8 MHz wide, with 4 at 2.3 GHz and 4 at 8.4 GHz for a total of 32 MHz in each frequency band. The data were correlated with the VLBA correlator at the Array Operations Center in Socorro, New Mexico. The correlated data were calibrated and corrected for time and frequency dependent phase variations using NRAO’s Astronomical Imaging Processing System (AIPS, @Greisen:98). Initial amplitude calibration for each IF was accomplished using system temperature measurements and gain curves generated during observations. Fringe-fitting was done in AIPS using a solution interval equal to the scan duration and a point source model. Amplitude calibration was improved in a second stage by using observations of sources whose core flux density is known to be $\ge 90$ per cent of the total flux density. For a precise definition of the core, refer to Section 3.1. To this end, a single amplitude gain correction factor was derived for each antenna based on fitting a Gaussian model to the core component. Gain correction factors were then calculated based on the difference between the observed and the model visibilities. Finally, the amplitude gain correction factors were applied to the target sources. This is a non-standard procedure that improves the overall amplitude calibration. The accuracy of the amplitude calibration determined in this way is conservatively estimated to be within 20 per cent. Data imaging and modelfitting were done using the Caltech Difference Mapping program ‘Difmap’ [@Shepherd:95], after inspecting the data and editing out obvious bad points. The data were imaged using Difmap in automatic mode. Generally, this mode fails for about one third of the sources which have structure that is complex or too extended for the automatic script to handle. These have to be redone by hand in interactive mode. For the southern sources in this sample, almost all had to be imaged by hand due to the poor  coverage. Difmap combines the visibilities in all four 2.3 GHz receiver IFs but does not correct for spectral index effects. It was assumed that the source structure variations across the IFs were negligible (6% variation in frequency at this band). Uniform weighting was used for the initial phase self-calibration before changing to natural weighting. For our arrays, uniform weighting gives more weighting to the longer baselines whilst natural weighting gives more weighting to the shorter baselines. Images of six of the 31 sources in our sample are shown in Figure \[images\] with the remainder available online. Images of additional epochs can be found in the form of contour plots at the RRFID and BVID websites. Generally, circular Gaussian models were used to fit the  data in order to parametrise the source morphology. Like imaging, modelfitting is an iterative process. Elliptical Gaussian components were used only to represent the core component or a very bright jet component if the residuals remaining from a circular Gaussian model were too large and made it difficult to continue modelfitting using the residual map. The modelfits generally describe the visibility data well but these models may not be unique because of incomplete sampling in the  plane. In order to determine suitability of a source as a calibrator, we then determined the amount of source structure, as well as its variation with time, using several different methods based on both the CLEAN components from the Difmap imaging as well as on the parameters of the fitted Gaussians. Analysis and Results ==================== Characterisation of the morphology of an AGN is very complex and no single metric is able to adequately define a good calibrator. Thus we have developed a number of complementary approaches that are described below. A weighted combination of all of these metrics is eventually used to classify the suitability of each source to be a high resolution phase calibrator. These are discussed below. ---------- -------- ----------- ----------------- ------------------ ----------- -------------- --------------- --------------- ------------------ ------------------- ---------------- ------------------- -------------------- Source Epochs Name $\bar{S}$ $\sigma_{core}$ $\sigma_{core}/$ $\bar{C}$ $\sigma_{C}$ $\sigma_{C}/$ $\bar{R}_{W}$ $\sigma_{R_{W}}$ $\sigma_{R_{W}}/$ $\bar{R}_{UW}$ $\sigma_{R_{UW}}$ $\sigma_{R_{UW}}/$ $\bar{S}$ $\bar{C}$ $\bar{R}_{W}$ $\bar{R}_{UW}$ 0003-066 24 1.55 0.31 0.20 0.64 0.07 0.11 1.67 0.40 0.24 6.17 2.22 0.36 0104-408 28 1.26 0.29 0.23 1.00 0.04 0.04 0.93 1.24 1.33 7.86 17.13 2.18 0238-084 16 0.91 0.20 0.22 0.53 0.10 0.19 0.37 0.10 0.27 17.34 16.13 0.93 0336-019 25 1.92 0.48 0.25 0.80 0.08 0.10 1.00 0.64 0.64 3.05 2.32 0.76 0402-362 17 1.22 0.11 0.09 1.00 0.03 0.03 1.18 0.26 0.22 8.11 4.54 0.56 0454-234 27 1.73 0.59 0.34 1.00 0.02 0.02 0.46 0.31 0.67 3.18 1.78 0.56 0458-020 26 0.78 0.31 0.40 0.68 0.15 0.22 2.22 1.02 0.46 16.30 17.93 1.10 0727-115 32 2.37 0.71 0.30 0.83 0.05 0.06 1.15 0.30 0.26 6.10 1.77 0.29 0919-260 17 1.32 0.83 0.63 0.73 0.11 0.15 1.70 0.34 0.20 6.46 1.55 0.24 0920-397 16 1.06 0.18 0.17 0.92 0.11 0.12 1.81 0.94 0.52 7.27 2.98 0.41 1034-293 27 1.11 0.30 0.27 0.88 0.21 0.24 0.91 0.68 0.75 5.83 8.22 1.41 1124-186 26 0.90 0.19 0.21 1.00 0.02 0.02 0.25 0.29 1.18 6.88 16.50 2.40 1144-379 23 1.34 0.39 0.29 0.83 0.05 0.06 0.44 1.79 4.10 9.61 20.66 2.15 1145-071 16 0.69 0.11 0.16 0.77 0.10 0.13 1.33 0.24 0.18 6.11 2.26 0.37 1253-055 3 6.75 0.27 0.04 0.65 0.15 0.23 2.27 0.25 0.11 12.5 0.50 0.04 1255-316 14 1.20 0.24 0.20 0.79 0.11 0.14 2.68 0.67 0.25 14.12 11.86 0.84 1313-333 17 0.80 0.28 0.35 0.71 0.05 0.07 1.29 0.40 0.31 7.82 3.05 0.39 1334-127 25 2.12 0.55 0.26 1.00 0.04 0.04 0.84 0.62 0.74 3.60 1.87 0.52 1351-018 13 0.78 0.25 0.32 1.00 0.03 0.03 1.51 2.93 1.94 2.55 1.86 0.73 1424-418 18 1.69 0.49 0.29 0.73 0.08 0.11 3.42 1.88 0.55 22.55 11.07 0.47 1451-375 14 1.11 0.30 0.27 0.88 0.07 0.08 0.37 3.71 10.08 9.31 5.40 0.58 1514-241 16 1.76 0.30 0.17 0.83 0.05 0.06 3.39 0.78 0.23 18.57 3.90 0.21 1622-253 24 1.27 0.38 0.30 0.83 0.05 0.06 1.26 0.72 0.57 6.35 7.37 1.16 1741-038 28 3.18 1.05 0.33 1.00 0.03 0.03 0.33 0.22 0.66 2.17 1.24 0.57 1908-201 23 1.95 0.45 0.23 0.77 0.07 0.09 1.50 0.45 0.30 5.60 1.68 0.30 1921-293 23 7.42 2.45 0.33 0.60 0.09 0.15 2.30 0.53 0.23 8.78 8.17 0.93 1954-388 21 2.16 0.54 0.25 1.00 0.05 0.05 0.56 0.52 0.93 3.52 2.71 0.77 1958-179 9 1.16 0.52 0.45 0.80 0.04 0.05 0.17 0.56 3.29 2.79 2.29 0.82 2052-474 10 1.36 0.34 0.25 0.81 0.29 0.36 1.33 2.09 1.56 10.45 10.66 1.02 2243-123 22 1.50 0.33 0.22 0.86 0.06 0.07 2.00 0.48 0.24 9.32 1.77 0.19 2255-282 20 0.95 1.11 1.17 0.77 0.10 0.13 2.30 2.05 0.89 6.58 3.49 0.53 ---------- -------- ----------- ----------------- ------------------ ----------- -------------- --------------- --------------- ------------------ ------------------- ---------------- ------------------- -------------------- Core Flux Density ----------------- A good calibrator should be relatively bright at the frequency of observation to be easily detectable. It should also be stable with minimal flux density variation over time. There is no precise definition of the AGN ‘core’ in the literature. In general the bright, compact flat spectrum feature is referred to as the core. We have confirmed our identification of the cores for our entire sample by establishing that these structures have flat or inverted spectra. This was done by using the 8.4GHz data that are observed simultaneously with the 2.3GHz observations presented here. Here we define a core flux density, $S_{core}$ for each source and for each epoch as the flux density of the Gaussian component fitted to the core. The latest core flux density as defined in this section is shown, for each source in Table \[tsouthernsample\]. From the core flux density, we computed the mean core flux density ($\bar{S}$) averaged over all epochs in which the source was observed. The extent to which the core flux density varies over time can be characterised by the core flux density variability index ($\sigma_{core}/\bar{S}$) where $\sigma_{core}$ is the core flux density standard deviation. A value of 0.0 indicates no variation over time (Table \[standarddev\], column 5). Figure \[corehist\] shows that 90 per cent of the sources have a core flux density variability index below 0.4 and 23 per cent have a variability index below 0.2 (which is below our estimated calibration uncertainities). These low variability indices indicate that most core fluxes in this sample are stable. Core Fraction ------------- Following @Ojha:04b we define a core fraction as: $$C = S_{core\_cc}/S_{total}$$ where $S_{core\_{cc}}$ is the sum of the flux densities of CLEAN components within one synthesised beam of the brightest pixel and $S_{total}$ is the sum of all the CLEAN component flux densities. It provides an indication of how point-like a source is and also provides a way to track source structure changes from epoch to epoch. The average core fraction, $\bar{C}$, was computed for each source over all epochs. As with the core flux density variability index, we also computed the source core fraction variability index ($\sigma_{C}/\bar{C}$) where $\sigma_{C}$ is the core fraction standard deviation. A value of 0.0 for the source core fraction variability index indicates no variation over time. The results are shown in Table \[standarddev\], column 8. The distribution of the core fraction variability index is shown in Figure \[corefrac\] while the distribution of the core fraction is shown in Figure \[compacthist\]. The mean core fraction for all the sources is 83 per cent with a standard deviation of 12 per cent. All sources have a mean core fraction variability index below 0.40. Twenty seven sources have a variability index between 0 and 0.2 and the remaining 4 between 0.21 and 0.40. In general, the southern sample sources are very compact with little variation. For this reason, we used the actual core fraction and not the variability index to classify the sources. Flux Weighted Radial Extent --------------------------- The previous two metrics are primarily measures of core-dominance. This and the following metric provide complementary information by quantifying the radial extent of the sources in the Southern sample. The flux density weighted radial extent [@Ojha:04b] is defined as: $$R_{W} = \frac{\sum_{i} {S_{i} r_{i}}}{\sum_{i}S_{i}}$$ where $R_{W}$ is in units of milliarcseconds and $r_{i}$ is the radius at which the $i$th CLEAN component has flux density $S_{i}$. The mean ($\bar{R}_{W}$) and standard deviation were also calculated as was the variability index ($\sigma_{R_{W}}/\bar{R}_{W}$) where $\sigma_{R_{W}}$ is the weighted radial extent standard deviation, see Table \[standarddev\], column 11. The weighted radial extent variability index is widely distributed as shown in Figure \[weightedhist\], with 10 per cent of the sources having an index between 0 and 0.2, showing minimal variation in the weighted radial extent. 32 per cent of the sources have an index between 0.21 and 0.40 while 13 per cent have an index between 0.41 and 0.60. 16 per cent of the sources have an index between 0.61 and 0.80 while the rest, (29 per cent) have an index greater than 0.81. This shows a high degree of variability in the weighted radial extent. Unweighted Radial Extent ------------------------ This is the radial extent within which 95 per cent of the source flux density is contained. This provides a measure of how extended a source is. The mean ($\bar{R}_{UW}$) and standard deviation were calculated as was the variability index ($\sigma_{R_{UW}}/\bar{R}_{UW}$) where $\sigma_{R_{UW}}$ is the unweighted radial extent standard deviation and $R_{UW}$ is in units of milliarcseconds, (Table \[standarddev\], column 14) Like the weighted radial extent variability index (see above), the unweighted radial extent variability index is also widely distributed. Figure \[unweighthist\] shows that 6 per cent of the sources have an index below 0.2, 23 per cent have an index between 0.21 and 0.40 while 29 per cent have an index between 0.41 and 0.60. About 10 per cent have an index between 0.61 and 0.80 while 35 per cent have an index greater than 0.80, showing very high variability in the unweighted radial extent. As discussed below, this high variability (in both the weighted and unweighted radial extent) is largely a result of low brightness ‘jet’ features that are detected in some epochs but not others, usually due to differing sensitivities of the observing array rather than a real variation in the source. Source Classification ===================== To capture the complexity of determining whether a radio source is a suitable phase calibrator for a radio interferometer, we have developed the set of complementary metrics described above. Here we discuss how we combine the information these metrics provide to arrive at a classification scheme for potential calibrators. We explain our reasons for weighting the different metrics as we have, and present our recommendations based on this weighting scheme. Since such a scheme is necessarily somewhat subjective it should be considered as reasonable and useful rather than a definitive classification. ------------------- ------- ------------------ ------- Core flux density Score Core fraction Score (%) variabilty Index $>1$ 0.00 0.91-1.00 3.5 91-100 35.0 0.81-0.90 7.0 81-90 31.5 0.71-0.80 10.5 71-80 28.0 0.61-0.70 14.0 61-70 24.5 0.51-0.60 17.5 51-60 21.0 0.41-0.50 21.0 41-50 17.5 0.31-0.40 24.5 31-40 14.0 0.21-0.30 28.0 21-30 10.5 0.11-0.20 31.5 11-20 7.0 0.01-0.10 35 1-10 3.5 ------------------- ------- ------------------ ------- : Source core fraction and core flux density variability index score distribution.[]{data-label="compact"} Index Score ----------- ------- $>1$ 0.00 0.91-1.00 1.5 0.81-0.90 3.0 0.71-0.80 4.5 0.61-0.70 6.0 0.51-0.60 7.5 0.41-0.50 9.0 0.31-0.40 10.5 0.21-0.30 12.0 0.11-0.20 13.5 0.01-0.10 15 : Distribution of scores for the weighted and unweighted radial extent variability index.[]{data-label="radialex"} The four metrics we used to classify the calibrator source quality were core fraction, core flux density variability index, weighted and unweighted radial extents. Each of these metrics is first assigned a score, with the scores as shown in Tables \[compact\] and \[radialex\]. The scores for the individual metrics were chosen so as to sum to 100 for a perfect calibrator source. We also chose to score the core fraction and core flux density variability so as to give them higher weight, each having a maximum score of 35, while the two radial extents have maximum scores of 15. We use a lower maximum score for the radial extents as they are highly sensitive to small epoch-to-epoch variations in the signal-to-noise ratio of the images. Faint extended features may be detected only at some epochs depending on the signal-to-noise, resulting in a large variation in the radial-extent measures. Such low surface brightness features generally have negligible impact on the usefulness of the source as a calibrator. In any case, for a lower resolution array like meerKAT these extended structures will be embedded in the main central component. Based on the overall score summed up from this weighting scheme each source falls into one of five classes ‘A’ through ‘E’. A source falls into class ‘A’ if its overall score is between 80 and 100 per cent, class ‘B’ if the score is between 60 and 80 per cent, class ‘C’ if it scores between 40 and 60 per cent, class ‘D’ for any score between 20 and 40 per cent and class ‘E’ for sources scoring between 0 and 20 per cent. We propose the following classifications: - - Excellent calibrator (score between 80 and 100). - - Very good calibrator (score between 60 and 80 ). - - Good calibrator (score between 40 and 60). - - Use with caution (score between 20 and 40). - - Unsuitable as calibrator (score below 20). In our sample (Table \[classification\]), we have 9 class ‘A’ sources, 19 class ‘B’ sources and 3 class ‘C’ sources. There are no sources in class ‘D’ and class ‘E’. Thus all the sources in our sample would be ‘Good’ calibrators and all but three sources are likely to be ‘Very Good’ calibrators. ---------- --------- ------------- --------- --------- ------ --- Source Total Score Class Name W$_{1}$ W$_{2}$ W$_{3}$ W$_{4}$ % 0003-066 31.5 31.5 10.5 12.0 85.5 A 0104-408 28.0 35.0 0.0 0.0 63.0 B 0238-084 28.0 31.5 1.5 12.0 73.0 B 0336-019 28.0 35.0 4.5 6.0 73.5 B 0402-362 35.0 35.0 7.5 12.0 89.5 A 0454-234 24.5 35.0 7.5 6.0 73.0 B 0458-020 24.5 31.5 0.0 9.0 65.0 B 0727-115 28.0 35.0 12.0 12.0 87.0 A 0919-260 14.0 31.5 12.0 13.5 71.0 B 0920-397 31.5 31.5 9.0 7.5 79.5 B 1034-293 28.0 28.0 0.0 4.5 60.5 B 1124-186 28.0 35.0 0.0 0.0 63.0 B 1144-379 28.0 35.0 0.0 0.0 63.0 B 1145-071 31.5 31.5 10.5 13.5 87.0 A 1253-055 35.0 28.0 15.0 13.5 91.5 A 1255-316 31.5 31.5 3.0 12.0 78.0 B 1313-333 24.5 35.0 10.5 10.5 80.5 A 1334-127 28.0 35.0 7.5 4.5 75.0 B 1351-018 24.5 35.0 4.5 0.0 64.0 B 1424-418 28.0 31.5 9.0 7.5 76.0 B 1451-375 28.0 35.0 7.5 0.0 70.5 B 1514-241 31.5 35.0 12.0 12.0 90.5 A 1622-253 28.0 35.0 0.0 7.5 70.5 B 1741-038 24.5 35.0 7.5 6.0 73.0 B 1908-201 28.0 35.0 12.0 12.0 87.0 A 1921-293 24.5 31.5 1.5 12.0 69.5 B 1954-388 28.0 35.0 4.5 1.5 69.0 B 1958-179 21.0 35.0 3.0 0.0 59.0 C 2052-474 28.0 24.5 0.0 0.0 52.5 C 2243-123 28.0 35.0 13.5 12.0 88.5 A 2255-282 00.0 31.5 7.5 3.0 42.0 C ---------- --------- ------------- --------- --------- ------ --- : Classification of the Sources. W$_{1}$ - Core Flux Density, W$_{2}$ - Core fraction, W$_{3}$ - Unweighted Radial Extent, W$_{4}$ - Weighted Radial Extent[]{data-label="classification"} Conclusions =========== We modelfitted up to 32 epochs of observations (average of 20 epochs) for each of the 31 sources in our sample which was selected from the RRFID Kinematic Survey [@Piner:07] and determined their suitability as phase calibrators. While the kinematic survey looks more at proper motion in the sources at 8.4 GHz, this paper concentrates on the morphological properties of the sources at 2.3 GHz, a frequency more relevant to emerging southern hemisphere arrays like meerKAT and ASKAP. We have developed a method to classify radio sources according to their suitability as phase calibrators for radio interferometers. We first characterize a source by calculating several metrics which give measures of the degree to which the core dominates the source, and the degree of variability, both in flux density, and the degree to which the source is extended. These metrics are then combined to give the source a total score, which is used to assign the source to one of five classes of suitability as a calibrator. All 31 sources in our sample were classified as ‘Good’ calibrators with 28 classified as ‘Very good’ or better. Acknowledgments {#acknowledgments .unnumbered} =============== FH is supported by a grant from the South African SKA project and HartRAO. FH thanks the United States Naval Observatory for their hospitality during two visits in 2008 and 2010. This research has made use of the United States Naval Observatory (USNO) Radio Reference Frame Image Database (RRFID) and Laboratoire d’Astrophysique de Bordeaux (LAB) Bordeaux VLBI Image Database (BVID). This research has made use of NASA’s Astrophysics Data System Bibliographic Services and the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. [99]{} Booth R.S., de Block W.J.G., Jonas J.L. and Fanaroff B., 2009, ArXiv e-prints, 0910.2935 Browne I.W.A., Savage A. and Bolton J.G., 1975, MNRAS, 173, 87P Carilli C.L. and Rawlings S., 2004, New Astronomy Review, 48, 979 Collioud A. and Charlot P., 2009, 19th European VLBI for Geodesy and Astrometry Working Meeting, 19 Denicol[ó]{} G., Terlevich R., Forbes D.A., Terlevich A. and Carrasco L., 2005, MNRAS, 356, 1440 di Serego-Alighieri S., Danziger I.J., Morghanti R., and Tadhunter C.N., 1994, MNRAS, 269, 998 Fey A.L. and Charlot P., 1997, ApJS., 111, 95 Fey A.L. and Charlot P., 2000, ApJS., 128,17 Fey A.L., Clegg A.W. and Fomalont E.B., 1996, ApJS., 105, 299 Fey A.L. et al., 2004, AJ., 127, 1791 Fey A.L., Ojha R., Reynolds J.E., Ellingsen S.P., McCulloch P.M., Jauncey D.L. and Johnston K.J., 2004, AJ., 128, 2593 Fey A.L. et al., 2006, AJ., 132, 1994 Fey A.L., Gordon D. and Jacobs S., 2009., IERS Technical Note 35, ISBN 3-89888-918-6, 204 Greisen E.W., 1998, Astronomical Society of the Pacific Conference Series, 145, 204 Halpern J.P., Eracleous M. and Mattox J.R., 2003, AJ, 125, 572 Hewitt A. and Burbidge G., 1989, A new catalog of quasi-stellar objects Jauncey D.L., Batty M.J., Gulkis S. and Savage A., 1982, AJ, 87, 763 Jauncey D.L., Batty M.J., Wright A.E., Peterson B.A., and Savage A., 1984, ApJ, 286, 498 Jones D.H. et al., 2004, The 6dF Galaxy Survey Data Release 1 Jones D.H. et al., 2009, The 6dF Galaxy Survey Data Release 3 Johnston S. et al., 2008, Experimental Astronomy, 22, 151 Linfield R.P. et al., 1989, ApJ, 336, 1105 Ma C. et al., 1998, AJ, 116, 516 Marziani P., Sulentic J.W., Dultzin-Hacyan D., Calvani M., and Moles M., 1996, ApJS, 104, 37 Ojha R. et al., 2004, AJ, 127, 3609 Ojha R., Fey A.L., Jauncey D.L., Lovell J.E., Johnston K.J., 2004, ApJ, 614, 607 Ojha R. et al., 2005, AJ, 130, 2529 Osmer P.S., Porter A.C. and Green R.F., 1994, ApJ, 436, 678 Peterson B.A., Jauncey D.L., Condon J.J., and Wright A.E., 1976, ApJ, 207, L5 Piner B.G., Mahmud M., Fey A.L. and Gospodinova K., 2007, AJ, 133, 2357 Rees M.J., 1997, Reviews in Modern Astronomy, 10, 179 Schilizzi, R. T. et al., 2007, SKA Memo, 100 Shepherd M.C., Pearson T.J. and Taylor G.B., 1995, BAAS, 27, 903 Stickel M., Fried J.W. and Kuehr H., 1989, A&AS, 80, 103 Stickel M., Kuehr H. and Fried J.W., 1993, A&AS, 97, 483 Strittmatter P.A., Carswell R.F., Gilbert G., and Burbidge E.M., 1974, ApJ, 190, 509 Veron-Cetty M.P., Veron P., 2003, Astron. Astrophys., 412, 399 White G.L., Jauncey D.L., Wright A.E., Batty M.J., Savage A., Peterson B.A., Gulkis S., 1988, ApJ, 327, 561 Wilkes B.J., 1986, MNRAS, 218, 331w Wills D., and Lynds R., 1978, ApJS, 36, 317 Wright A.E., Peterson B.A., Jauncey D.L., and Condon J.J., ApJ, 229, 73 Zensus J.A., Ros E., Kellerman K.I., Cohen M.H., Vermeulen R.C., and Kadler M., 2002, AJ, 124, 662 \[lastpage\] [^1]: E-mail: [email protected] (FH); [email protected] (RO) [^2]: The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc [^3]: The web site for the RRFID is located at http://rorf.usno.navy.mil/rrfid.shtml and the BVID at http://www.obs.u-bordeaux1.fr/BVID/
{ "pile_set_name": "ArXiv" }
--- abstract: 'A soft-wall warped extra dimension allows one to relax the tight constraints imposed by electroweak data in conventional Randall-Sundrum models. We investigate a setup, where the lepton flavour structure of the Standard Model is realised by split fermion locations. Bulk fermions with general locations are not analytically tractable in a soft-wall background, so we follow a numerical approach to perform the Kaluza-Klein reduction. Lepton flavour violation is induced by the exchange of Kaluza-Klein gauge bosons. We find that rates for processes such as muon-electron conversion are significantly reduced compared to hard-wall models, allowing for a Kaluza-Klein scale as low as 2 TeV. Accommodating small neutrino masses forces one to introduce a large hierarchy of scales into the model, making pressing the question of a suitable stabilisation mechanism.' bibliography: - 'bib.bib' --- $\mbox{ }$ \ Michael Atkins$^{a,}$[^1] and Stephan J. Huber$^{a,}$[^2]\ \ Introduction ============ Over the last ten years there has been a large increase in the study of extra dimensional models following the realisation that they could help explain some of the unresolved problems in the Standard Model (SM). In 1999, Randall and Sundrum showed that a warped extra dimension could offer a geometric solution to the gauge hierarchy problem [@Randall:1999ee]. In the original Randall-Sundrum (RS) model, the fifth dimension consists of a slice of AdS space bounded by ultraviolet (UV) and infrared (IR) branes. The warped space produces an exponential difference in energy scales between the two branes which solves the hierarchy problem. Matter fields were originally confined to the IR brane, however, it was soon realised that by allowing fermions to propagate in the extra dimension, the SM fermion mass hierarchy can be explained. By varying the location of the fermion wavefunctions in the fifth dimension, the full scale of fermion masses from neutrinos to the top quark can be generated using only order unity parameters [@Grossman:1999ra; @Gherghetta:2000qt; @Huber:2000ie]. This setup also contains a built in mechanism suppressing unobserved flavour changing processes that result from couplings between SM fermions and excited gauge bosons which appear in the model [@Gherghetta:2000qt; @Huber:2003tu; @Agashe:2004cp]. Further interest in warped extra dimensions was generated by the AdS/CFT conjecture, when it was realised that the RS scenario is holographically dual to strongly coupled 4D field theories [@ArkaniHamed:2000ds; @Rattazzi:2000hs; @PerezVictoria:2001pa]. It was in this context, studying AdS/QCD models, that the idea of a soft wall was first introduced [@Karch:2006pv]. The soft wall is realised by removing the IR brane so the extra dimension extends to infinity, and replacing it with a smoothly varying spacetime cut off. The original AdS/QCD motivation for this was to more faithfully reproduce the linear Regge-like mass squared spectrum of excited mesons as opposed to the usual quadratic spectrum found in hard wall RS models. Inspired by the possibility of qualitatively different phenomenology, the soft wall scenario was subsequently applied to modelling electroweak physics [@Falkowski:2008fz; @Batell:2008me]. These models successfully showed that a soft wall extra dimension is generally less constrained by electroweak precision observables than its hard wall counterpart, typically allowing Kaluza Klein (KK) modes with masses of a few TeV. An important issue is related to the stability of a soft-wall setup, which is an open question in the models discussed in Refs. [@Falkowski:2008fz; @Batell:2008me]. Such a mechanism was suggested in Ref. [@Cabrer:2009we], promising the soft-wall extra dimension to equally well resolve the gauge hierarchy problem. With the removal of the hard-wall brane the Standard Model matter fields must necessarily propagate in the bulk. Graviton fluctuations and gauge fields were successfully analysed in this background but it was found that fermions presented particular technical difficulties and only a simplified single generational model was developed. Later studies of fermions in a soft-wall extra dimension have developed solutions to the fermion problem [@Delgado:2009xb; @MertAybat:2009mk; @Gherghetta:2009qs] and have considered the experimental constraints imposed by the electroweak observables. However, the fermion flavour pattern of the SM has not been considered in much detail, in particular with respect to the generation of neutrino masses and the experimental bounds on lepton flavour violation. In this paper we present a numerical solution to analyse a single generation of fermions in the soft-wall extra dimension. We extend this solution to three generations by treating flavour mixings as perturbations to the original solutions, and apply it to the lepton sector of the SM. We construct a setup, where the lepton flavour pattern is accommodated by flavour dependent localisations. It is shown that in order generate small Dirac neutrino masses by this mechanism we need to introduce a hierarchy of scales of order $10^{15}$ into the model, making crucial the issue of a suitable stabilisation mechanism. We finally carry out an analysis of the constraints coming from various lepton flavour violating processes, averaging over random order unity Yukawa couplings, and find that models with only a modest hierarchy of scales are relatively mildly constrained, whereas the model with a large hierarchy allowing sub-eV neutrino masses lies well within current experimental constraints, even for a KK scale[^3] of 2 TeV. In the latter, flavour violation is considerably suppressed relative to its hard wall counterparts, such as the ones analysed in [@Huber:2003tu; @Moreau:2005kz], and the range of masses lies in the reach of the LHC experiment. At this stage we do not try to accommodate the flavour structure of the quark sector, which should be possible in a similar way. Also we reproduce the neutrino masses and mixings only at the qualitative level, which is sufficient to estimate the rates of lepton flavour violation. Bulk Fields in a Soft-Wall Extra Dimension ========================================== Our conventions follow most closely those laid down in Refs. [@Batell:2008me; @Gherghetta:2009qs]. The 5D spacetime has metric $$\label{swmetric}ds^2 = e^{-2A(y)}\eta_{MN}dx^Mdx^N,$$ where $y$ represents the extra spatial dimension and $\eta_{MN}=\mbox{diag}(+,-,-,-,-)$. We take a pure AdS metric, $A(y)=\log ky$, where $k$ is the AdS curvature scale. There is no IR brane, the extra dimension extends to infinity and the soft wall is introduced via a dilaton field $\Phi$ with the action describing gauge and matter fields given by $$\label{action} S = \int d^4x \int_{y_0}^\infty dy \: \sqrt{g} e^{-\Phi} \mathcal{L}.$$ Here $y_0=1/k$ is the location of the UV brane. The dilaton field is taken to have the following power law behaviour $$\label{dilaton}\Phi(y) = (\mu y)^2 .$$ The dimensionful parameter $\mu$ will set the mass scale of the lightest KK excitations. The behaviour of other powers has been discussed in detail in Ref. [@Delgado:2009xb]. It is shown in Ref. [@Batell:2008me] that an appropriate form for the Higgs VEV in such a background is given by $$\label{vev}h(y)=\eta k^{3/2} \mu^2 y^2$$ where $\eta$ is a dimensionless $\mathcal{O}(1)$ coefficient. Massive Gauge Fields -------------------- In Refs. [@Batell:2008me; @Gherghetta:2009qs] only massless gauge fields are considered as the gauge couplings being considered are assumed to be between quarks and massless gluons. The flavour changing neutral currents (FCNC) we will be considering here are mediated by the massive Z boson and so we first develop the solutions for such a gauge field. A massive gauge field propagating in the bulk has the Lagrangian $$\mathcal{L} = -\frac14 F_{MN}F^{MN} + \frac12 M_A^2 A_M A^M .$$ The mass term $M_A^2$ for the weak gauge bosons arises from spontaneous symmetry breaking and with the Higgs VEV as given in (\[vev\]) we have $M_A^2=\frac12 \: g_5^2 \: h(y)^2.$ Varying the action (\[action\]) we obtain the equation of motion $$\frac{1}{\sqrt{g}} \partial_M \left( \sqrt{g} \: g^{MN}g^{RS}F_{NS}\right) -g^{MN}g^{RS}F_{NS} \partial_R \Phi - M_A^2 G^{MN}A_M = 0.$$ Imposing the gauge $A_y=0$, inserting the Kaluza Klein (KK) reduction, $$\label{KKgSW}A_\mu(x,y)=\sum_{n=0}^\infty A_\mu^{(n)}(x) f_A^{(n)}(y),$$ and requiring the $A_\mu^{(n)}(x)$ to be mass eigenstates, we find the $f_A^n$ have to satisfy $$\label{gaugeeom}\left( \partial_y^2 -\left(\frac1y + \Phi'\right)\partial_y -\frac{1}{(ky)^2}M_A^2 + m_n^2 \right)f_A^{(n)}(y)=0$$ and are canonically normalised by $$\int_{y_0}^\infty dy \: e^{-(A+\Phi)}f_A^{(m)}(y)f_A^{(n)}(y) = \delta^{mn}.$$ A complicated analytic solution of the above equation of motion was developed in [@Delgado:2009xb] but for our purposes we find it more convenient to solve the equation of motion numerically. We apply Neumann boundary conditions to the wavefunctions at $y_0$ and vary $\eta g_5$ in order to find a normalisable solution with the appropriate 4D zero mode mass of $m_0 \simeq 91$ GeV for the Z boson. It is interesting to note that unlike in hard-wall models, the profile of a massive gauge boson is independent of the curvature scale $k$. This can easily be seen by looking at the equation of motion (\[gaugeeom\]) where $k$ only appears in the term involving the 5D mass $M_A$ which is the term that must be varied in order to generate the correct zero mode mass. The profile of the first few KK modes of the Z boson are plotted in Fig. 1. In Fig. \[fig:subfig1\] the zero mode and the first two KK modes are plotted with respect to a flat metric. Fig. \[fig:subfig2\] shows the profile of the zero mode in the form that it couples to the fermions (see later). The UV behaviour of the zero mode is less flat than in hard wall models and could possibly lead to large violations of universality of gauge couplings once fermions reside at different locations in the extra dimension. Note also that as is the case for massless gauge bosons [@Batell:2008me; @Gherghetta:2009qs] the higher KK modes become more and more IR localised, a fact that will be important when considering couplings between higher gauge modes and fermions. The KK spectrum for the Z boson for different values of $\mu$ with $k=10^7$ TeV is -------------------- ----------------- --------------- --------------- $\mu=\frac12$ TeV: $m_0=0.091$ TeV $m_1=1.3$ TeV $m_2=1.8$ TeV $\mu=1$ TeV: $m_0=0.091$ TeV $m_1=2.2$ TeV $m_2=3.0$ TeV $\mu=2$ TeV: $m_0=0.091$ TeV $m_1=4.1$ TeV $m_2=5.8$ TeV $\mu=4$ TeV: $m_0=0.091$ TeV $m_1=8.2$ TeV $m_2=12$ TeV -------------------- ----------------- --------------- --------------- Note that the mass of the first KK mode $m_1 \sim 2\mu$, and hence the KK scale, $M_{KK}$, scales with $\mu$. Higher modes follow a Regge like spectrum $m_n^2 \sim n$. \[fig:gaugeprofiles\] Fermions -------- We consider 5D Dirac spinors $\Psi_L$ and $\Psi_R$ which are components of doublets and singlets under $SU(2)_L$ respectively. Note that $L$ and $R$ do not denote chiralities, but are related to the charges under $SU(2)_L$. The chiral projections of these spinors are $\Psi_{L\pm}=\frac{1}{2}(1\mp\gamma^5)\Psi_L $, same for $\Psi_R$. The action for two free fermions in the bulk is $$\begin{aligned} S= \int d^4x \int_{y_0}^\infty dy \: \sqrt{g} e^{-\Phi} \left[ \frac{1}{2}\left( \: \overline\Psi_L i e_A^M \gamma^A D_M \Psi_L - D_M \overline\Psi_L i e_A^M \gamma^A \Psi_L\right) - M_L \overline\Psi_L \Psi_L\right. \\ \left.+ \frac{1}{2}\left( \: \overline\Psi_R i e_A^M \gamma^A D_M \Psi_R - D_M \overline\Psi_R i e_A^M \gamma^A \Psi_R\right) - M_R \overline\Psi_R \Psi_R\right]. \;\;\;\;\end{aligned}$$ The fünfbein and spin connection for the metric (\[swmetric\]) are $e_A^M=e^{A(y)}\delta_A^M$ and $\omega_M=\left( -\frac{A'}{2}\gamma_\mu \gamma^5,0\right) $ and the covariant derivative is then $D_M=\partial_M + \omega_M$. $M_{L,R}$ are the 5D Dirac masses related to $\Psi_{L,R}$. The difficulty of placing uncoupled fermions in the soft wall background is well documented in Ref. [@Batell:2008me], all solutions suffer from divergent gauge couplings for high enough KK modes. The underlying problems stem from the non compact nature of the extra dimension and it is shown that in order to find workable normalisable solutions, the Yukawa couplings between fermions and the Higgs must be taken into account. An alternative approach to introducing Yukawa couplings was presented in Ref. [@MertAybat:2009mk] where a $y$ dependent Dirac mass term is introduced, somewhat like the $y$ dependent bulk mass arising from the Higgs VEV in the case of the massive gauge boson above. Here, however we will stick to constant Dirac mass terms $M_L$ and $M_R$ and introduce Yukawa couplings into the action: $$S_{\rm Yuk}=-\int d^4x\int_{y_0}^\infty dy \sqrt{g} e^{-\Phi} \frac{\lambda_5}{\sqrt{k}}\left[\:\overline\Psi_L(x,y)h(y)\Psi_R(x,y) + \overline\Psi_R(x,y)h(y)\Psi_L(x,y) \right].$$ Defining $\Psi_{L,R} = e^{2A + \Phi/2}\psi_{L,R}$ and $m(y)=\frac{\lambda_5}{\sqrt{k}}h(y),$ the equations of motion are $$i \gamma^{\mu}\partial_{\mu} \psi_{L,R\pm} \pm \partial_y \psi_{L,R\mp} - e^{-A}M_{L,R} \psi_{L,R\mp} -e^{-A}m(y) \psi_{R,L\mp} =0.$$ Using the KK reduction $$\psi_{L,R\pm}(x,y)=\sum^\infty_{n=0}\; \psi_\pm^{(n)}(x)\;f_{L,R\pm}^{(n)}(y),$$ and requiring the $\psi_\pm^{(n)}(x)$ to be mass eigenstates, the $f^{(n)} $s will be given by $$\label{sweom}\pm \partial_y\left(\begin{array}{c}f^{(n)}_{L\pm}\\ f^{(n)}_{R\pm}\end{array}\right)+e^{-A}\left(\begin{array}{cc} M_L& m(y)\\ m(y)& M_R\end{array} \right) \left(\begin{array}{c}f^{(n)}_{L\pm}\\ f^{(n)}_{R\pm}\end{array}\right)=m_n \left(\begin{array}{c}f^{(n)}_{L\mp}\\ f^{(n)}_{R\mp}\end{array}\right).$$ The $\psi_\pm^{(n)}$ will be canonically normalised by $$\int_{z_0}^{\infty}\;dy\; \left( f^{(m)}_{L\pm}\: f^{(n)}_{L\pm} \;+\; f^{(m)}_{R\pm}\: f^{(n)}_{R\pm}\right) =\delta_{mn}.$$ We impose Dirichlet boundary conditions at $y_0$ for $f^{(n)}_{L-}$ and $f^{(n)}_{R+}$ in order to obtain a chiral 4D theory. Analytic solutions of (\[sweom\]) are only possible for a small set of Dirac mass terms, namely $M_L=M_R$ and $M_L+M_R \pm k = 0$ [@Gherghetta:2009qs]. As in hard-wall models, the Dirac mass parameters $M_{L,R}$ dictate how the fermion is localised in the extra dimension and it is convenient to parametrise these in terms of the AdS curvature, $M_{L,R}=c_{L,R}k$. Unfortunately the sets of parameters for which analytic solutions are available do not explore the full geography of possible mass parameters and as we shall see may lead to situations with unacceptably large rates of flavour violation. Ideally we would like to solve (\[sweom\]) for any set of Dirac masses and this requires a numerical approach. The numerical solution we have developed involves a shooting type method. In order for the solutions to be normalisable they must not diverge in the IR and this only occurs for the correct choice of $m_n$, thus generating the KK spectrum. We choose a suitably large distance $L$ into the IR and solve the equations of motion subject to the UV boundary conditions and a starting choice for $m_n$. We then iterate the solution using Newton’s method in order to find a value for $m_n$ such that two of the solutions e.g. $f^{(n)}_{L+}$ and $f^{(n)}_{R-}$ converge to zero at $y=L$. The equations of motion then automatically ensure that the other two solutions will also converge to zero for large $y$. Our solution has the advantage that it seems to be quite capable of finding solutions even for large values of the AdS curvature scale $k$, however it is not so suited to finding solutions for multiple generations of fermions as is done in Ref [@Gherghetta:2009qs]. Leptons ======= General Considerations {#sect.gc} ---------------------- Due to the presence of the extra KK states in extra dimensional models, couplings between SM particles and their KK excitations can potentially lead to conflict with experimental observations. In the SM, a tight set of constraints comes from the experimental bounds on flavour changing neutral currents (FCNC). In the hard-wall Randall Sundrum model these processes have been investigated and are shown to occur at rates that are dependent on the fermion locations. However, there are certain choices of fermion locations which provide almost universal gauge couplings and these almost universal gauge couplings are the source of the so called RS GIM mechanism which suppresses FCNC [@Gherghetta:2000qt; @Huber:2003tu; @Agashe:2004cp]. It was found in Ref. [@Batell:2008me] that for fermions in the soft-wall background, the analytic solution with $M_L=M_R$ can produce a large hierarchy of masses but only one of the fermion pair ($\psi_+$, $\psi_-$) could reside in an area of universal gauge couplings and it was thus assumed that dangerous rates of FCNC would be generated in such a situation. In fact this is one of the main motivations for finding a numerical solution to the fermion equations of motion, in the hope that one would be able to find fermion locations which can give a large mass hierarchy and yet simultaneously reside in an area of universal gauge couplings. The gauge interaction between bulk gauge bosons and fermions is given by $$S_{\rm{Gauge}}= g_5 \int d^4x \int_{y_0}^\infty dy \: \sqrt{g} e^{-\Phi} \left[\overline\Psi_L e_A^M \gamma^A A_M \Psi_L + \overline\Psi_R e_A^M \gamma^A A_M \Psi_R \right].$$ The couplings of $\psi^{(0)}_\pm$ to different KK gauge modes is then $$\label{gauge coup} g^n_\pm = g_5 \int_{y_0}^\infty dy \; f_A^{(n)} \left[ \left(f^{(0)}_{L\pm} \right)^2 + \left(f^{(0)}_{R\pm} \right)^2 \right].$$ The dependence of the gauge couplings on the fermion locations $c_{L,R}$ is shown in Fig. 2. It can be seen that for $c_L>1/2$ the couplings become universal. In the case where $c_L=c_R$, the couplings of one of the fermions would lie in the universal region the other would lie in the opposite part of the plot. However, with opposite Dirac masses, $c_L=-c_R$, we are able to place both fermions in a region of universal coupling at the same time and we would thus hope to suppress FCNC. In hard wall models the origin of the regions of universal couplings is quite clear and derives from the profile of the gauge field wavefunctions which are flat in the UV. Hence if fermion profiles are relatively UV localised the gauge couplings will be universal. However, in the soft wall model, looking at the profile of the zero mode of the massive gauge boson in Fig. \[fig:subfig2\] it is certainly not flat and one may wonder why we still find regions of universal gauge couplings. The explanation can be seen by considering the fermion profiles. Fig. \[fig:uvlocal\] shows the fermion wavefunctions contributing to the gauge coupling of $\psi_+$ for $c_L = - c_R = 0.7$ which are locations that live in an area of universal gauge couplings. Whilst $f^{(0)}_{L+}$ is heavily UV localised we see that $f^{(0)}_{R+}$ is actually peaked into the IR which we would expect to contribute to non-universal couplings. However, when one considers the relative size of the contributions of each of these wavefunctions to the gauge coupling as given by Eq. \[gauge coup\] we find $\int (f^{(0)}_{R+})^2/\int (f^{(0)}_{L+})^2 \sim 3 \times 10^{-4}$ i.e. almost the entire contribution to the gauge coupling comes from $f^{(0)}_{L+}$ which is heavily UV localised. Although the relative dominance of $f^{(0)}_{L+}$ is not that clear to see from Fig. \[fig:uvlocal\], it becomes obvious when one realises that it has a value at $\mu y_0$ of about 2000. Also, its extreme UV localisation can be seen by the fact that 99% of the area of $(f^{(0)}_{L+})^2$ lies in the region $\mu y < 0.01$. Hence the dominant contribution to the gauge coupling comes from a region in the extreme UV where the gauge profile is effectively given by its UV boundary value, thus producing universal couplings for fermions. \[fig:gaugecoup\] ![Fermion zero modes with $c_L=-c_R=0.7$ for $k=10^7$ TeV and $\mu=1$ TeV. $f^{(0)}_{L+}$ solid, $f^{(0)}_{R+}$ dashed. Note that $f^{(0)}_{L+}$ takes a value of around 2000 at $y_0$.[]{data-label="fig:uvlocal"}](uvlocal.eps) In order to generate the fermion mass hierarchy seen in the SM we have to carefully choose the $c$ parameters. The zero mode masses for different $c$ parameters can be seen in Fig. \[fig:contour\]. Unfortunately, the shape of the plot presents a problem for simultaneously generating a large hierarchy of masses and universal gauge couplings. Whilst it is easy to generate a large hierarchy of masses for the choice of parameters $c_L=c_R$, as has been stated above, this is likely to lead to high rates of FCNC. In order to avoid these unacceptable rates we would like both the fermions to reside in an area of universal gauge couplings, this corresponds to the top left corner of the contour plot where we have $c_L>1/2$ and $c_R<-1/2$. In this area the zero mode mass bottoms out at around $\mu^2 / k$ and it is not possible to create a large mass hierarchy. The solution to this is to simply increase the hierarchy of scales in the model. Keeping $\mu=1$ TeV we can see from Fig. \[fig:crosssection\] that with $k/ \mu = 10^{7}$ and $c_L=-c_R$ the zero mode masses could cover the full range of charged lepton masses whilst remaining in an area of universal gauge couplings. ![Contour plot of $\log_{10}{(m_0/\mu)}$ for the zero mode masses of fermions with $\mu=1$ TeV and $k=10^3$ TeV.[]{data-label="fig:contour"}](contour.eps) ![Fermion zero mode masses with $c_L=-c_R$, $\mu=1$ TeV and $k=10^7$ TeV.[]{data-label="fig:crosssection"}](crosssection.eps) The reason for the zero mode mass having a minimum value for $c_L=-c_R$ can be seen by considering the two different ways small masses are generated in such models. The zero mode masses are generated via Yukawa couplings which involve the overlap between $\Psi_L$, $\Psi_R$ and the Higgs VEV. In the case where $c_L=c_R$, the wavefunctions of the zero mode fermions become oppositely localised and can be arranged to have an arbitrarily small overlap with each other thus creating arbitrarily small masses, this is the mechanism used in the “split fermion” model [@ArkaniHamed:1999dc]. However when we take $c_L=-c_R$ the fermions completely overlap each other and the zero mode mass is then entirely determined by their overlap with the Higgs. In RS models where the Higgs resides on the IR brane, this overlap can be made arbitrarily small by heavily localising both the fermions at the UV brane. However in the soft-wall model where the Higgs must necessarily propagate in the bulk and has a non zero value at the UV brane, the fermion wavefunctions will always have a minimum overlap with the Higgs at $y_0=1/k$ however much they are UV localised. Neutrino Masses --------------- We can generate small Dirac neutrino masses by introducing right-handed neutrinos and allowing Yukawa couplings between them, the Higgs and the left-handed neutrinos. A similar approach has been taken in the hard-wall case in Refs. [@Grossman:1999ra; @Huber:2001ug]. The left-handed neutrinos will share the same $c_L$ parameters as the corresponding left-handed charged leptons since they are part of the same doublet under $SU(2)_L$. We are then free to place the right-handed neutrinos in a suitable location in order to generate sub-eV masses. However, since we still require $c_L > 1/2$ we again find that we are unable to generate such small masses without vastly increasing the overall hierarchy in the model. It seems necessary therefore to work with a hierarchy similar to that proposed in the original Randall Sundrum model. The issue of stabilising such a large hierarchy is an important question, and it would be very interesting to redo our analysis in context of the stabilised model proposed in Ref. [@Cabrer:2009we]. With this in mind we choose $k/\mu = 10^{15}$ and are able to produce neutrino masses of order $0.1$ eV by choosing $c_L=0.6$ and $c_R=-1.3$. Three Generations ----------------- When incorporating all three generations of leptons into our model the Dirac mass terms $M_L$ and $M_R$ and the Yukawa coupling constants are promoted to $3\times3$ matrices mixing the different generations. We assume that the basis of states in which the $M_L$ and $M_R$ are diagonal does not correspond to one in which the Yukawa couplings are diagonal. Rather than finding exact solutions for all three generations in such a scenario, our approach to this problem follows closely the method used in Ref. [@Huber:2003tu] for the Randall Sundrum model. We solve the equations of motion individually for each generation with a Yukawa coupling $\lambda_5=1$, excluding fermion mixing, and use these solutions as basis from which we treat the full matrix of Yukawa couplings including mixings between the generations as perturbations. We specifically choose a large number of random Yukawa couplings, taking $\frac12 < |\lambda_{5ij}|<2$ with random sign[^4], and require that the average zero mode masses reproduce the observed lepton masses. We also choose to locate the left-handed fields of each generation close to each other in order to generate large neutrino mixings, in the spirit of Ref. [@Huber:2001ug]. However, we do not aim at reproducing the neutrino masses and mixings precisely. All we arrange for is an overall neutrino mass scale of order $0.1$ eV. Using our tools, a full model of neutrino masses could be constructed. However, for the following estimate of lepton flavour violation these details are not needed. Also we use the fact that neutrino mixings are order unity. In the case where we are not interested in generating neutrino masses via locations, we take only a moderate hierarchy of scales, $k/ \mu = 10^{7}$. In this regime we choose the following three scenarios: ------ ------------------ ------------------ ------------------ (A): $c_{L1}=0.700$, $c_{L2}=0.700$, $ c_{L3}=0.700$, $c_{R1}=-1.376$, $c_{R2}=-0.903$, $c_{R3}=-0.703$, (B): $c_{L1}=0.720$, $c_{L2}=0.700$, $ c_{L3}=0.680$, $c_{R1}=-1.373$, $c_{R2}=-0.903$, $c_{R3}=-0.704$, (C): $c_{L1}=0.600$, $c_{L2}=0.600$, $ c_{L3}=0.600$, $c_{R1}=-1.430$, $c_{R2}=-0.980$, $c_{R3}=-0.790$. ------ ------------------ ------------------ ------------------ In the regime where we can also generate neutrino masses, $k/ \mu = 10^{15}$ we choose: ------ ----------------- ----------------- ----------------- (D): $c_{L1}=0.60$, $c_{L2}=0.60$, $ c_{L3}=0.60$, $c_{R1}=-0.82$, $c_{R2}=-0.64$, $c_{R3}=-0.55$. ------ ----------------- ----------------- ----------------- Our choices for the different scenarios (A), (B) and (C) are to demonstrate the effects of degenerate $c_L$ localisation (A), small separation in the $c_L$ to introduce some non-universality in the left handed sector (B), and placing the left handed fermions closer to the IR brane (C). We expect the behaviour to be quite general and thus only choose one scenario with a larger hierarchy. The mass of the first fermion KK states is about $1.5$ TeV. Flavour Violation ================= With a full three generations of leptons implemented as above, the transformation to fermion mass eigenstates will induce flavour violating couplings to gauge fields, in particular the Z boson and its KK excitations[^5] We define the neutral current gauge couplings in the basis of mass eigenstates as $$\mathcal{B}_{\pm}^{(n)} = \mathcal{U}_{\pm} \mathcal{G}_{\pm}^{(n)} \mathcal{U}_{\pm}^\dagger,$$ where the unitary matrices $\mathcal{U}_{\pm}$ diagonalise the full fermion mass matrices and $\mathcal{G}_{\pm}^{(n)}$ are diagonal matrices that contain the couplings of the $n$th KK state of the Z boson to each fermion state as derived from Eq. \[gauge coup\] and normalised to the coupling of the muon (see also Ref. [@Huber:2003tu]). Flavour violation induced by these couplings is dependent on the non-universality in the couplings of different flavour states and the mixing between the states. Different fermion locations increase the non-universality but at the same time lead to small mixing angles. Conversely similar fermion locations produce large mixing but this is compensated by universal couplings. As was done in Ref. [@Huber:2003tu] we calculate the rates of the various flavour violating processes using the techniques developed for family non-universal Z’ bosons [@Langacker:2000ju]. The main difference being that there is no mixing between the different KK states of the Z boson, while the zero mode also has flavour violating couplings. The first process we consider is the tree level exchange of a Z boson and its KK states mediating the process $l_j \rightarrow l_i l_i\bar{l}_i$. The rate for this process is given by [@Langacker:2000ju] $$\Gamma(l_j \rightarrow l_i l_i \bar{l}_i)=\frac{G_F^2m_{l_j}^5}{48 \pi^3}\left(2|C^+_{ij}|^2 + 2|C^-_{ij}|^2 + |D^+_{ij}|^2 + |D^-_{ij}|^2 \right)$$ where $$C^\pm_{ij}= \sum_n \frac{M_0^2}{M_n^2} (\mathcal{B}_{\pm}^{(n)})_{ij}(\mathcal{B}_{\pm}^{(n)})_{ii}$$ $$D^\pm_{ij}= \sum_n \frac{M_0^2}{M_n^2} (\mathcal{B}_{\pm}^{(n)})_{ij}(\mathcal{B}_{\mp}^{(n)})_{ii}$$ where we take the sum over the Z boson zero mode and the first two KK modes. $M_n$ is the mass of the $n$th KK mode of the Z boson. Due to the Regge type behaviour of the KK spectrum it is not clear that the above series should converge. However as noted in Ref. [@Gherghetta:2009qs], due to the increasing IR localisation of higher gauge boson KK modes, the couplings rapidly decrease and the series converges after only a few terms. In the fermion sector we take into account only the zero modes. Mixing with KK fermions is small, leading to negligible effects at the current precision. The branching ratios for the above processes in the different scenarios we consider are then found to be -------------------------------------------------- ------------------------- ------------------------- ------------------------- -------------------------- (A) (B) (C) (D) ${\rm Br}(\mu \rightarrow e e \bar{e}):$ $ 2.7 \times 10^{-14} $ $ 5.1 \times 10^{-12} $ $ 4.6 \times 10^{-12}$ $ 2.5 \times 10^{-15}$ ${\rm Br}(\tau \rightarrow \mu \mu \bar{\mu}): $ $2.4 \times 10^{-14} $ $ 1.8 \times 10^{-12} $ $ 7.0 \times 10^{-13} $ $ 2.7 \times 10^{-12}$ ${\rm Br}(\tau \rightarrow ee\bar{e}): $ $ 2.6 \times 10^{-15} $ $ 1.5 \times 10^{-12} $ $ 6.6 \times 10^{-13} $ $ 2.8 \times 10^{-16} $ -------------------------------------------------- ------------------------- ------------------------- ------------------------- -------------------------- These numbers are obtained for a KK scale of 2 TeV. They are the result of averaging over random Yukawa couplings in the range stated above. The experimental bound ${\rm Br}(\mu \rightarrow e e \bar{e})< 1.0 \times 10^{-12}$ [@Bellgardt:1987du] is satisfied in the cases (A) and (D). However, it appears that the couplings are not universal enough to allow for much separation between the left handed states (B), and placing the fermions too close to the IR brane (C) also exceeds the experimental bound. However, like in the hard-wall case, the rate for this process depends on the KK scale as $1/M_{KK}^4$. Thus with a KK scale of twice as big (i.e. take $\mu=2$ TeV while keeping $k=10^{7}$ TeV) scenarios (B) and (C) would also acquire an acceptable rate. The experimental bounds for the other two processes ${\rm Br}(\tau \rightarrow \mu \mu \bar{\mu})< 2.1 \times 10^{-8}$ and ${\rm Br}(\tau \rightarrow ee\bar{e})< 2.7 \times 10^{-8}$ [@Hayasaka:2010np] are well satisfied in all the scenarios. Note that in Ref. [@Huber:2003tu] in a case similar to (A), (C) and (D) a branching ratio ${\rm Br}(\mu \rightarrow e e \bar{e})=5\times10^{-14}$ has been found for a KK scale of 10 TeV, translating into ${\rm Br}(\mu \rightarrow e e \bar{e})=3\times10^{-11}$ for a KK scale of 2 TeV. This demonstrates that lepton flavour violation is suppressed by up to four additional orders of magnitude in the soft-wall case. We are also able to calculate the expected rate of $\mu \rightarrow e$ conversion in a muonic atom. The most stringent bound comes from the Sindrum-II Collaboration [@Wintz:1998] in $^{48}_{22}\mbox{Ti}$ where ${\rm Br}(\mu^{-}N \rightarrow e^{-}N)<6.1\times10^{-13}$. We can calculate the branching ratio for this process by [@Langacker:2000ju] $${\rm Br}(\mu^{-}N \rightarrow e^{-}N) = \frac{G_F^2 \alpha^3 m_{\mu}^5}{2 \pi^2 \Gamma_{\rm CAPT}}\frac{Z_{\rm eff}^4}{Z}|F_P|^2 \left(\left|B^-\right|^2+\left|B^+\right|^2\right)$$ where $$B^\pm=\sum_n\frac{M_0^2}{M_n^2}\:(\mathcal{B}_{\pm}^{(n)})_{12}\:\left[(2Z+N)(B_{u_L}^{(n)}+B_{u_R}^{(n)})+(Z+2N)(B_{d_L}^{(n)}+B_{d_R}^{(n)})\right]^2$$ and we take $$B_{\psi_{L,R}}^{(0)}=g_{\psi_{L,R}},\;\;\;B_{\psi_{L,R}}^{(1)}=0.19\: g_{\psi_{L,R}}\;\;\;\mbox{and}\;\;\;B_{\psi_{L,R}}^{(2)}=0.14 \:g_{\psi_{L,R}}.$$ Here, $g_{\psi_{L,R}}$ are the usual Standard Model quark couplings and we have taken the approximate values of the quark couplings to the higher KK gauge modes from the values derived in Ref. [@Gherghetta:2009qs]. We also take $Z_{\rm eff}=17.6$, $F_P=0.54$ and $\Gamma_{\rm CAPT}=2.59 \times 10^6 \: \rm{s}^{-1}$ [@Bernabeu:1993ta], where $Z_{\rm eff}$ is an effective atomic charge, $F_P$ is a nuclear matrix element and $\Gamma_{\rm CAPT}$ is the muon capture rate. We find branching ratios for the different scenarios of ------------------------------------------ ------------------------- ------------------------- ------------------------ ------------------------- (A) (B) (C) (D) ${\rm Br}(\mu^{-}N \rightarrow e^{-}N):$ $ 1.6 \times 10^{-13} $ $ 3.3 \times 10^{-11} $ $ 2.7 \times 10^{-11}$ $ 1.4 \times 10^{-14}$ ------------------------------------------ ------------------------- ------------------------- ------------------------ ------------------------- Again we find scenarios (A) and (D) lie within the experimental bounds but separating the states (B) or placing them too close to the IR brane (C) produces an unacceptable rate. Again we find flavour violation suppressed with respect to the hard-wall case [@Huber:2003tu], making a KK scale of 2 TeV consistent with observations. However, next-generation experiments, such as PRISM at JPARC with a reach of ${\rm Br}(\mu^{-}N \rightarrow e^{-}N)\sim 10^{-16}-10^{-18}$ could probe a KK scale of 6 - 20 TeV, i.e. the interesting parameter range of the present model. A third set of processes considered in Ref. [@Langacker:2000ju] are one-loop radiative lepton decays. Here the decay width is $$\Gamma(l_j\rightarrow l_i \gamma)=\frac{\alpha G_F^2 m_{l_j}^3}{8 \pi^4}\left(\left|\xi_+^{ij}\right|^2+\left|\xi_-^{ij}\right|^2\right),$$ where the dipole moment couplings of an on-shell photon to the chiral lepton currents are given by $$\xi_\pm^{ij}=\sum_n\frac{M_0^2}{M_n^2}\left(\mathcal{B}_\mp^{(n)} m_l \mathcal{B}_\pm^{(n)} \right)_{ij},$$ where $m_l$ is the charged lepton mass matrix. Using this we obtain the following rates for radiative decays ------------------------------------------- ------------------------- ------------------------- ------------------------- ------------------------- (A) (B) (C) (D) ${\rm Br}(\mu \rightarrow e\gamma):$ $ 2.0 \times 10^{-17} $ $ 3.7 \times 10^{-15} $ $ 3.4 \times 10^{-15}$ $ 1.2 \times 10^{-18}$ ${\rm Br}(\tau \rightarrow \mu \gamma): $ $5.2 \times 10^{-16} $ $ 1.9 \times 10^{-14} $ $ 9.4 \times 10^{-15} $ $ 4.1 \times 10^{-14}$ ${\rm Br}(\tau \rightarrow e \gamma): $ $ 3.7 \times 10^{-17} $ $ 1.5 \times 10^{-14} $ $ 9.2 \times 10^{-15} $ $ 2.6 \times 10^{-18}$ ------------------------------------------- ------------------------- ------------------------- ------------------------- ------------------------- All of these branching ratios lie well within the experimental bounds ${\rm Br}(\mu \rightarrow e\gamma)=1.2\times10^{-11}$ [@Edwards:1996te], ${\rm Br}(\tau \rightarrow \mu\gamma)=4.4\times10^{-8}$ and ${\rm Br}(\tau \rightarrow e\gamma)=3.3\times10^{-8}$ [@:2009tk]. Again these rates are suppressed relative to their hard-wall counterparts. In scenario (D), processes such as $\mu \rightarrow e \gamma$ can also be mediated by the KK states of the sterile neutrinos. This process was investigated for the Randall-Sundrum model in Ref. [@Kitano:2000wr]. We use the formalism developed there and we find that in our model the branching ratio for this process is given by the relative coupling strength of the muon, the W and the KK muon neutrino to the zero mode muon neutrino times a loop factor. Because of the large mass differences of the sterile KK neutrinos, the GIM mechanism breaks down. We assume that the neutrino mixing angles are large, not leading to any suppression of the rate. Given a relative coupling of $0.0057$ we find $\rm{Br}(\mu \rightarrow e\gamma)=1.5\times10^{-13}$ which again lies within the experimental bounds for a KK scale of 2 TeV. There are also contributions to $\mu \rightarrow e \gamma$ related to the exchange of KK fermions [@Agashe:2006iy; @Perez:2008ee], which were neither included in our estimate above nor in Ref. [@Kitano:2000wr]. These contributions dominate the rate of radiative lepton decays in hard-wall models. A similar behaviour is likely in the soft-wall model. However, given the suppressed rates for flavour violation in the latter, we expect that even including these extra contributions, the rate for $\mu \rightarrow e \gamma$ and a KK scale of 2 TeV should not exceed the experimental bound. We have shown that lepton flavour violation can be suppressed in the soft wall model and the amount of suppression depends on the fermion locations and the hierarchy of scales, $k/\mu$. The reason for this suppression comes from the ability to place the fermions in regions of universal gauge couplings i.e. heavily UV localised as explained in Section \[sect.gc\]. When the hierarchy of scales in the model is increased by fixing $\mu$ and increasing $k$ the gauge boson profiles remain unchanged whilst the fermion profiles become even more UV localised thus providing an even greater suppression of flavour violation. Conclusions =========== In this paper we have studied the lepton sector of the SM in a soft-wall extra dimension, applying flavour dependent fermion locations to accommodate the observed lepton flavour structure. The Higgs is a bulk field, with a VEV that increases near the soft wall. We have in particular considered the inclusion of small Dirac neutrino masses and investigated the constraints on the model from lepton flavour violation mediated by the Z boson and its KK states. In order to do so we first developed solutions for a massive gauge boson in the soft-wall background and found the profile is independent of the AdS curvature scale. In order to generate the masses of the charged leptons whilst keeping the fermions located in an area of almost universal gauge couplings we find that we need to increase the hierarchy of scales in the model to around $k /\mu = 10^{7}$. When incorporating sub-eV neutrino masses we need a much larger hierarchy, and we choose $k/\mu = 10^{15}$, similar to the hierarchy between the Planck and the electroweak scales. To incorporate three generations of leptons into our model we solve the fermion equations of motion numerically, including an order one flavour diagonal Yukawa coupling and use these solutions as a basis of states from which we treat off-diagonal Yukawa couplings, connecting different generations, as perturbations. The mass term related to the diagonal Yukawa coupling is necessary to generate a normalisable wavefunction and cannot be treated as a perturbation. We can construct the full lepton mass matrices, including KK states and diagonalise them to find the fermion masses and mixings. However, to our level of precision we can neglect the fermionic KK states. The locations of the left-handed fermions are dictated by the fact that we require large mixings in the neutrino sector. We take a large number of random Yukawa couplings and choose the locations of the right-handed fermions so that the averaged zero mode masses reproduce the SM charged lepton masses. With the inclusion of off-diagonal Yukawa couplings, the transformation to mass eigenstates produces flavour violating couplings. We calculated the expected rates for various flavour changing processes for a number of different scenarios. We found that the soft-wall model is in fact mildly constrained when we consider a scenario with a low hierarchy of scales such as $k /\mu = 10^{7}$. The most stringent constraint comes from $\mu \to e$ conversion in a muonic atom where we find that only the scenario where all the left-handed leptons have degenerate locations well toward the UV brane would occur at acceptable rates with a KK scale of 2 TeV. This is a considerable suppression of lepton flavour violation compared to hard-wall models, such as the one studied in Ref. [@Huber:2003tu]. Including a larger hierarchy of scales ($k /\mu = 10^{15}$), it is also possible to generate sub-eV Dirac neutrino masses. In this case the model is even less constrained and most of the FCNC processes would occur at rates well below the experimental bounds. The most stringent bounds are coming from radiative decays, such as $\mu \rightarrow e \gamma$. Again a KK scale of 2 TeV seems sufficient to keep the rate below the experimental bound. Our estimate for this rate does not include contributions from KK gauge bosons, and it would be interesting to include these in a more detailed analysis. Another obvious direction of research would be to extend the present setup to the quark sector, similar to an analysis that was performed recently in much detail for the hard-wall model in Ref. [@Bauer:2009cf]. The soft-wall extra dimension continues to offer a valid model for electroweak physics, with constraints from precision data relaxed compared to the hard-wall model. Having said this, we have found that with a (gauge boson) KK scale of 2 TeV the complete lepton flavour structure can be accommodated while keeping rare processes below experimental bounds. In our setup the KK states of fermions have masses around 1.5 TeV, within reach of the LHC experiment. Thus the soft-wall framework seems to offer an alternative when it comes to suppressing flavour violation to models relying on flavour symmetries [@Perez:2008ee; @Csaki:2008qq; @delAguila:2010vg], a bulk Higgs [@Agashe:2008fe] or to utilising non-minimal representations under the $SU(2)_R$ bulk gauge symmetry [@Agashe:2009tu]. The parameter range with a large hierarchy $k /\mu = 10^{15}$ is both attractive to further suppress flavour violation and necessary to accommodate neutrino masses. This rises the important question whether such a hierarchy can be stabilised, like in the way proposed in Ref. [@Cabrer:2009we]. It would be very interesting to extend our analysis to such a framework. Acknowledgements {#acknowledgements .unnumbered} ================ We would like to thank Sebastian Jäger for useful discussions. [^1]: [email protected] [^2]: [email protected] [^3]: When referring to the soft wall model we take the KK scale to be the mass of the first KK mode of the Z boson. [^4]: We do not consider CP violation, i.e. we take $\lambda_{5ij}$ to be real. [^5]: Note that we work in a gauge field basis, where the Z boson zero mode is massive. Its properly weighted wave function is not flat, resulting in non-universal couplings to fermions. Alternatively, one could work in a basis, where the zero mode is massless. Then only the KK states would couple non-universally.
{ "pile_set_name": "ArXiv" }
--- abstract: 'The Galaxy And Mass Assembly Survey (GAMA) covers five fields with highly complete spectroscopic coverage ($>95$ per cent) to intermediate depths ($r<19.8$ or $i < 19.0$ mag), and collectively spans 250 square degrees of Equatorial or Southern sky. Four of the GAMA fields (G09, G12, G15 and G23) reside in the ESO VST KiDS and ESO VISTA VIKING survey footprints, which combined with our GALEX, WISE and Herschel data provide deep uniform imaging in the $FUV\,NUV\,ugriZYJHK_s\,W1\,W2\,W3\,W4\,P100\,P160\,S250\,S350\,S500$ bands. Following the release of KiDS DR4, we describe the process by which we ingest the KiDS data into GAMA (replacing the SDSS data previously used for G09, G12 and G15), and redefine our core optical and near-IR catalogues to provide a complete and homogeneous dataset. The source extraction and analysis is based on the new <span style="font-variant:small-caps;">ProFound</span> image analysis package, providing matched-segment photometry across all bands. The data are classified into stars, galaxies, artefacts, and ambiguous objects, and objects are linked to the GAMA spectroscopic target catalogue. Additionally, a new technique is employed utilising <span style="font-variant:small-caps;">ProFound</span> to extract photometry in the unresolved MIR-FIR regime. The catalogues including the full FUV-FIR photometry are described and will be fully available as part of GAMA DR4. They are intended for both standalone science, selection for targeted follow-up with 4MOST, as well as an accompaniment to the upcoming and ongoing radio arrays now studying the GAMA $23^h$ field.' author: - | Sabine Bellstedt,$^1$[^1] Simon P. Driver,$^{1,2}$ Aaron S. G. Robotham,$^1$ Luke J. M. Davies,$^1$ Cameron R. J. Bogue,$^{1,3}$ Robin H. W. Cook,$^1$ Abdolhosein Hashemizadeh,$^1$ Soheil Koushan,$^1$ Edward N. Taylor,$^{4}$ Jessica E. Thorne,$^1$ Ryan J. Turner,$^1$ Angus H. Wright$^{5,6}$\ $^1$ICRAR, The University of Western Australia, 35 Stirling Highway, Crawley WA 6009, Australia\ $^2$SUPA, School of Physics & Astronomy, University of St Andrews, North Haugh, St Andrews, KY16 9SS, UK\ $^3$School of Physics and Astronomy, Cardiff University, Queens Buildings, The Parade, Cardiff CF24 3AA, UK\ $^4$Centre for Astrophysics and Supercomputing, Swinburne University of Technology, Hawthorn VIC 3122, Australia\ $^5$Astronomisches Institut, Ruhr-Universit[ä]{}t Bochum, Universit[ä]{}tsstr. 150, 44801 Bochum, Germany\ $^6$Argelander-Institut f[ü]{}r Astronomie, Auf dem H[ü]{}gel 71, 53121 Bonn, Germany bibliography: - 'BibLibrary.bib' title: 'Galaxy And Mass Assembly (GAMA): Assimilation of KiDS into the GAMA database' --- \[firstpage\] techniques: photometric; astronomical data bases: miscellaneous; catalogues; surveys; Introduction ============ The era of the modern wide-area imaging survey, i.e., those based on linear digital detectors and covering a sizeable portion of the sky, started in earnest with the 2MASS[^2] [@Skrutskie06], SDSS[^3] [@York00], and UKIDSS[^4] [@Lawrence07] surveys — although it would be remiss not to mention the equally transformational IRAS[^5] [@Neugebauer84] and ROSAT[^6] [@Voges99] space missions. These programs, as well as achieving transformational science from the Solar System to the distant Universe, have in turn motivated the emergence of a multitude of dedicated imaging facilities on the ground including, for example, SkyMapper [@Keller07]; VST[^7] [@Arnaboldi07]; VISTA[^8] [@Sutherland15]; and LSST[^9] [@Ivezic19], and in space, for example GALEX[^10] [@Martin05]; WISE[^11] [@Wright10]; Spitzer [@Werner04]; Herschel [@Pilbratt10]; Euclid [@Beaulieu10]; and WFIRST[^12] [@Gehrels15] to highlight a few. The further [*federation*]{} of these data streams with ground-based spectroscopy and other facilities, has allowed for the construction of a truly multi-wavelength and three-dimensional view of our Universe [e.g., @Jarrett17; @Driver18]. In particular, major advances have been made in quantifying: the spatial distribution of galaxies and their use for cosmology; the distribution of groups and clusters; studies of galaxy populations; galaxy merger rates; the assembly of mass (stellar, dust, gas and super-massive black holes); the transformation of mass; and identified the primary energy production pathways (star-formation and Active Nuclei); all as a function of look-back time and environment. The data and science from these surveys now dominate our knowledge of the near and intermediate Universe, and provide the vital zero redshift benchmark for studies of the distant and adolescent Universe. Not only has our knowledge and understanding been advanced, but also the way in which astronomy is conducted, shifting from an individual to team pursuit [@Milojevic14]. Collectively, these [*major*]{} endeavours have allowed us to start the process of [*comprehensively*]{} mapping the evolution of all mass, energy, and structure over all cosmic time and to build the scaffolding upon which the numerical N-body, hydrodynamic and semi-analytic models hang [e.g. @Lagos19]. In the coming years such comprehensive studies will be massively augmented with new wide-area optical/near-IR (LSST, Euclid, WFIRST), x-ray (eROSITA), and deep radio (MeerKAT, ASKAP, MWA and SKA) imaging and spectral surveys, taking us from a multi-wavelength outlook, to a truly panchromatic perspective. While acknowledging this impending paradigm shift from a mono- to pan- facility culture, it is worth noting that the majority of all photons produced, since mass-energy decoupling (by energy or number), arise in the ultra-violet, optical and near-IR regimes [see the recent summary of the extra-galactic background by @Hill18]. Half of these photons are predominantly produced by stars and through star-formation, and the other half are produced through the accretion of baryonic material onto supermassive black holes. One important caveat, is that almost half of these freshly minted photons [@Dunne03; @Driver16; @Driver16b], are almost immediately attenuated by dust grains, which reradiate the energy into the far-IR, before it emerges from the host galaxy. Including the shifting of wavelengths longward due to the expansion, the implication is that when building our panchromatic perspective, one might wish to start where photon production is dominant and readily detectable (i.e., the optical/near-IR) and where, arguably, the information content is highest. Here we describe the construction of a new deep optical/near-IR imaging dataset, built upon two ESO Public Surveys [VST KiDS[^13] and VISTA VIKING @deJong13; @Arnaboldi07] combined with the Galaxy And Mass Assembly (GAMA) panchromatic and spectroscopic survey [@Driver11; @Hopkins13; @Liske15; @Driver16b]. In particular a key 50 sq degree region [G23/WD23; see @Driver19], will be targeted for future high-density spectroscopic, x-ray spectral, and radio line and radio continuum observations. The dataset presented herein, and including all unique GAMA redshifts, therefore forms the basis upon which to grow our panchromatic perspective. At its core the Galaxy And Mass Assembly Survey [GAMA; @Driver11; @Liske15], spanning five fields, is a spectroscopic Legacy campaign using the Anglo Australian Telescope’s AAOmega wide-field facility [see @Hopkins13]. The five fields are each $\sim$50–60 sq degrees in extent, and located at: 2$^h$ (G02), 9$^h$ (G09), 12$^h$ (G12), 14.5$^h$ (G15), and 23$^h$ (G23). The G09, G12 and G15 fields lie in the equatorial North Galactic Cap region, and have similar properties in terms of depth of the spectroscopic follow-up ($r<19.8$ mag), area (60 sq deg), spectroscopic completeness (98 per cent), and panchromatic coverage [@Driver16b UV to far-IR;]. The bulk of the GAMA science to date is based on the analysis of these three fields. The original GAMA G02 field overlaps with the VIPERs and XMM-XXL equatorial field, and covers 55.7 sq degrees in extent. The field was not completed, however a 19.5 sq degree sub-region attained uniform 95.5 per cent spectroscopic completeness to $r<19.8$ mag [see @Baldry18]. The final GAMA field at 23$^h$ and $-32.5\deg$, lies in the Southern Galactic cap, with a spectroscopic survey limit of, $i < 19.0$ mag, but with a slightly lower completeness of 94 per cent [see @Liske15]. To date little science has been based on the G23 region [although see studies such as @Bilicki18; @Vakili19 for examples where these data have been used], however in due course it represents our premier field, because of its suitability for southern hemisphere follow up. In particular, this follow-up will be conducted by radio facilities, and a deep spectroscopic extension is planned as part of the Wide Area VISTA Extra-galactic Survey (WAVES); one of ten core surveys to be conducted by the 4MOST Consortium [see @Driver19; @deJong19]. This will extend the G23 region at high spectroscopic completeness ($>90$ per cent), to a limit of $m_Z \leq 21.2$. In addition WAVES will also survey the full KiDS region ($m_Z \leq 21.2$, $z_{\rm phot} < 0.2$), and the LSST Deep-Drill fields ($m_Z \leq 21.2$, $z_{\rm phot} < 0.8$). Note that object selection for the WAVES-wide survey will be conducted using joint KiDS and VIKING photometry. In preparation, the G23/WD23 region is being extensively observed by Southern Hemisphere located radio facilities including: the Australian Compact Array (ATCA) as part of the GAMA Legacy ATCA Sky Survey (GLASS; Hyunh et al. in prep.), the Australian Square Kilometre Array Pathfinder [@Leahy19] as part of the EMU[^14] [@Norris11], DINGO[^15] (Meyer et al. in prep.) and FLASH[^16] [@Allison20] surveys, and by the Murchison Wide-Field Array GOLD and MIDAS surveys (Seymour et al. in prep). The expectation is that the G23/WD23 region, with its exceptionally high-density and deep spectroscopic completeness, should be a suitable location for a medium-deep survey, with upcoming facilities such as the Vera Rubin Observatory[^17], the Square Kilometer Array (SKA), Euclid, and the Wide-Field InfraRed Space Telescope (WFIRST). In terms of panchromatic imaging, G23/WD23 currently has comparable coverage to the GAMA equatorial fields, with data arising from concerted GALEX (NUV), VST KiDS, VISTA VIKING, WISE, and Herschel imaging campaigns. This wealth of data, combined with radio observations, and future upcoming deep spectroscopic observations, makes G23 a field of interest in coming years for extensive follow-up of either the entire field, or well selected sub-samples. In Section \[sec:kidsdata\] we describe the assimilation of the KiDS data into the GAMA Panchromatic Database followed by the generation of the base source catalogues from FUV to W2 using the new ProFound image analysis package [@Robotham18]. In Section \[sec:FIR\] we use <span style="font-variant:small-caps;">ProFound</span> PSF-convolution mode to obtain photometry from W3 through to the PACS and SPIRE far-IR bands for objects brighter than $r \sim 20.5$ mag. In Section \[sec:verification\] we verify the zeropoints, astrometry and compare our revised photometry to our previous LAMBDAR-based photometry. This includes verification of the zeropoints and astrometry, star masking using GAIA DR2, Galactic extinction corrections using *Planck*, star-galaxy separation based on colour and size, extensive visual inspection, and comparisons to earlier data. In Section \[sec:catAccess\] we provide information on how to access the catalogues, and in particular provide some example extractions. Two companion papers describe the search for low surface brightness galaxies within the dataset (Turner et al. in prep), and the use of the panchromatic data to reconstruct the star-formation history of individual galaxies and sub-populations via a “forensic"-style analysis (Bellstedt et al. in prep). Further papers incorporating radio observations are in preparation. All magnitudes reported here are in the AB system and when necessary we assume a cosmology with $H_0 = 70\,\rm{km}\,\rm{s}^{-1}\,\rm{Mpc}^{-1}$, $\Omega_m = 0.3$ and $\Omega_{\Lambda} = 0.7$. Assimilation of VST KiDS into the GAMA Panchromatic Database {#sec:kidsdata} ============================================================ The target catalogues to the GAMA spectroscopic campaign were built upon three distinct optical surveys: the Sloan Digital Sky Survey (SDSS; G09, G12 and G15), the Canada-France Legacy Survey (CFHTLS; G02), and the European Southern Observatory’s VLT Survey Telescope’s Kilo-degree Suvey (KiDS; G23). For the equatorial regions the GAMA input catalogue is described in detail in [@Baldry10] and for the G02 region is described in [@Baldry18]. The G23 input catalogue, used for the GAMA spectroscopic survey, has not been described previously and in brief was constructed in 2014 based on initial pre-release VST KiDS data. These data have since undergone a number of revisions in terms of re-determination of the photometric zero-points, replacement of low quality data frames, and the filling in of data gaps as the KiDS team have honed their reduction and analysis pipelines. Nevertheless, our early KiDS analysis resulted in an $i$-band limited target catalogue ($i < 19.0$ mag), with star-galaxy separation based on table-matched near-infrared colours and size estimates, augmented with extensive and fairly ad hoc eyeball checks (based on selections designed to identify artefacts and ensure no galaxies were misclassified as stars). This initial input catalogue is available from the GAMA database and, while not ideal nor optimal, formed the basis for spectroscopic observations with the AAOmega facility on the AAT from 2014 – 2016 see @Liske15. Since this time the VST KiDS team has completed $ugri$ coverage of the three GAMA equatorial fields, and the G23/WD23 field. These data have recently been released as part of VST KiDS DR4[^18], [@Kuijken19] and provide near complete coverage in all bands across the four primary GAMA fields. In one region of G23 the DR4 data is missing, however data exists from the earlier DR3 release and so we include these three fields. We are hence now in a position to redefine the GAMA base optical/near-IR catalogues in a uniform manner across our four primary regions. In doing so we create both deeper, and higher resolution imaging, from which we can derive more robust flux, size measurements and derived parameters (e.g., stellar masses, star-formation rates, and photometric redshifts). The main purpose of this paper is to provide a record of this replacement process — a process akin to swapping the tablecloth on a fully laid table. KiDS DR4 data are downloadable from the ESO archive, and come pre-SWARPed [@Bertin10] into 1 sq degree tiles. These are astrometrically and photometrically calibrated by the KiDS team using, initially, the Sloan Digitial Sky Survey in the North and 2MASS in the South, and with further supplementary calibration to GAIA DR2 $g$, as part of the final DR4 calibration process. Note that the DR4 data tiles as released, contain both a zero-point for each tile ([PHOTZ]{}), reflecting the initial calibration, *and*, a further zero-point offset ([DMAG]{}) to adjust any derived flux measurements to the GAIA DR2 $g$-band system. Our initial action is therefore to modify all DR4tiles to absorb the [DMAG]{} correction into the specified zero-points, by scaling the data. This is done to legitimately mosaic tiles, allowing for the construction of KiDS maps at any location and any size within the KiDS footprint. We now follow the procedure outlined in the GAMA Panchromatic Data Release [PDR; @Driver16b] and build large single SWARP [@Bertin10] images for each GAMA region. These (very) large mosaics are available via the Panchromatic SWARP Imager[^19] [@Driver16b]. We discuss the revised panchromatic depth of this imaging in Section \[sec:verification\]. In total we SWARP 280 sq degree [tile]{}s from KiDS, and also take the opportunity to rebuild our VIKING SWARPs using additional data amounting to 129,869 VISTA detectors (see Koushan et al. in prep. for details on the VISTA VIKING data). This comprises a total data volume of 3.44TB and all mosaics are available via the URL indicated above and via the Public Data Central portal[^20]. Images showing a visual comparison of KiDS and SDSS data are provided in figure 1 of Turner et al. (in prep). The adoption of ProFound for source detection — a brief digression ------------------------------------------------------------------ In constructing the PDR [@Driver16], we made use of the original source detection as provided by the Sloan Digital Sky Survey Data Release 6, within our survey footprint. Following star-galaxy separation based on colour and size criteria [@Baldry10], these data were used to define the GAMA input catalogue for the GAMA spectroscopic survey of the three equatorial fields, undertaken on the Anglo-Australian Telescope from 2011-2016 [see @Liske15]. The original SDSS-derived equatorial input catalogue, was later table matched to our independent $r$-band catalogues, determined using [Source Extractor]{} [@Bertin96]. [Source Extractor]{} was then applied in dual band mode to determine forced aperture photometry from $u$ to $K_s$, using the elliptical apertures defined by [Source Extractor]{}. Further table matching to independent catalogues in GALEX, WISE, and Herschel bands resulted in the far-UV to far-IR publicly available PDR dataset[^21]. This catalogue was later superseded by flux measurements also based on our $r$-band defined apertures but now using the LAMBDAR in-house software [@Wright16] to measure forced photometric fluxes in all 21 bands (covering UV to IR wavelengths) following convolution of the initial aperture with the relevant facility point-spread function. Throughout this process, a number of important lessons related to galaxy photometry emerged. Firstly, the undesirable reliance on table-matching to connect the SDSS input catalogue to our [LAMBDAR]{} photometry, which inherently introduces errors due to different deblending outcomes between the [SDSS Imaging Pipeline]{} and [Source Extractor]{} methodologies. Secondly, issues arose around the integrity of the [Source Extractor]{} aperture definitions. In particular [Source Extractor]{}, like most detection algorithms, can be prone to bright galaxy fragmentation, and in some cases highly erroneous apertures, often due to the defined aperture following an isophotal bridge and looping round a nearby bright star — these issues arise because of the difficulty in simultaneously measuring fluxes for both very bright and very faint sources. Similar issues were also identified in our re-analysis of panchromatic photometry in the G10/COSMOS field [@Andrews17]. Following visual assessment of all GAMA apertures, via a citizen science project, it became apparent that typically 10 per cent of all apertures did not define the object to the desired level of accuracy. This led to the development of a new source finding code,[^22] <span style="font-variant:small-caps;">ProFound</span> [@Robotham18], with three important philosophical changes in source finding. Firstly, instead of using circular ([SDSS Imaging Pipeline]{}), or elliptical ([Source Extractor]{}) apertures, [ProFound]{} acknowledges that most galaxies have irregular shapes, particularly as one probes to higher redshifts, fainter isophotal levels, and closer to the confusion limit. The [ProFound]{} software identifies and preserves the initial isophote (or segment), which may be regular or irregular in shape. Secondly, we introduced the concept of segment-dilation (akin to a curve-of-growth) to obtain pseudo-total fluxes through the sequential addition of layers of pixels surrounding each segment. This process continues until the flux converges (in our case defined by a less than 5 per cent increase in flux), or the maximum number of allowable dilations is reached, resulting in pseudo-total magnitude estimates. Thirdly, [Source Extractor]{} uses a hierarchical or nested-deblend process, by which derived elliptical apertures may overlap, and hence where flux can be double counted if not managed appropriately in later analysis. In the [ProFound]{} software package, a watershed deblending approach is taken, where during the dilation process segments are not allowed to overlap — i.e., all the flux in any one pixel is allocated to one object only. One can argue in specific cases as to which deblend approach, nested or watershed, is more appropriate, e.g., a nested approach is better for a small satellite within a large halo, while a watershed approach is better for dense complexes, or as one approaches the confusion limit. Our experience is that the watershed approach behaves better when things go pathologically wrong — i.e., it is the least worst of the two approaches for difficult cases. For full details on [ProFound]{} see the code description paper @Robotham18, or a recent applications to the deep DEVILS imaging data from VISTA VIDEO [@Davies18]. In the sections that follow we will adopt and apply ProFound and develop a pipeline around it to manage the multitude of issues that arise with wide area data collected from multiple ground-based facilities. We show the adopted workflow for the pipeline presented within this paper in Fig. \[fig:flowchart\]. ![Workflow for the adopted pipeline. Inputs to the pipeline are shown in orange, steps of the pipeline in blue, and outputs in yellow. []{data-label="fig:flowchart"}](PhotometryWorkflow.pdf){width="45.00000%"} ![image](Photometry_example.png){width="\textwidth"} -------- -------------------- ------------ -------------------------- ----------- ----------------------------------- --------------------- -------------------- Band Central Wavelength Instrument Data set/survey Area Limits median seeing Zero-point (sq. deg) (AB mag) (arcseconds) (AB mag for 1 ADU) $FUV$ 1 539 Å GALEX MIS+GO $^{\star}$ 186.74 24.59 – 26.40 $^{\triangleright}$ N/A 18.82 $NUV$ 2 316 Å GALEX MIS+GO $^{\star}$ 204.3 23.64 – 24.07 $^{\triangleright}$ N/A 20.08 $u$ 3 582 Å VST KiDS $^{\ast}$ 211.21 24.8 $^{\ast}$ $0.9-1.1$ $^{\ast}$ 0 $g$ 4 760 Å VST KiDS $^{\ast}$ 211.21 25.4 $^{\ast}$ $0.7-0.9$ $^{\ast}$ 0 $r$ 6 326 Å VST KiDS $^{\ast}$ 211.21 25.2 $^{\ast}$ $<0.6$ $^{\ast}$ 0 $i$ 7 599 Å VST KiDS $^{\ast}$ 211.21 24.2 $^{\ast}$ $<1.1$ $^{\ast}$ 0 $Z$ 8 854 Å VISTA VIKING $^{\dagger}$ 211.21 23.04 – 23.19 $^{\triangleright}$ 1.0 $^{\ddagger}$ 30 $Y$ 10 229 Å VISTA VIKING $^{\dagger}$ 211.21 22.34 – 22.51 $^{\triangleright}$ 1.0 $^{\ddagger}$ 30 $J$ 12 556 Å VISTA VIKING $^{\dagger}$ 211.21 22.06 – 22.21 $^{\triangleright}$ 0.9 $^{\ddagger}$ 30 $H$ 16 499 Å VISTA VIKING $^{\dagger}$ 211.21 21.33 – 21.42 $^{\triangleright}$ 1.0 $^{\ddagger}$ 30 $K_S$ 21571 Å VISTA VIKING $^{\dagger}$ 211.21 21.30 – 21.48 $^{\triangleright}$ 0.9 $^{\ddagger}$ 30 $W1$ 3.40 $\mu$m WISE AllSky $^{\diamond}$ 211.21 21.09 – 21.41 $^{\triangleright}$ N/A 23.16 $W2$ 4.65 $\mu$m WISE AllSky $^{\diamond}$ 211.21 20.26 – 20.77 $^{\triangleright}$ N/A 22.82 $W3$ 12.8 $\mu$m WISE AllSky $^{\diamond}$ 211.21 18.44 – 18.89 $^{\triangleright}$ N/A 23.24 $W4$ 22.4 $\mu$m WISE AllSky $^{\diamond}$ 211.21 16.54 – 16.96 $^{\triangleright}$ N/A 19.6 $P100$ 98.9 $\mu$m PACS ATLAS $^{\triangleleft}$ 211.21 12.96 – 13.14 $^{\triangleright}$ N/A 8.9 $P160$ 156 $\mu$m PACS ATLAS $^{\triangleleft}$ 211.21 13.44 – 13.66 $^{\triangleright}$ N/A 8.9 $S250$ 249 $\mu$m SPIRE ATLAS $^{\triangleleft}$ 181.14 12.52 – 12.60 $^{\triangleright}$ N/A 11.68 $S350$ 350 $\mu$m SPIRE ATLAS $^{\triangleleft}$ 181.14 12.36 – 12.51 $^{\triangleright}$ N/A 11.67 $S500$ 504 $\mu$m SPIRE ATLAS $^{\triangleleft}$ 181.14 12.16 – 12.23 $^{\triangleright}$ N/A 11.62 -------- -------------------- ------------ -------------------------- ----------- ----------------------------------- --------------------- -------------------- $^{\star}$ @Martin05, $^{\ast}$ @deJong13 [@deJong13b], $^{\dagger}$ @Edge13, $^{\ddagger}$ @Venemans15, $^{\diamond}$ @Wright10, $^{\triangleleft}$ @Eales10, $^{\triangleright}$ @Driver16 Building 1.25 sq degree overlapping tiles with [SWARP]{} -------------------------------------------------------- Prior to running [ProFound]{} on KiDS and VIKING data, we first use the [SWARP]{} package [@Bertin10] to build slightly extended $1.12^{\deg} \times 1.12^{\deg}$ images in each band ($FUV$, $NUV$, $u$, $g$, $r$, $i$, $Z$, $Y$, $J$, $H$, $K_s$, $W1$, $W2$, $W3$, $W4$, $P100$, $P160$, $S250$, $S350$, $S500$) - essentially adding overlap regions. These revised tiles are centred on the rescaled KiDS 1 sq degree tiles sourced from the ESO archive. In [SWARP]{}ing the data we regrid all bands to a resolution of 0.339 arcsec, use a background smoothing mesh of $256 \times 256$ (pixels), and a background filter-size of $3 \times 3$ (background cells). The dat frames are combined using the MEDIAN combine option, which we have deemed to be the most stable option for regions with poor-quality data. Fig. \[fig:swarp\] shows an example of a SWARped tile indicating the depth and quality in each band. Visual inspection was made of all data frames, using images similar to Fig. \[fig:swarp\], to ensure each tile in each band was correctly built. In some cases it was noted that the individual tiles provided by KiDS were missing detectors in some bands, and in some tiles, which will leave gaps in the panchromatic coverage. Objects with missing coverage in a particular band will have fluxes and magnitudes set to either NA or -999. Based on a visual inspection of the image quality, we conclude that a [ProFound]{} analysis based on combined $r+Z$ stacks will provide near complete and contiguous coverage over all four GAMA regions. ![The above panel shows a section from the top mid region of Fig. \[fig:swarp\] VST KiDS $r$ frame (top) and a VISTA VIKING $Z$ frame (middle) and the combined stack weighted by inverse noise variance (bottom). A problem region is shown indicating how the use of two frames from different telescopes helps to mitigate issues due to stellar ghosting.[]{data-label="fig:zrzr"}](rZstack.png){width="47.00000%"} Source detection with [ProFound]{} {#sec:sourceDetection} ---------------------------------- Source detection is conducted via the [ProFound]{} package [@Robotham18]. This is based on <span style="font-variant:small-caps;">ProFound</span> version 1.10.8 which can be obtained from <https://github.com/asgr/ProFound>. Within [ProFound]{} numerous parameters exist that determine the manner in which sources are extracted from the image. For completeness we show the exact command we use in the Appendix using the [profoundMultiBand]{} command. The command allows the user to specify one or more bands to use for the detection pass as well as the bands for which measurements should be made. In the initial source detection phase, only the detection bands are provided. When multiple images are specified for detection they are combined in an inverse variance weighting based on the internal background assessment. Here we combine data from the KiDS r band and the VIKING Z band, i.e., $r+Z$. This has the distinct advantage of overcoming some artefact effects such as ghosting, satellite trails, and bad pixels. Fig. \[fig:zrzr\] shows a KiDS VST $r$ band image (upper), a VISTA VIKING image (middle), and the combined $rZ$ image (lower). ![image](sky.png){width="\textwidth"} Inherent to [ProFound]{}, is its robust modelling of the local background, which may include the sky, the haloes of bright objects, scattered light or artificially enhanced regions, through appropriate median filtering [see @Robotham18]. Fig.  \[fig:back\] shows an example [tile]{} with the original image (left), the derived background map (centre, where the background is seen to be elevated near the positions of bright objects, particularly bright ghosting), and the sky root-mean sky statistics (right, again indicating regions of heightened uncertainty in the background estimate). This information is used in determining flux errors, ensuring objects in noisier regions have appropriately derived errors. In examining the background in detail the genesis of the data is also apparent in the Sky Root Mean Square (SkyRMS) map (right panel). This is a common feature in surveys we have studied, and in this case highlights the varying noise characteristics of the individual detectors going into the initial [SWARP]{} [tile]{} image. Rebuilding fragmented galaxies {#sec:rebuildingFragmentedGalaxies} ------------------------------ A common problem in most automated detection algorithms is that of fragmenting of bright galaxies. To check this we select all known galaxies from the Third Reference Catalogue [RC3; @deVauc91] which lie within the GAMA regions, and produce cutout images with the derived segments overlain. There are 21, 144, 31, and 10 RC3 galaxies within the G09, G12, G15, and G23 regions respectively. Note that the G12 region includes the Virgo Southern Spur, and G23 includes a nearby void region. Initial investigations showed that [ProFound]{} also tended to overly fragment very bright galaxies. Two enhancements were implement in [ProFound]{} to assist with this. One new parameter on top of the standard <span style="font-variant:small-caps;">tolerance</span> threshold (which determines how much peak flux an object needs relative to neighbouring objects before being merged) is <span style="font-variant:small-caps;">reltol</span>. This modifies the <span style="font-variant:small-caps;">tolerance</span> by the ratio between the segment peak flux and the saddle point flux where it touches a neighbouring segment to the power of <span style="font-variant:small-caps;">reltol</span>. Since the default is <span style="font-variant:small-caps;">reltol</span>$=0$, this will in general have no effect. However, when it is made larger than 0 merging becomes more aggressive in the outskirts of galaxies where the peak flux will tend to be much larger than the saddle point flux. Subjectively, raising this above 0 tends to do a better job of keeping very extended and flocculent spiral galaxies intact, and it has little negative impact on the fainter source deblending that parameters will tend to be optimised for (since this is where most of our survey sources exist). The other new parameter to better control segmentation is <span style="font-variant:small-caps;">cliptol</span>. This specifies the saddle point flux above which segments are always merged, regardless of competing criteria. For very bright objects with complex image artefacts (e.g. around bright stars) this proves to be very successful at properly reconstructing sources which might otherwise be significantly fragmented due to the presence of spurious flux discontinuities. Given most applications will apply a bright star mask, this option is perhaps somewhat cosmetic, but it does mean <span style="font-variant:small-caps;">ProFound</span> will return reasonable photometry even for the brightest and most difficult sources. As this fragmentation was occurring despite the new <span style="font-variant:small-caps;">ProFound</span> parameters, a process was implemented to manually regroup segments that belong to a single object. To complete this task, an in-house tool was developed that allowed users to view a thumbnail of an object and click on segments to be regrouped. This tool is available through the `profoundSegimFix` function within <span style="font-variant:small-caps;">ProFound</span>. An example of an object whose segments have been merged in this way can be seen in Fig. \[fig:frag\], where the right panel shows the resulting segmentation map after merging. Per square degree, an output file was produced that recorded which segments (as determined by the detection phase of <span style="font-variant:small-caps;">ProFound</span>) needed to be regrouped. Rather than visually inspecting every object, only objects with three or more abutting segments, a total group[^23] magnitude $<20.5$, and groups not flagged as containing a star were selected to be visually checked. For the full GAMA sample, this resulted in 75,863 objects. Of these, 6,777 required manual intervention. In total, this task took a week, with seven authors (SB, SPD, AR, LJD, JT, RC, KB, HH) assisting in the regrouping process. Before running the multiband form of <span style="font-variant:small-caps;">ProFound</span>, the manual fixes to segments were applied to the detected segmentation map using the command `profoundSegimKeep`. This fixed segmentation map was used for the remainder of the photometry pipeline. ![image](Fragment_example_multiple_doubleColumn.png){width="\textwidth"} Multiband Photometry with [ProFound]{} {#sec:multibandPhotometry} -------------------------------------- After initial segments (isophotal outlines), are defined from the stacked $r$ and $Z$ image (as described in Section \[sec:sourceDetection\]) and have been fixed for fragmentation (as described in Section \[sec:rebuildingFragmentedGalaxies\]), the fixed segments then form the basis for subsequent measurements in the analysis bands ($FUV$, $NUV$, $u$, $g$, $r$, $i$, $Z$, $Y$, $J$, $H$, $K_s$, $W1$, $W2$). Within each band, flux measurements are presented in two different ways. In order to account for all of the flux of a single object, individual segments require dilation beyond the detected segment. Dilation is conducted iteratively, where the edges are extended until all object flux has been accounted for. The background sky estimate is made as both a *global* sky measurement, or a *local* sky measurement, where the sky is measured within the dilated annulus. The resulting flux when using the global sky measurement for sky subtraction is indicated as <span style="font-variant:small-caps;">flux\_t</span> in the catalogue, whereas the flux resulting from a local sky subtraction is indicated as <span style="font-variant:small-caps;">flux\_l</span>. For large objects with significant halo flux, a local sky subtraction is liable to subtracting off the halo light, and therefore the <span style="font-variant:small-caps;">flux\_t</span> measurement is expected to better represent the total galaxy flux. Conversely, for very faint galaxies (particularly those that are close to noisy regions), we expect that the <span style="font-variant:small-caps;">flux\_l</span> measurement will best represent the galaxy flux. For the sake of selection cuts in the remainder of the paper, we have utilised the <span style="font-variant:small-caps;">flux\_t</span> measurement. The flux errors derived by <span style="font-variant:small-caps;">ProFound</span> include the errors due to sky subtraction and sky RMS (as outlined in Section \[sec:sourceDetection\]). Whilst it is also possible for <span style="font-variant:small-caps;">ProFound</span> to derive the error contribution by pixel noise correlation (introduced through the SWARP pixel resampling process), we have not included this contribution in our analysis as it contributes only a small (often negligible) fraction of the error introduced by the sky and sky rms. Based on tests for a single square degree, we find that the median pixel correlation contribution to the flux error is highest in the $NUV$/$W1$/$W2$ bands with 11/12/8 per cent respectively, but overall it is much smaller, with a mean contribution of 3 per cent over all bands. The error due to pixel correlation is expensive to compute, therefore omitting this uncertainty contibution saves significant computational time. We refer the reader to @Robotham18 for the details on uncertainty derivation in <span style="font-variant:small-caps;">ProFound</span>. For completeness, we show the `profoundMultiBand` command used to derive our multi-band catalogues for each SWARped tile in the Appendix. This amounts to the production of 280 distinct catalogues each containing around 300,000 objects and taking about 12hrs to build the SWARPs, and a further 6hrs to process <span style="font-variant:small-caps;">ProFound</span>. The software is hence run on the Pawsey Supercomputing Centre’s Zeus machine, taking about 2 days to complete a full run across all tiles. Linking to GAMA objects {#sec:GAMAmatch} ----------------------- To link the new photometric catalogue to the existing GAMA target catalogue (which contains 1,468,620 objects across all four GAMA fields), we project the GAMA catalogue onto the <span style="font-variant:small-caps;">ProFound</span> segments. A successful projection will occur if a GAMA coordinate is encompassed by a corresponding <span style="font-variant:small-caps;">ProFound</span> segment. In some cases (44,766 instances over all four fields), a single <span style="font-variant:small-caps;">ProFound</span> segment is linked to more than one GAMA input object. This generally occurs if multiple GAMA targets were placed on a single object, but can also occur if two objects have not been appropriately deblended by <span style="font-variant:small-caps;">ProFound</span>, and hence share a segment. Two strategies are implemented in order to decide which GAMA ID should be assigned to the segment when a single segment coincides with multiple GAMA objects. If the GAMA objects have spectroscopically-measured redshifts, then the selected ID is taken to be the object whose redshift is closest to the flux-weighted mean redshift of all GAMA objects present in the segment. If redshift measurements do not exist, however, then the selected ID is taken from the object contributing the largest amount of flux to the segment. For individual cases where a segment contains multiple GAMA sources that have redshifts varying by more than 0.1, we allocate a flag <span style="font-variant:small-caps;">Z\_ConfusionFlag</span>$=1$. This assists in the identification of objects for which the redshift measurement is not indicative of all the flux in the segment. ### Objects in KiDS/VIKING not in GAMA After matching our final KiDS/VIKING catalogue to the GAMA input catalogue (which extends to $r_{\rm GAMA}$=21.0 mag), we cut the catalogue at $r_{\rm KiDS} < 20$ mag (just beyond the GAMA spectroscopic limit of $m_r = 19.8$). This identifies almost 11,000 galaxies not previously recorded. Visual inspection of all $\sim$11,000 reveals that $\sim$6,000 of these objects are artefacts not previously flagged, $\sim$3,000 are galaxies not previously identified, and the remaining objects are equally divided between stars or ambiguous objects. In total these objects represent $<1$ per cent of the galaxies above this flux limit but nevertheless we introduce an eyeball flag [eyeclass]{} so that these objects can be indicated. We also introduce an [uberclass]{} flag which takes as its value the [eyeclass]{} if known or the [class]{} flag if not known. We therefore recommend the [uberclass]{} flag be used to extract star, galaxy and/or ambiguous subsets. In addition to the previously missed galaxies we identify a mixture of low surface brightness systems, and objects that have been identified as the close pair of a previously-identified galaxy, but had not been separately resolved in the past. In a companion paper, Turner et al. (in prep), we provide more detail on these objects and discuss the implications for the stellar mass density. ### Objects in GAMA not in KiDS/VIKING Similarly we can also identify objects in the GAMA spectroscopic target catalogue ($r_{\rm GAMA}<19.8$mag) that are not matched in the new KiDS/VIKING catalogues. Either, a non-match arises from the fact that no object has been identified at the coordinate of the GAMA object, or because multiple GAMA objects have been engulfed by a single segment, resulting in only a portion of the GAMA objects appearing in the final catalogues. Within the four GAMA fields, 16,068 objects have been identified for which no object has been detected in the KiDS/VIKING photometry, and in almost all cases this is because the GAMA object points to a sky position in our updated photometry. This is likely an indication that the original SDSS photometry on which the GAMA input catalogue was based contained some sort of artefact at these coordinates. Only $\sim$150 of these GAMA targets have securely measured redshifts, and these objects fall within regions that are missing imaging in $r+Z$, and hence do not appear in the new KiDS/VIKING catalogues. Finally, 44,766 objects have been identified within the new catalogues that match back to two or more GAMA targets, corresponding to $\sim3$ per cent of the total sample. As a result, an additional 44,782 GAMA targets do not appear in the KiDS/VIKING catalogues. For those cases where multiple GAMA targets in a single <span style="font-variant:small-caps;">ProFound</span> segment had redshifts, we find that in 69 per cent of cases have $\Delta z<0.01$, implying that in the majority of cases where GAMA targets have been consolidated, these do in fact belong to a single object along the line-of-sight. This highlights that the original GAMA target catalogue had fragmented objects more often than <span style="font-variant:small-caps;">ProFound</span> has merged multiple objects into a single segment. Hence, a total of $\sim4$ per cent of the GAMA target objects do not appear in our updated catalogues. Extinction corrections using Planck E(B-V) maps {#sec:extinctionCorrection} ----------------------------------------------- ![image](Extinction.png){width="95.00000%"} To correct our magnitudes for the effects of Galactic extinction we use the Planck $E(B-V)$ map[^24] [@Planck13]. From this map we extract the $E(B-V)$ values, and convert the HEALpix values to RA and Dec, and identify the closest $E(B-V)$ value to each object in each of our catalogues. We then correct all magnitudes, magnitude errors, surface brightnesses, surface brightness errors, fluxes and flux-errors for [*all*]{} objects (i.e., stars, galaxies, artefacts etc). We determine the attenuation correction for each band in the normal way ($A_x = [A_x/E(B-V)] \times E(B-V)$) using the extinction co-efficients listed and cited in Table. \[tab:atten\] (which implicitly use the Galactic extinction law from @Schlafly11). [c|c|c]{} Filter (x) & \[$A_x/E(B-V)$\] & Vega to AB\ \ $FUV$ & 8.24152 & -\ $NUV$ & 8.20733 & -\ \ $u$ & 4.81139 & -\ $g$ & 3.66469 & -\ $r$ & 2.65460 & -\ $i$ & 2.07472 & -\ \ $Z$ & 1.55222 & 0.502\ $Y$ & 1.21291 & 0.600\ $J$ & 0.87624 & 0.916\ $H$ & 0.56580 & 1.366\ $K_s$ & 0.36888 & 1.827\ \ $W1$ & 0.20124 & -\ $W2$ & 0.13977 & -\ $W3$ & 0.05433 & -\ $W4$ & 0.02720 & -\ $^{\dagger}$ @Kuijken19 $^{\ddagger}$ @Gonzalez-Fernandez18 Fig. \[fig:extinction\] shows how the $E(B-V)$ values vary across the four GAMA fields highlighting the significant structure due to streaks of Galactic cirrus. However we note the maximum $E(B-V)$ shown in the plots reaches only to 0.07, hence amounting to $0.2$ mag of extinction in the $r$ band. Constructing the star-mask from GAIA DR2 {#sec:starMask} ---------------------------------------- Figs. \[fig:swarp\], \[fig:zrzr\] and  \[fig:back\] highlight the issue of ghosting around bright stars, and how the location of this ghosting is dependent on both the position within the focal plane, and the flux of these stars. Photometry of objects in these regions will be compromised, and for many purposes it will be necessary, or desirable, for these objects to be removed. To build a star-mask flag we elect to use the recently released GAIA DR2 catalogue [@Gaia18], which contains robust positions for all bright objects across the sky. However, first we need to remove any galaxies in the GAIA DR2 catalogue, as we do not wish to mask these objects. To do this we match to both the RC3 catalogue, and also our previous GAMA catalogue, which has been extensively visually inspected and for which most objects (98 per cent), have had redshifts measured to $r<19.8$ mag (or redshifts measured out to $r<19.2$ mag in the case of G23). Matching GAIA DR2 to RC3 results in 54 matches within the GAMA regions, while matching to GAMA identifies a further 684 objects to GAIA DR2 $g < 18.0$ mag. These are removed from our GAIA star-mask catalogue. We then extract cutouts of a random sample of GAIA stars and identify a $g$ mag-radius relation, as shown in Fig. \[fig:starmask\]. Inside the radius indicated (solid black line), the artefact rate is extremely high, and photometry will be compromised and classification problematic. Fig. \[fig:stars\] shows examples for four regions centred on four stars, showing firstly two extremely bright stars (which are relatively rare: upper), to two more typical regions with a smattering of masked stars (lower). The dotted circles indicate the masked region indicated by Eqn. \[eqn:starradius\], and all objects within these regions have their [starmask]{} flag set to 1 and are shown on these figures with yellow outlines. $$\begin{aligned} r[\mbox{arcmin}]=10^{(1.6-0.15g)} \mbox{ and } [r < 5.0', g < 16.0] \label{eqn:starradius}\end{aligned}$$ Note that we only define a starmask around stars brighter than 16$^{th}$ mag, as below this the ghosting appears to lie below the sky noise. Fig. \[fig:stars\] highlights the exclusion zones around stars of various magnitudes (as indicated by the yellow segments). One can see that by 16$^{th}$ magnitude there is no need for exclusion regions. Fig \[fig:tiles\] (centre panel) shows objects with [starmask]{}=1 in blue, highlighting the foreground coverage lost due to bright stars. To determine the reduction in area we create a grid of equally spaced points at $6''$ intervals and apply our starmask criteria. We sum the grid-points within our GAMA boundary for which [starmask]{}=0. This results in the areas as indicated in Table. \[tab:areas\]. ![The starmask exclusion radius (solid black line) was determined from eyeball measurements of randomly selected stars (blue data points) drawn from the four GAMA regions.[]{data-label="fig:starmask"}](ExclusionRadius.pdf){width="\columnwidth"} ------- --------------- ----------- ----------- ----------- ------------- -------------------------- --------------------------- GAMA RA range Dec range Full area Eff. area Masked area $\Delta RA$\[GAMA-GAIA\] $\Delta Dec$\[GAMA-GAIA\] field (deg) (deg) (sq. deg) (sq. deg) (sq. deg) (arcsec) (arcsec) G09 129.0 — 141.0 -2 — +3 59.97 54.93 4.91 0.056 0.066 G12 174.0 — 186.0 -3 — +2 59.97 57.44 2.39 0.134 0.106 G15 211.5 — 223.5 -2 — +3 59.97 56.93 2.90 0.101 0.098 G23 339.0 — 351.0 -35 — -30 50.58 48.24 2.28 -0.113 0.134 ------- --------------- ----------- ----------- ----------- ------------- -------------------------- --------------------------- ![Four panels showing our star mask regions, as indicated by the dotted lines. These lines are defined by Eqn. \[eqn:starradius\]. Objects within the star mask have their [starmask]{} flag set to 1 and are shown with yellow contours. Upper are two of our brightest stars with $g_{\rm GAIA DR2}$ mag = 6.6 )left) and 8.0 mag (right). The lower panels show two more typical regions centred on stars with $g_{\rm GAIA DR2}$ = $9^{th}$ mag (left) and $11^{th}$ mag (right) but also showing stars extending to our cutout limit of $g_{\rm GAIA DR2}=16^{th}$ mag[]{data-label="fig:stars"}](RGBx_222011083410393-min.png "fig:"){width="0.48\columnwidth"} ![Four panels showing our star mask regions, as indicated by the dotted lines. These lines are defined by Eqn. \[eqn:starradius\]. Objects within the star mask have their [starmask]{} flag set to 1 and are shown with yellow contours. Upper are two of our brightest stars with $g_{\rm GAIA DR2}$ mag = 6.6 )left) and 8.0 mag (right). The lower panels show two more typical regions centred on stars with $g_{\rm GAIA DR2}$ = $9^{th}$ mag (left) and $11^{th}$ mag (right) but also showing stars extending to our cutout limit of $g_{\rm GAIA DR2}=16^{th}$ mag[]{data-label="fig:stars"}](RGBx_184990972703151-min.png "fig:"){width="0.48\columnwidth"} ![Four panels showing our star mask regions, as indicated by the dotted lines. These lines are defined by Eqn. \[eqn:starradius\]. Objects within the star mask have their [starmask]{} flag set to 1 and are shown with yellow contours. Upper are two of our brightest stars with $g_{\rm GAIA DR2}$ mag = 6.6 )left) and 8.0 mag (right). The lower panels show two more typical regions centred on stars with $g_{\rm GAIA DR2}$ = $9^{th}$ mag (left) and $11^{th}$ mag (right) but also showing stars extending to our cutout limit of $g_{\rm GAIA DR2}=16^{th}$ mag[]{data-label="fig:stars"}](RGBx_178981141211076-min.png "fig:"){width="0.48\columnwidth"} ![Four panels showing our star mask regions, as indicated by the dotted lines. These lines are defined by Eqn. \[eqn:starradius\]. Objects within the star mask have their [starmask]{} flag set to 1 and are shown with yellow contours. Upper are two of our brightest stars with $g_{\rm GAIA DR2}$ mag = 6.6 )left) and 8.0 mag (right). The lower panels show two more typical regions centred on stars with $g_{\rm GAIA DR2}$ = $9^{th}$ mag (left) and $11^{th}$ mag (right) but also showing stars extending to our cutout limit of $g_{\rm GAIA DR2}=16^{th}$ mag[]{data-label="fig:stars"}](RGBx_138011112308615-min.png "fig:"){width="0.48\columnwidth"} Object classification {#sec:objectClassification} --------------------- ### Star, galaxy, and ambiguous classification Star-galaxy classification is performed in an initial phase using measured parameters, and then classifications are later overridden in a series of steps given prior knowledge (e.g., a known redshift, or direct eyeball classification). In the first round all objects are assigned a [class]{} flag that is initially set to `ambiguous`, and therefore those objects not reclassified in the latter stages will retain an `ambiguous` flag. We initially plot $(J-K_s)$ v $r_t$ and $\log{R_{50}}$ v $r_t$ where $J$ and $K_s$ are the extinction-corrected colour measurements, $r_t$ is the extinction-corrected total $r$-band magnitude, and $R_{50}$ is the effective half-light radius of the dilated segment. The latter is determined from the number of pixels within the segment. We then draw two lines on each plot to define the galaxy regions, stellar regions, and the ambiguous regions (see the solid lines on Fig. \[fig:stargal\] that divide the data into three regions). If an object is the same class in both planes, then this class is adopted. If it is `ambiguous` in only one plane, then it gains the `galaxy`/`star` class, and if it is a `star` in one plane and a `galaxy` in the other, then it gains the `ambiguous` class. The equations used to separate the parameter spaces in $(J-K_s)$ v $r_t$ space are given by: $$\begin{aligned} \begin{split} (J-K_s) = 0.025, & \qquad\text{if } r_t < 19.5\\ (J-K_s) = 0.025+0.025(r_t-19.5), & \qquad\text{if }r_t > 19.5\\ (J-K_s) = 0.025-0.1(r_t-19.5)^2, & \qquad\text{if }r_t > 19.5, \end{split}\end{aligned}$$ and in $\log(R_{50})$ v $r_t$ space: $$\begin{aligned} \begin{split} \log(R_{50}) = \Gamma+0.05-0.075(r_t-20.5), & \quad\text{any }r_t\\ \log(R_{50}) = \Gamma+0.05, & \quad\text{if }r_t > 20.5, \end{split}\end{aligned}$$ where $\Gamma$ is the median <span style="font-variant:small-caps;">log10seeing</span> value. Finally, based on the match to the GAMA redshift catalogue, we reassign any object with a confidently (<span style="font-variant:small-caps;">NQ</span> $> 2$) measured redshift above 0.002 to have a [class]{} flag of `galaxy`, and any object with a quality measured redshift of $-0.002 < z < 0.002$ to have a [class]{} flag set to `star`. The process hence starts with ambiguity and refines the classifications through a staged process using colour and size, then redshifts. Fig. \[fig:stargal\] shows the detected objects in the G23 region, coloured by their final [class]{} flag as indicated. Note that ambiguous objects are by definition those which reside in *both* the ill-defined regions, unless a redshift is known or the object has been visually inspected. ![Star-galaxy separation for the G23 field showing ($J-K_s$) colour versus magnitude (upper), and the measure half-light radius versus magnitude (lower) showing the stars, galaxies and ambiguous objects (in blue, red and green respectively). The solid lines denote the various cuts imposed (see text for full explanation), and the dashed lines show how these vary with seeing, with the 0.05-0.95 quantile range of the <span style="font-variant:small-caps;">log10seeing</span> value shown. []{data-label="fig:stargal"}](StarGalSep_G23.png){width="\columnwidth"} ### Cleaning spurious detections As with any dataset the VST and VISTA imaging contains a variety of spurious detection issues, with origins varying from diffraction spikes and offset ghosts, to baffling issues, noisy stacks, stack edge effects, transient objects (including Mars in one frame), and satellite trails etc. To identify artefacts we use a series of cuts to highlight objects in improbable parameter space. These have been arrived at through fairly extensive testing and visual checking and in particular viewing bright objects for which no match exists in the GAMA catalogue. From this process we arrive at a series of diagnostic flags and cuts as indicated in Table \[tab:artefactDiagnostics\]. Note that the table is progressed from top to bottom allowing overrides, hence we move from less certain to more certain classification markers. Fig. \[fig:crudplot\] shows a single sq degree for a problematic region indicating some of the issues: bright star ghosting, baffling issues, missing data, offset ghosting and frame edge effects. Overlain are objects with <span style="font-variant:small-caps;">starmask</span> flag set (orange), class set to artefact (cyan), and some remaining objects with bright fluxes not previously detected in the SDSS GAMA catalogue (purple). About half of the objects in this latter category represent new objects. An example of the object classifications is shown in Fig. \[fig:example\], where the segment of each object in this field is coloured by the corresponding classification, as indicated in the caption. The process is never going to be perfect but we believe the unclassified artefact rate is now well below 1 per cent of the galaxy population. We also note that no objects are removed and hence alternative or additional cleaning can be applied as we improve our understanding of the data. ![image](Classification.png){width="\textwidth"} ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Flag default setting criteria reason ----------------------------- ----------- ---------- -------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------- class ambiguous galaxy GAMA match with $Z > 0.002 \& NQ > 2$ Known GAMA galaxy class ambiguous star GAMA match with $-0.002 Z < 0.002 \& NQ > 2 $ Known GAMA star class ambiguous artefact $m_{\rm rt} - m_{\rm Zt}) < -0.75$ Improbable $(r-Z)$ colour class ambiguous artefact $log10(R50) < -0.4$ objects size is smaller than one pixel class ambiguous artefact No detection in two of $gri$ bands but optical data exists Only detecetd in 1 optical band class ambiguous artefact No detection in three of $ZYJHK_s$ bands but near-IR data exists Only detected in 1 or 2 near-IR bands class ambiguous artefact $\sigma_{\rm RMS} > median(\sigma_{\rm RMS})-5*{\rm st.dev.}(\sigma_{\rm RMS}$ Elevated skynoise class ambiguous star [starssize+starscol]{} $> 3.5 $ star class ambiguous galaxy [starssize+starscol]{} $< 0.5$ galaxy starmask 0 1 $r[\mbox{arcmin}]=10^{(1.6-0.15g)} object lies near a bright GAIA DR2 star \mbox{ and } [r < 5.0', g < 18.0]$ duplicate 0 1 Does not appear in all overlapping regions Any object that hasn’t been detected in all overlapping square degrees is spurious. duplicate 0 1 Duplicate identified by coordinate match with either `cen` or `max` coordinates, flag is assigned to the object with less flux Segment is incomplete, either due to frame edge, or insufficient segment rebuilding. mask 0 1 RA, Dec Inside the GAMA footprint Z\_confusionFlag 0 1 $\Delta z$ of multiple GAMA target matches $> 0.1$ Segment contains flux from multiple GAMA objects at different redshifts. \[tab:artefactDiagnostics\] ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Problematic sq degree region showing many of the issues we face. Overlain on the image are objects within starmask regions (orange) and objects labelled as artefacts (pale blue). Other objects that are bright but have no known GAMA counterpart are shown in purple. Finally, known GAMA galaxies are shown in green.[]{data-label="fig:crudplot"}](Crud1_lowres.png){width="\columnwidth"} Merging the [ProFound]{} catalogues {#sec:catMerging} ----------------------------------- Having generated 280 independent sq degree tiles across the four GAMA fields, we now need to combine these into four catalogues (one for each GAMA region), by merging while removing duplicates in the overlap regions. Fig. \[fig:tiles\] highlights the issue by showing the concatenated [tile]{} catalogues without consideration of duplicates. The over-density of the overlap regions manifests as a regular tartan-like pattern representing the double detection of objects in these regions. A simple coordinate match in the overlapping regions is insufficient to identify duplicates, as there are a number of reasons why the two sets of coordinates may not be identical: - An object in the overlapping region may have been visually regrouped in two slightly different ways. If so, some segments may exist in one field, but not the other. - Very noisy segments, due to slightly different <span style="font-variant:small-caps;">ProFound</span> sky solutions, may have different boundaries and hence different coordinates. - Segments that include multiple sources (which can occur if the saddle point between sources falls above the <span style="font-variant:small-caps;">cliptol</span> threshhold) may also have different <span style="font-variant:small-caps;">RAmax</span> and <span style="font-variant:small-caps;">Decmax</span> values[^25] if the sky solutions differ slightly. In all three of the above scenarios, a simple coordinate match will not identify “lone" objects that have not been exactly duplicated, and these objects will be counted twice. The following measures have been taken to account for these scenarios: - A check for duplicate objects is conducted exclusively in the overlapping regions. In these regions, any object that is *not* duplicated is immediately assumed to be spurious, and is assigned [duplicate]{}=1. In the scenario of differently merged objects, this ensures that only the “main" segment of the galaxy will be considered. Additionally, all noisy segments without stable coordinates will be given [duplicate]{}=1. - For each successfully duplicated object, priority is given to the duplicate with more flux. In the scenario of differently merged objects, this ensures that only the most aggressively regrouped version of an object will have [duplicate]{}=0. Additionally, if objects have been broken up due to proximity with the edge of the tile, then the version of the object that is most complete will be prioritised. - In the overlapping regions, a duplication is checked for with both the <span style="font-variant:small-caps;">RAmax</span>/<span style="font-variant:small-caps;">Decmax</span> coordinates, and also the <span style="font-variant:small-caps;">RAcen</span>/<span style="font-variant:small-caps;">Deccen</span> coordinates. If an object is flagged as being a duplicate in either one, then it will be assigned a value of [duplicate]{}=1. This has been done because even if a stable segment has differing <span style="font-variant:small-caps;">RAmax</span>/<span style="font-variant:small-caps;">Decmax</span> values due to a slightly different sky solution, the <span style="font-variant:small-caps;">RAcen</span>/<span style="font-variant:small-caps;">Deccen</span> values tend to be the same. As such, segments whose brightest pixel position fluctuate due to different sky solutions are adequately accounted for. Hence, [duplicate]{}=0 will produce a catalogue of unique objects, whereas [duplicate]{}=1 will produce a catalogue of redundant objects (see the blue-grey bands in the centre panel of Fig \[fig:tiles\]). Note than no objects are removed from the catalogue at any stage, instead we introduce flags to enable extraction of well-defined samples. ![The progression of the data flow for the G23 field. (top) [ProFound]{} is used to process each sq degree [tile]{} plus overlap to generate the star plus galaxy source catalogue shown (green points). (centre) the GAMA boundary is applied (red) and galaxies in the overlap region identified (grey), the star-mask is then applied (blue) and regions with missing photometry identified (cyan). (lower) the final galaxy catalogue (blue) showing only those galaxies in high quality regions away from bright stars or missing data.[]{data-label="fig:tiles"}](DataFlow_G23.png){width="47.50000%"} Extracting photometry in the mid- and far-IR {#sec:FIR} ============================================ Objects in the MIR-FIR are unresolved, unlike the FUV-MIR bands where objects are either fully or partially resolved. In addition, some of the brightest far-IR sources may have no or minimal optical fluxes, and vice-versa. Finally, the depth of the FIR imaging is also lower than optical images (as shown in Fig. \[fig:g23depth\]). We therefore run <span style="font-variant:small-caps;">ProFound</span> in a manner that utilises a different measurement technique in the MIR-FIR. To account for the above differences between the MIR-FIR and FUV-MIR, the FIR fluxes of optically-identified objects are iterated over by applying expectation maximisation (EM), as we describe in the following section. The philosophy of this measurement approach is to model the FIR flux of each optically-detected object, and then iterate over the flux of the object in each band, ensuring that all FIR flux is accounted for. In each band separately, the locations of optically-selected objects provide the coordinates at which to fit for objects. For the W3-W4 bands, the input objects are selected to be all objects from the optical catalogue with $m_r<20.5$ mag, excluding those objects with an <span style="font-variant:small-caps;">uberclass</span>$=$`artefact` flag. Because stars are still bright at these wavelengths, we make sure to model their flux contribution. In the $P100$-$S500$ bands, however, we also remove all objects with <span style="font-variant:small-caps;">uberclass</span>$=$`star` or `ambiguous` flag, as these objects are not expected to emit FIR flux. A magnitude guess for each object is determined by running an initial round <span style="font-variant:small-caps;">ProFound</span> at the input coordinates. Based on this initial magnitude guess, <span style="font-variant:small-caps;">ProFit</span>[^26] [@Robotham17] is applied to construct a model of each object given the PSF of each band and the magnitude guess of each object, to create a model of the FIR-emitting sources within the tile. This <span style="font-variant:small-caps;">ProFit</span> model is then subtracted from the image to produce a residual image. Iteratively, the model fluxes are modified using expectation maximisation in order to minimise the flux remaining in the residual image. After these initial iterations, any additional FIR-emitting sources that are not present in the original optical catalogue will be apparent. These are now identified using <span style="font-variant:small-caps;">ProFound</span> in source detection mode, and are included from this point onwards in the analysis as “additional sources". The inclusion of these additional sources mainly serves the purpose of accurate flux determination, to ensure that no background FIR flux is attributed to a foreground catalogue object[^27]. Once a catalogue of both optically-selected and additional sources has been determined, <span style="font-variant:small-caps;">ProFound</span> is run over the full image to determine the magnitudes of all objects, again in an iterative sense. Some of these objects will be marginalised out by <span style="font-variant:small-caps;">ProFound</span>, producing very low flux values, with correspondingly large uncertainties. This process is conducted using the `profoundFitMagPSF` command, which we show for completeness in the Appendix. To account for potential oversubtraction of the sky in the data reduction phase of the FIR imaging (possible due to the confused nature of sky pixels), we do an explicit sky subtraction in each band in a second phase. During the detection phase of additional sources, as described above, <span style="font-variant:small-caps;">ProFound</span> makes a measurement of the sky. In this second phase, we rerun the command as described above, in which the input image has undergone an explicit sky subtraction. In total over the four GAMA fields, photometry was measured for 822,326 objects with an <span style="font-variant:small-caps;">uberclass</span>=`galaxy` flag with $m_r<20.5$ mag. The fraction of these objects with detections in each band (defining a detection as having a measured magnitude $<30$) varied in each band, depending on the depth of the imaging. In the $W3$/$W4$/$P100$/$P160$/$S250$/$S350$/$S500$ bands, 62/42/38/43/50/42/32 per cent of objects had detections, respectively. 38,413 objects had detections in all seven bands, corresponding to just under 5 per cent of the $m_r<20.5$ mag galaxy sample. The resulting uncertainties, due to the EM mixture modelling process, accurately reflect the inherent uncertainties in this process. If two optically-detected sources are close in projection, then the uncertainty will reflect the potential confusion between these two sources. We show in Fig. \[fig:FIRphotometryresidual\] an example of the original image, final modelled image, and resulting residual of a specific square degree. Note that in this example, the main features in the residual image originate from resolved galaxies that are not well-modelled by the PSF. In addition to the objects detected in the FIR with optical counterparts, $4.5/43.4/0.3/0.4/23.6/19.6/4.9$ per cent of objects in the $W3$/$W4$/$P100$/$P160$/$S250$/$S350$/$S500$ bands were identified as “additional sources". Generally this percentage is reflective of the depth of the imaging. We note that the large number of additional objects detected in the $W4$ band indicates that the <span style="font-variant:small-caps;">ProFound</span> parameter used to identify these objects was likely too aggressive in this band. ![Example of the photometry extraction in the SPIRE 250 band for a single tile in the G09 region. The top panel shows the original image, and the middle panel shows the <span style="font-variant:small-caps;">ProFit</span>-produced model of the objects. The residual when subtracting the model off the image is shown in the bottom panel. []{data-label="fig:FIRphotometryresidual"}](FIR.png){width="\columnwidth"} The above process is conducted independently for each band in each individual tile. In total, it requires less than one hour to extract the W3-S500 photometry for a single square degree. The outputs for the additional sources are saved, however are not included in our final catalogue. An analysis of these sources is beyond the scope of this work. Verification and validation {#sec:verification} =========================== Comparison with previous photometry ----------------------------------- ### Comparison to SDSS/2MASS Fig. \[fig:photom\] shows a comparison of the [ProFound]{} measured photometry of stars (in selected magnitude ranges) to either SDSS DR13 ($ugriz$) or 2MASS ($JHK_S$). For the comparison to SDSS in the $u$ and $r$ bands we use the filter conversions as outlined by @Kuijken19. For the comparison to 2MASS, we include the appropriate corrections for the 2MASS filter shapes, colour terms, and AB to Vega corrections given in @Gonzalez-Fernandez18. The effect of saturation in the VST $gri$ and VISTA $Z$ bands is evident at the bright end, where <span style="font-variant:small-caps;">ProFound</span> recovers less flux for the bright stars than was measured by SDSS. Note that the SDSS data are *not* corrected for the known $u$ and $z$ offsets from the SDSS system to the AB system. In all bands the photometry agrees well with scatter about the equality line increasing slightly towards fainter fluxes. In all cases the offsets are $<0.06$ mag with the largest offset seen in the $K_S$ band. ![Comparison of the new [ProFound]{} photometry versus the SDSS/2MASS photometry, for the combined G09+G12+G15 data in the filters indicated for stars only. In each panel we show the data (grey dots for the full sample, and blue dots within the fitted magnitude range), the median offset (black points with 1$\sigma$ range), and the linear fit to the offset (orange line). The offset and the least squares fit to the median values are shown in the upper left of each panel.[]{data-label="fig:photom"}](PhotomComp.png){width="85mm"} ### Comparison to LAMBDAR We compare the colour distributions of the photometry presented in this paper against equivalent colour distributions produced by LAMBDAR [@Wright16] for the subset of objects with measured 20-band photometry, as shown in Fig. \[fig:ColourDist\]. The left panel of the plot features the relative colour distribution of the overlapping sample between LAMBDAR and the new photometry. Blue histograms refer to the new photometry presented in this paper, whereas mustard histograms refer to the LAMBDAR photometry. For each histogram, we show the 0.1-0.9 quartile range with a horizontal line, where the median value is shown. Note that the colour values have been shifted so that the peak in the distribution for the new photometry is at 0. The right panel of the plot shows the number of outliers in each sample for each band, where outliers are determined to be objects that have colours more than 0.5 mag outside the 0.1-0.9 quartile range. The new colour distributions are improved in the UV, optical and also (marginally) the FIR bands, however we note that the colour distributions in the NIR bands are better in LAMBDAR. This is likely the result of the generous segment dilation we have implemented in order to catch halo flux, adding more sky noise than in LAMBDAR. ![Colour distributions of all adjacent bands for the photometry presented in this paper (blue), as compared with LAMBDAR [@Wright16] (orange) for all matching galaxies. Individual colour distributions are shown in the left-hand panel, relative to the peak of the <span style="font-variant:small-caps;">ProFound</span> colour distribution. We show the 0.1-0.9 quantile ranges for each sample, as well as the median value. The right-hand panel shows the outlier percentage for each sample, quantified as the percentage of points that lie further than 0.5 mag away from the 0.1-0.9 quartile range. []{data-label="fig:ColourDist"}](ColourDist_summary_vertical.pdf){width="\columnwidth"} Seeing variations across tiles ------------------------------ By analysing the <span style="font-variant:small-caps;">R50</span> values of stars across the four GAMA fields, we are able to determine the extent to which the seeing varies across the tiles in each of the optical/NIR bands. The median <span style="font-variant:small-caps;">R50</span> value varies between $\sim$ 0.4 – 1 arcseconds over the four fields, in each of the different photometric bands. Due to the nature of the segment dilation within <span style="font-variant:small-caps;">ProFound</span>, the total flux within each object is accounted for despite potential variations in the PSF. To check that these PSF variations are not having an impact on the derived fluxes for galaxies in the sample, we assess how the minimum <span style="font-variant:small-caps;">R100</span> value for galaxies compares to the PSF in each tile. For galaxies down to an $r$-band magnitude of 23, we find that the smallest galaxies are on average larger than the PSF by factors of 2 – 4. This confirms that galaxy segments are consistently larger than the PSF, meaning that seeing variations across the tiles are not affecting the measured fluxes of galaxies in our photometry. Final astrometric accuracy -------------------------- To verify the astrometric accuracy we identify objects selected as stars in the range $ 14 < m_r < 16$ and match against GAIA DR2 [@Gaia18; @Lindegren18] taking the best match within 3 arcsec. Fig. \[fig:astrom\] shows the RA and Dec offset for each field, with the medians indicated (orange crosses) and the circles enclosing 1, 2, and 3$\sigma$. In all cases the 3$\sigma$ astrometry error is within $\pm0.75$ arcsec but some modest offset is seen. We therefore add two extra columns [RAGAIA]{} and [DecGAIA]{} where we correct the VST KiDS astrometry to the GAIA astrometric frame by implementing a simple RA and Dec offset appropriate to each region. The offsets are as shown in Table. \[tab:areas\] (columns 7 and 8). ![Astrometric accuracy of the native VST KiDS data against GAIA DR2 for each of the GAMA regions as indicated.[]{data-label="fig:astrom"}](AstrometricAccuracy.png){width="\columnwidth"} Star and galaxy counts in each region {#sec:counts} ------------------------------------- As an initial diagnostic we construct the galaxy number-counts for each region and compare to literature data [@Driver18] in the $r$ band (see Fig. \[fig:ncounts\]). Within each panel we show the galaxy counts (red) and the star-counts (blue), where the shaded regions indicate the range covered when ambiguous objects are included. Note that the ambiguous objects are not explicitly included in either main line. Literature counts are shown in grey and the predicted star-counts from the <span style="font-variant:small-caps;">TRILEGAL</span> v1.6 model [see @Girardi12], are shown with an orange dotted line. In all four regions the recovered $r$ band counts show broad agreement with the literature data. The galaxy number counts are slightly below the literature counts at the bright end in both the G09 and G23 fields. We note that the <span style="font-variant:small-caps;">TRILEGAL</span> models also agree well with our star-count data and suggest that the majority of the ambiguous detections are likely stars (as the upper bound of the star counts most closely matches the <span style="font-variant:small-caps;">TRILEGAL</span> prediction), and as also suggested by Fig. \[fig:stargal\]. More details will be discussed in Koushan et al. (in prep). ![image](Counts.pdf){width="\textwidth"} Revised panchromatic depth -------------------------- We show in Fig. \[fig:g23depth\] the revised depth of our imaging and catalogues in each of the 20 photometric bands. In addition to showing the stated 5$\sigma$ depth limits for each of the imaging bands in black, we make two separate measurements of the depth of our data. The first utilises the number counts of each band, as described in the above Section \[sec:counts\], to estimate the depth in each band as the magnitude at which the galaxy counts turn over (calculated as the magnitude at which the counts are lower than the counts in the previous magnitude bin). This measurement is shown in Fig. \[fig:g23depth\] as the dotted blue line. We note that this measurement is very close to the stated depths, except in the FUV, $g$, $r$ and $i$ bands where our depth measurement is shallower, and the Herschel bands in which our measurement is deeper. As an alternative way of indicating the image depth, we plot the surface brightness limits in each band as determined by the <span style="font-variant:small-caps;">ProFound</span>-measured skyRMS values in each band. These limits are shown in Fig. \[fig:g23depth\] as the yellow shaded region. We compare these revised depths with a <span style="font-variant:small-caps;">ProSpect</span>-derived SED for a $M_*=10^{10.5}\rm{M}_{\odot}$ galaxy with a constant SFH at redshifts of $z=$ 0.1, 0.3, and 0.5. This comparison highlights that, with our sensitivity, but $z=0.3$ we are no longer sensitive to a galaxy of this stellar mass in the FIR. ![image](G23depth.pdf){width="100.00000%"} Impact on existing GAMA studies ------------------------------- The main impact that this new catalogue will have on existing studies comes from the new segmentation maps. As discussed in Section \[sec:GAMAmatch\], not all objects in previous GAMA studies will be listed in this new catalogue, as they have been merged with neighbouring GAMA objects, or because the target was based on an artefact in SDSS imaging. This is the case for $\sim$ 4 per cent of GAMA targets. As a test of the impact of this on existing GAMA studies, we assess the number of objects in existing data products that no longer appear in the new catalogue. In the catalogue of visual morphologues (DMU *VisualMorphologyv3* @Kelvin14, as used in, for example, @Kelvin14b [@Moffett16; @Lange16; @Moffett16b; @Alpaslan15]), 0.81% of objects are not in the new catalogue. For the group catalogue (DMU *G3Galv10* @Robotham11, as used in for example @Alpaslan12 [@LaraLopez13; @Alpaslan14; @Robotham14; @Davies15; @Deeley17]), this number is only 0.26%. For the SED-fitting catalogue using <span style="font-variant:small-caps;">MagPhys</span> (DMU *MagPhysv06* @Driver16, as used in, for example, @Davies17 [@Mahajan18; @Driver18]) only 0.25% of objects are missing, and for the catalogue of Sérsic indices (DMU *SersicCatSDSSv09* @Kelvin12, as used in, for example, @Kelvin14 [@Deeley17; @Bremer18]) 0.87% of objects are missing. This corresponds to an absolute number of 314/470/487/1935 galaxies respectively. We highlight that this number is very small, and hence do not expect that this will have any impact on existing GAMA studies. The <span style="font-variant:small-caps;">GAMAKidsVikingCATAIDMatchv01</span> catalogue will aid in the bookkeeping of any such circumstances, by identifying for each GAMA target the corresponding <span style="font-variant:small-caps;">uberID</span> in the new catalogue. Access to catalogues {#sec:catAccess} ==================== The catalogues are available to any member of the public via a collaboration request[^28]. Included in the release are the Data Management Units (DMUs) as described in Table \[tab:DMUs\]. DMU Description ------------------------------ ------------------------------------------------------------------------------------------------- GAMAKidsVikingv01 <span style="font-variant:small-caps;">ProFound</span> FUV-NIR analysis described here. GAMAKidsVikingCATAIDMatchv01 CATAID match of legacy catalogues to new catalogue. GAMAKidsVikingFIRv01 <span style="font-variant:small-caps;">ProFound</span> FIR PSF-limited analysis described here. GAIADR2GAMAstarmaskv01 List of GAIA DR2 stars that define the starmask. \[tab:DMUs\] Each of these include DMU descriptions in the GAMA Schema Browser and tools to access the data via either the Single Object Viewer [@Liske15] or the Panchromatic SWARP Imager [@Driver16]. Note that the spectroscopic component is described in full in [@Liske15]. Note that version numbers may change as products are updated however older versions are available via the Schema Browser. Usage of DMUs ------------- In the analysis described within this paper, we have kept all measurements and introduced a series of flags (see Table. \[tab:flags\]) to indicate various issues. Hence, the extraction of a catalogue suitable for science requires the use of these flags. For example, to extract all galaxies in the GAMA G09 region outside the starmask with robust spectroscopic redshifts from the <span style="font-variant:small-caps;">KidsVikingGAMAv01</span> catalogue one must execute the following: `uberclass=galaxy & duplicate=0 & mask=0 & starmask=0 & region=G09 & NQ > 2` This results in 59930 galaxies covering 54.93 sq degress of sky (see Table \[tab:areas\]). The FIR catalogue (<span style="font-variant:small-caps;">KidsVikingGAMAFIRv01</span>) only includes a subset of the objects in <span style="font-variant:small-caps;">KidsVikingGAMAv01</span>, as per the description in Section \[sec:FIR\]. In all other respects, however, the structure of the catalogue is the same. The same set of flags to isolate catalogues has been implemented. Conducting the above command on this catalogue hence results in 58967 galaxies covering 54.93 sq degress of sky (see Table \[tab:areas\]). Summary {#sec:summary} ======= We present in this paper updated photometry of four GAMA fields in 20 wavelength bands, using deeper imaging from KiDS/VIKING and the code <span style="font-variant:small-caps;">ProFound</span>. <span style="font-variant:small-caps;">ProFound</span> has been separately run in first a resolved mode to extract photometry from first the $FUV$-$W2$ bands, and then in an unresolved to extract photometric measurements from the $W3$-$S500$ bands. In the resolved mode, sources are detected, fragmented sources are visually rebuilt, and then the corresponding fluxes in all bands are extracted, before correcting these values for dust exinction, defining star masks, and classifying the individual objects into classes of `star`, `galaxy`, or `ambiguous`. Spurious objects in the catalogue are assigned the class of `artefact`. The photometry in the unresolved regime is conducted using a subset of the resolved catalogue as an input, where fluxes are extracted in a Bayesian manner for objects that are identified as galaxies with $m_r < 20.5$. As part of this process, we identify additional sources that contribute flux in each FIR band (most probably background sources), however we leave an analysis of these objects to future work. As a verification of the new photometry, we have checked the astrometry of stars in our catalogue against their coordinates in GAIA, to identify our astrometric accuracy. A comparison of the galaxy and star number counts to literature data from @Driver16 show that our galaxy number counts are consistent with the literature in all fields, with only a slightly smaller count number at the bright end in the G09 and G23 fields. We identify that the colour distributions in our new photometry, as compared with the previous LAMBDAR photometry [@Wright16], are significantly tighter in the UV and optical bands, and also marginally better in the FIR bands. In future work (Bellstedt et al. in prep.), this photometry will be used to conduct SED-fitting with the newly-developed code <span style="font-variant:small-caps;">ProSpect</span> [@Robotham20] to measure stellar masses, star formation rates and star formation histories for individual galaxies. Acknowledgments {#acknowledgments .unnumbered} =============== We thank the anonymous referee for their careful reading of this work, and whose comments improved the paper. SPD and SB acknowledge support from the Australian Research Council under Discovery Project Discovery 180103740. GAMA is a joint European-Australasian project based around a spectroscopic campaign using the Anglo-Australian Telescope. The GAMA input catalogue is based on data taken from the Sloan Digital Sky Survey and the UKIRT Infrared Deep Sky Survey. Complementary imaging of the GAMA regions is being obtained by a number of independent survey programmes including GALEX MIS, VST KiDS, VISTA VIKING, WISE, Herschel-ATLAS, GMRT and ASKAP providing UV to radio coverage. GAMA is funded by the STFC (UK), the ARC (Australia), the AAO, and the participating institutions. The GAMA website is <http://www.gama-survey.org/>. This work is based on based on observations made with ESO Telescopes at the La Silla Paranal Observatory under programme IDs 177.A-3016, 177.A-3017, 177.A-3018 and 179.A-2004, and on data products produced by the KiDS consortium. The KiDS production team acknowledges support from: Deutsche Forschungsgemeinschaft, ERC, NOVA and NWO-M grants; Target; the University of Padova, and the University Federico II (Naples). This work was supported by resources provided by the Pawsey Supercomputing Centre with funding from the Australian Government and the Government of Western Australia. We have used <span style="font-variant:small-caps;">R</span> [@R] and <span style="font-variant:small-caps;">python</span> for our data analysis, and acknowledge the use of <span style="font-variant:small-caps;">Matplotlib</span> [@Hunter07] for the generation of plots in this paper. ProFound commands {#app:commands} ================= Here we show the command used to run <span style="font-variant:small-caps;">ProFound</span> in the multiband mode for the detection phase: `profoundMultiBand( ` `dir=`input directory, `skycut=2.0,` `pixcut=13, ` `ext=1,` `tolerance=15,` `reltol=-10,` `cliptol=100,` `detectbands=c(r,Z),` `multibands=c(r,Z), ` `keepsegims=TRUE,` `magzero=c(0,30),` `dotot=FALSE, ` `docol=FALSE, ` `dogrp=FALSE,` `verbose=TRUE,` `boxiters=4,` `grid=c(50,50),` `roughpedestal=TRUE,` `stats=FALSE,` `groupstats=TRUE,` `mask=0,` `app_diam=1.4,` `fluxtype=Jansky` `)` We also show the multiband call that was used to run the measurement phase after the regrouping of segments: `profoundMultiBand(` `segim=fixed_segim,` `dir=`input directory, `iters_det=6,` `iters_tot=c(3,3,2,2,2,2,2,2,2,2,2,3,3),` `totappend=t,` `sizes_tot=c(15,15,5,5,5,5,5,5,5,5,5,15,15),` `colappend=c,` `detectbands=c(r,Z),` `multibands=c(FUV,NUV,u,g,r,i,Z,` `Y,J,H,K,W1,W2),` `keepsegims=TRUE,` `magzero=c(18.82,20.08,0,0,0,0,30,30,30,30,30,` `23.16,22.82),` `dotot=TRUE,` `docol=TRUE,` `dogrp=TRUE,` `verbose=TRUE,` `box=c(200,200,100,100,100,100,100,100,100,100,` `100,200,200),` `boxiters=4,` `boxadd=c(50,50,50,50,50,50,50,50,50,50,50,50,` `50),` `grid=c(50,50,50,50,50,50,50,50,50,50,50,50,50),` `roughpedestal=TRUE,` `redosegim=FALSE,` `deblend=FALSE,` `groupstats=TRUE,` `mask=0,` `SBdilate=1.0,` `SBN100=100,` `app_diam=1.4, ` `fluxtype=Jansky` `)` The following commands presents the manner in which we have run proFound in the unresolved MIR-FIR regime, for each band separately: `profoundFitMagPSF(` `RAcen=`RA coordinates, `Deccen=`Dec coordinates, `image=`sky-subtracted image, `header=`image header, `psf=`band PSF, `magzero=`23.24/19.6/8.9/8.9/11.68/11.67/11.62[^29], `magdiff = 5, ` `fit_iters=5, ` `verbose = TRUE, ` `fluxtype=Jansky,` `doProFound=TRUE,` `findextra=TRUE, ` `itersub=TRUE,` `pixcut=3, ` `skycut=2, ` `ext=1, ` `redosky=FALSE,` `iters=4, ` `tolerance=0, ` `sigma=`2/2/2/2/0/0/0, `mask=0,` `psf_redosky=TRUE,` `boxiters=2` `)` \[lastpage\] [^1]: E-mail: [email protected] [^2]: Two Micron All-Sky Survey [^3]: Sloan Digital Sky Survey [^4]: UKIRT Infrared Deep Sky Survey [^5]: Infrared Astronomical Satellite [^6]: ROentgen SATellite [^7]: VLT Survey Telescope [^8]: Visible and Infrared Survey Telescope for Astronomy [^9]: Large Synoptic Survey Telescope [^10]: Galaxy Evolution Explorer [^11]: Wide-field Infrared Survey Explorer [^12]: Wide Field Infrared Survey Telescope [^13]: Kilo Degree Survey [^14]: Evolutionary Map of the Universe [^15]: Deep Investigation of Neutral Gas Origins [^16]: First Large Absorption Survey in H<span style="font-variant:small-caps;">i</span> [^17]: Formerly referred to as the Large Survey Synoptic Telescope (LSST) [^18]: <http://kids.strw.leidenuniv.nl/> [^19]: <https://datacentral.org.au/services/cutout/> [^20]: <http://datacentral.org.au/> [^21]: <https://datacentral.org.au/services/cutout/> [^22]: Available on Github: <https://github.com/asgr/ProFound> [^23]: In the context of the photometry, *group* refers to a collection of segments directly abutting each other. [^24]: <span style="font-variant:small-caps;">HFI\_CompMap\_ThermalDustModel\_2048\_R1.20.fits</span>, <https://irsa.ipac.caltech.edu/data/Planck/release_1/all-sky-maps/previews/HFI_CompMap_ThermalDustModel_2048_R1.20/index.html> [^25]: For each object, <span style="font-variant:small-caps;">ProFound</span> computes two sets of coordinates. <span style="font-variant:small-caps;">RAmax/Decmax</span> values indicate the coordinate of the brightest pixel within a segment, whereas <span style="font-variant:small-caps;">RAcen/Deccen</span> values show the flux-weighted central coordinate of the segment. [^26]: Available on Github: <https://github.com/ICRAR/ProFit> [^27]: Note that if a background source exists within the PSF of the foreground object, that this flux will no longer be separated. Such a scenario was presented by Allison et al. (submitted), where the SED of the target foreground object was likely polluted in the FIR through the contribution of a high-$z$ background object. [^28]: <http://www.gama-survey.org/collaborate/> [^29]: Zero-points as stated in table 3 of @Driver16.
{ "pile_set_name": "ArXiv" }
--- abstract: 'This paper studies how to design abstractions of large-scale combinatorial optimization problems that can leverage existing state-of-the-art solvers in general purpose ways, and that are amenable to data-driven design. The goal is to arrive at new approaches that can reliably outperform existing solvers in wall-clock time. We focus on solving integer programs, and ground our approach in the large neighborhood search (LNS) paradigm, which iteratively chooses a subset of variables to optimize while leaving the remainder fixed. The appeal of LNS is that it can easily use any existing solver as a subroutine, and thus can inherit the benefits of carefully engineered heuristic approaches and their software implementations. We also show that one can learn a good neighborhood selector from training data. Through an extensive empirical validation, we demonstrate that our LNS framework can significantly outperform, in wall-clock time, compared to state-of-the-art commercial solvers such as Gurobi.' author: - | **Jialin Song**^^ **Ravi Lanka**^^ **Yisong Yue**^^ **Bistra Dilkina**^^\ ^^ California Institute of Technology\ ^^Jet Propulsion Laboratory, California Institute of Technology\ ^^ University of Southern California bibliography: - 'reference.bib' title: A General Large Neighborhood Search Framework for Solving Integer Programs ---
{ "pile_set_name": "ArXiv" }
--- abstract: 'We prove that every unital $\mathrm{C}^*$-algebra $A$ has the Mazur–Ulam property. Namely, every surjective isometry from the unit sphere $S_A$ of $A$ onto the unit sphere $S_Y$ of another normed space $Y$ extends to a real linear map. This extends the result of A. M. Peralta and F. J. Fernández-Polo who have proved the same under the additional assumption that both $A$ and $Y$ are von Neumann algebras. In the course of the proof, we strengthen Mankiewicz’s theorem and prove that every surjective isometry from a closed unit ball with enough extreme points onto an arbitrary convex subset of a normed space is necessarily affine.' address: - 'Graduate School of Mathematical Sciences, The University of Tokyo, Japan' - 'RIMS, Kyoto University, Japan' author: - Michiya Mori - Narutaka Ozawa title: 'Mankiewicz’s theorem and the Mazur–Ulam property for $\mathrm{C}^*$-algebras' --- Introduction ============ The celebrated Mazur–Ulam theorem ([@mu]) asserts that every surjective isometry between normed spaces $X$ and $Y$ is necessarily affine. This was extended by P. Mankiewicz ([@mankiewicz]) to surjective isometries between the closed unit balls $B_X$ and $B_Y$. Motivated by these results, D. Tingley ([@tingley]) posed the following problem in 1987: Does every surjective isometry $T\colon S_X \to S_Y$ between the unit spheres of normed spaces $X$ and $Y$ extend to a real linear isometry between $X$ and $Y$? Currently, no counterexample to Tingley’s problem is known. A Banach space $X$ is said to have the *Mazur–Ulam property* ([@cd]) if Tingley’s problem has an affirmative answer for an arbitrary target $Y$. The main result of the present paper is the following. \[thm:mau\] Every unital complex $\mathrm{C}^*$-algebra (as a real Banach space) and every real von Neumann algebra has the Mazur–Ulam property. Tingley’s problem between von Neumann algebras has been solved earlier in [@pfp] (see also [@mori; @tanaka2]). See [@peralta] and the references herein for more information about Tingley’s problem for Banach spaces related to operator algebras. The Mazur–Ulam property for commutative $\mathrm{C}^*$-algebras has been proved in [@ding; @fw; @liu; @pca]. For more examples of Banach spaces with the Mazur–Ulam property, see [@thl] for example. The starting point of the present and many other works on Tingley’s problem for operator algebras is R. Tanaka’s observation ([@tanaka2]) that a surjective isometry from the unit sphere $S_A$ of a $\mathrm{C}^*$-algebra $A$ onto another unit sphere $S_Y$ maps closed faces onto closed faces. The unit sphere of a $\mathrm{C}^*$-algebra has many faces that are approximable by closed unit balls of $\mathrm{C}^*$-algebras. To exploit this, we revisit Mankiewicz’s theorem. Let $K$ be a convex subset in a normed space $X$. The subset $K$ is called a *convex body* if it has non-empty interior in $X$. Mankiewicz ([@mankiewicz]) has proved that any isometry between convex bodies is necessarily affine. We say $K$ has the *strong Mankiewicz property* if every surjective isometry $T$ from $K$ onto an arbitrary convex subset $L$ in a normed space $Y$ is affine. Every convex subset of a strictly convex normed space has this property because it is uniquely geodesic (cf. Lemma 6.1 in [@bfgm]), but some convex subset of $L^1[0,1]$ does not ([@schechtman]), see Example \[example\]. Every normed space also has this property by Figiel’s theorem ([@figiel]). This probably suggests that the same is true for every convex body, but this is not clear to the authors (since the range $L$ is not assumed to have non-empty interior). \[thm:man\] Let $X$ be a Banach space such that the closed convex hull of the extreme points $\operatorname{ext}B_X$ of the closed unit ball $B_X$ has non-empty interior in $X$. Then, every convex body $K\subset X$ has the strong Mankiewicz property. \[cor:man\] Let $A$ be a unital complex $\mathrm{C}^*$-algebra or a real von Neumann algebra. Then, every convex body in $A$ has the strong Mankiewicz property. Acknowledgment {#acknowledgment .unnumbered} -------------- We are grateful to Professor G. Schechtman for allowing us to include his example from [@schechtman]. The first author is partially supported by Leading Graduate Course for Frontiers of Mathematical Sciences and Physics, MEXT, Japan. The second author is partially supported by JSPS KAKENHI 17K05277 and 15H05739. Notations and Remarks on real $\mathrm{C}^*$-algebras {#notations-and-remarks-on-real-mathrmc-algebras .unnumbered} ----------------------------------------------------- Throughout this paper, $X$ and $Y$ are real normed (Banach) spaces and $A$ is a $\mathrm{C}^*$-algebra which may be real or complex unless otherwise stated. The unit sphere and the closed unit ball of $X$ are denoted respectively by $S_X$ and $B_X$. For any projection $p$ in a unital $\mathrm{C}^*$-algebra, we write $p^\perp:=1-p$. By definition, a real $\mathrm{C}^*$-algebra $A$ is the real part $\{ a \in A_{{\mathbb C}} : {\mathcal J}(a)=a\}$ of a complex $\mathrm{C}^*$-algebra $A_{{\mathbb C}}$ with respect to a conjugate-linear $*$-automorphism ${\mathcal J}$ such that ${\mathcal J}^2=\operatorname{id}$. Many of the standard operations in complex $\mathrm{C}^*$-algebras work equally well for real $\mathrm{C}^*$-algebras. For example, the modulus $|a|$ of an element $a$ in $A$ is firstly considered in the complexification $A_{{\mathbb C}}$ and by uniqueness one sees that $|a|$ belongs to the real part $A$. For any complex continuous function $f$ and any normal element $a$ in $A_{{\mathbb C}}$, one has ${\mathcal J}(f(a))=\bar{f}({\mathcal J}(a^*))$. A projection $p$ in a real von Neumann algebra $A$ is minimal if and only if $pAp={\mathbb R},\,{\mathbb C},\,{\mathbb H}$, in which case $p$ has rank at most $2$ in the complexification $A_{{\mathbb C}}$. See [@li] for more on real operator algebras. On the strong Mankiewicz property ================================= If $B_X$ has the strong Mankiewicz property, then every convex body $K\subset X$ has the strong Mankiewicz property. Let a surjective isometry $T$ be given. The assumption implies that every interior point $x$ in $K$ has a neighborhood on which $T$ is affine. By continuation (see Proof of Theorem 2 in [@mankiewicz]), one sees that $T$ is affine everywhere. By the above lemma, it suffices to show every surjective isometry $T\colon B_X\to L\subset Y$ is affine. We may assume that $T(0)=0$, which implies that $\|T(x)\|=\|x\|$ for all $x\in B_X$. Let $a\in\operatorname{ext}B_X$. Since the line segment $[-a,a]$ is the unique geodesic path between $-a$ and $a$, the map $T$ is affine (linear) on $[-a,a]$. We claim that if $x\in X$ and $\lambda\in{\mathbb R}$ are such that $\|x\|\le\frac{1}{2}$ and $\|x\|+|\lambda|\le1$, then $$\|T(x+\lambda a) - (T(x)+\lambda T(a))\|\le 4\|x\| |\lambda|.$$ For this, we may assume that $\lambda\geq0$ as $T(-a)=-T(a)$. Since $T$ is affine on $[x,x+(1-\|x\|) a]$ by the similar reason as above, one has $$\begin{aligned} T(x+\lambda a) - T(x) = \frac{\lambda}{1-\|x\|}(T(x+(1-\|x\|) a)-T(x)) \approx_{4\|x\|\lambda } \lambda T(a).\end{aligned}$$ Let $T_n\colon n B_X\to nB_Y$ be the map defined by $T_n(x)=nT(\frac{1}{n}x)$. By the previous inequality, for any $a_k\in\operatorname{ext}B_X$ and $\lambda_k\in{\mathbb R}$ such that $C:=\sum_k|\lambda_k|$, one has $$\tag{$\ast$} \| T_n(\sum_k\lambda_k a_k) - \sum_k \lambda_k T(a_k) \| \le 4n(\sum_k\frac{|\lambda_k|}{n})^2 \le \frac{4C^2}{n}$$ for all $n\geq 2C$. We consider the Banach space $Z_\infty:=\ell_\infty({\mathbb N};Z)/c_0({\mathbb N};Z)$ for $Z=X$ or $Y$ and define $\hat{T}\colon X_\infty\to Y_\infty$ by $\hat{T}([x_n]_n)=[T_n(x_n)]_n$. Here $[x_n]_n$ denotes the element in $X_\infty$ represented by $(x_n)_n \in \ell_\infty({\mathbb N};X)$. Observe that $\hat{T}$ is a well-defined isometry which is moreover linear by $(\ast)$ and the assumption on $X$. We claim that $$\delta B_{\hat{T}(X_\infty)}=\hat{T}(\delta B_{X_\infty})\subset L_\infty:=\{ [y_n]_n \in Y_\infty: y_n\in L\}\subset \hat{T}(X_\infty).$$ Here $\delta>0$ is such that $\delta B_X$ is contained in the closed convex hull of $\operatorname{ext}B_X$. Thus the first inclusion follows from $(\ast)$ and the fact that $\sum_k \lambda_k T(a_k) \in L$ for every $a_k\in \operatorname{ext}B_X$ and $\lambda_k\geq0$ with $\sum_k\lambda_k=1$. The second follows from the fact that if $y:=[y_n]_n\in L_\infty$, then for $x_n := n T^{-1}(\frac{1}{n}y_n)$, one has $[y_n]_n=\hat{T}([x_n]_n) \in\hat{T}(X_\infty)$. This claim implies that $L$ has non-empty interior in its linear span. Indeed, if $y\in L$ and $\lambda\in{\mathbb R}$ are such that $\|\lambda y\|\le \delta$, then the constant sequence $y$ belongs to $L_\infty$ and so is $\lambda y$, which means that there is a sequence $(z_n)_n$ in $L$ such that $\| \lambda y - z_n\|\to0$ and so $\lambda y\in L$. Therefore, we can apply Mankiewicz’s theorem ([@mankiewicz]) to $T$ and conclude that $T$ is affine. It follows from the Russo–Dye theorem (Theorem I.8.4 in [@davidson] for the complex case and Theorem 7.2.4 in [@li] for the real case) that the closed unit ball of a unital complex $\mathrm{C}^*$-algebra, as well as a real von Neumann algebra, coincides with the closed convex hull of its extreme points (unitary elements). Note that the real Russo–Dye theorem in [@li] states that $B_A$ is contained in the closed convex hull of $\{ \cos(h)e^k : h=h^*,k=-k^*\}$ and it is easily seen that $\cos(h)$ is contained in the closed convex hull of unitary elements in the real von Neumann algebra generated by $h$. The following beautiful example without the strong Mankiewicz property is provided for us by G. Schechtman ([@schechtman]). \[example\] Consider the set $K_0$ of all continuous strictly increasing functions $f$ from $[0,1]$ onto $[0,1]$ and put $K:=\overline{K_0}\subset L^1[0,1]$. Then, $K$ is compact and in particular it has empty interior. Let $T_0\colon K_0\to K_0$ be the map that sends $f$ to its inverse. Since the $L^1$ distance between two functions is the area enclosed by their graphs, $T$ is an isometry by the Fubini theorem. The continuous extension $T\colon K\to K$ of $T_0$ is a surjective isometry which is not affine. Hence, the compact convex subset $K\subset L^1[0,1]$ does not have the strong Mankiewicz property. Proof of Theorem \[thm:mau\] for ${\mathbb B}({\mathcal H})$ ============================================================ Since the proof of Theorem \[thm:mau\] for the type I factor ${\mathbb B}({\mathcal H})$, $\dim{\mathcal H}>2$, is much simpler than the general $\mathrm{C}^*$-case, we give it here as an appetizer. \[lem:cond\] Let $T\colon S_X\to S_Y$ be a surjective isometry. Assume that there are $\{{\varphi}_i\}_i\subset B_{X^*}$ and $\{\psi_i\}_i\subset B_{Y^*}$ such that ${\varphi}_i=\psi_i\circ T$ and that the family $\{ {\varphi}_i\}_i$ is norming for $X$. Then, $T$ extends to a linear isometry. By assumption, the linear map $U\colon X\to\ell_\infty$, defined by $(U x)(i)={\varphi}_i(x)$, is isometric. Since $\{\psi_i\}_i$ is also norming, the same holds true for $V\colon Y\to\ell_\infty$, which satisfies $U|_{S_X}=V\circ T$. Thus $T$ extends to a linear isometry (which is $V^{-1}\circ U$). We note that any real linear functional ${\varphi}$ on a complex Banach space $X_{{\mathbb C}}$ has the complexification ${\varphi}_{{\mathbb C}}$ on $X_{{\mathbb C}}$, which satisfies that ${\varphi}=\Re{\varphi}_{{\mathbb C}}$ and $\|{\varphi}\|=\|{\varphi}_{{\mathbb C}}\|$ (see Lemma III.6.3 in [@conway]). \[lem:arens\] Let ${\varphi}$ be a norm-one real or complex linear functional on a $\mathrm{C}^*$-algebra $A$. If $a\in B_A$ is such that ${\varphi}(a)=1$, then ${\varphi}(x)={\varphi}(aa^*xa^*a)$ for all $x\in A$. This is a simple consequence of the Arens trick. Since $$\|(1-aa^*)x+a\| = \| \left[\begin{smallmatrix} (1-aa^*) & a \end{smallmatrix}\right] \left[\begin{smallmatrix} x \\ 1 \end{smallmatrix}\right]\| \le \| \left[\begin{smallmatrix} (1-aa^*) & a \end{smallmatrix}\right]\|\, \|\left[\begin{smallmatrix} x \\ 1 \end{smallmatrix}\right]\| \le \sqrt{1+\|x\|^2},$$ one has $$| \lambda {\varphi}((1-aa^*)x)+1 |^2 = |{\varphi}(\lambda(1-aa^*)x+a)|^2 \le 1+|\lambda|^2\|x\|^2$$ for all $\lambda$. This is possible only if ${\varphi}((1-aa^*)x)=0$. The proof of the other side is similar. We call a closed face $F\subset S_X$ an *intersection face* if $$F=\bigcap\{ E : \mbox{$E\subset S_X$ a maximal face containing $F$}\}.$$ By Corollary 3.4 in [@tanaka2], every norm-closed face $F$ of the unit sphere $S_A$ of a complex $\mathrm{C}^*$-algebra $A$ is an intersection face. This fact persists for the real case. Indeed, if $A$ is a real $\mathrm{C}^*$-algebra, then $G:=\{ x \in S_{A_{{\mathbb C}}} : \frac{1}{2}(x+{\mathcal J}(x))\in F\}$ is an intersection face of $S_{A_{{\mathbb C}}}$ by Corollary 3.4 in [@tanaka2] and so is $F=G\cap S_A$. \[lem:intersection\] Let $T\colon S_X\to S_Y$ be a surjective isometry and $F$ be an intersection face. Then, $T(F)$ is an intersection face such that $T(-F)=-T(F)$. In particular, there is $\psi\in B_{Y^*}$ such that $\psi=1$ on $T(F)$. Recall from [@cd; @tingley] and Proposition 2.3 in [@mori] that for any maximal face $E\subset S_X$, its image $T(E)\subset S_Y$ is a maximal face such that $T(-E)=-T(E)$. Thus $T(F)=T(\bigcap E)=\bigcap T(E)$ is an intersection face such that $T(-F)=T(\bigcap -E)=-\bigcap T(E)=-T(F)$. The last assertion follows from Eidelheit’s separation theorem (see, e.g., Lemma 3.1 in [@tanaka2]). Let $\xi,\eta\in{\mathcal H}$ be unit vectors and ${\varphi}(\,\cdot\,)={\mathopen{\langle}\,\cdot\,\xi,\eta\mathclose{\rangle}}$ be the corresponding linear functional on $A:={\mathbb B}({\mathcal H})$. We consider the maximal face $$E_{\varphi}:=\{ x\in S_A : {\varphi}(x)=1\} = \{ x\in S_A : x\xi=\eta \}.$$ By Lemma \[lem:intersection\], there is $\psi\in B_{Y^*}$ such that $\psi=1$ on the face $T(E_{\varphi})$. We will show $\Re{\varphi}=\psi\circ T$, which along with Lemma \[lem:cond\] proves Theorem \[thm:mau\] for $A={\mathbb B}({\mathcal H})$. Let $v\in A$ denote the rank-one partial isometry such that $v\xi=\eta$. Let $u\in S_A$ be an arbitrary unitary element. Since $\dim{\mathcal H}>2$, there is a sub-partial isometry $w$ of $u$ such that $w\perp v$, for example, pick a unit vector $\zeta \in \{\xi,u^*\eta\}^\perp$ and set $w=up_\zeta$, where $p_\zeta$ is the rank-one projection corresponding to $\zeta$. Put $$F(w):=\{ x\in S_A : xw^*w=w \} = \{ w+ y : y\in B_{(1-ww^*)A(1-w^*w)} \}$$ the corresponding closed face, which contains $u$. Since $T|_{F(w)}$ is affine by Lemma \[lem:intersection\] and Corollary \[cor:man\], there is $\theta\in B_{A^*}$ such that $(\psi\circ T)(w+y)=(\psi\circ T)(w)+\Re\theta(y)$ for $y\in B_{(1-ww^*)A(1-w^*w)}$. Since $w\pm v \in F(w)\cap (\pm E_{{\varphi}})$, one has $\pm 1 = (\psi\circ T)(w\pm v)=(\psi\circ T)(w)\pm\Re\theta(v)$. This implies that $(\psi\circ T)(w)=0$ and $\theta(v)=1$, which means $\theta={\varphi}$ by Lemma \[lem:arens\] and the fact that $vv^*xv^*v={\varphi}(x)v$ for all $x\in A$. It follows that $(\psi\circ T)(u)=\Re{\varphi}(u)$ for every unitary element $u$. Now let $w'$ be an arbitrary rank-one partial isometry. Since $T|_{F(w')}$ is affine and $F(w')$ is the closed convex hull of the unitary elements in $F(w')$, the previous result implies that $\psi\circ T=\Re{\varphi}$ on $F(w')$. Since $\bigcup_{w'} F(w')$ is dense in $S_A$, we conclude by continuity that $\psi\circ T=\Re{\varphi}$. Convex combinations of a face and its opposite ============================================== For any face $E\subset S_X$ and $\lambda\in[-1,1]$, put $$E_\lambda := \{ x\in S_X : \operatorname{dist}(x,E)\le1-\lambda\mbox{ and }\operatorname{dist}(x,-E)\le1+\lambda\}.$$ Since $\operatorname{dist}(E,-E)=2$ for any convex subset $E\subset S_X$, the inequalities defining $E_\lambda$ are actually equalities. We upgrade Lemma \[lem:intersection\] as follows. \[lem:tflambda\] Let $T\colon S_X\to S_Y$ be an surjective isometry and $F\subset S_X$ be an intersection face. Then, $T(F)$ is an intersection face such that $T(F_\lambda)=T(F)_\lambda$ for every $\lambda\in[-1,1]$. \[lem:tflconv\] For any face $E\subset S_X$, $\lambda_1,\lambda_2\in [-1,1]$, and $\alpha\in[0,1]$, one has $$\bigl(\alpha E_{\lambda_1}+(1-\alpha)E_{\lambda_2}\bigr)\cap S_X \subset E_{\alpha\lambda_1+(1-\alpha)\lambda_2}.$$ For given $x_i\in E_{\lambda_i}$, put $x_3:= \alpha x_1+(1-\alpha)x_2$ and $\lambda_3:=\alpha\lambda_1+(1-\alpha)\lambda_2$. For any ${\varepsilon}>0$, take $y_i\in E$ such that $\| y_i - x_i \|\approx_{{\varepsilon}} 1-\lambda_i$. Then, $\alpha y_1+(1-\alpha)y_2\in E$ and $$\operatorname{dist}(x_3, E)\le \alpha\|x_1-y_1\|+(1-\alpha)\| x_2-y_2\| \approx_{\varepsilon}1-\lambda_3.$$ Since ${\varepsilon}>0$ was arbitrary, one has $\operatorname{dist}(x_3, E)\le 1-\lambda_3$. The proof of the other inequality is similar and one has $x_3\in E_{\lambda_3}$. Combining Lemmas \[lem:tflambda\] and \[lem:tflconv\], we obtain the following. \[lem:tfl\] Let $T\colon S_X\to S_Y$ be a surjective isometry and $E\subset S_X$ be an intersection face. Then, for every $\lambda_1,\lambda_2\in [-1,1]$ and $\alpha\in[0,1]$, one has $$\bigl( \alpha T(E_{\lambda_1})+(1-\alpha)T(E_{\lambda_2}) \bigr) \cap S_Y\subset T(E_{\alpha\lambda_1+(1-\alpha)\lambda_2}).$$ Facial structure of a $\mathrm{C}^*$-algebra ============================================ The following first three results are about Kadison’s transitivity theorem. They are all well-known in the complex case. \[lem:kadison\] Let $A$ be a $\mathrm{C}^*$-algebra and $p\in A^{**}$ be a finite rank projection. Then, for any norm-one (resp. self-adjoint) element $x\in pA^{**}p$, there is a norm-one (resp. self-adjoint) element $a\in A$ such that $ap= x = pa$. For any unitary element $x$ in $\pm{\mathcal U}_0(pA^{**}p)$, the connected component of $\pm1$, there is a unitary element $a$ in $A$ (or the unitization of, if $A$ is not unital) such that $ap= x = pa$. In the complex case, this follows from Theorem II.4.15 in [@takesaki]. We deal with the case of a real $\mathrm{C}^*$-algebra $A$. Let $(A_{{\mathbb C}},{\mathcal J})$ be its complexification. Then for every norm-one (resp. self-adjoint) element $x\in pA^{**}p$, there is a norm-one (resp. self-adjoint) element $a'\in A_{{\mathbb C}}$ such that $a'p= x = pa'$. Thus $a:=\frac{1}{2}(a'+{\mathcal J}(a'))\in A$ satisfies the desired condition. Now let $x\in \pm{\mathcal U}_0(pA^{**}p)$ be a unitary element. Since $x=\pm x_1\cdots x_n$ for some $x_k\in{\mathcal U}_0(pA^{**}p)$ with $\| x_k - p \|<2$, we may assume that $-1$ is not in the spectrum of $x$. Let $\log\exp(\sqrt{-1}\lambda)=\sqrt{-1}\lambda$ for $\lambda\in(-\pi,\pi)$. Then, $h:=\frac{1}{\sqrt{-1}}\log (x+p^\perp) \in A_{{\mathbb C}}^{**}$ is a self-adjoint element such that $hp=ph$ and ${\mathcal J}(h)=-\frac{1}{\sqrt{-1}}\overline{\log} (x^*+p^\perp)=-h$. Similarly as above, there is a self-adjoint element $b \in A_{{\mathbb C}}$ such that $bp = php = pb$ and ${\mathcal J}(b)=-b$. It follows that $a=\exp(\sqrt{-1} b)$ is a unitary element in (the unitization of) $A$ such that $ap= x = pa$. The assumption on the unitary element $x$ cannot be removed in general; E.g., there is no unitary element $f$ in $\{ f\in C([0,1],{\mathbb M}_2({\mathbb R})) : f(0)\in{\mathbb R}1\}$ such that $f(1)=[\begin{smallmatrix} 1 & 0 \\ 0 & -1 \end{smallmatrix}]$. However, this is a rather special case. The finite dimensional real von Neumann algebra, $pA^{**}p$ in the above lemma, is a direct sum of ${\mathbb M}_k({\mathbb F})$’s, where ${\mathbb F}\in\{{\mathbb R},{\mathbb C},{\mathbb H}\}$. Among them, the unitary groups of ${\mathbb M}_k({\mathbb C})$ and ${\mathbb M}_k({\mathbb H})$ are connected, while the unitary group of ${\mathbb M}_k({\mathbb R})$ has two connected components according to the sign of the determinant. Thus, unless there is a nonzero central projection $z\le p$ in $A^{**}$ such that $zA^{**}\cong{\mathbb M}_k({\mathbb R})$, one can inflate $p$ to a finite rank projection $p_0$ and $x$ to a unitary element in ${\mathcal U}_0(p_0A^{**}p_0)$. Recall that a linear functional ${\varphi}$ on a $\mathrm{C}^*$-algebra $A$ is called a *state* if it is positive and has norm one. It is said to be *pure* if it is an extreme point of the state space. \[lem:pure\] Let ${\varphi}$ be a pure state on a $\mathrm{C}^*$-algebra $A$. Then, $p:=\operatorname{supp}({\varphi})$ is a minimal projection in $A^{**}$. Let $L:=\{ a\in A : {\varphi}(a^*a)=0\}$ be the corresponding left ideal and $(e_n)_n$ be an approximate unit for the $\mathrm{C}^*$-subalgebra $L\cap L^*$. Then, $e_n\to p^\perp$ ultrastrongly. In particular, any $\theta\in B_{A^*}$ such that $\lim_n\theta(1-e_n)=1$ coincides with ${\varphi}$. We view ${\varphi}$ as a normal state on the second dual $A^{**}$. Recall that $p:=\operatorname{supp}({\varphi})$ is the smallest projection in $A^{**}$ such that ${\varphi}(p)=1$. In the complex case, it is well-known that $p$ has rank one. In the real case, the set $\Omega$ of states ${\varphi}_0$ on $A_{{\mathbb C}}$ such that $\Re{\varphi}_0={\varphi}$ on $A$ is a weak$^*$-closed face of the state space of $A_{{\mathbb C}}$ and any pure state ${\varphi}_0\in\operatorname{ext}\Omega$ satisfies ${\varphi}_{{\mathbb C}}=\frac{1}{2}({\varphi}_0+\bar{{\varphi}}_0)$ on $A_{{\mathbb C}}$, where $\bar{{\varphi}}_0(a)=\overline{{\varphi}_0({\mathcal J}(a))}$ and ${\varphi}_{{\mathbb C}}$ is the complexification of ${\varphi}$, which is the unique state extension of ${\varphi}$ on $A_{{\mathbb C}}$ such that ${\varphi}_{{\mathbb C}}=\bar{{\varphi}}_{{\mathbb C}}$. It follows that $p=\operatorname{supp}({\varphi}_{{\mathbb C}})$ has rank at most two in $A_{{\mathbb C}}^{**}$. Hence, by compactness, for any nonzero projection $q\le p$ there is ${\varepsilon}>0$ such that ${\varepsilon}{\varphi}(q\,\cdot\,q)\le{\varphi}(p\,\cdot\,p)$. Since ${\varphi}(p\,\cdot\,p)={\varphi}(\,\cdot\,)$ is a pure state, this implies $q=p$, proving that the projection $p$ is minimal. We note that there is only one state on $pA^{**}p\cong{\mathbb R},\,{\mathbb C},\,{\mathbb H}$. Since $e_ne_m \to e_m$ as $n\to\infty$, the ultrastrong limit $e:=\sup e_n\in A^{**}$ is a projection such that $e^\perp\geq p$. By Kadison’s transitivity theorem (Lemma \[lem:kadison\]), one has $$\{ x \in A : pxp=0 \} = L + L^* \subset \{ x\in A : e^\perp x e^\perp=0\}.$$ Indeed, for any $x\in A$ such that $pxp=0$, there is a finite rank projection $q\geq p$ in $A^{**}$ such that $px=pxq$ and there is $a\in A$ such that $aq = pxq = qa$. Since $ap=pxp=0$ and $p(x-a)=pxq-pqa=0$, one has $a\in L$ and $x-a\in L^*$. The right inclusion is obvious. Hence the map $pAp \ni pxp\mapsto e^\perp x e^\perp$ is a well-defined continuous linear map. Since $p$ is of finite rank, this map is ultraweakly continuous on $pA^{**}p$ and so $e^\perp\le p$. This proves $e=p^\perp$. If $\theta\in B_{A^*}$ is such that $\lim_n\theta(1-e_n)=1$, then $\theta(e^\perp)=1$, which implies that $\theta$ is a state and $\theta={\varphi}$ by Lemma \[lem:arens\] and the uniqueness of the state on $e^\perp A^{**} e^\perp$. Recall that for any maximal face $E\subset S_X$ there is ${\varphi}\in \operatorname{ext}B_{X^*}$ such that $$E=E_{\varphi}:=\{ x\in S_X : {\varphi}(x)=1\}.$$ \[lem:maxface\] Let $A$ be a $\mathrm{C}^*$-algebra and ${\varphi}\in \operatorname{ext}B_{A^*}$. Then, $|{\varphi}|$ is a pure state and there is a unitary element $u$ in the unitization of $A$ such that ${\varphi}(\,\cdot\,)=|{\varphi}|(u^*\,\cdot\,)$ and $E_{{\varphi}} =u E_{|{\varphi}|}$. Every ${\varphi}\in S_{A^*}$ has a polar decomposition ${\varphi}(\,\cdot\,)=|{\varphi}|(v^*\,\cdot\,)$ (see Section III.4 in [@takesaki]), where $v$ is a partial isometry in $A^{**}$. By Lemma \[lem:arens\], one sees that the state $|{\varphi}|$ is pure if (and only if) ${\varphi}\in\operatorname{ext}B_{A^*}$. Hence $p:=v^*v$ and $q:=vv^*$ are minimal projections in $A^{**}$. By Lemma \[lem:arens\], it suffices to show that there is a unitary element $u$ in (the unitization of) $A$ such that $up=v$. If $p=q$, then $v$ is a unitary element in $pA^{**}p\cong{\mathbb R},{\mathbb C},{\mathbb H}$ and so there is a unitary element $u$ such that $up=v$ by Kadison’s transitivity theorem (Lemma \[lem:kadison\]). From now on, we assume that $p\neq q$. It follows that $p\vee q - p \sim q - p\wedge q = q$ (Proposition V.1.6 in [@takesaki]) is a minimal projection and $(p\vee q - p)v(p\vee q - q)v^*(p\vee q - p) = \alpha(p\vee q - p)$ for some $\alpha>0$. Hence $w:=\alpha^{-1/2}(p\vee q - q)v^*(p\vee q - p)$ is a minimal partial isometry in $(p\vee q)A^{**}(p\vee q)$ such that $w\perp v$. We claim that at least one of the two unitary elements $v \pm w$ in $(p\vee q)A^{**}(p\vee q)$ does not have both $\pm1$ in its spectrum. If this was not the case, then $v+w=e-e^\perp$ and $v-w=f-f^\perp$ for some minimal projections $e$ and $f$, but since $v=e+f-1$ is a partial isometry, this implies $e=f$ or $e\perp f$, which is absurd. Hence, by Kadison’s transitivity theorem (see Proof of Lemma \[lem:kadison\]), there is a unitary element $u$ in (the unitization of) $A$ such that $up=v$. \[lem:realvna\] Let $A$ be a von Neumann algebra and $E\subset S_A$ be a maximal face. Then, $E$ coincides with the closed convex hull of unitary elements in $E$. Moreover, there is a net $(\Theta_n)_n$ of affine contractions from $E$ onto ultraweakly-closed face $E_n\subset E$ such that $\Theta_n\to\operatorname{id}_E$ in the point-norm topology and each $E_n$ is affinely isometrically isomorphic to the closed unit ball of a real von Neumann algebra. In particular, any surjective isometry $T\colon S_A\to S_Y$ is affine on $E$. By Lemma \[lem:maxface\], we may assume that $E=E_{\varphi}$ for some pure state ${\varphi}$ on $A$. Let $L:=\{ a\in A : {\varphi}(a^*a)=0\}$ and take an approximate unit $(e_n)_n$ for $L\cap L^*$. By enlarging the index set if necessary, we may find ${\varepsilon}(n)>0$ such that ${\varepsilon}(n)\to0$ and put $q_n := \chi_{({\varepsilon}(n),1]}(e_n)$. Then, $(q_n)_n$ is again an approximate unit for the hereditary $\mathrm{C}^*$-subalgebra $L\cap L^*$ and so any $x\in E$ satisfies $1-x\in L\cap L^*$ and $q_n(1-x)\approx 1-x \approx (1-x)q_n$. This implies $(1-q_n)x \approx 1-q_n \approx x(1-q_n)$ and $q_n x\approx xq_n$. Therefore $\Theta_n(x) := q_n^\perp + q_n x q_n \in q_n^\perp+B_{q_nAq_n}$ satisfies $\Theta_n(x)\to x$ in norm. We put $E_n:= q_n^\perp+B_{q_nAq_n} \subset E_{\varphi}$. Then, by the Russo–Dye theorem (Theorem 7.2.4 in [@li]) $E_n$ coincides with the closed convex hull of unitary elements in $E_n$. Also, $T|_{E_n}$ is affine by Lemma \[lem:intersection\] and Corollary \[cor:man\], and so is $T|_E$. This is enough for the proof of Theorem \[thm:mau\] for von Neumann algebras. More technical results below are needed to deal with $\mathrm{C}^*$-algebras. Let $A$ be a $\mathrm{C}^*$-algebra and $A^{**}$ be its second dual. A projection $p$ in $A^{**}$ is said to be *compact* if it is the ultrastrong limit of an decreasing net of norm-one positive elements in $A$. (See [@ap] for more detail.) For any nonzero compact projection $p\in A^{**}$, denote the corresponding face by $$F(p) := \{ x\in S_A : xp = p = px \} = A\cap \{ p+y : y\in B_{p^\perp A^{**}p^\perp}\}$$ and put for $\lambda\in[-1,1]$, $$\begin{aligned} F(p,\lambda) := \{ x\in S_A : xp=\lambda p = px \} = S_A\cap\{ \lambda p+ y : y\in B_{p^\perp A^{**}p^\perp} \}.\end{aligned}$$ \[lem:fwc\] One has $\overline{F(p)}^{\sigma(A^{**},A^*)}=\{ p+y : y\in B_{p^\perp A^{**}p^\perp}\}$. The inclusion $\subset$ is clear. For the other inclusion, take $y\in B_{p^\perp A^{**}p^\perp}$ arbitrary. By Kaplansky’s density theorem, there is a net $y_n\in B_A$ such that $y_n\to y$ ultrastrongly. Since $p$ is compact, there is a net $p_n\in A$ such that $0\le p_n\le 1$ and $p_n\searrow p$. (We may assume that these nets are indexed by the same directed set.) Then, $p_np=p=pp_n$ and hence the net $$z_n:=p_n+(1-p_n)^{1/2}y_n(1-p_n)^{1/2} =[\begin{smallmatrix} p_n^{1/2} & (1-p_n)^{1/2} \end{smallmatrix}] [\begin{smallmatrix} 1 & 0 \\ 0 & y_n\end{smallmatrix}] [\begin{smallmatrix} p_n^{1/2} \\ (1-p_n)^{1/2}\end{smallmatrix}]$$ belongs to $F(p)$ (the latter expression shows $\|z_n\|\le1$). Since it converges ultrastrongly to $p+y$, we are done. \[lem:fvlambda\] One has $F(p)_\lambda=F(p,\lambda)$. We view $A^{**}\subset{\mathbb B}({\mathcal H})$. Let $x\in F(p)_\lambda$ be given and take a unit vector $\xi\in p{\mathcal H}$. For every ${\varepsilon}>0$, there are $y\in F(p)$ and $z\in -F(p)$ such that $\| y-x \|\approx_{\varepsilon}1-\lambda$ and $\|z-x\|\approx_{\varepsilon}1+\lambda$. Since one has $y\xi=\xi=-z\xi$ and $$2\le\| \xi-x\xi\|+\|\xi+x\xi\|\le \|y-x\|+\|-z+x\|\approx_{2{\varepsilon}} 2,$$ $\xi-x\xi$ and $\xi+x\xi$ are almost parallel. Since ${\varepsilon}>0$ was arbitrary, this means $x\xi$ is parallel to $\xi$ and $x\xi=\lambda \xi$, i.e., $xp=\lambda p$. The proof of $px=\lambda p$ is similar. This proves $F(p)_\lambda\subset F(p,\lambda)$. For the converse inclusion, let $x=\lambda p+x'\in F(p,\lambda)$ be given. Then, $y := p+x'\in \overline{F(p)}^{\sigma(A^{**},A^*)}$ by Lemma \[lem:fwc\] and $\|x-y\|=1-\lambda$. By the Hahn–Banach separation theorem, one has $$\operatorname{dist}(x,F(p))=\operatorname{dist}(x,\overline{F(p)}^{\sigma(A^{**},A^*)})\le 1-\lambda.$$ The proof of the other inequality is similar and $F(p)_\lambda\supset F(p,\lambda)$. The following is formally stronger than the Russo–Dye theorem (Theorem I.8.4 in [@davidson]), but it can be proved by adapting the standard proof. \[lem:rd\] Let $A$ be a unital complex $\mathrm{C}^*$-algebra and $E\subset S_A$ be a maximal face. Then, $E$ coincides with the closed convex hull of unitary elements in $E$. By Lemma \[lem:maxface\], we may assume that $E=E_{\varphi}$ for some pure state ${\varphi}$ on $A$. Let $p:=\operatorname{supp}({\varphi})$ and view elements in $A$ as operator valued $2 \times 2$ matrix in accordance with $p\oplus p^\perp$. Thus, $E=\{[\begin{smallmatrix} 1 & 0 \\ 0 & \ast \end{smallmatrix}]\}$. Let $x\in E$ and ${\varepsilon}>0$ be given. We set $v_0:=1$ and choose $u_k,v_k\in E\cap{\mathcal U}(A)$ inductively as follows. Fix $\delta>0$ very small and let $(1+\delta)v_{k-1}+x=w_k|(1+\delta)v_{k-1}+x|$ be the polar decomposition. Note that $w_k\in E\cap{\mathcal U}(A)$, as it is easily seen from the operator valued matrix viewpoint. We define $u_k,v_k \in E\cap{\mathcal U}(A)$ to be $w_k(|\frac{v_{k-1}+x}{2}|\pm\sqrt{-1}(1-|\frac{v_{k-1}+x}{2}|^2)^{1/2})$. Then, $$u_k+v_k=w_k|v_{k-1}+x|\approx_{\delta'}w_k|(1+\delta)v_{k-1}+x| =(1+\delta)v_{k-1}+x\approx_{\delta}v_{k-1}+x,$$ where $\delta'>0$ depends only on $\delta$ and converges to $0$ as $\delta$ converges to $0$. Thus, by choosing $\delta>0$ small enough, one has $v_{k-1}+x\approx_{{\varepsilon}/3}u_k+v_k$. It follows that $$v_0+nx \approx_{{\varepsilon}/3} u_1+v_1+(n-1)x \approx_{{\varepsilon}/3}\cdots\approx_{{\varepsilon}/3} u_1+\cdots+u_n+v_n$$ and $\| x-\frac{1}{n}\sum_{k=1}^n u_k \|<{\varepsilon}$ for $n\geq3/{\varepsilon}$. We will need the following ad hoc result. \[lem:adhoc\] Let $A$ be a unital complex $\mathrm{C}^*$-algebra and $\psi\colon A\to{\mathbb C}$ be a nonzero multiplicative linear functional. Then the closed unit ball of the real $\mathrm{C}^*$-algebra $A^\psi_{{\mathbb R}} := \psi^{-1}({\mathbb R})$ has the strong Mankiewicz property. Let $\bar{A}=\{\bar{a}:a\in A\}$ denote the complex conjugate $\mathrm{C}^*$-algebra of $A$. Then, $A^\psi_{{\mathbb R}}$ is the real part of the complex $\mathrm{C}^*$-algebra $A^\psi := \{ a\oplus \bar{b} \in A\oplus \bar{A} : \psi(a)=\overline{\psi(b)}\}$ with respect to the conjugate linear automorphism ${\mathcal J}(a\oplus \bar{b})=b\oplus\bar{a}$. By Theorem \[thm:man\], it suffices to show that $B_{A^\psi_{{\mathbb R}}}$ coincides with the closed convex hull of unitary elements in $A^\psi_{{\mathbb R}}$. Let $x\in B_{A^\psi_{{\mathbb R}}}$ and ${\varepsilon}>0$ be given. We consider the second dual von Neumann algebra $A^{**}$ and the ultraweakly continuous extension $\psi\colon A^{**}\to{\mathbb C}$. Then, $M:=\ker\psi$ is a von Neumann subalgebra such that $A^{**}={\mathbb C}\oplus M$ as a von Neumann algebra. Hence there are unitary elements $u_k$ in $(A^{**})^\psi_{{\mathbb R}}={\mathbb R}\oplus M$ such that $x\approx_{\varepsilon}\frac{1}{n}\sum_{k=1}^n u_k$. Let $h_k:=\frac{1}{\sqrt{-1}}\log u_k \in (A^{**})_{\mathrm{s.a.}}$, where $\log \exp(\sqrt{-1}\lambda)=\sqrt{-1}\lambda$ for $\lambda\in[-\pi,\pi)$. Since $\psi(u_k)\in\{1,-1\}$, one has $\psi(h_k)=\frac{1}{\sqrt{-1}}\log \psi(u_k)\in\{0,-\pi\}$. By Kaplansky’s density theorem, there are bounded nets $(h_{k,i})_i$ in $A_{\mathrm{s.a.}}$ such that $h_{k,i}\to h_k$ ultrastrongly. We may assume that $\psi(h_{k,i})\in\{0,-\pi\}$. It follows that $u_{k,i}:=\exp(\sqrt{-1}h_{k,i})$ are unitary elements in $A^\psi_{{\mathbb R}}$ such that $\frac{1}{n}\sum_{k=1}^n u_{k,i} \to \frac{1}{n}\sum_{k=1}^n u_k$ ultrastrongly. Hence, by the Hahn–Banach separation theorem, one has $\operatorname{dist}(x, \operatorname{conv}\{ u_{k,i} : k,i\}) = \operatorname{dist}(x, \overline{\operatorname{conv}\{ u_{k,i} : k,i\}}^{\sigma(A^{**},A^*)}) < {\varepsilon}.$ The following is the main technical result. \[prop:tech\] Let $A$ be a $\mathrm{C}^*$-algebra, $T\colon S_A\to S_Y$ be a surjective isometry, and $E\subset S_A$ be a maximal face. Then, there is a net $(\Theta_n)_n$ of affine contractions from $E$ into closed convex subsets $K_n\subset E$ such that $\Theta_n\to\operatorname{id}_E$ in the point-norm topology, each $K_n$ is affinely isometrically isomorphic to the closed unit ball of a real $\mathrm{C}^*$-algebra $C_n$, and each $T(K_n)$ is convex. In the case $A$ is a unital complex $\mathrm{C}^*$-algebra, the above $C_n$ can be taken so that $B_{C_n}$ has the strong Mankiewicz property and so $T|_E$ is affine. By Lemma \[lem:maxface\], we may assume that $E=E_{\varphi}$ for some pure state ${\varphi}$ on $A$. We denote by $\tilde{A}$ the unitization of $A$ if $A$ is not unital, else $\tilde{A}=A$. We consider ${\varphi}$ a pure state on $\tilde{A}$ and consider $L:=\{ a\in \tilde{A} : {\varphi}(a^*a)=0\}$. One can skip this paragraph if $A$ is unital. In case $A$ is not unital, let $\sigma$ denote the character on $\tilde{A}$ corresponding to the unitization. We claim that there is an approximate unit $(e_n)_n$ for $L\cap L^*$ such that $\sigma(e_n)=1$. Let $(e_n)_n$ be any approximate unit. Then, by Lemma \[lem:pure\], we may assume $\lambda:=\inf\sigma(e_n)>0$. We consider the continuous function $h(t)=\min(\lambda^{-1}t,1)$. Then, $h(e_n)$ is an approximate unit such that $\sigma(h(e_n))=h(\sigma(e_n))=1$ for all $n$. Let $(e_n)_n$ be an approximate unit for $L\cap L^*$ such that $1-e_n\in A$ for all $n$ (which is equivalent to $\sigma(e_n)=1$). By perturbation using functional calculus, we may assume that there is $f_n\in L\cap L^*$ such that $0\le e_n\le f_n\le1$ and $e_nf_n=e_n=f_ne_n$. (Although $(e_n)_n$ may not be increasing anymore, this does not matter for the following.) Since $1-x \in L \cap L^*$ for every $x\in E_{\varphi}$, one has $e_n(1-x)\approx 1-x \approx (1-x)e_n$. This implies $(1-e_n)x \approx 1-e_n \approx x(1-e_n)$ and $e_n x\approx xe_n$. Therefore $$x=((1-e_n)+e_n)x\approx (1-e_n) + e_n^{1/2} x e_n^{1/2} =: \Theta_n(x) \in E_{\varphi}.$$ See Proof of Lemma \[lem:fwc\] for the proof that $\Theta_n$ is contractive and $\|\Theta_n(x)\|\le1$ for every $x\in E_{\varphi}$. To ease notation, we fix $n$ and write $f:=f_n$. We consider $s:=1-f \in E_{\varphi}$, its support projection $p:=\chi_{(0,1]}(s)\in A^{**}$, and the closed face $$F:=\{ x\in S_A : xs = s = sx\} = \{ x\in S_A : xp=p=px\} \subset E_{\varphi}$$ (although $p$ is not a compact projection). Since $(1-e_n)s=s=s(1-e_n)$, one has $\Theta_n(E_{\varphi})\subset F$. If $p\in A$, then the face $K_n:=F=p+B_{p^\perp A p^\perp}$ satisfies the desired property with $C_n=p^\perp A p^\perp$. Let’s assume $p\notin A$ and consider the $\mathrm{C}^*$-subalgebra $$C:=\{ a\in A : ap=\gamma p = pa \mbox{ for some scalar }\gamma\}.$$ For the following, it is probably easier to digest if one views elements in $A$ as operator valued $2\times2$ matrices in accordance with $p\oplus p^{\perp}$. Thus $s=[\begin{smallmatrix} s & 0 \\ 0 & 0\end{smallmatrix}]$ (slightly abusing the notation), $p=[\begin{smallmatrix} 1 & 0 \\ 0 & 0\end{smallmatrix}]$, $F=\{ [\begin{smallmatrix} 1 & 0 \\ 0 & \ast\end{smallmatrix}]\}$, and $C=\{ [\begin{smallmatrix} \gamma & 0 \\ 0 & \ast\end{smallmatrix}]\}$. Since $p\notin A$, one has $\|px\|\le\|p^\perp x\|$ for all $x\in C$. Hence, there is a norm-one (multiplicative) linear functional $\psi$ on $p^\perp C$ such that $\psi(p^\perp x)p=px$ for all $x\in C$, or equivalently $x=[\begin{smallmatrix} \psi(y) & 0 \\ 0 & y\end{smallmatrix}]$ for all $x\in C$ and $y:=p^\perp x$. Hence, $\| x \| = \|(1-s)x\|$ for all $x\in C$ and one has $$F\subset K_n := s+(1-s)B_{C^\psi_{{\mathbb R}}} = S_A \cap \{ [\begin{smallmatrix} s+\gamma (1-s) & 0 \\ 0 & \ast \end{smallmatrix}] : \gamma\in[-1,1]\} \subset E_{\varphi}.$$ By Lemma \[lem:adhoc\], the closed unit ball $B_{C^\psi_{{\mathbb R}}}$ satisfies the strong Mankiewicz property provided that $A$ (and hence $C$) is a unital complex $\mathrm{C}^*$-algebra. It is left to show that $T(K_n)$ is convex. For $\gamma\in[-1,1]$, put $h_\gamma(\lambda):=\lambda+\gamma(1-\lambda)$. For $i=1,2$, put $$G^i_m(\gamma):=E_{\varphi}\cap{\bigcap}_k\, F\bigl(\chi_{[\frac{2k-2+i}{2m},\frac{2k-1+i}{2m}]}(s),h_\gamma(\frac{k}{m})\bigr) \mbox{ and }H^i_m(\gamma) := E_{\varphi}\cap{\mathcal N}_{\frac{1}{m}}(G^i_m(\gamma)),$$ where the intersection is over $k=1,2,\ldots,m$ for which $\chi_{[\frac{2k-2+i}{2m},\frac{2k-1+i}{2m}]}(s)\neq0$ and ${\mathcal N}_\delta$ means the $\delta$-neighborhood in $A$. Note that $\chi_{[\frac{2k-2+i}{2m},\frac{2k-1+i}{2m}]}(s)\le p$ for all $i$, $m$, and $k$. By Lemmas \[lem:fvlambda\] and \[lem:tflconv\], one has $$\alpha G^i_m(\gamma_1)+(1-\alpha) G^i_m(\gamma_2) \subset G^i_m(\gamma_3) \mbox{ and } \alpha H^i_m(\gamma_1)+(1-\alpha) H^i_m(\gamma_2) \subset H^i_m(\gamma_3)$$ for every $\gamma_1,\gamma_2\in[-1,1]$, $\alpha\in[0,1]$, and $\gamma_3:=\alpha\gamma_1+(1-\alpha)\gamma_2$. We claim that $$\begin{aligned} K(\gamma):=\{ x\in E_{\varphi}: px=h_\gamma(s)=xp\} = {\bigcap}_{m\in{\mathbb N}}\, (H^1_m(\gamma)\cap H^2_m(\gamma)).\end{aligned}$$ To prove the inclusion $\subset$, we define $g_m$ to be the continuous function such that $g_m(0)=\gamma$, $g_m(\lambda)=h_\gamma(\frac{k}{m})$ for $\lambda\in[\frac{2k-1}{2m},\frac{2k}{2m}]$, and linear on $[\frac{2k-2}{2m},\frac{2k-1}{2m}]$ for $k=1,\ldots,m$. Then, $\|g_m-h_\gamma\|_\infty\le \frac{1}{m}$ and $(g_m-h_\gamma)(0)=0$. It follows that $(g_m-h_\gamma)(s)\in A\cap pAp$ and for any $x\in K(\gamma)$, one has $x+(g_m-h_\gamma)(s)\in G^1_m(\gamma)$. This proves $x\in H^1_m(\gamma)$. The proof of $x\in H^2_m(\gamma)$ is similar. For the converse inclusion, take $x$ from the RHS of the claimed equality. Since $x\in H^1_m(\gamma)\cap H^2_m(\gamma)$, there are $y^i_m\in G^i_m(\gamma)$ such that $\| x - y^i_m \|\le\frac{1}{m}$. For the projection $p^i_m:=\sum_{k=1}^m \chi_{[\frac{2k-2+i}{2m},\frac{2k-1+i}{2m}]}(s)$ in $A^{**}$, one has $\|h_\gamma(s)p^i_m - y^i_m p^i_m\|\le\frac{1}{m}$. Hence, $\| (h_\gamma(s) - x)(p^1_m\vee p^2_m)\|\le\frac{2}{m}$. Since $p^1_m\vee p^2_m\to p$ ultrastrongly, one sees $h_\gamma(s)=xp$. The proof of $h_\gamma(s)=px$ is similar. Now, since $K_n=\bigcup_{\gamma\in[-1,1]} K(\gamma)$ and $$T^{-1}(\alpha T(K(\gamma_1)) + (1-\alpha)T(K(\gamma_2))) \subset {\bigcap}_{m\in{\mathbb N}}\, (H^1_m(\gamma_3)\cap H^2_m(\gamma_3))=K(\gamma_3)$$ by Lemma \[lem:tfl\], one concludes that $T(K_n)$ is convex. Proof of Theorem \[thm:mau\] ============================ We first give the proof of Theorem \[thm:mau\] for the case where $A$ is not a type $\mathrm{I}_k$ factor with $k=1,2$. The $\mathrm{I}_1$ factors ${\mathbb R},\,{\mathbb C},\,{\mathbb H}$ are real Hilbert spaces and the Mazur–Ulam property for them is already known (see [@day; @cd]). The case of $\mathrm{I}_2$ factor is dealt with separately. We will show that for any surjective isometry $T\colon S_A\to S_Y$ and any pure state ${\varphi}$ on $A$, there is $\psi\in B_{Y^*}$ such that $\Re{\varphi}=\psi\circ T$. This yields the assertion by Lemmas \[lem:cond\] and \[lem:maxface\]. Let $p:=\operatorname{supp}({\varphi}) \in A^{**}$ and consider the maximal face $$E_{\varphi}:=\{ x\in S_A : {\varphi}(x)=1\} = \{ x\in S_A : xp = p = px \}.$$ By Lemma \[lem:intersection\], there is $\psi\in B_{Y^*}$ such that $\psi=1$ on the face $T(E_{\varphi})$. We will show $\Re{\varphi}=\psi\circ T$. Let $u\in S_A$ be an arbitrary unitary element. By the assumption that $A$ is not a type $\mathrm{I}_k$ factor with $k=1, 2$, there is a minimal projection $q\in A^{**}$ such that $q\perp p\vee u^*pu$. We consider the face $$F:=\{ x\in S_A : xq = uq \} = S_A\cap (uq + B_{(uqu^*)^\perp Aq^\perp}).$$ Since $T|_F$ is affine by Proposition \[prop:tech\] or Lemma \[lem:realvna\], there are $\gamma\in{\mathbb R}$ and $\theta\in B_{A^*}$ such that $(\psi\circ T)(x)=\gamma+\Re\theta(x)$ for $x\in F$. By Kadison’s transitivity theorem (Lemma \[lem:kadison\]), there is $x_0\in B_A$ such that $x_0q=uq$ and $x_0p=0=px_0$. Then, $x_0\in F\cap L\cap L^*$, where $L:=\{ a \in A : {\varphi}(a^*a)=0 \}$. Let $(e_n)_n$ be an approximate unit for the $\mathrm{C}^*$-subalgebra $L\cap L^*$ such that $e_n\geq |x_0|$ for all $n$. Since $e_n\geq |x_0|$, one has $e_n q = q$ and $x_0 \pm (1-e_n) \in F\cap (\pm E_{\varphi})$ for all $n$. Thus $$\pm 1 = (\psi\circ T)(x_0\pm(1-e_n)) = (\psi\circ T)(x_0) \pm \theta(1-e_n).$$ This implies that $(\psi\circ T)(x_0)=0$ and $\theta(1-e_n)=1$ for all $n$. By Lemma \[lem:pure\], one sees $\theta={\varphi}$ and so $(\psi\circ T)(u)=\Re{\varphi}(u)$. Now $\Re{\varphi}=\psi\circ T$ follows from Lemma \[lem:rd\] or \[lem:realvna\]. For the rest of the paper, we put $A={\mathbb M}_2({\mathbb F})$, where ${\mathbb F}\in\{{\mathbb R},{\mathbb C},{\mathbb H}\}$. \[lem:asab\] Let $\operatorname{tr}$ denote the tracial state on $A$ and put ${\mathcal H}:=A_{\mathrm{sa}}\cap\ker\operatorname{tr}$. Then, ${\mathcal H}$ is a real Hilbert space and the self-adjoint part $A_{\mathrm{sa}}$ of $A$ is isometrically isomorphic to the $\ell_1$-direct sum ${\mathbb R}1\oplus_1{\mathcal H}$, via $A_{\mathrm{sa}} \ni a \mapsto \operatorname{tr}(a)1 \oplus (a-\operatorname{tr}(a)1)$. Moreover, if $x\in S_A$ satisfies $\|1\pm x\|=2$, then $x\in{\mathcal H}$. Every $b \in {\mathcal H}$ is of the form $\mu p - \mu p^{\perp}$ for some $\mu\in{\mathbb R}$ and a minimal projection $p$. It follows that ${\mathcal H}$ is a real Hilbert space as $\| b \|=\operatorname{tr}(b^*b)^{1/2}$ and that $a=\lambda1+b$ satisfies $$\|a\|=\max\{|\lambda+\mu|,|\lambda-\mu|\}=|\lambda|+|\mu|=|\operatorname{tr}(a)| + \| a-\operatorname{tr}(a)1\|.$$ Now suppose $x\in S_A$ satisfies $\|1\pm x\|=2$. Then, there are unit vectors $\xi_+$ and $\xi_-$ such that $x\xi_{\pm}=\pm\xi_{\pm}$. It follows that the minimal projections $p_{\pm}$ that satisfy $p_{\pm}\xi_{\pm}=\xi_{\pm}$ satisfies $p_{\pm}xp_{\pm}=\pm p_{\pm}$. Since $A={\mathbb M}_2({\mathbb F})$, one concludes that $x=p_+-p_- \in {\mathcal H}$. \[lem:asaext\] Let $T\colon S_A\to S_Y$ be any surjective isometry. Then, $T|_{S_A\cap A_{\mathrm{sa}}}$ admits a linear extension. Let $\tilde{T}\colon A\to Y$ denote the homogeneous extension of $T$, which is given by $\tilde{T}(a)=\|a\|T(\frac{a}{\|a\|})$ for $a\neq0$ and $\tilde{T}(0)=0$. For any $b\in S_{{\mathcal H}}$, the convex hull of $1$ and $b$ is contained in $S_A$ by Lemma \[lem:asab\] and hence $T$ is affine there. Moreover, since $T$ preserves antipodal points ([@tingley]), $\tilde{T}$ is linear on the linear span of $1$ and $b$. It remains to show that $\tilde{T}$ is linear on ${\mathcal H}$. For this, we first prove that $\|T(b)+T(c)\| = \| b + c \|$ and $T(\frac{1}{\|b+c\|}(b+c))=\frac{1}{\|b+c\|}(T(b)+T(c))$ for every $b,c\in S_{\mathcal H}$. Let $b,c\in S_{\mathcal H}$ be given. We may assume that they are not parallel. For any $\lambda\in[-1,1]$, one has $$\begin{aligned} \| 2(1-|\lambda|) 1 +\lambda (b + c) \| &= \|((1-|\lambda|)1 + \lambda b) - (-(1-|\lambda|)1 - \lambda c) \| \\ &= \|T((1-|\lambda|)1 + \lambda b) - T(-(1-|\lambda|)1 - \lambda c) \| \\ &= \| 2(1-|\lambda|) T(1) +\lambda (T(b) + T(c)) \|.\end{aligned}$$ In particular, $\mu:=\|b + c \| = \| T(b) + T(c) \|>0$ and put $\lambda:=\frac{2}{2+\mu}$. Then, $$\begin{aligned} \| 1 \pm T^{-1}(\frac{1}{\mu}(T(b) + T(c))) \| &=\frac{1}{\lambda\mu}\| 2(1-\lambda)T(1) \pm \lambda( T(b) + T(c))\| \\ &=\| 1 \pm \frac{1}{\mu}(b + c) \| = 2.\end{aligned}$$ By Lemma \[lem:asab\], this implies that $x:=T^{-1}(\frac{1}{\mu}(T(b) + T(c))) \in S_{{\mathcal H}}$. If $\mu\le 1$, then $$\begin{aligned} 1-\mu+\|\mu x - b \| &= \|( (1-\mu)1 + \mu x ) - b \| \\ &= \| (1-\mu)T(1)+\mu T(x) - T(b) \| =1-\mu + \|T(c)\|\end{aligned}$$ and hence $\|\mu x- b \|=1$. On the other hand, if $\mu\geq 1$, then $$\begin{aligned} 1-\frac{1}{\mu}+\|x - \frac{1}{\mu}b \| = \| x - ((1-\frac{1}{\mu})1 + \frac{1}{\mu}b) \| = 1-\frac{1}{\mu}+ \frac{1}{\mu}\|T(c)\|\end{aligned}$$ and hence $\|\mu x - b \|=1$ again. Thus in any case, one has $\|\mu x - b \|=1$, and similarly $\|\mu x - c \|=1$. Also $\|\mu x - 0\|=\mu$. Therefore by trilateration, one has $\mu x=b+c$. It follows that $$\tilde{T}(\frac{\| c \| }{ \|b\|+\|c\|}b+ \frac{\|b \|}{ \|b\|+\|c\|}c) =\frac{\| c \|}{ \|b\|+\|c\|}\tilde{T}(b) + \frac{\|b \|}{ \|b\|+\|c\|}\tilde{T}(c)$$ for every $b,c\in{\mathcal H}$. Thus the bisection method and continuity of $\tilde{T}$ imply that $\tilde{T}$ is affine on the segment $[b_0,c_0]$ for any $b_0,c_0\in {\mathcal H}$. Note that the bisection process works whenever $b_0$ and $c_0$ are not parallel, because in which case the norm on $[b_0,c_0]$ is bounded above and away from zero. This proves that $\tilde{T}$ is linear on ${\mathcal H}$. We consider the unitary group ${\mathcal U}$ of $A$ and the diagonal subgroup $${\mathcal D}=\{[\begin{smallmatrix} \alpha & \\ & \beta \end{smallmatrix}] : \alpha,\beta \in{\mathbb F},\ |\alpha|=1=|\beta|\}\subset{\mathcal U}.$$ We claim that $${\mathcal U}= {\mathcal D}\cdot\{[\begin{smallmatrix} \lambda & \sqrt{1-\lambda^2} \\ \sqrt{1-\lambda^2} & -\lambda \end{smallmatrix}] : \lambda\in[0,1]\}\cdot{\mathcal D}.$$ Indeed, for any $u\in{\mathcal U}$, it is obvious that ${\mathcal D}\cdot u \ni [\begin{smallmatrix} \lambda & \gamma \\ \delta & -\mu \end{smallmatrix}]$ for some $\lambda,\mu\in[0,1]$. But since the latter is unitary, one must have $\lambda=\mu$ and $|\gamma|=|\delta|$. If $\lambda=\mu>0$, then one moreover has $\gamma=\delta^*$ and so by conjugating $[\begin{smallmatrix} 1 & \\ & \delta \end{smallmatrix}]$, one obtains the claim. One obtains the claim in the case $\lambda=\mu=0$ also. \[lem:2by2\] Let $E=\{ [\begin{smallmatrix} 1 & \\ & \ast \end{smallmatrix}]\}$ be the maximal face of $S_A$ and ${\varphi}_0$ be the corresponding pure state such that ${\varphi}_0=1$ on $E$. Let $T\colon S_A\to S_Y$ be a surjective isometry and $\psi\in S_{Y^*}$ be such that $\psi\circ T=1$ on $E$. Then, one has $\psi\circ T={\varphi}_0$ on $S_A\cap A_{\mathrm{sa}}$ as well as on ${\mathcal D}$. By Lemma \[lem:asaext\], the map ${\varphi}:=\psi\circ T$ admits a linear extension on $A_{\mathrm{sa}}$. Since $A_{\mathrm{sa}}\cong {\mathbb R}\oplus_1{\mathcal H}$ by Lemma \[lem:asab\], the linear functional ${\varphi}$ corresponds to a norm one element in ${\mathbb R}\oplus_\infty{\mathcal H}$, which is easily seen to be $1\oplus [\begin{smallmatrix} 1 & \\ & -1 \end{smallmatrix}]$. It follows that ${\varphi}(a)=\operatorname{tr}(a)+\operatorname{tr}([\begin{smallmatrix} 1 & \\ & -1 \end{smallmatrix}]a)={\varphi}_0(a)$ for $a\in S_A\cap A_{\mathrm{sa}}$. Next, let $\beta\in{\mathbb F}$ be such that $|\beta|=1$. Then, ${\varphi}$ is affine on $\{ [\begin{smallmatrix} \ast & \\ & \beta \end{smallmatrix}]\}$. Since ${\varphi}$ is a contractive map such that ${\varphi}( [\begin{smallmatrix} \pm 1 & \\ & \beta \end{smallmatrix}])=\pm1$, one obtains ${\varphi}( [\begin{smallmatrix} \alpha & \\ & \beta \end{smallmatrix}])= \Re\alpha$. Let $T\colon S_A\to S_Y$ be given and $\psi\in S_{Y^*}$ be given such that ${\varphi}:=\psi\circ T$ satisfies ${\varphi}={\varphi}_0$ on $E$. It suffices to show ${\varphi}={\varphi}_0$ everywhere. Let $x\in S_A$ be given and consider the polar decomposition $x=v|x|$. We may assume that $v$ is a unitary element. Since the surjective isometry $T(v\,\cdot\,)$ admits a linear extension on $A_{\mathrm{sa}}$ by Lemma \[lem:asaext\] and $\operatorname{ext}A_{\mathrm{sa}}\subset{\mathcal U}$, to prove ${\varphi}(x)={\varphi}_0(x)$, it suffices to show ${\varphi}={\varphi}_0$ on ${\mathcal U}$. Let $u\in{\mathcal U}$ be given and write it as $u=[\begin{smallmatrix} \alpha_1 & \\ & \alpha_2 \end{smallmatrix}][\begin{smallmatrix} \lambda & \sqrt{1-\lambda^2} \\ \sqrt{1-\lambda^2} & -\lambda \end{smallmatrix}][\begin{smallmatrix} \beta_1 & \\ & \beta_2 \end{smallmatrix}]=:axb$. Since the surjective isometry $S_A\ni z\mapsto T(azb) \in S_Y$ admits a linear extension on $A_{\mathrm{sa}}$, one has $${\varphi}(u)={\varphi}(axb)=\lambda{\varphi}(a[\begin{smallmatrix} 1 & \\ & -1\end{smallmatrix}]b)+ \sqrt{1-\lambda^2}{\varphi}(a[\begin{smallmatrix} & 1 \\ 1 & \end{smallmatrix}]b).$$ Since $a[\begin{smallmatrix} 1 & \\ & -1\end{smallmatrix}]b\in{\mathcal D}$, one has ${\varphi}(a[\begin{smallmatrix} 1 & \\ & -1\end{smallmatrix}]b)={\varphi}_0(a[\begin{smallmatrix} 1 & \\ & -1\end{smallmatrix}]b)$ by Lemma \[lem:2by2\]. On the other hand, since $a[\begin{smallmatrix} & 1 \\ 1 & \end{smallmatrix}]b= [\begin{smallmatrix} 1 & \\ & \alpha_2\beta_1 \end{smallmatrix}][\begin{smallmatrix} & 1 \\ 1 & \end{smallmatrix}][\begin{smallmatrix} 1 & \\ & \alpha_1\beta_2 \end{smallmatrix}]$ and $T'(\,\cdot\,)=T([\begin{smallmatrix} 1 & \\ & \alpha_2\beta_1 \end{smallmatrix}]\,\cdot\,[\begin{smallmatrix} 1 & \\ & \alpha_1\beta_2 \end{smallmatrix}])$ is a surjective isometry such that ${\varphi}':=\psi\circ T'$ satisfies ${\varphi}'={\varphi}_0$ on $E$, Lemma \[lem:2by2\] implies that ${\varphi}(a[\begin{smallmatrix} & 1 \\ 1 & \end{smallmatrix}]b) ={\varphi}'([\begin{smallmatrix} & 1 \\ 1 & \end{smallmatrix}])={\varphi}_0([\begin{smallmatrix} & 1 \\ 1 & \end{smallmatrix}])=0$. These two imply ${\varphi}(u)={\varphi}_0(u)$. This finishes the proof of Theorem \[thm:mau\]. We note that a similar proof yields the Mazur–Ulam property for the self-adjoint part $A_{\mathrm{sa}}$ of any real von Neumann algebra $A$. [BFGM]{} C. A. Akemann and G. K. Pedersen; Facial structure in operator algebra theory. *Proc. Lond. Math. Soc. (3)* **64** (1992), 418–448. U. Bader, A. Furman, T. Gelander, and N. Monod; Property (T) and rigidity for actions on Banach spaces. *Acta Math.* **198** (2007), 57–105. L. Cheng and Y. Dong; On a generalized Mazur-Ulam question: extension of isometries between unit spheres of Banach spaces. *J. Math. Anal. Appl.* **377** (2011), 464–470. J. B. Conway; A course in functional analysis. Graduate Texts in Mathematics, 96. Springer-Verlag, New York, 1990. xvi+399 pp. A. M. Peralta and M. Cueto-Avellaneda; The Mazur-Ulam property for commutative von Neumann algebras. *Linear Multilinear Algebra*, to appear. K. R. Davidson; $\mathrm{C}^*$-algebras by example. Fields Institute Monographs, 6. American Mathematical Society, Providence, RI, 1996. xiv+309 pp. M. M. Day; Some characterizations of inner-product spaces. *Trans. Amer. Math. Soc.* **62** (1947), 320–337. G.-G. Ding; The isometric extension of the into mapping from a ${\mathcal L}^\infty(\Gamma)$-type space to some Banach space. *Illinois J. Math.* **51** (2007), 445–453. X. N. Fang and J. H. Wang; Extension of isometries between the unit spheres of normed space $E$ and $C(\Omega)$. *Acta Math. Sin. (Engl. Ser.)* **22** (2006), 1819–1824. T. Figiel; On nonlinear isometric embeddings of normed linear spaces. *Bull. Acad. Polon. Sci. Sér. Sci. Math. Astronom. Phys.* **16** (1968), 185–188. B. Li; Real operator algebras. World Scientific Publishing Co., Inc., River Edge, NJ, 2003. xiv+241 pp. R. Liu; On extension of isometries between unit spheres of ${\mathcal L}^\infty(\Gamma)$-type space and a Banach space $E$. *J. Math. Anal. Appl.* **333** (2007), 959–970. P. Mankiewicz; On extension of isometries in normed linear spaces. *Bull. Acad. Polon. Sci. Sér. Sci. Math. Astronom. Phys.* **20** (1972), 367–371. S. Mazur and S. Ulam; Sur les transformationes isométriques d’espaces vectoriels normés. *C. R. Acad. Sci. Paris* **194** (1932), 946–948. M. Mori; Tingley’s problem through the facial structure of operator algebras. *J. Math. Anal. Appl.*, **466** (2018), 1281–1298. A. M. Peralta; A survey on Tingley’s problem for operator algebras. *Acta Sci. Math. (Szeged)* **84** (2018), 81–123. A. M. Peralta and F. J. Fernández-Polo; On the extension of isometries between the unit spheres of von Neumann algebras. *J. Math. Anal. Appl.* **466** (2018), 127–143. G. Schechtman; Generalizing the Mazur-Ulam theorem to convex sets with empty interior in Banach spaces. *mathoverflow.* [https://mathoverflow.net/questions/225597/]{} D. Tan, X. Huang, and R. Liu; Generalized-lush spaces and the Mazur-Ulam property. *Studia Math.* **219** (2013), 139–153. M. Takesaki; Theory of operator algebras. I. Springer-Verlag, New York-Heidelberg, 1979. vii+415 pp. R. Tanaka; Tingley’s problem on finite von Neumann algebras. *J. Math. Anal. Appl.* **451** (2017), 319–326. D. Tingley; Isometries of the unit sphere. *Geom. Dedicata* **22** (1987), 371–378.
{ "pile_set_name": "ArXiv" }
--- author: - | Blair D. Sullivan\ Dept. of Mathematics, Princeton University\ [email protected] bibliography: - 'CaccettaHaggkvist.bib' title: 'A Summary of Results and Problems Related to the Caccetta-Häggkvist Conjecture' --- \[section\] \[thm\][Lemma]{} \[thm\][Conjecture]{} Introduction ============ This paper is an attempt to survey the current state of our knowledge on the Caccetta-Häggkvist conjecture and related questions. In January 2006 there was a workshop hosted by the American Institute of Mathematics in Palo Alto, on the Caccetta-Häggkvist conjecture, and this paper partly originated there, as a summary of the open problems and partial results presented at the workshop. Our thanks to the many participants who helped with this paper. Notations & Definitions {#notations-definitions .unnumbered} ----------------------- - A [*graph*]{} $G = (V(G), E(G))$ is a collection of vertices $V(G)$ and edges $E(G)$, where each edge is an unordered pair of distinct vertices ${u,v}$. - A [*digraph*]{} $D = (V(D), E(D))$ is a collection of vertices $V(D)$ and edges $E(D)$, along with two incidence relations $h, t: V(D) \times E(D) \rightarrow {0,1}$. We let $t(u,e) = 1$ iff $u$ is the tail of $e$ $e$ (i.e. the edge is directed from $u$ to another vertex $v$), and similarly $h(v,e) = 1$ iff $v$ is the head of the edge. Note this allows multiple edges from $u$ to $v$. We will assume that $|V(D)|$ and $|E(D)|$ are finite throughout this paper, unless explicitly stated otherwise. - A [*simple digraph*]{} is a directed graph $G$ such that for all $u,v \in V(G)$, at most one edge from $u$ to $v$ appears in $E(G)$ (i.e. no parallel directed edges). - In a digraph $G$, for $u,v \in V(G)$, the distance [ $d(u,v)$]{} from $u$ to $v$ is the length of the shortest directed path from $u$ to $v$. We say $v$ is at [*out-distance*]{} $d(u,v)$ from $u$, and $u$ at [*in-distance*]{} $d(u,v)$ from $v$. - For integers $j > 0$, [ $N_j^+(v)$]{} is the set of vertices at out-distance exactly $j$ from $v$, and [ $N_j^-(v)$]{} is the set of vertices at in-distance exactly $j$ from $v$. We may abbreviate $N_1^+(v)$ and $N_1^-(v)$ to $N^+(v)$ and $N^-(v)$, respectively. - In a digraph $G$, [ $\delta_G^+$ and $\delta_G^-$ ]{} denote the minimum out-degree and in-degree of $G$, respectively. For a given vertex $v$, [ $\delta^+_G(v), \delta^-_G(v)$]{} denote the out-degree and in-degree of the vertex $v$. In cases where the graph being referenced is clear, we may write $\delta^+(v)$ and $\delta^-(v)$. The Caccetta-Häggkvist Conjecture ================================= \[CH\] (L. Caccetta, R. Häggkvist  [@caccettahaggkvist]) Every simple $n$-vertex digraph with minimum out-degree at least $r$ has a cycle with length at most $\lceil \frac{n}{r} \rceil$. This can be restated in the following way: Let $A$ be an $n \times n$ $0$-$1$ matrix such that $a_{ij} = 1$ implies $a_{ji} \neq 1$ for all $i \neq j$. Let $a_{ii}=1$ for all $i$. If the sum of every row in $A$ is at least $r+1$, $A^{\lceil n/r \rceil}$ has trace greater than $n$. Partial Results --------------- The C-H conjecture has been proved for: - $r = 2$ by Caccetta and Häggkvist  [@caccettahaggkvist] - $r = 3$ by Hamidoune  [@hamidouneCH] - $r = 4$ and $r = 5$ by Hoáng and Reed  [@hoangreed] - $r \leq \sqrt{n/2}$ by Shen  [@shen:dm2000]. For the exact statement of his result, see Theorem \[shensqrt\]. This shows that for any given $r$, the number of counterexamples to the conjecture (if any) is finite. - Cayley graphs (which implies all vertex transitive graphs using coset representations) by Hamidoune  [@hamidouneCayley]. This proof uses a lemma of Kemperman  [@kemperman] (Lemma  \[kempermanlemma\]). Also, Shen  [@shen:ejc2003] proved that if $deg^+(u)+deg^+(v) \geq 4$ for all $(u,v) \in E(G)$, then $g \leq \lceil n/2 \rceil$, where $g$ denotes the girth of $G$. This is an average local outdegree version for the $r=2$ case of the Caccetta-Häggkvist conjecture. Approximate Results I - Additive Constant ----------------------------------------- Another approach is to show that if $\delta^+_G \geq r$, then there is a cycle of length at most $\frac{n}{r} + c$ for some small $c$. This has been proved for some values of $c$, as follows: - $c = 2500$ by Chvátal and Szemerédi  [@chvatalszemeredi] - $c = 304$ by Nishimura  [@nishimura] - $c = 73$ by Shen  [@shen:gc2002]. Approximate Results II - Special Case n/3 ----------------------------------------- The case $r = n/2$ is trivial, but $r = n/3$ has received much attention. Research has sought the minimum constant $c$ such that $\delta^+_G \geq cn$ in an $n$-vertex simple digraph $G$ forces a directed cycle of length at most 3. The conjecture is that $c = 1/3$, and the current results are: - $c \leq (3-\sqrt{5})/2 = 0.382$ by Caccetta and Häggkvist  [@caccettahaggkvist] - $c \leq (2\sqrt{6}-3)/5 = 0.3797$ by Bondy in a neat subgraph counting argument  [@bondy] - $c \leq 3 - \sqrt{7} = 0.3542$ by  [@shen:jct1998] Similarly, Seymour, Graaf, and Schrijver  [@seymourgraafschrijver] asked for the minimum value of $\beta$ so that when the minimum in- and out-degrees of $G$ are at least $\beta n$, $G$ has a directed cycle of length at most $3$. They proved that $\beta \leq 0.3487$ and gave a formula relating $\beta$ and $c$. Shen applied this formula to his 1998 result to get a slight improvement to $\beta \leq 0.3477$  [@shen:jct1998]. Seymour’s Second Neighborhood Conjecture ======================================== This conjecture implies the special case of Caccetta-Häggkvist when both in- and out-degrees are at least $n/3$, and has received much attention of its own. (Seymour) Any simple digraph with no loops or digons has a vertex $v$ whose second neighborhood is at least as big as its first neighborhood, i.e. $|N^+_2(v)| \geq |N^+(v)|$. The following is known for Seymour’s second neighborhood conjecture: - When G is a tournament, this is Dean’s conjecture, and was proved by Fisher  [@fisher] using probabilistic methods. There is also a combinatorial proof by Havet and Thomassé  [@havetthomasse]. - It was proved for digraphs with minimum outdegree $\leq 6$ by Kaneko and Locke  [@kanekolocke]. - There is a vertex $v$ where $|N^+_2(v)| \geq \gamma|N^+(v)|$ and $\gamma = 0.657298...$ is the unique real root of $2x^3 + x^2 -1=0$. (Note the conjecture is that $\gamma = 1$) by Chen, Shen, and Yuster  [@chenshenyuster]. They also claim a slight improvement to $\gamma = 0.67815$ (proof unpublished). - Godbole, Cole, and Wright  [@godbolecolewright] showed that the conjecture holds for almost all digraphs. r-Regular Digraphs ================== A digraph $G$ is [*$r$-regular* ]{} if every vertex $v$ has $\delta^+_G(v)=\delta^-_G(v) = r$. (Behzad, Chartrand, Wall  [@bhc]) The minimum number of vertices in an $r$-regular digraph $G$ with girth $g$ is $r(g -1)+1$. Behzad, Chartrand and Wall give an example achieving this by placing $r(g-1)+1$ vertices on a circle with each vertex having edges to the next $r$ vertices in clockwise order. The Caccetta-Häggkvist Conjecture is a generalization of this earlier conjecture. The Behzad-Chartrand-Wall conjecture was proved for the following special cases: - $r = 2$ by Behzad  [@behzad] - $r = 3$ by Bermond  [@bermond] - Vertex-transitive graphs by Hamidoune  [@hamidouneVT] - If $\delta^+_G \geq r$, then $g \leq 3\lceil \frac{n}{r}\ln(\frac{2+\sqrt{7}}{3})\rceil \approx \frac{1.312n}{r}$ by Shen  [@shen:gc2002]. Related Results =============== \[shensqrt\] (Shen  [@shen:dm2000]) For a digraph $G$ on $n$ vertices, if $\delta^+_G \geq r$, and $n \geq 2r^2 - 3r + 1$, then $G$ has a cycle of length at most $\lceil \frac{n}{r} \rceil$. In a graph $G$, for $u,v \in V(G)$, $\kappa(u,v)$ denotes the maximum number of internally disjoint paths between $u$ and $v$. If $G$ is a digraph, $\kappa$ counts the maximum number of internally disjoint directed paths from $u$ to $v$. In a graph $G$, for $u,v \in V(G)$, [ $\lambda(u,v)$]{} denotes the maximum number of edge-disjoint paths between $u$ and $v$. If $G$ is a digraph, $\lambda$ counts the maximum number of edge-disjoint directed paths from $u$ to $v$. \[Thom85\] (Thomassen  [@thomassen]) For all positive integers $r$, there is a digraph $D$ without digons with $\delta^+_D \geq r$ and $\delta^-_D \geq r$ such that: 1. no vertex $v \in V(D)$ is contained in three openly disjoint circuits (that is, three circuits which pairwise share only $v$) 2. no edge $(x,y) \in E(D)$ has $\kappa(y,x) \geq 3$. (Mader  [@mader95]) For every integer $k \geq 0$, If $G$ has $|V(G)| > k^2(k+1)$ and at most $k^2(k+1)$ vertices of $G$ have out-degree at most $k^3(k+1)$, then there are vertices $x \neq y$ so that $\kappa(x,y) > k$. Undirected Graph Theorems ------------------------- (Mader  [@mader71]) Every graph $G$ with $\delta_G \geq r$ contains vertices $x,y$ with $\kappa(x,y) \geq r$ when $r \geq 1$. (Mader  [@mader74]) Every graph $G$ with $\delta_G \geq r$ contains $r+1$ vertices $v_1, \dots, v_{r+1}$ with $\lambda(v_i, v_j) \geq r$ for all $i \neq j$. Additive Number Theory Results ------------------------------ Given an additive group $\Gamma$, and sets $A,B \subseteq \Gamma$, let $A+B$ $ := \{a + b \;|\; a \in A, b \in B\}$, and $A \hat{+} B$ $ := \{a + b \;|\; a \in A, b \in B, a\neq b\}$. Finally, for a positive integer $r$, let $rB$ $:= \{b_1 + \cdots + b_h |$ all $ b_i \in B$, not necessarily distinct$\}$. (Cauchy  [@cauchy] and Davenport  [@davenport], [@davenport2]) Let $p$ be a prime, and $A,B \subseteq \Z/p\Z$ be nonempty. Then $|A+B| \geq \min(p, |A|+|B|-1)$. (I. Chowla  [@chowla]) Let $m$ be a positive integer, and $A,B \subseteq \Z/m\Z$ such that $0 \in B$ and $gcd(b,m) = 1$ for all nonzero $b \in B$. Then $|A+B| \geq \min(m, |A|+|B|-1)$. (Dias de Silva and Hamidoune  [@desilvahamidoune]) \[erdosheilbronn\] The Erdös-Heilbronn Conjecture: Let $A \subseteq \Z/p\Z$, with $p$ prime. Then $|A \hat{+} A| \geq \min(2|A|-3, p)$. For a multiplicative group $\Gamma$ and sets $A,B \subseteq \Gamma$, let [ $AB$]{} $:= \{ab \;|\; a\in A, b\in B\}$. \[kempermanlemma\] (Kemperman  [@kemperman]) Given a group $\Gamma$ and finite non-empty subsets $A,B \subseteq \Gamma$, if $1 \in A, B$ but $(1,1)$ is the only pair $(a,b)$ with $a \in A, b \in B$ such that $ab=1$, then $|AB| \geq |A| + |B| -1$. We say $G$ is a [*layered digraph*]{} if $G$ is a digraph with $V(G) = \cup_{i=0}^{h} V_i$ with $V_i \neq \emptyset$, and $V_i \cap V_j = \emptyset$ for all $i \neq j $, and $(u,v) \in E(G)$ implies $u \in V_{i-1}, v \in V_i$ for some $i \in \{1,\dots,h\}$. A [*Plünnecke graph*]{} is a layered digraph $G$ with the following two properties: 1. If $u,v,w_1, \dots, w_k$ are vertices of $G$ with $(u,v), (v,w_1), \dots, (v,w_k) \in E(G)$, then there are distinct vertices $v_1, \dots, v_k$ so that $(u,v_i), (v_i, w_i) \in E(G)$ for $i=1,\dots,k$. 2. If $v,w, u_1, \dots, u_k$ are vertices of $G$ with $(v,w), (u_1,v), \dots, (u_k,v) \in E(G)$, then there are distinct vertices $v_1, \dots, v_k$ so that $(u_i,v_i), (v_i, w) \in E(G)$ for $i=1,\dots,k$. Let $G$ be a digraph, and $X,Y$ nonempty subsets of $V(G)$. Then\ [*Im(X,Y)*]{} $:=\{y \in Y |$ there is a directed path from $X$ to $y\}$. The [*magnification ratio D(X,Y)*]{} is $$D(X,Y) := \min_{Z\subseteq X, Z \neq \emptyset} \left\{\frac{|Im(Z,Y)|}{|Z|}\right\}.$$ (Plünnecke  [@plunnecke]) In a Plünnecke graph, let $D_i = D(V_0, V_i)$. Then $$D_1 \geq D_2^{\frac{1}{2}} \geq \cdots \geq D_h^{\frac{1}{h}}.$$ The following are consequences of applying Plünnecke’s Inequalities to a special graph created from subsets $A,B$ of a group: For sets $A,B \subset \Gamma$: 1. $|iB|^{\frac{1}{i}} \geq |hB|^{\frac{1}{h}}$ for all $0 \leq i < h$. 2. If $|B| = k$, and $|B+B| \leq ck$, then $|hB| \leq c^hk$. 3. If $|A| = n$, and $|A+B| < cn$ then for all $k, \ell \in \Z^+$, we have that $|kB-\ell B| \leq c^{k+\ell}n$ where $kB-\ell B$ denotes the set of all elements expressable as $(b_1 + \cdots + b_k) - (b_1'+ \cdots + b_{\ell}')$ where all $b_i, b_i'$ are in $B$. (generalization of Erdös-Heilbronn (Thm. \[erdosheilbronn\])) Let $A,B \subseteq \Z/p\Z$, with $p$ prime and $|A| \neq |B|$. Let $C = A \hat{+} B$. Then $|C| \geq \min(|A| + |B| - 2, p)$. Open Problems and Conjectures ============================= Rainbow Conjectures ------------------- ### A Colored Generalization of Seymour’s Second Neighborhood Given a digraph $G = (V,E)$ with each edge $e \in E$ having a set $S_e$ of labels in $\{1,2,\dots,k_G\}$, a [*rainbow*]{} structure $H$ in $G$ (such as a path or cycle) means one in which there is a way to assign each edge $e \in E(H)$ a label $\ell (e) \in S_e$ so that $ell (e) \neq \ell (f)$ for all edges $e \neq f$ in $E(H)$. \[colorconj\] (Seymour, Sullivan) Let $G$ be a simple digraph on the vertex set $V$, and $E_1, \dots, E_k \subseteq E(G)$. Say an edge $e \in E(G)$ has label set $S_e \subseteq \{1,\dots,k\}$ where $i \in S_e$ if and only if $e \in E_i$. Finally, let $G_i = (V, E_i)$. 1. There exists a rainbow (di)cycle in $G$ or 2. There exists a vertex $v$ such that $|\{w \,|\,$there exists a rainbow path from $v$ to $w\}| \geq \sum_{i=1}^{k} \delta_{G_i}^+(v)$. This is false if you require that the colors appear in an increasing order (cyclic on the cycle). We have been able to show that this conjecture holds when $G_1,\dots, G_k$ are Cayley graphs on a common group $\Gamma$ (using induction on Lemma  \[kempermanlemma\]), and when $\delta_{G_i}^+(v) \leq 1$ for all $v$ and all $i$ except $i = 1$, where we allow the outdegrees to be unbounded (but finite). ### Implications of Conjecture \[colorconj\] Seymour’s Second Neighborhood Conjecture To see this, for a digraph $H$, let $k=2, G=H$, and $E_1=E_2=E(H)$. Caccetta-Häggkvist Conjecture (general case) For a digraph $H$, take $G=H$, $k = \lceil \frac{n}{\delta^+_H} \rceil$, and $E_1=\dots=E_k=E(H)$. We must get a rainbow cycle because the sum of the outdegrees at each vertex is $\geq n$. This corresponds to a dicycle of length at most $\lceil \frac{n}{\delta^+_H} \rceil$ in $H$, as desired. \[secondinconj\] Any simple digraph with no loops or digons has a vertex $v$ such that $|N^+_2(v)| + |N^{+}(v)| \geq 2|N^{-}(v)|$. Recall, for comparison, SSN can be written as $|N^+_2(v)| + |N^{+}(v)| \geq 2|N^{+}(v)|$. To see how Conjecture \[colorconj\] implies \[secondinconj\], take a simple digraph $H$ with no loops or digons, and set $G=H$ and $E_1 = E_2 = E(H)$. $G$ cannot have a rainbow cycle by definition of $H$. Define $N^{+*}_G(u) = \{$vertices you can reach by a rainbow path in $G$ from $u\}$, and $N^{-*}_G(u) = \{$vertices that have a rainbow path in $G$ to $u\}$. Let $E_3$ be the edges $\{(u,v)\;|\; v$ is not in the set $N^{-*}_G(u)\}$. Let $G' = H$, and have subsets $E_1, E_2,E_3 \subseteq E(G')$ giving rise to label sets $S'_e \subseteq \{1,2,3\}$ for $e \in G'$. We can see from these definitions that for any $u$ in $V(G')$, $$\sum_{i=1}^{3} \delta_{G_i}^+(u) = 2|N^+(u)| + ((n-1) - |N^-(u)| - |N^-_2(u)|),$$ where all neighborhoods referenced on the RHS are in $H$. We can rewrite this as: $$\sum_{i=1}^{3} \delta_{G_i}^+(u) = ((n-1) - (|N^-(u)| + |N^-_2(u)|-2|N^+(u)|)).$$ Then $\sum_{i=1}^{3} \delta_{G_i}^+(u) \geq n$ whenever $|N^-(u)| + |N^-_2(u)| < 2|N^+(u)|.$ If this were true for all vertices $u$, then no vertex could have $|N^{+*}_{G'}(u)| \geq \sum_{i=1}^{3} \delta_{G_i}^+(u) \geq n$, so we must have a rainbow cycle in $G'$, by Conjecture \[colorconj\]. However, by construction, since $G$ has no rainbow cycle, $G'$ has no rainbow cycle. Thus there is a vertex $v \in V(H)$ so that $|N^-(u)| + |N^-_2(u)| \geq 2|N^+(u)|.$ If we reverse all edges in $H$, this gives $|N^+(u)| + |N^+_2(u)| \geq 2|N^-(u)|,$ as claimed. \[forcecycleconj\] (Seymour) Under the hypotheses of Conjecture \[colorconj\], if $|V| = d$ and $\sum_{i=1}^{k} \delta_{G_i}(v) \geq d$ for all $v$, $G$ must have a rainbow cycle. This conjecture is false if $|V| =d$ is replaced by $|V|=d+1$. ### Other Conjectures Inspired by (or related to) Conjecture \[colorconj\] If we believe Seymour’s second neighborhood conjecture and Conjecture \[colorconj\] (specifically \[secondinconj\]), we might be led to ask if the following holds: (“Compromise Conjecture”) Any simple digraph with no loops or digons has a vertex $v$ such that $|N^+_2(v)| \geq |N^{-}(v)|$. Any simple digraph with no loops or digons has a vertex $v$ such that $|N^+_2(v)|+|N^+(v)| \geq 2\min(|N^{-}(v)|, |N^{+}(v)|)$. Under the hypotheses of Conjecture \[colorconj\], if $\delta^+_{G_i} \geq r_i$, and $\sum_{i=1}^{t} r_i \geq |V|$, there is a rainbow cycle in $G$. Under the hypotheses of Conjecture \[colorconj\], if $\sum_{i=1}^{t} \delta^+_{G_i}(v) \geq |V|$ for all vertices $v$, there is a rainbow cycle in $G$. (Devos) Under the hypotheses of Conjecture  \[colorconj\], if there is no rainbow cycle in $G$ with strictly increasing edge labels, then the average number of vertices reachable from a fixed vertex $v$ by label-increasing (possibly trivial) paths is at least $1+\sum_{i=1}^k \delta^+_{G_i}$. This can be proved when $G_i =$ Cayley$(\Gamma,A_i)$ for some group $\Gamma$, using Lemma  \[kempermanlemma\] Second & $K^{th}$ Neighborhood Conjectures ------------------------------------------ Is Seymour’s Second Neighborhood true for locally finite digraphs? What if we just require the outdegrees to be finite? (Thomassé) Let $G$ be a digraph with no directed cycle of length at most three. Then there is a vertex $v \in V(G)$ with $\delta^+(v)$ at most the number of non-neighbors of $v$. The following generalization of second neighborhood to $k$th neighborhood was taken from a (unpublished) paper of Serge Burckel: Any simple digraph with no directed cycles of length at most $k$ has a vertex $v$ such that $|N^+_k(v)| \geq |N^+_{k-1}(v)|$. Serge Burckel also asked the following structural question: For any $k$ and any digraph $G$, define $G^*$ to be those vertices with $|N^+_2(v)| \geq |N^+(v)|$. Then any vertex of out-degree $k$ is at distance at most $k$ from a vertex in $G^*$. He motivates this with the following remark: “If a vertex $x$ has one successor $y$, then if $y$ has no successors, $y \in G^*$, otherwise $x \in G^*$. This property seems to generalize for any out-degree, and if it is true, is optimal considering ’pyramids’ where any vertex of out-degree $k$ is at distance exactly $k$ from the (unique) solution.” His ’pyramids’ are formed by placing one vertex, then two in the row beneath it, and so forth ($i$ vertices in row $i$) to form a triangle. The vertices in row $i$ are then completely joined to all vertices in row $i-1$ for $i \geq 2$. (Seymour &/or Jackson) If $G$ is an Eulerian digraph with no loops or digons, then $$\sum_{v\in V(G)} |N^+_2(v)| \geq \sum_{v \in V(G)} |N^+(v)|.$$ (Thomassé & Kral) Let $G$ be an Eulerian digraph on $n$ vertices, so that $|E(G)| \geq n^2/3$. Then $G$ has a directed cycle of length at most three. Matrices -------- The following were all presented by Seymour, with no other attributions given: For the following questions, matrices are assumed to be $n \times n$ $0$-$1$ matrices such that $a_{ij} = 1$ implies $a_{ji} \neq 1$, and all diagonal elements equal to $1$. 1. Let $A_1, A_2, \dots, A_{\lceil n/r \rceil}$ be matrices (not necessarily distinct) so that the row sums of $A_i$ are at least $r+1$ for all $i$. Does $A_1A_2\cdots A_{\lceil n/r \rceil}$ have trace $>n$? This is a special case of Conjecture \[forcecycleconj\]. 2. Let $A_1, A_2, \dots, A_t$ be matrices (not necessarily distinct) so $A_i$ has row sums at least $r_i + 1$ and $\sum_{i=1}^t r_i \geq n$. Does $A_1A_2\cdots A_t$ have trace $>n$? This is equivalent to Conjecture \[forcecycleconj\]. 3. Form a digraph from the matrices $A_1, \dots, A_t$ by putting $t+1$ copies of the vertex set $V$ in a row, connecting copy $k$ of $v$ to copy $k+1$ of $v$ with a horizontal edge for all vertices $v$ and all $k$, and then from copy $k$ to copy $k+1$ put in the edge from copy $k$ of $v_i$ to copy $k+1$ of $v_j$ precisely when $a_{ij} = 1$ in matrix $A_k$ for $k=1,\dots,t$. The question now is whether there exists a vertex $u$ so that there is a non-trivial (i.e. not all horizontal edges) path from copy $1$ of $u$ to copy $t+1$ of $u$. 4. If we “squish” all the bipartite graphs from the previous items so they live on a single copy of $V$, marking edges from copy $i$ to copy $i+1$ with color $i$, then we have the sum of the (colored) outdegrees at each vertex is at least $n$, and we’re asking for a non-trivial rainbow cycle which has the colors appearing in increasing order. For a matrix $A$, the [*spectral radius of A*]{} is defined to be $\max \{ |\lambda| \; : \; \lambda$ an eigenvalue of $A\}$. (Charbit) Let $A$ be the adjacency matrix of a digraph $G$ with zero on the diagonal and $a_{ij}=1$ if and only if the edge $(i,j) \in E(G)$. If the spectral radius of $A$ is at least $n/k$, then $G$ has a cycle of length $\leq k$. Disjoint Cycles --------------- \[bermondthomassen\] (Bermond-Thomassen) In a digraph $D$ with $\delta^+_D \geq 2k-1$, there are $k$ vertex disjoint cycles. Open for $k \geq 3$, and the proof for $k=2$ by Thomassen is not intuitive. (Hoáng & Reed) \[hoangreed\] If $G$ is a digraph with minimum outdegree $r$, then there are directed cycles $C_1, \dots, C_r$ such that for all $\ell$, $$|V(C_{\ell}) \cap (\cup_{i=1}^{\ell-1}V(C_i))| \leq 1.$$ The related conjecture that given minimum outdegree at least $r$, there should be a vertex $v$ with $r$ cycles through $v$ which are otherwise vertex-disjoint is false. A counterexample for $r=3$ was given by Thomassen in 1985 (see Theorem \[Thom85\] for his complete result). Adding the condition that the minimum indegree is also at least $r$ does not improve the veracity of the conjecture, though it is still open when indegree and outdegree are identically $r$ everywhere. Connectivity ------------ (Hamidoune, 1981) Let $D$ be a digraph with $\delta^+_D \geq r$, $\delta^-_D \geq r$, and $r \geq 1$. Then there is an edge $(x,y)$ such that $\kappa(y,x) \geq r$. Note that a counterexample to the above conjecture for all $r \geq 3$ is given by Thomassen in Theorem \[Thom85\]. (Mader) If a digraph $D$ has $\delta^+_D \geq r$, then there are vertices $x \neq y$ such that $\lambda(x,y) \geq r$. Also, there is an edge $(x,y)$ so that $\lambda(x,y) \geq r$. The first part of this conjecture was proven by Mader [@mader85] for $\lambda(x,y) \geq r-1$. Weighted Versions ----------------- (Bollobás & Scott) \[bollobasscott\] Let $p:E(G) \rightarrow [0,1]$. If $\sum_{v\in N^+_G(u)} p(uv) \geq 1$ and $\sum_{v\in N^-_G(u)} p(vu) \geq 1$ for all $u \in V(G)$, there is a directed cycle in $G$ of total weight $\geq 1$. There is a nice proof that there is a dipath of total weight at least $1$. (Zhang) Let $G$ be a digraph on $n$ vertices, and $f: E(G) \rightarrow \{0,1,\dots\}$. If $\sum_{e \in E^+(v)} f(e) \geq n/k$ for all $v \in V(G)$, then there is a directed cycle $C$ such that $\sum_{e \in C} \frac{1}{f(e)} \leq k$. (Charbit) Take the directed Cayley graph $G$ with group $G = \Z_8$, and generators $\{1,2\}$. Let the weight function $f$ be $1$ on 2-edges, and $2$ on 1-edges (where r-edges are those coming from the generator r). Now let $k =3$. We can calculate $\sum_{e \in E^+(v)} f(e) = 4 \geq 8/3$ for all vertices $v$, but there is no directed cycle with $\sum_{e \in C} \frac{1}{f(e)} \leq 3$. Averaged Outdegree Conditions ----------------------------- If $D$ is a digraph on $n$ vertices with $$\sum_{v\in V(D)} \log(1+\frac{1}{\delta^+_D(v)}) \geq n\log(1+\frac{1}{r}),$$ then $D$ has a cycle of length at most $\lceil n/r \rceil$. Take a transitive tournament of size $n-1$, and replace the edge from the vertex of out-degree $n-2$ to the vertex of out-degree zero with a path of length 2 in the opposite direction (thus increasing the number of vertices to $n$). (Shen  [@shen:ejc2003]) Let $G$ be a digraph with $n$ vertices and minimum outdegree at least one. If $\delta^+_G(v)+ \delta^+_G(u) \geq 2r$ for every edge $(u,v)$ in $G$, then the girth $g$ of $G$ is at most $\lceil n/r \rceil$. This was proved by Shen for $r=2$ in  [@shen:ejc2003]. UnCategorized ------------- The following conjecture would imply that in a counterexample for Caccetta-Häggkvist for $r = n/3$, one can order the vertices so that at least $75\%$ of the edges go from left to right: (Chudnovsky, Seymour, Sullivan) Let $G$ be a simple digraph with $k$ non-edges (unordered pairs $\{u,v\}$ where both $uv$ and $vu$ are not in $E(G)$). If $G$ has no directed cycle of length at most $3$, one can delete at most $k/2$ edges from $G$ and obtain a graph with no directed cycle. We know that there are tight examples for this conjecture (transitive tournaments, $C_4$, and products of these). It is also known that a minimal counterexample has no source or sink vertex, and no directed cut. Kostochka recently proved all vertices in a minimal counterexample have at least $3$ and at most $(n-1)/2$ non-neighbors. He has an argument using these facts to show the conjecture for all $k \leq 14$. (Devos) For any digraph $G$ with no directed cycles of length at most three, there is a probability distribution $p$ on $V(G)$ such that at every vertex $v$, $p(N^+(v)) \geq p(N^-(v))$ and $p(N^-_2(v)) \geq p(N^-(v))$, where $p(S) := \sum_{s\in S} p(s)$ for a set $S \subseteq V(G)$. For tournaments, such a distribution exists, and is unique. Its existence for a general digraph would imply Seymour’s Second Neighborhood Conjecture, and actually it would suffice to just have a probability distribution $p$ so that $p(N^-_2(v)) \geq p(N^-(v))$ on average in $G$. Given a digraph $G = (V,E)$, we say $F \subseteq E(G)$ is a [*feedback arc set*]{} if the digraph $G' = (V, E-F)$ has no directed cycles. (Lichiardopol’s Conjecture) Every digraph $D$ has some minimal feedback arc set $F$ which contains a path of length $\delta^+_D$. This implies Hoáng & Reed (\[hoangreed\]), Caccetta-Häggkvist (\[CH\]), Bermond-Thomassen (\[bermondthomassen\]), and Thomassé (\[thomassegirth\]). (Mader, 1985) For all $k \in \Z^+$, there exists $r \in \Z^+$ such that every $r$-out-regular digraph contains a subdivision of the transitive tournament on $k$ vertices. This conjecture is known for $k = 3$ (where $r = 2$), and $k=4$ ($r=3$ proven by Mader in 1996  [@mader96]). The existence of $r$ for $k =5$ is still not known, though $r=6$ has been conjectured. (Shen) For a digraph $G$ on $n$ vertices of girth $g$, define $$t(G,r) = \sum_{u: \delta^+_G(u) < r} (r-\delta^+_G(u)).$$ If $\delta^+_G \geq 1$, then $n \geq r(g-1) + 1 - t(G,r)$. (Thomassé) If $G$ is a loopless digon-free digraph the maximum number of induced directed 2-edge paths is $n^3/15 + \mathcal{O}(n^2)$. First, note that $n^3/15 +\mathcal{O}(n^2)$ can be obtained by substituting $C_4$’s inside $C_4$’s. Next, given a digraph $G$, let $\tau$ be the number of induced directed 2-paths, $\eta$ the number of induced edges, and $\theta$ the number of cyclic triangles. Then if $|V(G)| = n$, $$\tau + \eta/3 + 2\theta = n^3/12 + \mathcal{O}(n^2) - \frac{1}{6}\sum_{v\in V(G)}\left( (\delta^+(v) - \delta^-(v))^2 + (\frac{n}{2}-\delta^+(v))^2 + (\frac{n}{2} -\delta^-(v))^2 \right).$$ Füredi rewrote the terms on the right hand side in terms of $\tau, \eta,$ and $\theta$, and showed $$\tau \leq \frac{n^3}{12} + \mathcal{O}(n^2).$$ Bondy has a slight improvement of this result, proving that $$\tau \leq \frac{2n^3}{25}.$$ Define an [ *$\alpha$/$\beta$-digraph* ]{} to be a digraph $D$ on vertex set $V$ and edges defined by $\sigma_1,\dots,\sigma_{\beta}$ permutations (linear orders) on $V$ where the edge $(i,j) \in E \iff i<j$ in at least $\alpha$ of the $\sigma_i$. We say $G$ is a [*majority digraph*]{} if $\frac{\alpha}{\beta} > 2/3$. (Thomassé & Charbit) Caccetta-Häggkvist holds for 3/4-digraphs. Majority digraphs have no cycles of length at most three. The class of 3/4-digraphs is stable under substitution, and contains the circular interval graph on $3k+1$ vertices with outdegree $k$ going clockwise. The class includes all known extremal examples for C-H, yet this class of 3/4-digraphs seems manageably small. It is still open whether or not 3/4 digraphs must have vertices $x$ of outdegree less than $n/3$. One can more generally ask if Caccetta-Häggkvist holds for larger classes of majority digraphs. (Thomassé) \[thomassegirth\] Every digraph $D$ has a path of length $\delta^+_D(g-1)$, where $g$ is the girth of $D$. This is open for $g = 3$, and implies Caccetta-Häggkvist. (Thomassé) In a digraph $D$, $$\sum_{v\in V(D)} |\delta^+_D(v)-\delta^-_D(v)| + |\{(u,v) | d(u,v) \leq 2\}| \geq 2|E(D)| + |\{v| \delta^+_D(v) > \delta^-_D(v)\}|.$$ This is exact for transitive tournaments. (Thomassé) Let $G$ be a digraph on $n$ vertices with minimum outdegree at least $4n/15$ so that $G$ is maximal with no cycles of length at most three, and has no homogeneous set (in other words, G cannot be obtained by substitution). Then $G$ is a Cayley graph on $3k+1$ vertices with $1,\dots,k$ as generators for some $k$ (i.e. the circular interval graph with everyone joined to next $k$ clockwise). $\Z/15\Z$ with generators $S=\{1,2,4,8\}$ gives exactly $4n/15$. Acknowledgements {#acknowledgements .unnumbered} ---------------- This research was performed while on appointment as a U.S. Department of Homeland Security (DHS) Fellow under the DHS Scholarship and Fellowship Program, a program administered by the Oak Ridge Institute for Science and Education (ORISE) for DHS through an interagency agreement with the U.S Department of Energy (DOE). ORISE is managed by Oak Ridge Associated Universities under DOE contract number DE-AC05-06OR23100. All opinions expressed in this paper are the author’s and do not necessarily reflect the policies and views of DHS, DOE, or ORISE.
{ "pile_set_name": "ArXiv" }
--- abstract: 'To a Coxeter group $W$ one associates a quandle $X_W$ from which one constructs a group ${\mathrm{Ad}}(X_W)$. This group turns out to be an intermediate object between $W$ and the associated Artin group. By using a result of Akita, we prove that ${\mathrm{Ad}}(X_W)$ is given by a pullback involving $W$, and by using this pullback, we show that the classifying space of ${\mathrm{Ad}}(X_W)$ is given by a space called a polyhedral product whenever $W$ is right-angled. Two applications of this description of the classifying space are given.' address: 'Department of Mathematics, Kyoto University, Kyoto, 606-8502, Japan' author: - Daisuke Kishimoto title: | Right-angled Coxeter quandles\ and polyhedral products --- .525cm Introduction ============ A quandle is a set with a binary operation satisfying three conditions. Since these three conditions are thought of as algebraic abstraction of the Reidemeister moves of knots, a quandle has been intensively studied in low dimensional topology. The three conditions are also thought of as axiomatization of conjugation in a group, so it has been studied in representation theory as well. In this paper, we comprehend a quandle in the latter sense. We look at the following connection between groups and quandles. Any conjugation closed subset $X$ of a group $G$ can be regarded as a quandle by a binary operation given by conjugation, and to any quandle $Y$ one associates a group ${\mathrm{Ad}}(Y)$ which is called the adjoint group. Thus any conjugation closed subset $X$ of a group yields a new group ${\mathrm{Ad}}(X)$. We now consider a Coxeter group $W$. The set of reflections $X_W$ of $W$ is closed under conjugation, so $X_W$ is a quandle which we call the Coxeter quandle associated with $W$. Then we get a group ${\mathrm{Ad}}(X_W)$ which is the object to study in this paper. The adjoint group ${\mathrm{Ad}}(X_W)$ has been studied mainly in connection with representation theory [@AFGV; @E], and there are few results on its topology [@N]. This paper studies the topology of ${\mathrm{Ad}}(X_W)$ by applying the recent result of Akita [@A], and in particular, we show that if $W$ is right-angled, then the classifying space of ${\mathrm{Ad}}(X_W)$ is given by a space which is called a polyhedral product. The following fundamental property of ${\mathrm{Ad}}(X_W)$ suggests that ${\mathrm{Ad}}(X_W)$ possibly gives a new direction for the study of Coxeter groups and Artin groups. The symmetric group $\Sigma_n$ of $n$ letters is a Coxeter group and its associated Artin group is the braid group $B_n$ of $n$ strands. Then there is a natural epimorphism $B_n\to\Sigma_n$. In [@AFGV], it is shown that ${\mathrm{Ad}}(X_{\Sigma_n})$ is an intermediate object between $\Sigma_n$ and $B_n$ in the sense that the epimorphism $B_n\to\Sigma_n$ factors as the composite of epimorphisms $B_n\to{\mathrm{Ad}}(X_{\Sigma_n})\to\Sigma_n$. Akita generalized this result to an arbitrary Coxeter group. Let $A_W$ be the Artin group associated with a Coxeter group $W$. \[pi\] For an arbitrary Coxeter group $W$, the epimorphism $A_W\to W$ factors as the composite of epimorphisms $$A_W\to{\mathrm{Ad}}(X_W)\to W.$$ Akita [@A] studied further structures of ${\mathrm{Ad}}(X_W)$ and generalized Eisermann’s result [@E] on ${\mathrm{Ad}}(X_{\Sigma_n})$ concerning abelianization. From this, we will deduce that ${\mathrm{Ad}}(X_W)$ is given by a certain pullback involving $W$ and its abelianization. When $W$ is right-angled, it is known that the classifying spaces of $W$ and $A_W$ are given by polyhedral products. Then it is natural to ask whether the classifying space of ${\mathrm{Ad}}(X_W)$ is a polyhedral product or not, if $W$ is right-angled. We will give an affirmative answer to this question by using the above pullback description of ${\mathrm{Ad}}(X_W)$. Then we will showe two application of this description of the classifying space: a stable splitting of the classifying space of ${\mathrm{Ad}}(X_W)$ and calculation of the mod 2 cohomology of ${\mathrm{Ad}}(X_W)$. *Acknowledgement:* The author is grateful to Toshiyuki Akito for explaining his work [@A] and giving useful comments. Thanks also go to Takefumi Nosaka for comments and to Ye Liu and Mentor Stafa for careful reading of the first draft. The author was partly supported by JSPS KAKENHI (No.17K05248). Recollection on Coxeter groups ============================== Recall that a pair $(W,S)$ of a group $W$ and a set $S$ is called a Coxeter system if there is given a map $m\colon S\times S\to\mathbb{N}\cup\{\infty\}$, called the Coxeter matrix, satisfying the following conditions: 1. $m(s,t)=m(t,s)$ for any $s,t$; 2. $m(s,t)=1$ if and only if $s=t$; 3. $W$ is defined by the presentation $$W=\langle s\in S\,\vert\,(st)^{m(s,t)}=1\text{ for }m(s,t)<\infty\rangle.$$ We call the group $W$ a Coxeter group, by which we often mean the Coxeter system $(W,S)$ too. The Artin group associated with a Coxeter system $(W,S)$ is defined by $$A_W=\langle a_s\;(s\in S)\,\vert\,\underbrace{a_sa_ta_s\cdots}_{m(s,t)}=\underbrace{a_ta_sa_t\cdots}_{m(t.s)}\text{ for }2\le m(s,t)<\infty\rangle.$$ Since the Coxeter group is alnternatively presented as $$W=\langle s\in S\,\vert\,\underbrace{sts\cdots}_{m(s,t)}=\underbrace{tst\cdots}_{m(t.s)}\text{ for }m(s,t)<\infty\rangle,$$ one gets: For a Coxeter system $(W,S)$, the assignment $$\pi\colon A_W\to W,\quad a_s\mapsto s\quad(s\in S)$$ is a well-defined epimorphism. The symmetric group of $n$ letters $\Sigma_n$ is a Coxeter group with the generating set $\{\sigma_1,\ldots,\sigma_{n-1}\}$, where $\sigma_i$ is the transposition $(i\;i+1)$. Indeed, $\Sigma_n$ is generated by $\sigma_1,\ldots,\sigma_{n-1}$ subject to the relations $\sigma_i^2=1$, $\sigma_i\sigma_{i+1}\sigma_i=\sigma_{i+1}\sigma_i\sigma_{i+1}$ for $i=1,\ldots,n-1$ and $\sigma_i\sigma_j=\sigma_j\sigma_i$ for $|i-j|\ge 2$. Then the associated Artin group is the braid group with $n$ strands $B_n$ and the map $\pi\colon B_n\to\Sigma_n$ maps each braid to a permutation given by its ends. Let $\mathcal{R}_W$ be the complete set of representatives of $S/\sim$, where $\sim$ is given by conjugation by elements of $W$, and let $c(W)$ be the cardinality of $\mathcal{R}_W$. For instance, we have $c(\Sigma_n)=1$. Let $G_\mathrm{ab}$ denote the abelianization of a group $G$. As in [@BMMN], one has: \[abelianization\] There are isomorphisms $W_\mathrm{ab}\cong({\mathbb{Z}}/2)^{c(W)}$ and $(A_W)_\mathrm{ab}\cong{\mathbb{Z}}^{c(W)}$. A Coxeter system $(W,S)$ is called right-angled if the Coxeter matrix $m$ satisfies $m(s,t)=1,2,\infty$ for any $s,t\in S$. By the definition of $c(W)$, one gets: If a Coxeter system $(W,S)$ is right-angled, then $c(W)=|S|$. For a Coxeter system $(W,S)$, we define a graph $\Gamma_W$ such that the vertex set is $S$ and vertices $s,t\in S$ are joined by an edge whenever $2\le m(s,t)<\infty$. Then if a Coxeter group $W$ is right-angled, all the information of $W$ is included in the graph $\Gamma_W$. We will see this more precisely below in terms of a graph product of groups. The adjoint group of a Coxeter quandle ====================================== A quandle is a set $X$ with a binary operation $*\colon X\times X\to X$ satisfying the three conditions: 1. $x*x=x$; 2. $(x*y)*z=(x*z)*(y*z)$; 3. the map $X\to X$, $x\mapsto x*y$ is bijective for any $y\in X$. A quandle is related with group theory (and representation theory) as follows. For a group $G$, we put $x*y=y^{-1}xy$ for $x,y\in G$. Then one can easily check the three conditions of a quandle so that $G$ is a quandle with this binary operation. More generally, any conjugation closed subset of a group can be regarded as a quandle in the same way. Motivated by the above construction of a quandle from a group, for a quandle $X$, we define a group $${\mathrm{Ad}}(X)=\langle e_x\;(x\in X)\,\vert\,e_{x*y}=e_y^{-1}e_xe_y\rangle$$ which is called the adjoint group of $X$. When $X$ is a conjugation closed subset of a group $G$ regarded as a quandle, ${\mathrm{Ad}}(X)$ is directly related with $G$. \[phi\] For a conjugation closed subset $X$ of a group $G$, the assignment $$\phi\colon{\mathrm{Ad}}(X)\to G,\quad e_x\mapsto x\quad(x\in X)$$ is a well-defined homomorphism. Moreover, if $X$ generates $G$, then $\phi$ is surjective. For $x,y\in X$, we have $\phi(e_{x*y})=x*y=y^{-1}xy=\phi(e_y^{-1}e_xe_y)$, implying that the map $\phi$ is a well-defined homomorphism. The remaining statement is obvious. Let $(W,S)$ be a Coxeter system. An element of $W$ of the form $w^{-1}sw$ for $w\in W$ and $s\in S$ is called a reflection of $W$. Then the set of reflections of $W$, denoted $X_W$, is a conjugation closed subset of $W$, so $X_W$ is a quandle which is called the Coxeter quandle associated with a Coxeter system $(W,S)$. Since $W$ is generated by $X_W$, we have the following by Proposition \[phi\]. \[phi\_Coxeter\] For a Coxeter group $W$, the map $\phi\colon{\mathrm{Ad}}(X_W)\to W$ is an epimorphism. Akita [@A] showed that ${\mathrm{Ad}}(X_W)$ is related also with the Artin group $A_W$, where we reproduce its proof. \[Phi\] For an arbitrary Coxeter system $(W,S)$, the assignment $$\Phi\colon A_W\to{\mathrm{Ad}}(X_W),\quad a_s\mapsto e_s\quad(s\in S)$$ is a well-defined epimorphism. Suppose that $m(s,t)=k$ with $2\le k<\infty$ for $s,t\in S$. Then we have $$\underbrace{(\cdots(s*t)*s\cdots)}_k=\underbrace{\cdots s^{-1}t^{-1}}_{k-1}\underbrace{sts\cdots}_k=\underbrace{\cdots s^{-1}t^{-1}}_{k-1}\underbrace{tst\cdots}_k=u$$ where $u=s$ for $n$ even and $u=t$ for $n$ odd, that is, $u$ is the last letter of the word $\underbrace{tst\cdots}_k$. Then it follows that $$\Phi(\underbrace{a_sa_ta_s\cdots}_k)=\underbrace{e_se_te_s\cdots}_k=e_te_{s*t}\underbrace{e_te_s\cdots}_{k-1}=\underbrace{e_te_se_t\cdots}_{k-1}e_{\scriptsize\underbrace{\cdots((s*t)*s)*\cdots}_k}=\underbrace{e_te_se_t\cdots}_k,$$ implying that $\Phi$ is a well-defined homomorphism. To see that $\Phi$ is surjective, we shall show that ${\mathrm{Ad}}(X_W)$ is generated by $e_s$ for $s\in S$. For $s_1,\ldots,s_n,s\in S$ and $w=s_1\cdots s_n$, one has $w^{-1}sw=(\cdots(s*s_1)*s_2\cdots)*s_n$, implying that $e_{w^{-1}sw}=e_{\cdots(s*s_1)*s_2\cdots)*s_n}=e_{s_n}^{-1}\cdots e_{s_1}^{-1}e_se_{s_1}\cdots e_{s_n}$. Thus the proof is completed. Combining Corollary \[phi\_Coxeter\] and Proposition \[Phi\], we obtain Theorem \[pi\]. We now recall the structure theorem of ${\mathrm{Ad}}(X_W)$ due to Akita [@A]. Eisermann [@E] showed that there is a short exact sequence $1\to A_n\to{\mathrm{Ad}}(X_{\Sigma_n})\to{\mathbb{Z}}\to 0$, where $A_n$ is the alternating group of $n$ letters. Note that $A_n\cong[\Sigma_n,\Sigma_n]$ and $c(\Sigma_n)=1$. Akita [@A] generalized this result to an arbitrary Coxeter group. \[Akita\] For any Coxeter group $W$, the following hold: 1. there is an isomorphism ${\mathrm{Ad}}(X_W)_\mathrm{ab}\cong{\mathbb{Z}}^{c(W)}$; 2. the map $\phi\colon{\mathrm{Ad}}(X_W)\to W$ induces an isomorphism $[{\mathrm{Ad}}(X_W),{\mathrm{Ad}}(X_W)]\cong[W,W]$. For any Coxeter group $W$, there is a short exact sequence $$1\to[W,W]\to{\mathrm{Ad}}(X_W)\xrightarrow{\mathrm{ab}}{\mathbb{Z}}^{c(W)}\to 0.$$ By [@BMMN], $W_\mathrm{ab}$ is a ${\mathbb{Z}}/2$-vector space with a basis $\{[x]\in W_\mathrm{ab}\,\vert\,x\in\mathcal{R}_W\}$, from which the first isomorphism of Proposition \[abelianization\] follows. On the other hand, Akita [@A] showed that ${\mathrm{Ad}}(X_W)_\mathrm{ab}$ is a free abelian group with a basis $\{[e_x]\in{\mathrm{Ad}}(X_W)_\mathrm{ab}\,\vert\,x\in\mathcal{R}_W\}$, which yields Theorem \[Akita\] (1). Then we identify the abelianization of the map $\phi\colon{\mathrm{Ad}}(X_W)\to W$ as: The map $\phi_\mathrm{ab}\colon{\mathrm{Ad}}(X_W)_\mathrm{ab}\to W_\mathrm{ab}$ is identified with the canonical projection ${\mathbb{Z}}^{c(W)}\to({\mathbb{Z}}/2)^{c(W)}$. Thus we get a commutative diagram $$\label{pb_Ad} \xymatrix{{\mathrm{Ad}}(X_W)\ar[r]^{\mathrm{ab}}\ar[d]^\phi&{\mathbb{Z}}^{c(W)}\ar[d]^{\mathrm{proj}}\\ W\ar[r]^{\mathrm{ab}}&({\mathbb{Z}}/2)^{c(W)}.}$$ We show that this diagram is a pullback. \[pullback\_lem\] Suppose that there is a commutative square of groups $$\label{pb} \xymatrix{G_1\ar[r]^{f_1}\ar[d]^g&H_1\ar[d]^h\\ G_2\ar[r]^{f_2}&H_2}$$ where $f_1$ are surjective. Then the square is a pullback if and only if the canonical map $\mathrm{Ker}\,f_1\to\mathrm{Ker}\,f_2$ is an isomorphism. Let $G=\{(x,y)\in G_2\times H_1\,\vert\,f_2(x)=h(y)\}$ which is the pullback of the triad $G_2\xrightarrow{f_2}H_2\xleftarrow{h}H_1$. Let $p_1\colon G\to G_2$ and $p_2\colon G\to H_1$ be the projections, and define a map $e\colon G_1\to G$ by $e(x)=(g(x),f_1(x))$ for $x\in G_1$. Then we have $p_1\circ e=g$ and since $f_1$ is surjective, the projection $p_2$ is surjective too. Moreover, the commutative square extends to the following commutative diagram with exact columns and rows $$\xymatrix{1\ar[r]&\mathrm{Ker}\,f_1\ar[r]\ar[d]^{\bar{e}}&G_1\ar[r]^{f_1}\ar[d]^e&H_1\ar[r]\ar@{=}[d]&1\\ 1\ar[r]&\mathrm{Ker}\,p_2\ar[r]\ar[d]^{\bar{p}_1}&G\ar[r]^{p_2}\ar[d]^{p_1}&H_1\ar[r]\ar[d]^h&1\\ 1\ar[r]&\mathrm{Ker}\,f_2\ar[r]&G_2\ar[r]^{f_2}&H_2}$$ The square diagram is a pullback if and only if the map $e\colon G_1\to G$ is an isomorphism. By the above diagram, the latter is equivalent to that the map $\bar{e}$ is an isomorphism. One sees that $\mathrm{Ker}\,p_2\cong\mathrm{Ker}\,f_2$ and the map $\bar{p}_1$ is identified with the identity map. Thus the map $\bar{e}$ is identified with the canonical map $\mathrm{Ker}\,f_1\to\mathrm{Ker}\,f_2$, completing the proof. \[pullback\_thm\] The commutative square is a pullback. Combine Theorem \[Akita\] (2) and Lemma \[pullback\_lem\]. In [@AFGV] it is proved that there is a short exact sequence $0\to{\mathbb{Z}}\to{\mathrm{Ad}}(X_{\Sigma_n})\to\Sigma_n\to 1$, and this was also generalized by Akita [@A] to an arbitrary Coxeter group. We can reprove this by applying Theorem \[pullback\_thm\] \[exact\] For any Coxeter group $W$, there is a short exact sequence $$0\to{\mathbb{Z}}^{c(W)}\to{\mathrm{Ad}}(X_W)\xrightarrow{\phi}W\to 1.$$ By Theorem \[pullback\_thm\], the kernel of $\phi$ is isomorphic to the kernel of the projection ${\mathbb{Z}}^{c(W)}\to({\mathbb{Z}}/2)^{c(W)}$, which is isomorphic to ${\mathbb{Z}}^{c(W)}$. Thus the proof is done. Classifying spaces and polyhedral products ========================================== Let $K$ be a simplicial complex on the vertex set $[m]=\{1,\ldots,m\}$ and $(X,A)$ be topological pair. The polyhedral product of $(X,A)$ with respect to $K$ is defined by $$Z(K;(X,A))=\bigcup_{\sigma\in K}(X,A)^\sigma$$ where $(X,A)^\sigma=Y_1\times\cdots\times Y_m$ such that $Y_i=X,A$ according to $i\in\sigma$ and $i\not\in\sigma$. Note that $Z(K;(X,A))$ is natural with respect to $(X,A)$ and inclusions of simplicial complexes. Polyhedral products were introduced as a generalisation of the moment-angle complex and the Davis-Januszkiewicz space which are fundamental in toric topology [@BBCG; @DJ], and connect algebraic geometry, combinatorics, commutative algebra, geometry, group theory, and topology. Among others, their homotopy theory is rapidly developing (cf. [@IK2]). We will use the following property of polyhedral products, which is immediately deduced from the definition. For $\emptyset\ne I\subset[m]$, the full subcomplex of $K$ on $I$ is defined by $K_I=\{\sigma\in K\,\vert\,\sigma\subset I\}$. \[retract\] For $\emptyset\ne I\subset[m]$, $Z(K_I;(X,A))$ is a retract of $Z(K;(X,A))$. For an acyclic space $X$, the acyclicity of $Z(K;(X,*))$ is completely characterized in terms of $K$. The necessity of the acyclicity of $Z(K;(X,*))$ has been often referred to [@DO] although it is an easy consequence of the old result of Whitehead [@W]. So we here give a simple proof using the result of Whitehead. A simplicial complex $K$ is called flag if $\sigma\subset[m]$ is a simplex of $K$ whenever any two elements of $\sigma$ are joined by an edge of $K$. \[acyclic\] For an acyclic space $X$, $Z(K;(X,*))$ is acyclic if and only if $K$ is flag. Whitehead [@W] proved the following. Suppose that there is a homotopy pushout of path-connected spaces $$\xymatrix{X_{12}\ar[r]^{\alpha_1}\ar[d]^{\alpha_2}&X_1\ar[d]\\ X_2\ar[r]&X}$$ in which $X_1,X_2,X_{12}$ are acyclic and $\alpha_1,\alpha_2$ are injective in $\pi_1$. Then $X$ is acyclic. We apply this to prove the if part of the proposition. For a vertex $v$ of $K$, let $\mathrm{lk}(v)$ and $\mathrm{dl}(v)$ denote the link and the deletion of a vertex $v$ in $K$. Then there is a pushout of simplicial complexes $$\xymatrix{\mathrm{lk}(v)\ar[r]\ar[d]&\mathrm{lk}(v)*\{v\}\ar[d]\\ \mathrm{dl}(v)\ar[r]&K}$$ which induces a homotopy pushout $$\label{po_Z} \xymatrix{Z(\mathrm{lk}(v);(X,*))\ar[r]\ar[d]&Z(\mathrm{lk}(v);(X,*))\times X\ar[d]\\ Z(\mathrm{dl}(v);(X,*))\ar[r]&Z(K;(X,*)).}$$ The upper horizontal arrow is obviously injective in $\pi_1$. $K$ is flag if and only if $K_V=\mathrm{lk}(v)$ for any vertex $v$, where $V$ is the vertex set of $\mathrm{lk}(v)$. Then it follows from Proposition \[retract\] that if $K$ is flag, the left vertical arrow of is injective in $\pi_1$. Moreover, if $K$ is flag, then $K_I$ is flag too for any $\emptyset\ne I\subset[m]$. Then we can apply the above result of Whitehead to inductively on the number of vertices. Thus we obtain that if $K$ is flag, then $Z(K;(X,*))$ is acyclic for an acyclic space $X$. Next we conversely suppose that $Z(K;(X,*))$ is acyclic for an acyclic space $X$. Assume that $K$ is not flag. Then there is $I\subset[m]$ with $|I|\ge 3$ such that $K_I$ is the boundary of the $(|I|-1)$-dimensional simplex. Recall from [@IK2] that there are a homotopy fibration $Z(K_I;(C\Omega X,\Omega X))\to Z(K_I;(X,*))\to X^{|I|}$ and a homotopy equivalence $Z(K_I;(C\Omega X,\Omega X))\simeq\Sigma^{|I|-1}\Omega X\wedge\cdots\wedge\Omega X$, where the number of $\Omega X$ in the wedge is $|I|$. Since $\Omega X$ is discrete, $Z(K_I;(C\Omega X,\Omega X))$ is a wedge of spheres of dimension $|I|-1\ge 2$. Then $Z(K_I;(X,*))$ is not acyclic. On the other hand, by Proposition \[retract\], $Z(K_I;(X,*))$ is a retract of $Z(K;(X,*))$, implying that $Z(K_I;(X,*))$ is acyclic. This is a contradiction, so $K$ is flag. Therefore the proof is completed. Let $G$ be a group and $\{G_s\}_{s\in S}$ be a family of groups such that $G_s=G$ for all $s\in S$. We denote the free product of $G_s$ for $s\in S$ by $F_S(G)$. Let $\Gamma$ be a simple graph (a graph without loops and multiple edges) with the vertex set $S$. The graph product of $G$ with respect to $\Gamma$, denoted $G^\Gamma$, is defined by dividing out $F_S(G)$ by the commuting relations $[G_s,G_t]=1$ for edges $\{s,t\}$ of $\Gamma$. Note that $G^\Gamma$ is natural with respect to homomorphisms of groups and inclusions of graphs. \[pi\_1\] For a path-connected space $X$, there is an isomorphism $$\pi_1(Z(K;(X,*)))\cong\pi_1(X)^{K^{(1)}}$$ where $K^{(n)}$ denotes the $n$-skeleton of $K$. By the cellular approximation theorem, the inclusion $Z(K^{(1)};(X,*))\to Z(K;(X,*))$ is an isomorphism in $\pi_1$. Since $Z(K^{(0)};(X,*))$ is a wedge of $m$ copies of $X$, we have $\pi_1(Z(K^{(0)};(X,*)))\cong F_{[m]}(\pi_1(X))$. By the van Kampen theorem, attaching the edge $\{i,j\}$ adds the commutator relation of $i$-th and $j$-th $\pi_1(X)$ in $F_{[m]}(\pi_1(X))$. Thus we have proved the lemma. For a simple graph $\Gamma$, let $C(\Gamma)$ be the flag complex whose 1-skeleton is $\Gamma$. \[BG\_Z\] For a group $G$ and a finite simple graph $\Gamma$, there is a homotopy equivalence $$B(G^\Gamma)\simeq Z(C(\Gamma);(BG,*))$$ which is natural with respect to $G$ and $\Gamma$. The homotopy equivalence is obtained by Proposition \[acyclic\] and Lemma \[pi\_1\], and the naturality is obvious by the construction. We now consider Coxeter groups. If a Coxeter group $W$ is right-angled, there are isomorphisms $W\cong({\mathbb{Z}}/2)^{\Gamma_W}$ and $A_W\cong{\mathbb{Z}}^{\Gamma_W}$. Then by Proposition \[BG\_Z\], one gets: \[W\_Z\] If a Coxeter system $(W,S)$ is finitely generated and right-angled, then there are homotopy equivalences $$BW\simeq Z(C(\Gamma_W);({\mathbb{R}}P^\infty,*))\quad\text{and}\quad BA_W\simeq Z(C(\Gamma_W);(S^1,*)).$$ We are going to show that ${\mathrm{Ad}}(X_W)$ is also given by a certain polyhedral product when $W$ is right-angled. To this end, we need several lemmas. The following lemma is well known and useful, which is proved, for example, in [@F Proposition, pp.180]. \[hocolim\] Let $\{F_i\to E_i\to B\}_{i\in I}$ be an $I$-diagram of homotopy fibration with a fixed base $B$. Then the sequence $$\underset{I}{\mathrm{hocolim}}\,F_i\to\underset{I}{\mathrm{hocolim}}\,E_i\to B$$ is a homotopy fibration. The following is an up-to-homotopy version of [@DS Lemma 2.3.1]. \[homotopy\_fib\] Let $(F,F')\to(E,E')\to(B,B)$ be a pair of homotopy fibrations such that $(F,F'),(E,E')$ are NDR pairs. Then $$Z(K;(F,F'))\to Z(K;(E,E'))\to B^m$$ is a homotopy fibration. For any $\sigma\subset[m]$, the sequence $(F,F')^\sigma\to(E,E')^\sigma\to B^m$ is a homotopy fibration, where $(X,A)^\sigma$ is as in the definition of a polyhedral product in the previous section. Since this homotopy fibration is natural with respect to $\sigma$, we get a $F(K)$-diagram of homotopy fibrations $\{(F,F')^\sigma\to(E,E')^\sigma\to B^m\}_{\sigma\in F(K)}$, where $F(K)$ is the face poset of $K$. Then it follows from Lemma \[hocolim\] that the sequence $$\underset{F(K)}{\mathrm{hocolim}}\,(F,F')^\sigma\to \underset{F(K)}{\mathrm{hocolim}}\,(E,E')^\sigma\to B^m$$ is a homotopy fibration. Since $(F,F')$ is an NDR pair, the projection $\mathrm{hocolim}\,(F,F')^\sigma\to\mathrm{colim}\,(F,F')^\sigma=Z(K;(F,F'))$ is a homotopy equivalence. Similarly we get a natural homotopy equivalence $\mathrm{hocolim}\,(E,E')^\sigma\simeq Z(K;(E,E'))$. Thus the proof is completed. \[pb\_lem\] Let $F\to E\to B$ be a homotopy fibration such that $F\to E$ is a cofibration. Then the following commutative square is a homotopy pullback. $$\label{pb_fib} \xymatrix{Z(K;(E,F))\ar[r]\ar[d]&E^m\ar[d]\\ Z(K;(B,*))\ar[r]&B^m}$$ To see that is a homotopy pullback, it is sufficient to show that the natural map between the homotopy fibers of the horizontal arrows is a homotopy equivalence. By Lemma \[homotopy\_fib\], the homotopy fibers of the both horizontal arrows are homotopy equivalent to $Z(K;(C\Omega B,\Omega B))$ and the natural map between them is identified with the identity map. Thus is a homotopy pullback. \[pb\_group\] Suppose that there is a pullback of groups $$\xymatrix{G_1\ar[r]^g\ar[d]&G_2\ar[d]\\ H_1\ar[r]^h&H_2}$$ where $g,h$ are surjective. Then the induced square $$\xymatrix{BG_1\ar[r]\ar[d]&BG_2\ar[d]\\ BH_1\ar[r]&BH_2}$$ is a homotopy pullback. It is sufficient to show that the natural map between the homotopy fibers of the horizontal arrows in the second square is a homotopy equivalence. Since $g,h$ are surjective, the homotopy fibers are $B\mathrm{Ker}\,g$ and $B\mathrm{Ker}\,h$, respectively, and the natural map between them is induced from the canonical map $\mathrm{Ker}\,g\to\mathrm{Ker}\,h$ which is an isomorphism by Lemma \[pullback\_lem\]. Thus the proof is completed. Let $M$ be the closed Möbius band and $(M,S^1)$ be the pair of $M$ and its boundary circle. \[main\] For a finitely generated right-angled Coxeter group $W$, there is a homotopy equivalence $$B{\mathrm{Ad}}(X_W)\simeq Z(C(\Gamma_W);(M,S^1)).$$ By Theorem \[pullback\_thm\] and Lemma \[pb\_group\], for any Coxeter group $W$, we have a homotopy pullback $$\xymatrix{B{\mathrm{Ad}}(X_W)\ar[r]\ar[d]&(S^1)^{c(W)}\ar[d]\\ BW\ar[r]&({\mathbb{R}}P^\infty)^{c(W)}.}$$ Consider a homotopy fibration $S^1\to M\to{\mathbb{R}}P^\infty$ where the first map is the boundary inclusion and the second map is equivalent to the bottom cell inclusion. Then by Lemma \[pb\_lem\], there is a homotopy pullback $$\xymatrix{Z(K;(M,S^1))\ar[r]\ar[d]&(S^1)^m\ar[d]\\ Z(K;({\mathbb{R}}P^\infty,*))\ar[r]&({\mathbb{R}}P^\infty)^m.}$$ Thus the proof is done by Corollary \[W\_Z\]. By Corollary \[W\_Z\] and Theorem \[main\] together with the naturality of Proposition \[BG\_Z\], one gets: \[phi\_Phi\] Let $W$ be a finitely generated right-angled Coxeter group. 1. The map $\Phi\colon BA_W\to B{\mathrm{Ad}}(X_W)$ is identified with $$Z(C(\Gamma_W);(S^1,*))\to Z(C(\Gamma_W);(M,S^1))$$ which is induced from the composite $(S^1,*)\simeq(M,*)\to(M,S^1)$. 2. The map $\phi\colon B{\mathrm{Ad}}(X_W)\to BW$ is identified with $$Z(C(\Gamma_W);(M,S^1))\to Z(C(\Gamma_W);({\mathbb{R}}P^\infty,*))$$ which is induced from the composite $(M,S^1)\to(M/S^1,*)=({\mathbb{R}}P^2,*)\to({\mathbb{R}}P^\infty,*)$. Applications ============ We give two applications of Theorem \[main\] and first show a stable splitting of $B{\mathrm{Ad}}(X_W)$. We will use the following stable splitting of a polyhedral product, which was proved in [@BBCG] (See [@IK1] for a more precise proof of the naturality). Define the polyhedral smash product $\widehat{Z}(K;(X,A))$ in the same way as $Z(K;(X,A))$ by replacing the direct product with the smash product. \[BBCG\] There is a homotopy equivalence $$\Sigma Z(K;(X,A))\simeq\Sigma\bigvee_{\emptyset\ne I\subset[m]}\widehat{Z}(K_I;(X,A))$$ which is natural with respect to $(X,A)$. \[split\] The inclusion $Z(K;(X,*))\to Z(K;(X,A))$ has a left homotopy inverse after a suspension. By definition, if $L$ is a full simplex with $n$ vertices, then $\widehat{Z}(L;(Y,B))$ is the smash product of $n$ copies of $Y$. Then $\widehat{Z}(K_I;(X,*))=\widehat{Z}(K_I;(X,A))$ if $I\subset[m]$ is a simplex of $K$. On the other hand, if $I$ is not a simplex of $K$, we have $\widehat{Z}(K_I;(X,*))=*$. Thus by Theorem \[BBCG\], $\Sigma Z(K;(X,*))$ is a wedge summand of $\Sigma Z(K;(X,A))$, up to homotopy, completing the proof. For a finitely generated right-angle Coxeter group W, the map $\Phi\colon BA_W\to B{\mathrm{Ad}}(X_W)$ has a left homotopy inverse after a suspension. In particular, there is a simply connected space $X$ such that $$\Sigma B{\mathrm{Ad}}(X_W)\simeq\Sigma BA_W\vee\Sigma X.$$ By Corollary \[phi\_Phi\] and Lemma \[split\], the first assertion follows. If we put $X$ to be the cofiber of $\Phi\colon BA_W\to B{\mathrm{Ad}}(X_W)$, then the second assertion follows from the first one, where $X$ is simply connected by the van Kampen theorem. Let $W$ be a right-angled Coxeter group with the generating set $[m]$. We next calculate the mod 2 cohomology of ${\mathrm{Ad}}(X_W)$. By Corollary \[exact\], there is a homotopy fibration $$\label{fibration_W} (S^1)^{m}\to B{\mathrm{Ad}}(X_W)\to BW.$$ We calculate the Serre spectral sequence of , where we denote its $E_r$ by $E_r(W)$. For a subset $T\subset[m]$, let $W_T$ be the subgroup of $W$ generated by $T$. Then $W_T$ is a right-angled Coxeter group with the generating set $T$ such that $\Gamma_{W_T}=(\Gamma_W)_T$. By the naturality of Theorem \[pullback\_thm\], one sees that for $V\subset U\subset[m]$, for $W_V$ is a retract of for $W_U$, implying the following. \[retract\] For $V\subset U\subset[m]$, $E_r(W_V)$ is a retract of $E_r(W_U)$. By Corollary \[W\_Z\], one has $BW\simeq Z(C(\Gamma_W);({\mathbb{R}}P^\infty,*))$. Then quite similarly to the calculation of the cohomology of $Z(K;({\mathbb{C}}P^\infty,*))$ in [@DJ], we see that $$H^*(BW;{\mathbb{F}}_2)={\mathbb{F}}_2[x_1,\ldots,x_m]/(x_ix_j\,\vert\,\{i,j\}\not\in E(\Gamma_W)),\quad|x_i|=1$$ where $E(\Theta)$ denotes the edge set of a graph $\Theta$. By Theorem \[pullback\_thm\], the homotopy fibration is a homotopy pullback of the homotopy fibration $(S^1)^m\xrightarrow{2}(S^1)^m\to({\mathbb{R}}P^\infty)^m$, implying that its local system of coefficients is trivial. Then one gets $$E_2(W)\cong {\mathbb{F}}_2[x_1,\ldots,x_m]/(x_ix_j\,\vert\,\{i,j\}\not\in E(\Gamma_W))\otimes\Lambda(y_1,\ldots,y_m)$$ such that $d_2x_i=0$ and $d_2y_i=x_i^2$, where $|y_i|=1$. For $I\subset[m]$, put $z_{i,I}=x_iy_I$, where $y_I=\prod_{i\in I}y_i$ for $I\ne\emptyset$ and $y_\emptyset=1$. Then for $I\subset N_i$, we have $d_2z_{i,I}=0$ and $E_3(W)$ is generated by such $z_{i,I}$ as an algebra, where $N_i=\{j\,\vert\,\{i,j\}\not\in E(\Gamma_W)\}$. By definition, we have $$\begin{aligned} \label{rel1} z_{i,I}z_{j,J}&=\begin{cases}0&\{i,j\}\not\in E(\Gamma_W)\text{ or }I\cap J\ne\emptyset\\z_{i,I-k}z_{j,J\sqcup\{k\}}&\{i,j\}\in E(\Gamma_W),\;I\cap J=\emptyset,\;k\in N_j\end{cases}\end{aligned}$$ in $E_2(W)$. Moreover, since $d_2y_I=\sum_{i\in I}z_{i,\emptyset}z_{i,I-i}$, we have $$\label{rel2} \sum_{i\in I}z_{i,\emptyset}z_{i,I-i}=0\quad\text{if }I-i\subset N_i\text{ for any }i\in I$$ in $E_3(W)$. Since these are all relations that $z_{i,I}\in E_3(W)$ satisfy, one gets: \[E\_3\] $E_3(W)$ is an algebra generated by $z_{i,I}$ for $i=1,\ldots,m$ and $I\subset N_i$ subject to the relations and , where $|z_{i,I}|=i+|I|$. We show that $E_r(W)$ collapses for $r=3$ by considering the special case that $\Gamma_W=K_{m-1}\sqcup\{m\}$, where $K_n$ is the complete graph with $n$ vertices. If $\Gamma_W=K_{m-1}\sqcup\{m\}$, then $C(\Gamma_W)=\Delta^{[m-1]}\sqcup\{m\}$, where $\Delta^S$ is the full simplex on the vertex set $S$. \[dim\] For a right-angled Coxeter group $W$ with $\Gamma_W=K_{m-1}\sqcup\{m\}$, the dimension of $H^m({\mathrm{Ad}}(X_W);{\mathbb{F}}_2)$ is $m+1$. By Theorems \[main\] and \[BBCG\], one has $$\Sigma B{\mathrm{Ad}}(X_W)\simeq\Sigma\bigvee_{\emptyset\ne I\subset[m]}\widehat{Z}(C(\Gamma_W)_I;(M,S^1)).$$ Note that $C(\Gamma_W)_I=\Delta^I$ for $m\not\in I$ and $C(\Gamma_W)_I=\Delta^{I-m}\sqcup\{m\}$ for $m\in I$. By definition, $\widehat{Z}(\Delta^I;(M,S^1))=\bigwedge_{i\in I}M\simeq S^{|I|}$. When $m\in I$, we have $\widehat{Z}(\Delta^{I-m}\sqcup\{m\};(M,S^1))=U\cup V$ for $U=(\bigwedge_{i\in I-m}M)\wedge S^1$ and $V=(\bigwedge_{i\in I-m}S^1)\wedge M$ such that $U\cap V=\bigwedge^{|I|}S^1=S^{|I|}$. Since the inclusions of $U\cap V$ into $U$ and $V$ are trivial in the mod 2 cohomology, by the Mayer-Vietoris sequence of this cover, we get $$H^*(\widehat{Z}(\Delta^{I-m}\sqcup\{m\};(M,S^1));{\mathbb{F}}_2)\cong\begin{cases}{\mathbb{F}}_2&*=0,|I|+1\\{\mathbb{F}}_2\oplus{\mathbb{F}}_2&*=|I|\\0&\text{otherwise}.\end{cases}$$ Thus $H^m({\mathrm{Ad}}(X_W);{\mathbb{F}}_2)\cong\bigoplus_{J\subset[m-1],\,|J|\ge m-2}H^m(\widehat{Z}(\Delta^J\sqcup\{m\};(M,S^1));{\mathbb{F}}_2)\cong{\mathbb{F}}_2^{m+1}$, completing the proof. Suppose that $\Gamma_W=\Delta^{[m-1]}\sqcup\{m\}$. By Lemma \[E\_3\], $E_3^m(W)$ is spanned by $z_{m,[m-1]}$, $z_{1,\emptyset}\cdots z_{m-2,\emptyset}z_{m-1,\{m\}}$ and $z_{m,\emptyset}z_{m,[m-1]-i}$ for $i=1,\ldots,m-1$. Then it follows from Lemma \[dim\] that these elements are permanent cycles. Thus one gets: \[permanent\] If $\Gamma_W=\Delta^{[m-1]}\sqcup\{m\}$, then $z_{m,[m-1]}\in E_3(W)$ is a permanent cycle. Let $W$ be an arbitrary right-angled Coxeter group on the vertex set $[m]$. Let $I\subset N_i$ and $W'$ be the right-angled Coxeter group with the generating set $I\sqcup\{i\}$ such that $\Gamma_{W'}=K_{|I|}\sqcup\{i\}$. Since $\Gamma_{W_{I\sqcup\{i\}}}\subset\Gamma_{W'}$, by the naturality of graph products of groups, there is a homotopy commutative diagram $$\xymatrix{(S^1)^{|I|+1}\ar[r]\ar@{=}[d]&B{\mathrm{Ad}}(X_{W_{I\sqcup\{i\}}})\ar[r]\ar[d]&BW_{I\sqcup\{i\}}\ar[d]\\ (S^1)^{|I|+1}\ar[r]&B{\mathrm{Ad}}(X_{W'})\ar[r]&BW'.}$$ Then we get a map $E_r(W')\to E_r(W_{I\sqcup\{i\}})$ which maps $z_{i,I}\in E_3(W')$ to $z_{i,I}\in E_3(W_{I\sqcup\{i\}})$, so by Lemma \[permanent\], $z_{i,I}\in E_3(W_{I\sqcup\{i\}})$ is a permanent cycle. Thus by Lemma \[retract\], $z_{i,I}\in E_3(W)$ is a permanent-cycle too and we obtain: \[collapse\] The Serre spectral sequence of collapses at $E_3$. For a degree reason, the extension of $E_\infty(W)$ to $H^*({\mathrm{Ad}}(X_W);{\mathbb{F}}_2)$ is trivial. Therefore by Lemmas \[E\_3\] and \[collapse\], we finally obtain: For a right-angled Coxeter group $W$ with the generating set $[m]$, the mod 2 cohomology of ${\mathrm{Ad}}(X_W)$ is an algebra generated by $z_{i,I}$ for $i=1,\ldots,m$ and $I\subset N_i$ subject to the relations and , where $|z_{i,I}|=i+|I|$. [W]{} T. Akita, *The adjoint group of a Coxeter quandle*, <arXiv:1702.07104v2>. N. Andruskiewitsch, F. Fantino, G. A. García, and L. Vendramin, *On Nichols algebras associated to simple racks, Groups, algebras and applications*, Contemp. Math. **537**, Amer. Math. Soc., Providence, RI, 2011, pp. 31-56. A. Bahri, M. Bendersky, F.R. Cohen, and S. Gitler, *The polyhedral product functor: a method of decomposition for moment-angle complexes, arrangements and related spaces*, Advances in Math. **225** (2010), 1634-1668. N. Brady, J.P. McCammond, B. Mühlherr, and W.D. Neumann, *Rigidity of Coxeter groups and Artin groups*, Geom. Dedicata **94** (2002), 91-109. M.W. Davis and T. Januszkiewicz, *Convex polytopes, Coxeter orbifolds and torus actions*, Duke Math. J. **62** (1991), 417-451. M. Davis and B. Okun, *Cohomology computations for Artin groups, Bestvina-Brady groups, and graph products*, Groups Geom. Dyn. **6** (3) (2012), 485-531. G. Denham and A. Suciu, *Moment-angle complexes, monomial ideals, and Massey products*, Pure Appl. Math. **3** (2007), 25-60. M. Eisermann, *Quandle coverings and their Galois correspondence*, Fund. Math. **225** (2014), no. 1, 103-168. E.D. Farjoun, *Cellular spaces, null spaces and homotopy localization*, Lecture Notes in Mathematics [**1622**]{}, Springer-Verlag, Berlin, 1996. K. Iriye and D. Kishimoto *Decompositions of suspensions of spaces involving polyhedral products*, Algebr. Geom. Topol. **16** (2016), 825-841. K. Iriye and D. Kishimoto, *Fat wedge filtrations and decomposition of polyhedral products*, to appear in Kyoto J. Math. T. Nosaka, *Central extensions of groups and adjoint groups of quandles*, <arXiv:1505.03077>. J.H.C. Whitehead, *On the asphericity of regions in a $3$-sphere*, Fund. Math. **32** (1939), no. 1, 149-166.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Cold atoms provide a unique arena to study many-body systems far from equilibrium. Furthermore, novel phases in cold atom systems are conveniently investigated by dynamical probes pushing the system out of equilibrium. Here, we discuss the pumping of doubly-occupied sites in a fermionic Mott insulator by a periodic modulation of the hopping amplitude. We show that deep in the insulating phase the many-body system can be mapped onto an effective two-level system which performs coherent Rabi oscillations due to the driving. Coupling of the two-level system to the remaining degrees of freedom renders the Rabi oscillations damped. We compare this scheme to an alternative description where the particles are incoherently pumped into a broad continuum.' author: - 'F. Hassler' - 'S. D. Huber' title: 'Coherent pumping of a Mott insulator: Fermi golden rule versus Rabi oscillations' --- Cold atoms in optical lattices provided us with numerous insights into the paradigmatic quantum phase transition displayed by interacting lattice bosons [@fisher:89; @greiner:02]. Progress in the same setup with fermionic atoms has now led to new, equally-exciting results in the study of the Hubbard model [@scarola:08; @leo:08; @helmes:08; @huber:09; @kollath:06]. A decrease of the compressibility when crossing over from the weakly to the strongly interacting region was reported [@jordens:08; @schneider:08]. Furthermore, in the experiment by Jördens [*et al.*]{} [@jordens:08], the strongly correlated regime of fermions in an optical lattice has been investigated by harmonically modulating the strength of the light beams forming the optical lattice. The modulation of the lattice generates doubly-occupied sites (doublons). This generation of doublons deep in the Mott insulator, where the interaction $U$ dominates over the hopping $t_{0}$, is the focus of our present work. Experimentally, the fraction of particles on doubly-occupied sites (or double occupancy $\mathcal{D}$) was measured after modulating the optical lattice. For $U/6t_0=13.6$, a [*saturation*]{} was observed at around $\mathcal{D}_\text{sat}=30\%$ after a modulation time $\tau_\text{mod} \approx 50 h /U$ (50 “cycles”) [@jordens:08] which sets an upper bound on the saturation time $\tau_\text{sat}\leq \tau_\text{mod}$ \[see Fig. \[fig:experiment\](a)\]. ![(color online) (a) Sketch of an experimental sequence for the generation of double occupancy by modulating the optical lattice during a time $\tau$. After an initial buildup of doubly-occupied sites with a rate $\Gamma_\text{up}$, the fraction of particles residing in doubly-occupied sites saturates at a value $\mathcal D_\text{sat}$ after a time $\tau_\text{sat}$. (b) Illustration of the effective Hilbert space $\mathcal H_\text{eff}$ consisting of one bond embedded in an array of sites. The bath degrees of freedom consist of all doublon-hole configurations in which the pair resides on a different bond than where it was created. []{data-label="fig:experiment"}](experiment) We consider fermionic atoms, such as $^{40}$K, subject to an optical lattice in the framework of the Hubbard model $$H=-t_{0}\sum_{\langle i,j\rangle,\sigma} \bigl( c_{i\sigma}^{\dag}c_{j\sigma}^{{\vphantom \dagger}}+ {\rm H.c.} \bigr) + U\sum_{i} c_{i\uparrow}^{\dag}c_{i\uparrow}^{{\vphantom \dagger}}c_{i\downarrow}^{\dag}c_{i\downarrow}^{{\vphantom \dagger}},$$ at half filling and with the interaction $U$ dominating over the hopping amplitude $t_{0}$. Here, the fermionic operators $c_{i\sigma}^{\dag}$ create an atom in one of two possible hyperfine states $\sigma=\uparrow,\downarrow$ at site $i$ and the sum $\langle i,j\rangle$ runs over nearest neighbors. The ground state of the system consists of one particle per site with the hopping $t_0$ suppressed due to the strong on-site repulsion $U$. As the temperature $T$ in the experiment is larger than the super-exchange coupling $J=4t_0^2/U$ [@jordens:08], no significant quantum correlations over the bonds are present and the ground state of the system is largely composed of singly-occupied sites ($T\ll U$) with a random spin-direction on each site. A harmonic modulation of the lattice with frequency $\omega \approx U/\hbar$ transfers energy to the system and (possibly) promotes a particle across the bond, thereby creating a doubly-occupied site next to an empty site, a doublon-hole pair. Despite the fact that the Hubbard model represents a many-body problem in a huge Hilbert space, we adopt a simple description via modeling a single (typical) bond \[Fig. \[fig:experiment\](b)\], and taking the remaining bonds into account as an effective bath. Two different ideas lend themselves to modeling the process of pumping doublon-hole pairs \[see Fig. \[fig:comparison\]\]: (a) the direct excitation of the bond via *incoherent* transitions into the doublon-hole continuum with a rate given by the Fermi golden rule $$\label{eq:fgr} \Gamma^{{\scriptscriptstyle\rm}FGR}_\text{mod} = \frac{2\pi}{\hbar} |\langle f | H_\text{mod} | g \rangle |^2 \rho(E_g+\hbar \omega);$$ here, $E_g$ denotes the energy of the ground state, $\rho$ is the density of states of the doublon-hole continuum, and $\langle f| H_\text{mod}|g\rangle$ is the matrix element arising from the modulation of the lattice. Taking the modulation strength to be given by $\delta t$ and estimating the density of states as $\rho \propto 1/t_0$, we obtain the scaling of the pump rate $\Gamma^{{\scriptscriptstyle\rm}FGR}_\text{mod} \propto \delta t^2/t_0$ as a function of the hopping amplitude $t_0$ and the strength of the modulation $\delta t$. (b) *Coherent* Rabi oscillations in the isolated bond between the ground state of the bond and the doubly-occupied excited state with a Rabi frequency $\Omega_{{\scriptscriptstyle\rm}R}\propto \delta t$ proportional to the modulation strength. The Rabi oscillations generate doublon-hole pairs and lead to a periodic modulation (with frequency $\Omega_{{\scriptscriptstyle\rm}R}$) of the double occupancy. Note that the rates of the two (extreme) processes, incoherent pumping into a band (a) (with rate $\Gamma_\text{mod}^{{\scriptscriptstyle\rm}FGR}$) and coherent excitation of doublon-hole pairs (b) (with rate $\hbar \Omega_{{\scriptscriptstyle\rm}R}$), have a different dependence on the modulation strength $\delta t$ and can therefore be discriminated experimentally. Here, we propose a third way where the coherent Rabi oscillations produce a doublon-hole pair and, subsequently, hopping of the doublon (or the hole) to the neighboring sites leads to an incoherent decay of the excitation into the doublon-hole continuum. The process is governed by two time scales, the Rabi frequency $\Omega_{{\scriptscriptstyle\rm}R} \propto \delta t$ which periodically modulates the double occupancy and the decay $\Gamma_\text{hop}\propto t_0$ which leads to a saturation. A similar system, where a two-level atom is coupled simultaneously to a laser field and to vacuum modes of the radiation field, is well-studied in quantum optics [@tannoudji:77]. The two-level atom corresponds to the two states on the bond and the laser plays the role of the harmonic modulation of the lattice. However, in the atomic example, the spontaneously emitted photon takes the atom back into the ground state where it continues with the Rabi cycles, while in the present case the doublon (or hole) itself [*exits*]{} the system. The present process then resembles more strongly that of optical pumping into a third nondecaying level, with the accumulation in the pumped level playing the role of the buildup in double occupancy. ![Comparison of two possible descriptions: (a) A rate equation description where an [*incoherent*]{} pump rate \[$\Gamma_\text{up}^{{\scriptscriptstyle\rm}FGR}$\] into a broad doublon-hole continuum and subsequent decay \[$\Gamma_\text{down}^{{\scriptscriptstyle\rm}decay}$\] or incoherent stimulated emission \[$\Gamma_\text{up}^{{\scriptscriptstyle\rm}FGR}$\] lead to a saturation. (b) [*Coherent*]{} Rabi oscillations \[with frequency $\Omega_{{\scriptscriptstyle\rm}R}$\] between the ground state and a single exited state where decay out of the effective Hilbert space is introduced via coupling to a bath \[$\Gamma_\text{in(out)}$\]. []{data-label="fig:comparison"}](comparison) In the following, we first describe the effective model in the single bond Hilbert space driven by a harmonic modulation of the lattice and derive the frequency of the Rabi oscillations. We then account for the bath consisting of the surrounding sites via a master equation approach and analyze it in different limits. We compare this approach to the situation where the bond is directly excited into the doublon-hole continuum; see (a) above. It turns out that the time scale ($\tau_{{\scriptscriptstyle\rm}R}$) for the Rabi cycle is in rough agreement with the experimental value of the saturation time, whereas the treatment via a direct excitation into the doublon-hole continuum leads to a time scale ($\tau_{{\scriptscriptstyle\rm}FGR}$) which is an order of magnitude larger ($\tau_{{\scriptscriptstyle\rm}sat}^{{\scriptscriptstyle\rm}exp}\lesssim \tau_{{\scriptscriptstyle\rm}R} \ll \tau_{{\scriptscriptstyle\rm}FGR}$). For a quantitative agreement, one has to go beyond a single bond in the description of the pumping process. Different bonds are coupled via the pumping field, leading to a Dicke-like description [@dicke:54], but, most likely, many-body effects dominate the coupling of different bonds; a systematic study of larger clusters deserves future efforts. In deriving the effective Hamiltonian governing the dynamics in the Hilbert space of a single bond (the sites are labeled by L and R), we neglect the confining potential and focus on a homogeneous three dimensional setup. The ground state at half-filling involves one particle per site without any spin order, i.e., the particles on the bond occupy the three triplet states $ |T^{\pm 1}\rangle= c_{{{{\scriptscriptstyle\rm}L}}\uparrow(\downarrow)}^{\dag} c_{{{{\scriptscriptstyle\rm}R}}\uparrow(\downarrow)}^{\dag}|0\rangle,\quad |T^{0}\rangle= \frac{1}{\sqrt{2}} \bigl( c_{{{{\scriptscriptstyle\rm}L}}\uparrow}^{\dag}c_{{{{\scriptscriptstyle\rm}R}}\downarrow}^{\dag}+ c_{{{{\scriptscriptstyle\rm}L}}\uparrow}^{\dag}c_{{{{\scriptscriptstyle\rm}R}}\downarrow}^{\dag} \bigr) |0\rangle, $ and the singlet state $ |S\,\rangle= \frac{1}{\sqrt{2}} \bigl( c_{{{{\scriptscriptstyle\rm}L}}\uparrow}^{\dag}c_{{{{\scriptscriptstyle\rm}R}}\downarrow}^{\dag}- c_{{{{\scriptscriptstyle\rm}L}}\uparrow}^{\dag}c_{{{{\scriptscriptstyle\rm}R}}\downarrow}^{\dag} \bigr) |0\rangle. $ These states constitute the four degenerate ground states of a single bond and are realized with equal probability of $1/4$ each. A harmonic modulation of the lattice beam with $V=V_{0}+\delta V\cos(\omega\tau)$ leads to a time-dependent hopping Hamiltonian $$\label{eqn:hopp} H_\text{hop}= \delta t \cos(\omega\tau) \sum_\sigma \bigl( c_{{{{\scriptscriptstyle\rm}L}}\sigma}^{\dag}c_{{{{\scriptscriptstyle\rm}R}}\sigma}^{{{\vphantom \dagger}}}+{\rm H.c.} \bigr),$$ with $\delta t/t_{0}=\bigl(3/4-\sqrt{V_{0}/E_{r}}\bigr)\delta V/V_{0}$ and $\tau$ denoting the time. This expression is valid for $\delta V/V_{0}\ll 1$ and $\delta V/V_{0} \ll \sqrt{E_{r}/V_{0}}$ [@hopping]. For current experiments with fermions [@jordens:08], where $\delta V/V_{0} = 0.1$ and $V_{0}$ does not exceed $10 E_{r}$, both conditions are safely fulfilled. In the derivation of Eq. (\[eqn:hopp\]), we have dropped the constant hopping amplitude proportional to $t_0$, as hopping is blocked in the Mott insulator at half-filling. The hopping amplitude $t_0$ will be reintroduced later as it is responsible for the decay of the doublon-hole pair by coupling the bond to the surrounding lattice. The modulation couples the ground states to the doublon-hole states $ |d_{{\scriptscriptstyle\rm}L(R)} \rangle = c^\dagger_{{{\scriptscriptstyle\rm}L(R)}\uparrow} c^\dagger_{{{\scriptscriptstyle\rm}L(R)}\downarrow} |0\rangle, $ with the doublon located on the left (right) site. These states exhibit an energy offset $U$ with respect to the energy of the ground states. The total Hamiltonian in the Hilbert space spanned by the four ground states (triplet and singlet) and the two doublon-hole pairs is given by $$\begin{aligned} H_\text{eff}&= \sum_{\alpha={{\scriptscriptstyle\rm}L, R}} U |d_{\alpha}\rangle\langle d_{\alpha} | + H_\text{hop} \approx \sum_{\alpha={{\scriptscriptstyle\rm}L, R}} U|d_{\alpha}\rangle\langle d_{\alpha} | \nonumber \\ \label{eq:ham} &\qquad \qquad+ \delta t \Bigl[ e^{i\omega\tau/\hbar} | d_\text{a}\rangle\langle S| + e^{-i\omega\tau/\hbar} |S\rangle\langle d_\text{a}| \Bigr] ,\end{aligned}$$ where the first term describes the energy offset by $U$ and the second term the driving due to the lattice modulation (\[eqn:hopp\]). In going from the first to the second equation in (\[eq:ham\]), the rotating wave approximation has been used and we have introduced the [*active*]{} and [*passive*]{} doublon-hole states given by $|d_\text{a(p)}\rangle=(|d_{{\scriptscriptstyle\rm}L}\rangle \pm |d_{{\scriptscriptstyle\rm}R}\rangle)/\sqrt{2}$; these are related to the left (right) doublon-hole state via a basis change. Driving the system on resonance with $\hbar \omega=U$, the singlet state is coupled to the excited state via Rabi oscillations. Within the present description, off-resonant coupling can be described in the standard manner. As can be seen from Eq. (\[eq:ham\]), the triplet states are unaffected by the lattice modulations. Only the singlet state is coupled to the active doublon-hole pairs $|d_\text{a}\rangle$. Going over to the interaction representation with the reference dynamics given by the first term in (\[eq:ham\]), the relevant Hamiltonian in the effective two-level system $\mathcal{H}_\text{eff}= \text{span}\{ |d_\text{a}\rangle,|S\rangle\}$ can be expressed as a matrix $$\label{eqn:hammatrix} \mathsf H = \begin{pmatrix} 0 & \delta t \\ \delta t & 0 \end{pmatrix},$$ which describes the coherent oscillations in the two-level system with the Rabi frequency $$\Omega_{{\scriptscriptstyle\rm}R}=\frac{2\delta t}{\hbar}.$$ The triplet states and the passive doublon-hole pair remain eigenstates of the time-dependent Hamiltonian and therefore particles in such states are not driven. So far, we have neglected the effect of $t_{0}$ on the excited state $|d_\text{a}\rangle$. Provided the bond is in a doublon-hole state, we can lose the state from our single bond description via the hopping of the doublon (or the hole) out of the bond; cf. Fig. \[fig:experiment\](b) [@vacuum]. This effect is captured by a bath representing the states with the doublon and hole separated from each other. Note that the bath states still contribute to the double occupancy (as there is still a doublon present) but do not participate in Rabi oscillations (as the doublon and the hole are not located on the same bond). We denote the rate at which the doublon-hole bond decays into a separated doublon and a hole by $\Gamma_\text{out}$, and the reverse rate at which a doublon and a hole collide to form a doublon-hole pair by $\Gamma_\text{in}$. Taking the coupling to the bath into account and performing a standard analysis similar to [@gardiner:5] leads to an incoherent term (besides the standard term proportional to $[\mathsf{H},\rho]$) in the evolution of the reduced density matrix $\rho$ in the effective Hilbert space $\mathcal{H}_\text{eff}= \text{span}\{ |d_\text{a}\rangle,|S\rangle\}$ of the form $$\label{eq:incoh} \frac{\partial\rho}{\partial\tau} \bigg|_\text{inc} = -\frac{1}{2} \begin{pmatrix} 2(\Gamma_\text{out} \rho_{11} - \Gamma_\text{in} \rho_0) & \Gamma_\text{out} \rho_{12} \\ \Gamma_\text{out}\rho_{21} & 0 \end{pmatrix}.$$ Different from a standard master equation [@gardiner:5], Eq. (\[eq:incoh\]) is not probability preserving, since with probability $\rho_0 = 1-\rho_{11} -\rho_{22}$ the doublon and the hole have separated and thus have left the effective Hilbert space $\mathcal{H}_\text{eff}$. In the steady state situation with $\partial_\tau \rho=0$, the probability that the system contributes to the double occupancy is given by $$\label{eq:sat} \rho_{{\scriptscriptstyle\rm}D}^\text{sat}= \rho_0^\text{sat} + \rho_{11}^\text{sat} = 1-\rho^\text{sat}_{22} = \frac{\Gamma_\text{out}+ \Gamma_\text{in}} {\Gamma_\text{out} + 2 \Gamma_\text{in}},$$ the sum of the probabilities that the system is in the doublon-hole pair state ($\rho_{11}$) or in the bath ($\rho_0$). Note that atoms which left the effective Hilbert space still contribute to the double occupancy. The bath can be pictured as a gas of doublons and holes (with equal numbers). Initially, the bath does not contain any particles such that $\Gamma_\text{in}=0$ (this approximation remains valid as long as the density of particle is low); furthermore, we estimate the escape rate by $\Gamma_\text{out}=\Gamma_\text{hop} \sim t_{0}/\hbar$. For $\Gamma_\text{in}=0$, Eq. (\[eq:sat\]) provides the saturation value $\rho_{{\scriptscriptstyle\rm}D}^\text{sat}=1$, i.e., the singlet is certainly transferred to one of the doublon-hole states. ![Fraction of particles residing in doubly-occupied sites after modulation time $\tau$. After a buildup time corresponding to approximately one $\pi$-pulse \[$\tau_{{\scriptscriptstyle\rm}R}=\pi/\Omega_{{\scriptscriptstyle\rm}R}=\hbar\pi/2\delta t$\], the decay, with (weak) strength $\Gamma_\text{out} = \delta t/\hbar \approx 0.24 t_0/\hbar$ and $\Gamma_\text{in}=0$, leads to a saturation at $\mathcal D_\text{sat}=58\%$ corresponding to the fact that all “active” bonds, singlets, are excited in the system (solid line). A more realistic situation where the damping is increased by a factor of 2 renders the oscillations barely visible (dashed line). Taking into account a finite value of $\Gamma_\text{in}$ leads to a decrease of the saturation value. Note that, irrespective of the saturation value, the time scale needed to pump the system into a doubly occupied state is given by the Rabi frequency. []{data-label="fig:solutions"}](solutions) Using the dilute gas approximation, $\Gamma_\text{in}=0$, we can solve the equation of motion for the density matrix $$\label{eq:time_evol} \partial_{\tau}\rho=\frac{1}{i\hbar}[\mathsf H, \rho]- \frac{\Gamma_\text{out}}{2} \begin{pmatrix} 2 \rho_{11} & \rho_{12} \\ \rho_{21} & 0 \end{pmatrix}.$$ Assuming that initially the bond is in a singlet state $$\rho(\tau=0)= \begin{pmatrix} 0 & 0 \\ 0 & 1 \end{pmatrix},$$ Eq. (\[eq:time\_evol\]) gives the time evolution of the density matrix of the bond. Qualitatively, the weight is transferred to the exited state $|d_\text{a}\rangle$ at a rate $\hbar \Omega_{{\scriptscriptstyle\rm}R}$ and then the decay into other states occurs with a rate $\Gamma_\text{out}$ (cf.Fig. \[fig:solutions\]). To calculate the fraction of particles in a doublon state, we need to know how many bonds are initially in a singlet configuration. We estimate the probability that a specific site is part of a singlet as $P_S = 1- (3/4)^{d} \approx 0.58 $, here $d=3$ is the “dimension” of the lattice modulation (it counts the number of bonds per site) and we have assumed that the singlets are independent of each other and occupied with a probability $1/4$. The fraction of particles in a doublon state is given by $$\label{eq:doublon} \mathcal{D}=P_S \rho_{{\scriptscriptstyle\rm}D}(\tau)$$ with $\rho_{{\scriptscriptstyle\rm}D}(\tau)=1-\rho_{22}(\tau)$ the probability that this singlet is transferred to a doublon state. The system quickly reaches a steady state with $\mathcal{D}_\text{sat}=P_S$ (cf.Fig. \[fig:solutions\]). In three dimensions, we reach a maximal value $\mathcal D_\text{sat}=58\%$. Note that at this value of double occupancy the dilute gas approximation $\Gamma_\text{in}=0$ becomes invalid. We expect that $\Gamma_\text{in}$ increases for increasing double occupancy and, therefore, the double occupancy saturates at a lower value. A theory that aims at a quantitative result for the saturation value should take such effects into account and solve the system self-consistently. Assuming that for long times the density of particles in the bath is high such that $\Gamma_\text{in} \gg \Gamma_\text{out}$, we obtain the result that $\rho_{{\scriptscriptstyle\rm}D}^\text{sat} = 50\%$ \[cf. Eq. (\[eq:sat\])\] and $\mathcal{D}_\text{sat}=29\%$, comparable to experimental results [@jordens:08]. The time scale needed to excite the system is approximately given by the time $\tau_{{\scriptscriptstyle\rm}R} = \pi/\Omega_{{\scriptscriptstyle\rm}R}$ of a $\pi$ pulse, irrespective of the saturation value obtained. Note that the finite line-width of the perturbing laser, the scattering between doublons, and an inhomogeneous environment for a single bond due to the trap are other processes which lead to a dephasing of the coherent single bond dynamics. As the effect of $t_{0}$ on the excited state leads already to an over-damped dynamics, these other effects are neglected here. Another approach to describe the saturation in the double occupancy is based on a rate equation for the doublon fraction. The [*rates*]{} at which excitations are generated are taken into account via a rate equation for the doublon concentration $$\label{eqn:rateequation} \partial_{\tau} {\mathcal D} = (1-{\mathcal D})\Gamma_\text{up} -{\mathcal D} \Gamma_\text{down}.$$ Within this description two independent processes are assumed. First, the system is excited with a rate $\Gamma_\text{up}$. Second, at a later uncorrelated stage, these excitations are removed with a rate $\Gamma_\text{down}$ \[cf. Fig. \[fig:comparison\](a)\]. In order to compare the two schemes, the rate equation and the Rabi oscillation, we extract the time scales of both processes. The Fermi golden rule matrix element is approximately given by $\delta t^{2}$ [@huber:09]. Additionally, we need an estimate for the density of states. Once excited, the doublon as well as the hole can move freely through the lattice. The corresponding doublon-hole continuum acquires a width of approximately $24t_{0}$ (in three dimensions). A crude estimate for the density of states is then given by $\rho\approx1/24t_{0}$. We then arrive at the two time scales $\tau_{{\scriptscriptstyle\rm}R} = \pi/\Omega_{{\scriptscriptstyle\rm}R}=\pi\hbar/2\delta t$ and $\tau_{{\scriptscriptstyle\rm}FGR} = (\hbar/2 \pi) 24 t_{0}/\delta t^{2}= (24 t_0/\pi^{2}\delta t)\tau_{{\scriptscriptstyle\rm}R} \approx 10 \tau_{{\scriptscriptstyle\rm}R}$, associated with a Rabi oscillation and the rate equation, respectively. For the last step, we have used $\delta t/t_{0}\approx 0.24$ corresponding to current experiments. Note that the time scale to reach saturation in the incoherent Fermi golden rule approach is an order of magnitude larger than the one predicted by the Rabi oscillation picture. In Ref. [@jordens:08] the saturation sets in before 50 “cycles”, which is even smaller than $\tau_{{\scriptscriptstyle\rm}R}$ \[$\tau_\text{mod} = 50 h/U = (200 \delta t/U)\tau_{{\scriptscriptstyle\rm}R} \approx 0.6 \tau_{{\scriptscriptstyle\rm}R}$ for $U/6t_0=13.6$ and $\delta t/t_{0}\approx 0.24$\], promoting the picture of the coherent Rabi oscillations. In an experiment, one can test one theory against the other by changing the coupling strength $\delta t$ while keeping all other parameters, such as the bandwidth $1/24t_0$, constant. The time scale for the saturation scales as $\delta t^{-2}$ in the Fermi golden rule picture, whereas Rabi oscillations produce a saturation time inversely proportional to $\delta t$. In summary, we have discussed two alternative explanations for the buildup and the saturation of the double occupancy in the modulation experiments of Ref. [@jordens:08] deep in the Mott insulating phase. The comparison to experiment favors the description in terms of an effective two-level system with coherent driving, leading to the characteristic time scale of $\tau_{{\scriptscriptstyle\rm}R}\approx \pi/\Omega_{{\scriptscriptstyle\rm}R}=\hbar\pi/2\delta t$ for half a Rabi cycle. Given the large decay rate proportional to $t_{0}>\hbar \Omega_{{\scriptscriptstyle\rm}R} \approx \delta t$, only one $\pi$-pulse is observable. The predictions can be experimentally tested by studying the dependence of the saturation time on various system parameters as $V_{0}/E_{r}$, $\delta V/V_{0}$, and $t_{0}/U$. We acknowledge fruitful discussions with E. Altman, G. Blatter, H. P.Büchler, the group of T. Esslinger, S. Fölling, L. Goren, and A.Rüegg. This work was supported by the Swiss National Foundation through the NCCR MaNEP. [10]{} M.P.A. Fisher, P.B. Weichman, G. Grinstein, and D.S. Fisher, Phys. Rev. B [**40**]{}, 546 (1989). M. Greiner, O. Mandel, T. Esslinger, T.W. Hänsch, and I. Bloch, Nature [**415**]{}, 39 (2002). V.W. Scarola, L. Pollet, J. Oitmaa, and M. Troyer, arXiv:0809.3239 (2008). L. De Leo, C. Kollath, A. Georges, M. Ferrero, and O. Parcollet, Phys. Rev. Lett. [**101**]{}, 210403 (2008). R.W. Helmes, T.A. Costi, and A. Rosch, Phys. Rev. Lett. [**100**]{}, 056403 (2008). S.D. Huber and A. Rüegg, Phys. Rev. Lett. [**102**]{}, 065301 (2009). C. Kollath, A. Iucci, I.P. McCulloch, and T. Giamarchi, Phys. Rev. A [**74**]{}, 041604(R) (2006). R. Jördens, N. Strohmaier, K. Günter, H. Moritz, and T. Esslinger, Nature [**455**]{}, 204 (2008). U. Schneider, L. Hackermuller, S. Will, T. Best, I. Bloch, T.A. Costi, R.W. Helmes, D. Rasch, and A. Rosch, Science [**322**]{}, 1520 (2008). C. Cohen-Tannoudji, in [*Frontiers in Laser Spectroscopy*]{}, edited by R. Balian, S. Haroche, and S. Liberman, Proceedings of the Les Houches Summer School of Theoretical Physics, 1975 (North-Holland, Amsterdam, 1977). R.H. Dicke, Phys. Rev. [**93**]{}, 99 (1954). When deriving (\[eqn:hopp\]), care has to be taken in expanding the exponential part of the hopping amplitude [@zwerger:03]. We think about the singly-occupied lattice as being the vacuum with excitations given by doublons and holes. C.W. Gardiner and P. Zoller, [*Quantum Noise*]{}, chapter 5.1 (Springer, Berlin, 2000). W. Zwerger, J. Opt. B: Quantum Semiclassical Opt. [**5**]{}, S9 (2003).
{ "pile_set_name": "ArXiv" }
--- abstract: | We derive an analytic model for nonlinear “photon bubble" wave trains driven by buoyancy forces in magnetized, radiation pressure-dominated atmospheres. Continuous, periodic wave solutions exist when radiative diffusion is slow compared to the dynamical timescale of the atmosphere. We identify these waves with the saturation of a linear instability discovered by Arons — therefore, these wave trains should develop spontaneously. The buoyancy-driven waves are physically distinct from photon bubbles in the presence of rapid diffusion, which evolve into trains of gas pressure-dominated shocks as they become nonlinear. Like the gas pressure-driven shock trains, buoyancy-driven photon bubbles can exhibit very large density contrasts, which greatly enhance the flow of radiation through the atmosphere. However, steady-state solutions for buoyancy-driven photon bubbles exist only when an extra source of radiation is added to the energy equation, in the form of a flux divergence. We argue that this term is required to compensate for the radiation flux lost via the bubbles, which increases with height. We speculate that an atmosphere subject to buoyancy-driven photon bubbles, but lacking this compensating energy source, would lose pressure support and collapse on a timescale much shorter than the radiative diffusion time in the equivalent homogeneous atmosphere. author: - 'Mitchell C. Begelman' title: Nonlinear Photon Bubbles Driven by Buoyancy --- Introduction ============ The tendency of “photon bubbles" to form in magnetized, radiation pressure-supported atmospheres is well-established. In photon bubble instability, the atmosphere spontaneously forms a propagating pattern of low-density channels separated by regions of high density. Radiation tends to diffuse through the underdense regions, avoiding the regions of high density. Both analytic calculations (Begelman 2001) and numerical simulations (Turner et al. 2005; Hsu, Arons & Klein 1997) indicate that the instability can lead to very large and persistent density contrasts. In this case, the atmosphere becomes “leaky" and the net radiation flux can greatly exceed the equilibrium value for a homogeneous atmosphere. Two families of linear photon bubble instabilities were discovered by Arons (1992) and Gammie (1998; see also Blaes & Socrates 2003), respectively. The Gammie modes depend on the interplay of gas pressure and radiation pressure and are destabilized slow magnetosonic waves. In the nonlinear limit they become trains of gas pressure-dominated shocks. The Arons modes are an outgrowth of internal entropy modes, and are related to previously-discovered instabilities driven by convection in a strong magnetic field, in the presence of heat conduction (Syrovatskii & Zhugzhda 1968). They are driven by buoyancy, are governed by the rate of radiative diffusion, and do not depend on the presence of gas pressure. The Arons modes occur where the parameter $M_0$, denoting the ratio of dynamical time to radiation diffusion time across the atmosphere, is smaller than one, indicating that radiation is relatively well-coupled to the gas. In contrast, the Gammie modes occur where radiation can diffuse across the atmosphere faster than the sound crossing time in the gas, a reflection of weak coupling — the relevant condition is $M_0 > \beta^{1/2}$, where $\beta \ll 1 $ is the ratio of gas pressure to radiation pressure. Both modes depend for their existence on the presence of a magnetic field stiff enough to enforce approximately one-dimensional motion. The purpose of this paper is to study the nonlinear development of the buoyancy-driven waves discovered by Arons (1992), and to assess their possible effects on radiation-dominated atmospheres. As in our earlier analysis of the nonlinear development of Gammie modes, we will use analytic techniques based on the assumption of periodic wave trains. In keeping with the essential physical characteristics of the Arons modes, we will ignore gas pressure in our analysis. This eliminates the possibility of obtaining a series of shock fronts, which prove to be the dominant feature of nonlinear Gammie waves. Instead, we will seek — and find — continuous periodic solutions. Because the waves are driven by buoyancy (and in contrast to our earlier analysis), we must carefully account for the vertical structure of the atmosphere. We do this by expanding our solution in powers of the vertical coordinate $z$, in addition to using a similarity variable to describe the waves. By focusing on wavelengths that are much shorter than the scale height of the atmosphere, we are able to perform a multi-scale perturbation analysis in the radiation pressure (analogous to a WKB approximation, but nonlinear) to obtain a model for the nonlinear Arons waves. The plan of the paper is as follows. In § 2 we describe the basic equations and approximations relevant to steady-state, nonlinear waves driven by buoyancy in a radiation-dominated atmosphere. The section concludes with a derivation of the basic equation describing the nonlinear waves. We solve this equation in § 3, where we also discuss the main features of the waves. In § 4 we interpret our results and discuss the waves’ global effects on atmospheres. We emphasize that the existence of steady-state modes requires the presence of a distributed radiation source throughout the atmosphere. Absent this source, the atmosphere is expected to collapse due to the increasing leakage of radiation with height. We attribute the collapse of the atmospheres in the Hsu et al. (1997) simulations to this effect, and speculate that atmospheres subject to buoyancy-driven photon bubble instability will generally lose pressure support and collapse on a timescale short compared to the radiation diffusion time in the equivalent homogeneous atmosphere. We summarize our results and present our conclusions in § 5. Equations ========= We base our model on a simplified set of radiation-hydrodynamical equations, in which the magnetic field is stiff and vertical. The latter was assumed in Arons (1992), although Hsu et al. (1997) performed simulations for an inclined magnetic field. All results below can be readily generalized to an arbitrary (uniform) magnetic field direction. We assume that all vectors lie in the $x-z$ plane, and that there is a uniform gravitational field $- g \hat z$. The five equations that must be satisfied describe continuity, radiative diffusion in both the $x$ and $z$ directions, momentum conservation parallel to ${\bf B}$, and conservation of radiation energy. Eliminating the radiation flux through the diffusion equation ${\bf F} = - (c/\kappa) \nabla p / \rho $, where $\kappa$ is the opacity (assumed constant) and $p$ is the radiation pressure, we obtain (for a vertical magnetic field) $$\label{cont} \rho_{\tilde t} + (\rho v)_{\tilde z} = 0 ,$$ $$\label{mom} v_{\tilde t} + v v_{\tilde z} = - {p_{\tilde z} \over \rho} - g ,$$ $$\label{en} 3 p_{\tilde t} + 3 v p_{\tilde z} + 4 p v_{\tilde z} = {c\over \kappa} \left[ \left({p_{\tilde z} \over \rho}\right)_{\tilde z} + \left({p_{\tilde x} \over \rho}\right)_{\tilde x} \right] - S_0 ,$$ where subscripts denote partial differentiation and dimensional coordinates are surmounted by a tilde. The source term $-S_0$ in the energy equation represents an assumed divergence $S_0 = \nabla\cdot {\bf F_0}$ of radiation flux that is intrinsic to the fluid. As we will show in § 3, this added energy flux is necessary in order to obtain steady-state wave trains — otherwise the atmosphere collapses and no steady solution is possible. For simplicity, we will assume that $S_0$ is a constant, corresponding to a linear variation of injected radiation flux with height. This is an arbitrary choice; steady wave train solutions may exist for various dependences of $S_0$ on local state variables. We nondimensionalize the equations by normalizing the pressure, density, and velocity to fiducial values: $p \equiv p_0 \nu$, $\rho \equiv \rho_0 \eta$, $v \equiv v_p u$. We then define a pressure scale height $H = p_0/(\rho_0 g) \equiv c_0^2 / g$, where $c_0$ is the isothermal sound speed associated with the radiation pressure. Dimensionless coordinates are given by $x \equiv \tilde x /H$, $z \equiv \tilde z / H$, $t \equiv v_p \tilde t / H$, and a characteristic Mach number is defined by $m_p \equiv v_p / c_0$. The diffusivity parameter from Arons (1992) is given by $$\label{mzero} M_0 \equiv {c \over \kappa \rho_0 H c_0}.$$ Finally, we define a dimensionless radiation source term by $\hat \sigma \equiv H S_0/(4 m_p p_0 c_0)$. The dimensionless equations are then $$\label{contnon} \eta_{t} + (\eta u)_{z} = 0 ,$$ $$\label{momnon} m_p^2 (u_{t} + u u_{z} ) = - \left( 1 + {\nu_{z} \over \eta} \right) ,$$ $$\label{ennon} 3 \nu_{t} + 3 u \nu_{z} + 4 \nu u_{z} = {M_0\over m_p} \left[ \left({\nu_{z} \over \eta}\right)_{z} + \left({\nu_{x} \over \eta}\right)_{x} \right] - 4\hat \sigma .$$ We seek periodic “plane-wave" solutions with wavelength $\lambda$, in which quantities associated with the wave depend on position and time through the combination $$\label{sdef} s = x \tan \theta + z + t ;$$ we henceforth denote differentiation with respect to $s$ by a prime. Thus, the wavevector makes an angle $\theta$ with respect to the $z-$axis, as in the analysis by Arons (1992). The quantity $-v_p$ is then the $z-$component of the phase velocity. The speed of the intersection of a wave front with the $x-$axis is given by $-v_p \cot\theta$, while the overall phase speed of the waves is $v_p \cos \theta$. In addition to the periodic dependence on $s$ there must be a secular dependence on $z$. We will seek solutions in which the wavelength is short compared to the scale height, hence we can expand $\eta$, $\nu$ and $u$ as a Taylor series in $z$, with $|z| \ll 1$: $$\label{Taylor} \eta = \eta_0 (s) + z \eta_1 (s) + {z^2 \over 2} \eta_2 (s) + \dots \ ,$$ with similar relations for $\nu$ and $u$. Since we demand that all functions of $s$ be periodic, we can define $z$-independent wave-averaged quantities by the integral $$\label{meandens} \langle A \rangle \equiv \lambda^{-1} \int^{s+\lambda}_s A(s) ds ;$$ we are free to choose the density normalization so that $\langle \eta_0 \rangle = 1$. We seek steady-state solutions with no secular mass flow, implying that $$\label{meanmassflux} \lim_{n\rightarrow \infty} {1\over n\lambda} \int^{n\lambda}_0 \eta u (ads + bdz) = 0 ,$$ for all $a$ and $b$, where $$\label{massflux} \eta u = \eta_0 u_0 + z (\eta_1 u_0 + \eta_0 u_1) + {z^2\over 2} (\eta_2 u_0 + 2 \eta_1 u_1 + \eta_0 u_2) + \dots$$ This implies $\langle \eta_0 u_0 \rangle = 0 $ and $\eta_1 u_0 + \eta_0 u_1 = \eta_2 u_0 + 2 \eta_1 u_1 + \eta_0 u_2 =0 $. Similarly, all higher order terms in the Taylor expansion for $\eta u$ must vanish. Let us now consider the continuity equation. To $O(z^0)$ we have $$\label{contzero} \eta_0' + (\eta_0 u_0)' + (\eta_1 u_0 + \eta_0 u_1) = \eta_0' + (\eta_0 u_0)' = 0 ,$$ which is easily integrated to yield $$\label{uzero} u_0 = {1 - \eta_0 \over \eta_0} .$$ To $O(z^1)$ we have $$\label{contone} \eta_1' + (\eta_1 u_0 + \eta_0 u_1)' + (\eta_2 u_0 + 2 \eta_1 u_1 + \eta_0 u_2) = \eta_1' = 0 ,$$ implying $\eta_1 = -\alpha =$ const., and $$\label{uone} u_1 = - {\eta_1 u_0 \over \eta_0} = \alpha {1 - \eta_0 \over \eta_0^2 } .$$ Note that the density scale height is $H/\alpha$, with $\alpha =1$ corresponding to the $p \propto \rho \propto e^{-z}$ atmosphere considered by Arons (1992). It is straightforward to carry the analysis of the continuity equation to higher orders in $z$ with, for example, $\eta_2= \beta =$ const. and $u_2 = (2 \alpha^2 - \beta \eta_0)(1 - \eta_0)/\eta_0^3 $. However, these higher-order terms are not required for our analysis. We next consider the energy equation. To $O(z^0)$ we have $$\label{enzero} 3 (1 + u_0) \nu_0' + 3 u_0 \nu_1 + 4 \nu_0 (u_0' + u_1) = {M_0 \over m_p } \left[ \left( {\nu_z \over \eta }\right)_0' + \left( {\nu_z \over \eta }\right)_1 + \tan^2\theta \left( {\nu_0' \over \eta_0 }\right) '\right] - 4\hat \sigma,$$ while to $O(z^1)$ the equation is $$\begin{aligned} \label{enone} 3 (1 + u_0) \nu_1' + 3 u_1 (\nu_0' + \nu_1) + 3 u_0 \nu_2 + 4 \nu_0 (u_1' + u_2) + 4 \nu_1 (u_0' + u_1) = \qquad\qquad\qquad \nonumber \\ {M_0 \over m_p } \left[ \left( {\nu_z \over \eta }\right)_1' + \left( {\nu_z \over \eta }\right)_2 + \tan^2\theta \left( {\nu_1' \over \eta_0 } - {\eta_1 \nu_0'\over \eta_0^2 }\right) '\right] , \qquad\end{aligned}$$ where the terms in the expansion $$\label{etanuexp} {\nu_z \over \eta} = \left( {\nu_z \over \eta }\right)_0 + z \left( {\nu_z \over \eta }\right)_1 + {z^2\over 2}\left( {\nu_z \over \eta }\right)_2 + \dots$$ are obtained from the momentum equation (\[momnon\]): $$\label{etanuzero} \left( {\nu_z \over \eta }\right)_0 = -1 - m_p^2 \left[ (1 + u_0)u_0' + u_0 u_1 \right]$$ $$\label{etanuone} \left( {\nu_z \over \eta }\right)_1 = - m_p^2 \left[ (1 + u_0)u_1' + u_0 u_2 + u_1 (u_0' + u_1) \right] .$$ We argue below that $(\nu_z / \eta)_2$ is not needed at the desired level of approximation. Multiplying eq. (\[momnon\]) by $\eta$, we also obtain the useful results: $$\label{etazero} \nu_0' + \nu_1 = -\eta_0 - m_p^2 \left[ u_0' + (1 - \eta_0) u_1 \right]$$ $$\label{etaone} \nu_1' + \nu_2 = -\eta_1 - m_p^2 \left[ u_1' + \eta_1 u_0' + \eta_0 u_0 u_2 \right] .$$ Our expansion in $z$ is valid only if the wavelength projected on the $z-$axis is short compared to the scale height, $\tilde\lambda \ll H \cos\theta$. (Note that this approximation just breaks down for the fastest growing linear waves, according to Arons \[1992\].) In this limit, the inertial terms of the momentum equation — those on the left-hand side of eq. (\[momnon\])— should be small compared to the terms on the right-hand side. This ordering goes along with the assumption that the phase speed of the waves projected along the $z-$axis should be very subsonic, i.e., $|m_p| \ll 1$. In this spirit, we introduce $m_p$ as a second expansion parameter for the dimensionless pressure, so that $$\label{nuexp} \nu = \nu_0 + z \nu_1 + \dots = (\nu_{00} + m_p\nu_{01} + m_p^2 \nu_{02} + \dots ) + z(\nu_{10} + m_p \nu_{11} + \dots ) + \dots$$ We are free to set $\nu_{00} = 1$, $\nu_{10} = -1$, in order to mimic an exponential atmosphere with $p = p_0 \exp (- z/H)$, to lowest order. By renormalizing $s$ to $$\label{wdef} \hat w \equiv m_p^{-1} s$$ (where a prime henceforth denotes differentiation with respect to $\hat w$), and treating all orders of $u$ and $\eta$ and their derivatives with respect to $w$ as $O(1)$, we can expand equations (\[enzero\]) and (\[enone\]) in powers of $m_p $. In these expansions we also treat $M_0/m_p$ and $\hat \sigma$ as $\sim O(1)$. The leading terms in both equations are then $\sim O(m_p^{-1})$; in order to describe periodic wave trains we will need the $O(m_p^{-1})$ and $O(1)$ solutions to eq. (\[enzero\]), and the $O(m_p^{-1})$ solution to eq. (\[enone\]). Carrying out this program, we first note that the terms proportional to $(\nu_z/\eta)_1$ in eq. (\[enzero\]) and $(\nu_z/\eta)_2$ in eq. (\[enone\]) are $\sim O(m_p)$, so they can be ignored. Next, by expanding equations (\[etazero\]) and (\[etaone\]) we obtain: $$\label{nuexp2} \nu_{01}' = 1 - \eta_0; \qquad\qquad \nu_{02}' = - (\nu_{11} + u_0'); \qquad\qquad \nu_{11}' = - (\eta_1 + \nu_{20}) .$$ These results allow us to write down eq. (\[enzero\]) to $O(m_p^{-1})$: $$\label{enzeromp0} 4 u_0' = {M_0\over m_p} \tan^2\theta \left( {\nu_{01}'\over \eta_0}\right)' = {M_0\over m_p} \tan^2\theta \left( - {\eta_0'\over \eta_0^2}\right) = {M_0\over m_p} \tan^2\theta \ u_0'$$ (where we have used eq. \[\[uzero\]\]), implying $$\label{enzeromp1} m_p = {1\over 4} M_0 \tan^2\theta .$$ Now consider eq. (\[enone\]) to $O(m_p^{-1})$: $$\label{enonemp1} 4 (u_1' - u_0') = {M_0\over m_p} \tan^2\theta \left( {\nu_{11}'\over \eta_0} - {\eta_1\nu_{01}'\over \eta_0^2}\right)' .$$ Substituting $\eta_1 = -\alpha$ and using equations (\[uzero\]), (\[uone\]), and (\[enzeromp1\]), we obtain $\nu_{11}' = \eta_{0} - 1 = - \nu_{01}'$, where we have imposed the condition $\langle \nu_{11}' \rangle = 0$ for a periodic solution. We may then set $\nu_{11} = - \nu_{01}$. Finally, we consider the $O(1)$ terms in eq. (\[enzero\]). Using previous results, we find that $3 (1+ u_0) \nu_{01}' + 3 u_0 \nu_{10} = 0$. Therefore, the surviving terms are $$\label{enzerozero} 4 \nu_{01}u_0' + 4 u_1 = {M_0\over m_p} \left[ \left( {\eta_0'\over \eta_0^3}\right)' + \tan^2\theta \left( {\nu_{02}'\over \eta_0}\right)' \right] - 4\hat \sigma = {M_0\over m_p} \left[ \left( {\eta_0'\over \eta_0^3}\right)' + \tan^2\theta \left( {\nu_{01} - u_0' \over \eta_0}\right)' \right] - 4\hat \sigma ,$$ where we have used eq. (\[nuexp2\]) and substituted $\nu_{11} = - \nu_{01}$. Using the result that $\nu_{01}'/\eta_0 = u_0$ and expressing $u_0$ and $u_1$ in terms of $\eta_0$ using equations (\[uzero\]) and (\[uone\]), we obtain $$\label{etaeq1} \left( {\eta_0'\over \eta_0^3}\right)' = \sin^2\theta {(1- \eta_0)(\alpha - \eta_0 ) \over \eta_0^2} - \hat \sigma ,$$ which is the equation describing steady wave trains. To simplify our study of solutions, we renormalize the independent variable to $w = \hat w \sin\theta $ and set $\sigma \equiv \hat\sigma / \sin^2\theta$, thus eliminating the factor $\sin^2\theta$ from eq. (\[etaeq1\]). The basic equation for our study of wave solutions is therefore $$\label{etaeq} \left( {\eta'\over \eta^3}\right)' = {(1- \eta)(\alpha - \eta) \over \eta^2} - \sigma ,$$ where a prime now denotes differentiation with respect to $w$ and we have dropped the subscript “0" from $\eta_0$. Solutions ========= We seek continuous periodic solutions of eq. (\[etaeq\]), subject to the constraint that $\langle \eta \rangle = 1$. That such solutions might exist is evident from the fact that eq. (\[etaeq\]) can be written in the form $$\label{genericODE} {1\over 2} {d\over dy} \left( y'\right)^2 + {df(y) \over dy} = 0$$ where $y$ is a function of $\eta$ and $y' = \eta' dy/d\eta$. $f(y)$ is also an implicit function of $\eta$ through $y$. This is, of course, just a generalized version of the Hamiltonian equation of energy conservation in a potential. In order for eq. (\[genericODE\]) to have periodic solutions, $f(\eta)$ must have two real roots, $f(\eta_+) = f(\eta_-) = 0$, with $f < 0$ for $\eta_- < \eta < \eta_+$, and the integral $$\label{genericint} \int^{\eta_+}_{\eta_-} {dy\over d\eta} {d\eta\over \left( - 2f \right)^{1/2}} ,$$ which equals half the wavelength, must be finite. The requirement of an intrinsic energy source in the atmosphere, $\sigma \neq 0$, is also apparent from the form of eq. (\[etaeq\]). If we choose $\alpha = 1$, for example, then the first term on the right-hand side of eq. (\[etaeq\]) is $\geq 0$, precluding a periodic solution unless $\sigma > 0$. In general, for a given $\alpha$ and wavelength, $\sigma$ is an eigenvalue of the problem. Equation (\[etaeq\]) is most easily solved by choosing a maximum density, $\eta_+$, then integrating to find the wavelength, minimum density $\eta_-$, and $\sigma$, subject to the mean density constraint. It is better to start integrating from the maximum density rather than the minimum density because, in solutions with large density contrasts, $\eta_+$ is exponentially sensitive to $\eta_-$. Figure 1 shows the density profile of the waves for $\eta_+ =$ 1.1, 2, and 11, with $\alpha = 1$. Parameters for a wider range of solutions are shown in Table 1. Even for wave trains with very low amplitudes, $\eta_+ - 1 \ll 1$ (Fig. 1a), the density pattern is far from sinusoidal. At low amplitudes, $\eta_+ - 1 \la 1$, the depression in the low-density portion of the wave is sharper and deeper than the enhancement in the dense region. This behavior switches markedly for $\eta_+ - 1 \ga 1$, and for high density contrasts (Fig. 1b) the dense regions are extremely sharp and narrow, while the low density regions are broad and flat. Figure 2 and Table 1 illustrate the variation of minimum density $\eta_-$, wavelength $\lambda$ (in units of $w$), and intrinsic flux divergence $\sigma$ with $\log_{10}(\eta_+ - 1)$. The minimum density varies in the opposite sense to the maximum density, but at large density contrasts this variation is slower than logarithmic. The wavelength decreases with increasing density contrast, also slower than logarithmically, while the intrinsic radiation source $\sigma$, required for a steady-state solution, increases slightly faster than logarithmically, although still very slowly. Interestingly, the inverse relationship between density contrast and wavelength is opposite to that found in the nonlinear shock train solutions (Begelman 2001). Numerical simulations (Turner et al. 2005) indicate that the shock train solutions evolve toward longer wavelengths, higher density contrasts, and larger luminosities. If higher density contrasts and larger luminosities are also energetically favored in the case of buoyancy-driven photon bubbles, then we would expect evolution toward [*shorter*]{} wavelengths. Indeed, the fastest-growing Arons modes have long wavelengths, close to the pressure scale height (see § 4.2), whereas the fastest-growing Gammie modes are short.    $\alpha = 1$ $\alpha=0.5$ $\alpha=2$ ------------------------- -- ---------- -------------- ---------- -- ---------- -------------- ---------- -- ---------- -------------- ---------- $\log_{10}(\eta_+ - 1)$ $\eta_-$ $\lambda[w]$ $\sigma$ $\eta_-$ $\lambda[w]$ $\sigma$ $\eta_-$ $\lambda[w]$ $\sigma$ -2 0.979 82.2 0.000111 — — — 0.990 6.29 0.000146 -1 0.839 25.9 0.00979 — — — 0.909 6.24 0.0136 0 0.478 9.44 0.395 0.308 17.2 0.386 0.547 5.18 0.722 1 0.270 6.11 2.94 0.225 10.1 1.49 0.287 3.97 6.16 2 0.186 5.39 8.30 0.170 8.44 3.79 0.193 3.62 17.6 3 0.143 5.11 16.2 0.134 7.79 7.43 0.146 3.49 34.1 4 0.116 4.96 26.55 0.111 7.44 12.2 0.118 3.41 55.5 Dependence on Density Gradient ------------------------------ The parameter $\alpha$ represents the underlying density gradient in the atmosphere, with $\alpha = 1$ corresponding to the “isothermal" atmosphere $p\propto\rho$ assumed by Arons (1992). The radiation entropy gradient $\nabla \ln (p/\rho^{4/3}) \sim 4\alpha/3 - 1$, so all atmospheres with $\alpha > 0.75$ are convectively stable according to the Schwarzschild criterion. Note that, because $M_0 < 1 $ for atmospheres subject to buoyancy-driven photon bubbles, it may be possible for large convective cells to transport radiation energy without excessive leakage. (Recall that $M_0$ is the ratio of the dynamical time to the diffusion time across a scale height.) The models with $\alpha = 1$ and $\alpha =2$ correspond to convectively stable atmospheres, while the model with $\alpha = 0.5$ is strongly unstable to convection. From Fig. 2b, it is apparent that density contrasts are relatively insensitive to $\alpha$. The one glaring exception to this statement is the fact that no periodic solutions exist for low density contrasts, in the case of $\alpha = 0.5$. This is easily understood by considering the right-hand side of eq. (\[etaeq\]). In order to have a periodic solution, the right-hand side must be negative at the upper turning point ($\eta_+$) and positive at the lower turning point ($\eta_-$). For sufficiently weak density contrasts and $\alpha < 1$, it is possible to have $\alpha < \eta_- < 1$. In this case, the first term on the right-hand side is negative, implying that $\sigma$ must be negative in order to make the right-hand side positive. But at $\eta_+$ the first term is positive implying that $\sigma $ must be positive, leading to a contradiction. The quantities $\sigma$ and $\lambda$ are more sensitive to the density gradient than is $\eta_-$. At fixed $\eta_+$, $\sigma$ increases and $\lambda$ decreases with $\alpha$. Since larger $\alpha$ means a steeper density gradient, these trends might reflect the increasing effectiveness of buoyancy in driving radiation through the atmosphere via the low-density channels. Larger $\sigma$ means that more energy is being transported by diffusion through the photon bubbles, while a smaller wavelength could mean that a smaller fraction of a pressure scale is sufficient to get the required driving force. Note that this is opposite to the trend for ordinary convection, in which energy is assumed to be advected by fluid motions, and energy transport is expected to be larger for smaller $\alpha$. In buoyancy-driven photon bubble instability, diffusion rather than advection transports the energy. Interpretation ============== Physical Properties of Solutions -------------------------------- The dimensional wavelength of solutions, $\tilde \lambda$, is related to the dimensionless wavelength $\lambda$ (in units of $w$), by $$\label{lambdatilde} \tilde \lambda = m_p H \cot\theta \ \lambda .$$ Combining this equation with eq. (\[enzeromp1\]), we obtain $$\label{kh} {\tilde\lambda \over H } = {\lambda\over 4} M_0 \tan\theta .$$ From our numerical solutions (Table 1) we have determined that $ \lambda /4 \sim O(1)$ for the cases with high density contrast. In order for our analysis to be valid, however, it is not sufficient for the wavelength merely to be smaller than the pressure scale height. A stronger condition — required by our expansion in powers of $z$ — is that the wavelength projected onto the $z-$axis, $\tilde\lambda_z = \tilde\lambda / \cos\theta$, must be smaller than $H$. Thus, a necessary condition for the validity of our analysis is that $$\label{kzh} M_0 < {\cos^2\theta \over \sin\theta } .$$ From eq. (\[enzeromp1\]) this condition implies that $m_p < \sin\theta / 4 < 1$, thus ensuring the validity of our expansion in $m_p$. In all cases $\sigma$ is positive, implying that the atmosphere itself must supply extra energy at all heights, in order to support a steady-state wave solution. In physical units, the intrinsic flux radiated by the atmosphere is $$\label{Szero} F_0 \sim S_0 H = 4 \sigma \sin^2 \theta \ m_p p_0 c_0$$ Since $\sigma \ga O(10)$ for the high contrast solutions, these atmospheres must radiate furiously. By definition, the energy loss rate due to diffusion in the equivalent homogeneous atmosphere is $\sim M_0 p_0 c_0$. The inhomogeneous atmospheres in our model are losing energy $\sigma \sin^2 \theta \tan^2\theta $ times faster — possibly leaking energy on less than a dynamical time. It is important to remember that our adoption of an intrinsic flux divergence $S_0$ was arbitrary — we took this step only because a steady-state wave solution could not be obtained without it. What would happen if, as is more likely, such a powerful intrinsic source term is absent? Although this case is not covered explicitly by our calculations, the only sensible deduction would seem to be that the atmosphere loses radiation pressure support and collapses on the leakage time or the dynamical time, whichever is longer. Indeed, since the flux used in our radiation hydrodynamical equations is defined relative to the fluid frame, one way for the atmosphere to generate a flux divergence is for the gas to sink in the gravitational potential. We precluded this possibility by demanding $\langle \rho v \rangle = 0$, for mathematical reasons, but this outcome cannot be excluded physically. Relationship to Arons Instability --------------------------------- While we have not proven that our nonlinear wave trains represent the outgrowth of Arons’s (1992) instability, there are a number of aspects which make this identification plausible. First, our wave trains are driven by buoyancy forces — we guaranteed this by setting the gas pressure to zero (as did Arons in his linear analysis) and treating inertial terms as higher order than buoyancy terms. Second, we exploited the same near-cancellation in the flux-divergence as Arons did. In his case, a slight imbalance drove the instability by diffusing more energy into the lower parts of bubbles than is removed at the top. In our case, it supplied the crucial driving term $(\eta'/\eta^3)'$ in the wave equation (\[etaeq\]). Third, our waves only appear in the $O(z)$ terms of an expansion in height, further illustrating that they are driven by buoyancy effects. Fourth, the vertical component of the wave phase velocity is negative, in agreement with the linear instability (Arons 1992, equations \[19\] and \[42\]; in contrast to the group velocity of the linear waves, which is positive — see Arons eq. \[22\]). Finally, the wave speeds in our model scale $\propto M_0$, indicating that the wave speed is set by the diffusion of radiation across the bubble, as in the linear instability. According to the linear analysis, the most rapidly growing modes have $kH \cos\theta = 0.66$ (Arons 1992, eq. \[44\]), where $k$ is the wavenumber. Projected on the $z-$axis, the wavelength is about 10 times longer than the scale height, putting it outside the limit of validity for our analysis. The fastest growing linear mode that is (marginally) consistent with our analysis has $\tilde \lambda_z \sim H$. If we assume that this mode determines the orientation of the nonlinear wave pattern, then the inequality in eq. (\[kzh\]) is replaced by an approximate equality. In the limit $M_0 \ll 1$, we then have $\cos\theta \sim \sqrt M_0 \ll 1$, and the dominant density contours will tend to align with the magnetic field. Thus, our nonlinear analysis reproduces another key feature of the linear instability, as well as the numerical simulations by Hsu et al. (1997). Finally, we note that the only existing numerical simulations of buoyancy-driven photon bubbles (Hsu et al. 1997), while reproducing the linear growth rates adequately, are plagued by the rapid collapse of the simulated atmosphere as the waves saturate. We conjecture that this outcome is real and reflects the lack of a compensating intrinsic source of radiation in the simulations. The fastest growing modes grow on a dynamical timescale (Arons 1992, eq. \[45\]), which is similar to the collapse time. Minimum Magnetic Field Strength and Effects of Gas Pressure ----------------------------------------------------------- The radiation pressure perturbations associated with the waves is $\Delta p \sim (\tilde\lambda / H) p_0 \tan\theta \sim M_0 p_0 \tan\theta $. These pressure fluctuations will be largely compensated by fluctuations in the magnetic field strength, in order to maintain overall pressure balance in the $x-$direction. The change in the magnetic pressure $p_B$ is related to sideways compressional or expansional motion by $\Delta p_B \sim (\Delta \tilde x / \tilde \lambda) p_B$, where $\Delta \tilde x$ is the sideways displacement caused by the passage of the wave. Now, in order for the magnetic field to enforce approximately one-dimensional motion, we require that the sideways displacement be much smaller than the parallel displacement of a typical fluid element, $\Delta \tilde x \ll \tilde\lambda / \cos\theta$. This implies $$\label{maglimit} {p_B \over p_0} \gg {\tilde \lambda \over H} \sin\theta = {\lambda\over 4} M_0 {\sin^2\theta \over \cos\theta}.$$ Equation (\[maglimit\]) appears somewhat different from the analogous criterion presented by Arons (1992) in connection with the linear instability. It is beyond the scope of this paper to carry out a fully magnetohydrodynamical analysis. The effects of gas pressure will be most important in the densest regions of the waves, where $\eta \approx \eta_+$. Rapid energy exchange between matter and radiation will enforce approximately isothermal conditions across the wavefronts, implying that the gas pressure is proportional to the local density. If the mean gas pressure is $\beta_g p_0$, where $\beta_g \ll 1$, then gas pressure will affect the wave dynamics when $\beta_g \eta_+ p_0$ approaches $\Delta p$, or equivalently, when $$\label{gaslimit} \eta_+ \sim \beta_g ^{-1} {\tilde \lambda \over H} \tan\theta .$$ In such cases, gas pressure may limit the maximum density but is unlikely to have much effect on the values of $\lambda$ or $\sigma$. Discussion and Conclusions ========================== We have produced an analytic model for periodic, nonlinear waves driven by buoyancy in a radiation pressure-dominated atmosphere with a stiff magnetic field. To isolate the effects of buoyancy, we ignored gas pressure and performed a multi-scale expansion in powers of the vertical coordinate $z$ and wave Mach number $m_p$. All solutions for steady-state, periodic wave structures require the existence of a large radiation source term in the atmosphere, which we modeled as a uniform, positive flux divergence term in the energy equation. Physically, we believe that this source term compensates for the enhanced loss of radiation via the waves. Without it the atmosphere would lose pressure support and collapse, on a timescale much shorter than the diffusion time associated with the equivalent homogeneous atmosphere. Based on several similarities of generic properties and physical mechanisms, we identify our wave solutions with the nonlinear development of the photon bubble instability discovered by Arons (1992). Simulations of the linear growth and saturation of this instability (Hsu et al. 1997) were plagued by the rapid collapse of the simulated atmosphere shortly after the waves reached nonlinear amplitudes. We suggest that this collapse resulted from the rapid release of energy via the waves, as indicated by our analytic model. Buoyancy-driven photon bubbles occur primarily when the radiation diffusion timescale across a pressure scale height is longer than the dynamical time, corresponding to the parameter $M_0$ being smaller than one. (It may be possible for buoyancy-driven photon bubbles to develop in atmospheres with $M_0 \ga 1$, provided that the wavefronts are sufficiently oblique to the magnetic field.) This condition is satisfied in a number of astrophysical systems, including the equatorial regions of radiation-dominated accretion disks (where $M_0 \sim \alpha$, the viscosity parameter), accretion columns onto neutron stars in X-ray binaries, massive stellar envelopes well below the photosphere, and gamma-ray burst fireballs. Whether the linear instability develops and saturates in any of these systems is not clear, because the growth of the instability is rather slow. For example, in accretion disks subject to the magnetorotational instability (MRI), the magnetic field structure is expected to change on a timescale comparable to the dynamical (Keplerian) time. This is shorter than the growth time of all but (possibly) the fastest growing buoyancy-driven photon bubbles. Hence, it would not be surprising if buoyancy-driven photon bubbles were suppressed in accretion disks. (Gas pressure-driven photon bubbles, on the other hand, can grow faster than MRI, and are expected to be important in accretion disks \[Begelman 2002, 2005; Turner et al. 2005\].) In accretion columns onto a neutron star polar cap, however, the magnetic field is anchored into the neutron star crust and its structure is fixed. We would expect buoyancy-driven photon bubbles to grow to nonlinear amplitudes under these circumstances, perhaps leading to the collapse of the accretion column. What would happen to an atmosphere that collapsed under the action of buoyancy-driven photon bubbles? If we treat the collapsing atmosphere as a sequence of hydrostatic models, then both the diffusion parameter, $M_0$ ($\propto c_0^{-1}$ for a fixed column density), and the ratio of gas pressure to radiation pressure, $\beta_g$, are expected to increase. If $M_0$ became large, then the atmosphere would enter the strong-diffusion regime and gas pressure-driven shock trains would become the dominant form of photon bubble. Alternatively, $\beta_g$ could increase until gas pressure halted the collapse. A third possibility, which may be relevant in the case of neutron star accretion columns, is that the gravitational binding energy liberated by the accretion flow provides at least part of the compensating flux divergence. It is not clear, however, whether this would be adequate to prevent the collapse of the column. I am grateful to Jon Arons for pointing out the atmospheric collapse problem in his 1997 simulations with Hsu and Klein. This research was supported in part by NASA Astrophysical Theory Program grant NAG5-12035 and by the National Science Foundation under Grants AST-0307502 and PHY-990794. Part of this work was carried out at the Kavli Institute for Theoretical Physics at the University of California, Santa Barbara; I thank the members of KITP for their hospitality. Arons, J. 1992, ApJ, 388, 561 Begelman, M. C. 2001, ApJ, 551, 897 . 2002, ApJ, 568, L97 . 2005, ApJ, submitted Blaes, O. M., & Socrates, A. 2003, ApJ, 596, 509 Gammie, C. F. 1998, MNRAS, 297, 929 Hsu, J. J.-L., Arons, J., & Klein, R. I. 1997, ApJ, 478, 663 Syrovatskii, S. I., & Zhugzhda, Yu. D. 1968, –AJ, 11, 945 Turner, N. J., Blaes, O. M., Socrates, A., Begelman, M. C., Davis, S. W. 2005, ApJ, 624, 267
{ "pile_set_name": "ArXiv" }
--- abstract: 'We illustrate and study the evolution of reported infections over the month from March 1st to April 1st in the New York State as a whole, as well as in each individual county in the state. We check for exponential trends, and try to understand whether there is any correlation between the timing and dynamics of these trends and state-wide mandated measures on testing and social distancing. We conclude that the reports on April 1st may be dramatically under-representing the actual number of state-wide infections, and we propose reassessment of the data over the coming weeks using the increasing number of casualties as a validating measure. The will also allow to monitor for effects of the PAUSE directive.' --- [**Course of the first month of the COVID 19 outbreak in the New York State counties**]{} *Anca Rǎdulescu, Mathematics, SUNY New Paltz* Introduction ============ The study of early epidemic growth has historically revealed different patterns, depending on the particular pathogen, and even on the particular outbreak. While other growth patterns have also been found (e.g., polynomial), exponential growth has been a seemingly ubiquitous trend detected in data of early outbreaks of influenza, Ebola, foot-and-mouth disease, plague, measles, smallpox. That is somewhat motivated by the free spread of the pathogen in the first stages of the epidemic. Exponential growth patterns appear to also be representative for the current COVID-19 pandemic, over its first months of development in the US. In this study, we investigate these patterns specifically within the New York State, in which the epidemic has caught the widest proportions. Since its first confirmed cases in the US, it has become clear that the COVID-19 outbreak will globally affect all US states and territories. However, there have been essential between-state differences in the timeline and magnitude of the epidemic. It is likely that these differences are based on a variety of factors, from timing of the first contamination (earlier states were caught unprepared), to population density and social dynamics, to timing and efficiency of state-wide mandated directives for travel bans and social distancing, to timing and availability of testing. We will first illustrate such differences between three early states with most significant spread: California (9,816 infection by April 1st) , Washington State (5,588 infections) and New York State (83948 infections). Then we will focus more specifically on the New York State dynamics, trying to understand both the specific and the unifying trends in the contamination data from different counties, and to correlate these trends with state-mandated measures on testing and social distancing. Modeling methods ================ We will use time series estimating the confirmed number of infections with COVID-19, as reported in the public domain by the COVID Tracking Project [@testing] and by the Johns Hopkins Center for Systems Science and Engineering [@COVID_archive]. In order to more easily detect potential exponential growth $N(t) = N_0 e^{\alpha t}$ (where $t$ is time in days, and $N$ is number of infected individuals), we will be considering the logarithm of the time series $\ln N = \ln N_0 + \alpha t$. This will allow to test for linearity, and – if linear – compute the exponential growth as the slope $\alpha$. What one would generally expect to see in the epidemic development is an initial period with high growth rate (potentially exponential). As the growth rate subsides, the time series should move to a slower growing curve, then finally reach a peak, and eventually transition to a decreasing trend (as the epidemic is dying out). Figure \[comparison\] illustrates a comparison between the data for the states of California, Washington and New York. These are all states with early contamination compared to other states (January 1st in Washington, January 25th in California, March 1st in New York). The figure shows the logarithm of the raw time series, with the day of progression labeled from the start of the outbreak in the respective state. The unifying pattern between the three states illustrated here is that there are almost linear time windows in the evolution of each, ultimately suggesting that all three raw time series were still growing exponentially for the time interval preceding April 1st. Table \[intro\_table\] describes more throughly these time windows, showing the corresponding slope $\alpha$ and the goodness of fit. There are, however, significant differences in the dynamic patterns for each case – reflected in the length, succession and slope of the linear pieces. According to this data, the number of confirmed cases in California has been on an exponential rise with slope $~ 0.18$ since day 36 (February 29th), with a slight sub-exponential tendency towards the very end. Since day 38 (February 27th) Washington has been showing a pattern which can be explained by successively moving (approximately every week) to slightly lower slope exponentials, which have flattened out to a rate close to zero by the end of March. In turn, New York State shows a sequence of oscillations in rate, with a large exponential rate over the first few days (March 4-7), even after the initial transient spike in contamination (March 1-4), followed by a reduction to a new approximately constant rate over the next 10 days (March 7-16), succeeded by another rate increase (March 16-23), and then a seeming reduction in rate (although the behavior still exhibits exponential growth, with a significant goodness of the linear fit to the data from days 23-32). We are interested to explore whether these fluctuations are significant to the intrinsic disease dynamics, or whether they are confounded by other factors. The primary causes that would reflect in actual fluctuations of the infection rate per se are the population raised awareness to the epidemic, followed by the transportation bans and state-wide mandated closure of different venues (schools, campuses, churches, restaurants), followed by the directive to exercise social distancing. A potential confound is the limitation in testing ability, which, only indirectly related to the state-wide size of the outbreak, could lead to a significant under-representation of the real number of infections. While at this point there is no perfect way to discriminate between the two types of causes just based on the data, we will explore here some possibilities. One is to investigate if the trends in the infection time series synchronize with the timeline of different state-wide measures on both testing and social distancing. We work under the assumption that different New York State counties have very different social dynamics, which should reflect into a variety of patterns in the corresponding time series of county-wise confirmed infections. Along the same line, a synchronized trend among the time series of many counties is likely to be triggered by a state-wide mandated measure. Our study will therefore analyze the dynamic for each county, look for unifying trends, and attempt to interpret them in conjunction with state-wide control of these dynamics. First, we examine state wide data on testing, for the period from March 4th to April 1st, and look for trends that may underlie the patterns in confirmed infections. Efficiency of social distancing is harder to assess directly. We need to keep in mind that, while the results of a state government directed reduction in testing can be observed immediately, other factors, like social distancing measures, may need two weeks or longer to take effect. ![*[**Log time series**]{} for the states of California (left), Washington (center) and New York (right), illustrating the portions of the time series which justify a linear fit. The best fit line is shown as a dashed curve in each case, and the values of the slope and goodness of fit are shown in Table \[intro\_table\].*[]{data-label="comparison"}](Figure1.png){width="\textwidth"} [**County**]{} [**Interval**]{} [**Slope $\alpha$**]{} [**Sum of squares**]{} [**Pearson $\chi^2$**]{} ---------------- ------------------ ------------------------ ------------------------ -------------------------- California 36-68 0.1898 0.1528 0.0231 Washington 38-45 0.3842 0.0181 0.0064 45-52 0.2609 0.0334 0.0067 52-69 0.1416 0.0146 0.0020 69-72 0.0418 $<10^{-4}$ $<10^{-4}$ New York 4-7 0.6965 $<10^{-4}$ $<10^{-4}$ 7-16 0.2739 0.0357 0.0067 16-23 0.4593 0.0691 0.0079 23-32 0.1518 0.0347 0.0033 : *[**Rate of exponential growth**]{} calculated as a the slope of the linear fit to the log plot. The Pearson goodness of fit statistic is also provided in each case.*[]{data-label="intro_table"} For the county-wide analysis, we use archived data on the COVID 19 outbreak evolution in the US [@COVID_archive], between March 1st (when the first case was reported in the State of New York), and April 1st (the date this study was initiated). The information on population estimates and population density in each New York county was obtained from the Department of Health web page [@population]. For the counties which estimate over 100 cases by April 1st (to permit an adequate assessment), we study during which time windows the behavior is exponential, and calculated the exponential growth rate. As before, we consider for each county the logarithm of its time series, and search for portions that could be well approximated by linear behavior. We perform a best linear fit (using the traditional least square method), and compute the slope and goodness of fit for each almost linear portion. We illustrate separately the counties which had an early infection start (as listed in the introduction) and the counties with later starts (after March 10th). Results ======= State-wide trends ----------------- We performed a comparison of the evolution of the confirmed infection count versus the state-wide number of individuals tested for COVID-19, the number of COVID-associated hospitalizations and deaths. We used data from the COVID-19 Tracking repository [@testing], with raw time series represented in Figure \[NYS\_combined\]a and b. ![*[**Testing, hospitalization and casualties time series for the New York State,**]{} over the period of March 4th to April 1st. [**A.**]{} Total number of COVID-19 tests administered in blue, and number of positive tests (i.e., number of confirmed infections) in red. [**B.**]{} Number of COVID-19 diagnosed hospitalizations in yellow (data missing prior to March 20th), and number of COVID-19 associated deaths in black. [**C.**]{} Fraction of confirmed positive diagnoses out of the total number of tests performed state-wide.*[]{data-label="NYS_combined"}](Figure2.png){width="\textwidth"} ![*[**Title.**]{} [**A.**]{} Comparison of the log time series of the confirmed infections, as provided by the data source use for this section [@testing] (in red) and by the date source used in the next section [@COVID_archive] (in pink). [**B.**]{} Logarithmic time series corresponding to the raw time series in (A) and (B), shown with the same color coding. The illustration shows that all these functions are close to linear over a significant time period before April 1st. The slopes and goodness of fit for these time windows are shown in Table \[testing\_table\].* []{data-label="NYS_2"}](Figure3.png){width="65.00000%"} In Figure \[NYS\_2\]a, we show for cross-validation the time series of the state-wide infection count, as provided by two different sources in the public domain: the source that we used in conjunction with the county-wise infection breakdown [@COVID_archive] (same as in Figure\[intro\_table\]), covering the period from March 1st to April 1st; and the source that we used in conjunction with the evolution of state-wide testing [@testing], covering the period from March 4th to April 1st. On the common portion, the two are identical, via local minor details, and exhibit the same increasing patterns at different piece-wise slopes. [**Curve significance**]{} [**Interval**]{} [**Slope $\alpha$**]{} [**Sum of squares**]{} [**Pearson $\chi^2$**]{} ---------------------------- ------------------ ------------------------ ------------------------ -------------------------- Total tests 23-32 0.1155 0.0236 0.0020 Positive tests 23-32 0.1545 0.0257 0.0024 Hospitalizations 26-32 0.1614 0.0048 0.0005 Fatalities 24-32 0.2817 0.0193 0.0029 : *[**Rate of exponential growth**]{} calculated as a the slope of the linear fit to the log plot. The Pearson goodness of fit statistic is also provided in each case.*[]{data-label="testing_table"} County-wide behavior -------------------- In the State of New York, the first signs of contagion were detected and recorded in New York City (on March 1st), then a few days later in Westchester (March 4th), Nassau (March 5th), Rockland (March 6th), Saratoga (March 7th), Suffolk and Ulster Counties (March 8th). One month later, on April 1st, the leading five counties in infection counts, with over 3,000 infected individuals, were among this list of early foci: New York City (with 47,439 reported infections), Westchester County (10,683 infections), Nassau County (9,554 infections), Suffolk County (7,605 infections), Rockland County (3,321 infections). On the other hand, Saratoga and Ulster Counties were only counting 122 and respectively 222 infections by April 1st. We first illustrate and analyze data in these counties with confirmed early infection (before March 10th). These are networks of communities in which the infection has already propagated for a long enough time to provide more substantial data that may allow us to understand the mechanics of this propagation. We will illustrate separately the five counties which had over 3,000 confirmed infections by April 1st, and the two counties which, despite an early start, were an order of magnitude lower in the number of infections by the same date. While by April 1st New York City was transcending all other counties by a factor of at least five, and Rockland County showed the lowest numbers (Figure \[raw1\]a and b), in a normalized representation of infections per 1,000 individuals, New York City came fourth, following in order Westchester, Rockland and Nassau (Figure \[raw1\]c. It is also relatively easy to see that in different counties , confirmed infections appear to be increasing at different rates at the end of the time window: while Westchester is leading in terms of the proportion of individuals with confirmed diagnoses, Rockland is increasing at a higher rate. In order to better represent the evolution of the rate of change, and test for exponential trends, we again considered the time series in logarithmic form. ![*[**Time series showing the progression in the county-wide number of infected individuals**]{} up to April 1st, for the counties with early confirmed contamination (before March 10th) and over 3000 confirmed infections. [**A.**]{} New York City (red curve). [**B.**]{} Westchester County (green curve), Nassau County (blue curve), Rockland County (yellow curve) and Suffolk County (purple curve). [**C.**]{} Normalized time series showing the number of infected individuals per thousand of county residents, with the same color coding as in (B).*[]{data-label="raw1"}](Figure4.png){width="\textwidth"} ![*[**Log time series**]{} corresponding to the data shown in Figure \[raw1\], together with the best linear fit (shown as a dashed curve in each case): for the New York City population (left panel, black curve), Westchester (green curve), Nassau (blue curve), Rockland (brown curve) and Suffolk (purple curve), in the right panel.* []{data-label="log1"}](Figure5.png){width="65.00000%"} ![*[**Log time series**]{} together with the best linear fit (shown as a dashed curve in each case): for Saratoga County (left) and Ulster County (right).*[]{data-label="log2"}](Figure6.png){width="65.00000%"} We performed linear fitting for each county, starting with the day when the first case was reported and detecting best fit linear intervals throughout the whole evolution. The slopes $\alpha$ (corresponding to the exponential growth rates) and the goodness of fit statistics are shown in Table \[table1\]. [**County**]{} [**Interval**]{} [**Slope $\alpha$**]{} [**Sum of squares**]{} [**Pearson $\chi^2$**]{} ---------------- ------------------ ------------------------ ------------------------ -------------------------- New York City 1-17 0.4345 0.4935 0.0078 17-22 0.5253 0.1001 0.0128 22-32 0.1579 0.0911 0.0093 Westchester 7-16 0.1269 0.0116 0.0024 16-26 0.3239 0.1104 0.0165 26-32 0.0922 0.0080 0.0009 Nassau 9-18 0.2712 0.1019 0.0256 18-22 0.5853 0.0462 0.072 22-32 0.1600 0.0123 0.0015 Suffolk 12-18 0.2744 0.0137 0.0037 18-22 0.5689 0.0237 0.0041 22-32 0.1952 0.0543 0.0070 Rockland 9-18 0.2077 0.0747 0.0356 18-22 0.7036 0.0408 0.0088 22-32 0.1991 0.0923 0.0123 Saratoga 12-15 0 0 0 15-21 0.3910 0.1306 0.0629 21-32 0.1144 0.0829 0.0194 Ulster 12-20 0.1444 0.0618 0.0339 20-26 0.3275 0.0086 0.0029 26-32 0.1641 0.0254 0.0050 : *[**Rate of exponential growth**]{} calculated as a the slope of the linear fit to the log plot. The Pearson goodness of fit statistic is also provided in each case (one outlier – March 24 – was left out when computing the slopes for Rockland County).*[]{data-label="table1"} Notice that (after an initial transient of a few days in some cases) all these counties show a similar piece-wise linear pattern to that pointed out in the state-wide logarithmic time series. In each case, we were able to identify three pieces: a milder increasing segment ending between March 16-18, followed by a steeper segment ending on March 22 (26 in the case of Westchester), followed again by a linear segment with more relaxed slope, and very significant goodness of fit. While the numbers were an order of magnitude lower for Saratoga and Ulster Counties, the same three-piece pattern was surprisingly apparent in their time series as well, suggesting that the piece-wise linear effect with alternating steepness which we observed in the state wide dynamics was not an average effect, but is rather present in every one of the counties with early contamination. This suggests that the effect is either a reflection of specific New York social dynamics across counties – leading to a subsequently specific spread pattern and speed – or more likely the effect of state-wide unifying controls. We want to further investigate whether the transitions between linear pieces use as temporal reference the time of the original confirmed infection in the corresponding county, or if they are synchronized at a central level. Since the counties represented so far had similar times of original infection, we need to compare with the similar data in counties with a different timeline. The only other nine New York State counties which had a count of over 100 infection cases by April 1st are as follows, in order of the number: Orange (1,756), Erie (582), Dutchess (547), Monroe (349), Onondaga (277), Albany (240), Ulster (222), Putnam (207), Saratoga (122) and Sullivan (121). Of these, as specified before, Saratoga and Ulster had confirmed cases by March 10th; the other counties have later contamination dates. Their logarithmic time series, together with the linear fits, are shown in Figure \[log3\]; the slopes and goodness of fit statistics are described in Table \[table3\]. The time series of these counties all exhibit only two linear segments (the first with steeper slope, followed by one with flatter slope) with the transition occurring over the same short time window (March 22-26) as the similar transition in the counties with early starts. This confirms the idea that the evolution in this later batch is not replicating the evolution of the earlier counties from the very start, but rather is correlating with their current evolution (further suggesting that the reason behind these trends is centralized control). ![*[**Log time series**]{} corresponding to the data shown in Figure \[raw1\], together with the best linear fit (shown as a dashed curve in each case): for the New York City population (left panel, black curve), Westchester (green curve), Nassau (blue curve), Rockland (brown curve) and Suffolk (purple curve), in the right panel.* []{data-label="log3"}](Figure7.png){width="\textwidth"} [**County**]{} [**Interval**]{} [**Slope $\alpha$**]{} [**Sum of squares**]{} [**Pearson $\chi^2$**]{} ---------------- ------------------ ------------------------ ------------------------ -------------------------- Orange 15-23 0.5293 0.0527 0.0155 23-32 0.1655 0.0499 0.0074 Albany 16-22 0.3640 0.0810 0.0268 22-32 0.0688 0.0116 0.0023 Dutchess 13-22 0.3583 0.1050 0.0439 22-32 0.1913 0.0073 0.0014 Onondaga 20-23 0.6149 0.0108 0.0042 23-32 0.2061 0.0850 0.0174 Putnam 19-22 0.9898 0.1477 0.0707 22-32 0.1566 0.1164 0.0271 Sullivan 20-26 0.3092 0.0113 0.0042 26-32 0.1366 0.0105 0.0024 Erie 15-19 0.4621 0.4982 0.2044 19-32 0.2537 0.2827 0.0553 Monroe 14-19 0.5465 0.6212 0.4150 19-32 0.1928 0.2246 0.0490 : *[**Rate of exponential growth**]{} calculated as a the slope of the linear fit to the log plot. The Pearson goodness of fit statistic is also provided in each case.*[]{data-label="table3"} Discussion ========== In this study, we focused on the dynamics of the COVID-19 epidemic in the state of New York for the first month of the outbreak (March 1st to April 1st), and we analyzed state-wide data on confirmed infections, testing, hospitalizations and deaths, as well as county-wide data on infection rates, for the counties in the state that were reporting over 100 infections by April 1st. Our primary goal was to determine whether it is possible to dissociate between the real, infection-related effects of social distancing measures, and the reporting effects produced by the limitations in testing. In particular, we identified a signature in the evolution of reported infections, specific to the state on New York, characterized by following exponential growths with rates oscillating between higher and lower values, for different time windows, eventually settling to a relatively low exponential rate. We aimed understand the factors behind the transitions, and whether the trend shown at the start of April corresponds to a damping in infections, or is an artifact. When looking at the state-wide patterns of testing, we observed a substantial spike between March 12th and March 13th. While this is not obvious in the raw data in Figure \[NYS\_combined\]a, it is clear in the logarithmic plot in Figure \[NYS\_2\]b, as well as in the sudden drop in the fraction of positive tests shown in Figure \[NYS\_combined\]c (higher availability of tests allowed for tests on more susceptible individuals who were not necessarily infected). In fact, notice that before March 12th, this ratio had steadily increased to the point where over 70% of tests were positive; on March 13th, this ratio dropped to under 15%, but promptly started creeping up again, having climbed back to almost 40% by April 1st. This is likely the promoter for the ramping up in the slope of detected infections (a comparison between the blue and red curves in Figure \[NYS\_2\] suggests that the two effects occurred simultaneously at the overall state level). For specific counties, the distribution of this transition point in the slope ranged from March 12 to March 18, but can still be viewed as a potential effect of the increase in testing (especially since the counties with later infection onset do not show this transition, and start off with a segment of higher exponential rate directly). In this light, these slopes are expected to show a more faithful (if not perfect) representation of the actual growth in infections, since it is less restricted by testing limitations. Interestingly enough, there does not seem to be any disruption to this existing trend in positive versus total number of tests around March 20th, in conjunction with the state government directive to refocus testing primarily around individuals with high hospitalization risk. However, this date coincides with the last transition in the infection curve for New York State, from a higher exponential diagnosis rate $\alpha \sim 0.46$ to a lower rate $\alpha \sim 0.15$. In support of this observation is that all counties exhibit a similar transition at dates distributed between March 19 and March 23, from a higher exponential rate (with mean $\mu_\alpha = 0.51$ and standard deviation $\sigma_\alpha = 0.18$ between the 15 counties examined) to a milder rate (with mean $\mu_\alpha = 0.16$ and standard deviation $\sigma_\alpha = 0.04$). This interval appears too early to reflect any of the state mandated initial closures (which were initiated around March 12th, would need at least two weeks to take effect and would likely not be as tightly synchronized between counties). This is also too early to reflect the PAUSE directive, which could produce a highly synchronized response between counties, but the effects of which are not expected to be immediate. A strong hypothesis to consider is that the number of infections (with a rate of exponential increase possibly even transcending the reported $\alpha \sim 0.5$ up to March 20th) overtook the sampling rate provided by the testing. In order to maintain realistic reporting, testing availability would have had to also increase at a comparable pace. The state governor directive on March 20th allegedly had as an effect restricting testing on potentially positive and negative COVID-19 diagnoses at the same extent (people being instructed to refrain from testing unless in very serious condition). While there was no visible cut in testing, or additional imbalance in positive versus negative tests, the directive may have effectively capped the reported infections to the point where the rate of the new infections ($\alpha =0.15$) settled only a little higher than the overall rate of testing ($\alpha = 0.115$) after March 20th. Let’s assume here that the lowering of the exponential rate is only a testing-driven artifact, and the infections were still occurring up to April 1st at the same, higher rate $\alpha = 0.46$ shown prior to the transition in slope on March 23rd. Then the actual number of infections in New York State would be an estimate of 1,302,800 infected individuals by April 1st, compared to the reported 83,712, and the projected 80,524 when using the $\alpha = 0.15$ lower rate. Additional information will contribute over the next days and weeks to assessing whether this is the case or not. One option could be provided by monitoring whether the changes expected to follow PAUSE will actually appear in the infection time series. Another, potentially more reliable assessment will come from the evolution of casualties, evolving prior to April 1st at an exponential rate $\alpha = 0.28$ higher than the infection rate estimated from the confirmed cases, and which is expected to increase (assuming that COVID-19 susceptible deaths are actually tested). When considering the number of confirmed infections for each county on April 1st, one simple observation is that they did not correlate with the county total population, but rather with the population density, as reported in each county by the NYS Health Department (correlation coefficient $R=0.4195$, significance value $p=0.0007$). This suggests that these numbers may still reflect, at least to some extent, the type of social dynamics associated to the population density distribution in each county. Our future work involves quantifying county specific evolution of social dynamics (from how it reflected into travel patterns to different destinations within the community), and understanding whether these dynamics contributed to the slopes for each county, and to the transitions between regimes. [99]{} <https://github.com/CSSEGISandData/COVID-19/tree/master/csse_covid_19_data/csse_covid_19_daily_reports>. Accessed 04/02/2020. <https://www.health.ny.gov/statistics/vital_statistics/2010/table02.htm>. Accessed 04/02/2020. <https://covidtracking.com/data/state/new-york#historical>. Accessed 04/04/2020.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Relations between two classes of Hilbert spaces of entire functions, de Branges spaces and Fock-type spaces with non-radial weights, are studied. It is shown that any de Branges space can be realized as a Fock-type space with equivalent area norm, and several constructions of a representing weight are suggested. For some special classes of weights (e.g., weights depending on the imaginary part only) the corresponding de Branges spaces are explicitly described.' address: 'Anton Baranov, Department of Mathematics and Mechanics, St. Petersburg State University, St. Petersburg, Russia, National Research University Higher School of Economics, St. Petersburg, Russia, Hélène Bommier-Hato, I2M, Aix-Marseille Université, CNRS, Marseille, France, Faculty of Mathematics, University of Vienna, Vienna, Austria, ' author: - 'Anton Baranov, Hélène Bommier-Hato' title: De Branges spaces and Fock spaces --- [^1] Introduction ============ Fock-type spaces and de Branges spaces are arguably two main examples of reproducing kernel Hilbert spaces (RKHS) of entire functions. They include as special cases the Bargmann–Segal–Fock space (famous for its connections with quantum mechanics and time-frequency analysis) and the Paley–Wiener space of bandlimited functions. The aim of the present paper is to establish a link between these two classes of spaces. Fock and de Branges spaces -------------------------- Let us give the necessary definitions. For a measurable function $W: {{\mathbb C}}\rightarrow(0,\infty)$, one defines the [*Fock-type space*]{} $${{\mathcal F}}_W:=\left\{F\text{ entire}: \ \left\|F\right\|^{2}_{{{\mathcal F}}_W}=\int_{{{\mathbb C}}}\left|F(z)W(z)\right|^2dm(z)<\infty\right\},$$ where $dm$ stands for the area Lebesgue measure in the complex plane $\mathbb{C}$. A special attention in the literature was addressed to radial weights (i.e., $W(z) = W(|z|)$; see, e.g., [@S; @bdk; @bom] and references therein. The classical Bargmann–Segal–Fock space corresponds to $W(z)= e^{-\frac{|z|^2}{2}}$. However, in the present paper we will be mainly interested in non-radial weights. An entire function $E$ is said to be in the [*Hermite–Biehler class*]{} if it satisfies $\left|E(z)\right|>\left|E^{\sharp}(z)\right|$ for $z\in{{\mathbb C}}^+$. Here ${{\mathbb C}}^+$ denotes the upper half-plane and $E^\sharp(z) = \overline{E(\overline{z})}$. We also always assume in what follows that $E$ does not vanish on $\mathbb{R}$. Any such function $E$ determines a [*de Branges space*]{} (see [@db]) $${{\mathcal H}}(E)=\left\{F\text{ entire}:\ \frac{F}{E},\frac{F^{\sharp}}{E}\in H^{2}({{\mathbb C}}^+)\right\},$$ ($H^2= H^2({{\mathbb C}}^+)$ being the standard Hardy space in the upper half-plane) which is a Hilbert space when equipped with the norm $$\left\|F\right\|^{2}_{E}=\int_{{{\mathbb R}}} \left|\frac{F(x)}{E(x)}\right|^{2}dx.$$ There exist alternative descriptions of de Branges spaces: - Axiomatic (see [@db Theorem 23]) – any reproducing kernel Hilbert space of entire functions $\mathcal{H}$ such that the mapping $F\mapsto F^\sharp$ preserves the norm in $\mathcal{H}$ and the mapping $F\mapsto \frac{z-\overline w}{z-w}F(z)$ is an isometry in $\mathcal{H}$ whenever $w\in{{\mathbb C}}\setminus\mathbb{R}$, $F(w) = 0$, is of the form ${{\mathcal H}}(E)$ for some $E$ in the Hermite–Biehler class); - Via the [spectral data]{} (see Section \[spect\]). Spaces ${{\mathcal H}}(E)$ were introduced by L. de Branges in his celebrated solution of the inverse spectral problem for canonical systems of differential equations. At the same time, these spaces are a very interesting object from the function theory point of view (see, e.g., [@ls; @os; @hm1; @hm2; @kw; @bbb]). The Paley–Wiener space $PW_a$ (which consists of entire functions of exponential type at most $a$ which are square integrable on $\mathbb{R}$) is a de Branges space corresponding to $E(z) = e^{-iaz}$. Any de Branges space is a RKHS (i.e., evaluation functionals are bounded). The Fock space ${{\mathcal F}}_W$ also is a RKHS if we assume, e.g., that $$\label{bab1} \inf_{z\in K} |W(z)| >0$$ for any compact $K\subset {{\mathbb C}}$. In what follows we always impose this restriction on $W$. Note that if two RKHS coincide as sets then, by the Closed Graph Theorem, their norms are automatically equivalent. Recently, in [@bbb Theorem 1.2] a complete description was found of those de Branges spaces which coincide with [*radial*]{} Fock spaces as sets, [*with equivalence of norms*]{}. (It is clear that a Fock space can not isometrically coincide with a de Branges space since division by the Blaschke factor $\frac{z-w}{z-\overline w}$ is not an isometry.) These turn out to be very small spaces generated by functions $E$ of slow growth (in particular, satisfying $\log|E(z)| =O(\log^2 |z|)$, $|z|\to \infty$). Interestingly, this is exactly the class of de Branges spaces where any complete and minimal system of reproducing kernels (RK) admits the so-called spectral synthesis (see [@bbb]). Also, in [@bbb1] the inverse problem was considered: given a radial Fock space ${{\mathcal F}}_W$ with $W(z) = \exp(-{\varphi}(|z|))$ for some increasing ${\varphi}$, when does it coincide with some de Branges space? Under some regularity conditions on ${\varphi}$, the same growth gives a sharp threshold here: if ${\varphi}(r) = O(\log^2 r)$, then ${{\mathcal F}}_W$ has a Riesz basis of (normalized) RK corresponding to real points and, thus, coincides with a de Branges space (this is a characteristic property of de Branges spaces that they have orthogonal bases of RK). If, on the contrary, $\log^2 r = o({\varphi}(r))$, $r\to\infty$, then ${{\mathcal F}}_W$ has no Riesz bases of RK and so it does not coincide with any de Branges space. In view of these results one may ask whether it is possible to represent any de Branges space as a Fock-type space ${{\mathcal F}}_W$ with a [*non-radial*]{} weight, i.e., to replace the norm given by an integral over $\mathbb{R}$ by an equivalent area integral. One of the main results of the present paper says that it is always possible. However, such weight $W$ (we call it a [*representing Fock weight for ${{\mathcal H}}(E)$*]{}) is by no means unique, and there arises a question how to choose it in a “canonical” way so that it encompasses in the most economical way the properties of the space. This is the question we address in the present paper. Main results ============ Natural weight and the Paley–Wiener spaces ------------------------------------------ We start with the simple special case of the problem: [*find a weight $W$ such that $PW_a = {{\mathcal F}}_W$*]{}. Recall that all equalities of spaces are understood as equalities of sets with equivalence (but not equality) of norms. In what follows we write $U(x)\lesssim V(x)$ (or, equivalently, $V(x)\gtrsim U(x)$) if there is a constant $C$ such that $U(x)\leq CV(x)$ holds for all values of the parameter $x$. We write $U(x)\asymp V(x)$ if both $U(x)\lesssim V(x)$ and $V(x)\lesssim U(x)$. Note that for an arbitrary de Branges space ${{\mathcal H}}(E)$ we have $F/E\in H^2$ for any $F\in {{\mathcal H}}(E)$ and so $$\sup_{y>0} \int_{{\mathbb R}}\bigg|\frac{F(x+iy)}{E(x+iy)}\bigg|^2dx = \|F\|^2_E.$$ Hence, $$\int_{{{\mathbb C}}^+} \frac{1}{(y+1)^2} \bigg|\frac{F(x+iy)}{E(x+iy)}\bigg|^2 dxdy \le \|F\|^2_E.$$ This leads to the following “natural” candidate for the weight $W$: $$\label{wb} \begin{aligned} W_0(z) & :=\frac{1}{\left|E(z)\right|\left(1+{{\rm Im}\,}z\right)}, \qquad & z\in {{\mathbb C}}^+\cup{{\mathbb R}}, \\ W_0(\overline z) & := W(z), & z\in {{\mathbb C}}^-. \end{aligned}$$ Since $F^\sharp/E$ also is in $H^2$ we always have ${{\mathcal H}}(E) \subset {{\mathcal F}}_{W_0}$. The weight $W_0$ depends on $E$, but usually it will be clear from the context for which space it is constructed and so we do not introduce $E$ into the notation. Clearly, $(1+y)^{-1}$ in the definition of the weight can be replaced by $(1+y)^{-\alpha}$ with any $\alpha\in (1/2, \infty)$, as well as by an arbitrary positive function from $L^2(0, \infty)$ with at most power decay. The choice of this function is of no importance. If $E(z) = e^{-iaz}$, then $W_0(z) = (1+|{{\rm Im}\,}z|)^{-1}e^{-a|{{\rm Im}\,}z|}$. It turns out that this weight will define an equivalent norm on $PW_a$. Indeed, let $F\in {{\mathcal F}}_W$ and $z = x+iy \in{{\mathbb C}}$. We denote by $D(z, r)$ the open disc with the center $z$ of radius $r$. Applying the subharmonicity property to $|F|^2$ in the disc $D(z, |y|+1)$, we get $$|F(z)|^2 \le \frac{1}{\pi(|y|+1)^2} \int_{D(z, |y|+1)} |F(\zeta)|^2 dm(\zeta) \lesssim e^{2a |y|} \int_{D(z, |y|+1)} |F(\zeta)W(\zeta)|^2 dm(\zeta).$$ with a constant independent on $F$ and $z$. Hence, $|F(x)|^2 \lesssim \int_{D(x,1)} |F(\zeta)W(\zeta)|^2 dm(\zeta)$ and integrating over $x$ we see that $F\in L^2({{\mathbb R}})$. Also, by the above estimates, $F$ is of exponential type at most $a$ and so $F\in PW_a$. Thus, ${{\mathcal F}}_{W_0} = PW_a$. It is interesting to note that for any weight of the form $W(z) = (1+|{{\rm Im}\,}z|)^{\alpha} e^{-a|{{\rm Im}\,}z|}$, $\alpha>0$, the corresponding Fock space does not coincide with a de Branges space. The reason for that is that the space ${{\mathcal F}}_W$ has no Riesz bases of normalized RK. This result was communicated to us by A. Borichev. It follows also from our Theorem \[depy\] which says that the Paley–Wiener spaces are the only de Branges space which have a representing weight depending on ${{\rm Im}\,}z$. As we will see, the above simple argument can not be extended to general de Branges spaces: it may happen that ${{\mathcal H}}(E) \ne {{\mathcal F}}_{W_0}$. However, we are able to describe the class of de Branges spaces for which the “natural” weight $W_0$ generates the same space. Recall that for any Hermite–Biehler function $E$ one can find a decreasing branch of the argument on ${{\mathbb R}}$ (see [@db]), that is there exists an increasing continuous function $\varphi$ (so-called [*phase function*]{}) such that $$\label{phase} E(x)=\left|E(x)\right|e^{-i\varphi(x)},\qquad x\in {{\mathbb R}}.$$ If $Z_E=\left\{\overline z_n=x_n-iy_n\right\}\subset {{\mathbb C}}^-$ denotes the zero set of $E$ (note that in our notation $z_n \in {{\mathbb C}}^+$), then there exists $a \geq 0$ such that $$\label{def phi'} \varphi'(x)=a+\sum_{n}{{\rm Im}\,}\frac{1}{x- z_n} = a+\sum_{n} \frac{y_n}{|x-\overline z_n|^2}, \qquad x\in {{\mathbb R}}.$$ Now we can state our first main result. \[F Wb\] Let ${{\mathcal H}}(E)$ be a de Branges space and let $W_0$ be defined by . Then the following are equivalent: 1. The spaces ${{\mathcal H}}(E)$ and ${{\mathcal F}}_{W_0}$ coincide as sets; 2. $\varphi' \in L^{\infty}({{\mathbb R}})$. De Branges spaces with the property $\varphi' \in L^{\infty}({{\mathbb R}})$ (sublinear growth of the argument) have many nice properties similar to the classical Paley–Wiener spaces (see [@dy] or [@hm2]). General de Branges spaces ------------------------- We have seen that the weight $W_0$ solves the problem only for a special class of de Branges spaces. In general we need to modify it to encompass the subtle behavior of the elements of ${{\mathcal H}}(E)$ near ${{\mathbb R}}$ in the case when zeros of $E$ approach the real axis. We will do this using the level sets of inner functions. Note that any Hermite–Biehler function $E$ gives rise to a meromorphic inner function in ${{\mathbb C}}^+$, namely $$\label{def Theta} \Theta=\frac{E^{\sharp}}{E},$$ which has the form $$\label{merom} \Theta(z)=e^{i\tau z}\prod_{n}{\varepsilon}_n \frac{z-z_n}{z-\overline z_n},\qquad z\in{{\mathbb C}},$$ where $\overline{z}_n$ are zeros of $E$, ${\varepsilon}_n=\frac{\left|z_n^2+1\right|}{z_n^2+1}$ (for $z_n=i$, ${\varepsilon}_n=1$) and $\tau$ is a nonnegative constant. Note that $|\Theta'(x)| = 2{\varphi}'(x)$, $x\in {{\mathbb R}}$. With each inner function $\Theta$ one can associate the [*model*]{} ([*backward shift invariant*]{}) subspace $K_\Theta = H^2\ominus \Theta H^2$ of the Hardy space $H^2$. The map $F\mapsto F/E$ is a canonical unitary map from ${{\mathcal H}}(E)$ onto $K_\Theta$ where $\Theta$ is defined by (this is easy to verify; see, e.g., [@hm1 Theorem 2.10]). The reproducing kernel of ${{\mathcal H}}(E)$ is given by $$K_w(z)=\frac{i}{2\pi}\frac{E(z)\overline{E(w)}-E^{\sharp}(z) \overline{E^{\sharp}(w)}}{z-\overline w},\qquad z,w\in {{\mathbb C}};$$ in particular, we have $\|K_x\|^2_E = K_x(x) = |E(x)|^2 \varphi'(x)/\pi$, $x\in {{\mathbb R}}$. Note also that $K_w(z) = E(z)\overline{E(w)} k_w(z)$ where $$k_w(z) = \frac{i}{2\pi} \frac{1-\overline{\Theta(w)}\Theta(z)}{z-\overline{w}}$$ is the reproducing kernel of $K_\Theta$ and $$\|k_z\|_2^2 = \frac{1-|\Theta(z)|^2}{4\pi {{\rm Im}\,}z}, \quad z\in {{\mathbb C}}^+,\qquad \|k_x\|_2^2 = |\Theta'(x)|/2\pi, \quad x\in {{\mathbb R}};$$ here $\|f\|_2$ denotes the standard $L^2({{\mathbb R}})$-norm. Though the function $\Theta$ associated to a de Branges space is meromorphic in ${{\mathbb C}}$, we prefer now to work in a slightly more general setting. Let $\Theta$ be an arbitrary inner function in ${{\mathbb C}}^+$ (see, e.g., [@ga]). Let $\sigma(\Theta)$ be the so-called [*spectrum*]{} of $\Theta$, that is, the set of all $\zeta\in\overline {\mathbb{C^+}}$ such that $\liminf\limits_{z\to\zeta,\, z\in\mathbb{C^+}}|\Theta(z)|=0$. Equivalently, $\sigma(\Theta)$ is the smallest closed subset of ${\mathbb{C^+}}\cup {{\mathbb R}}$ containing the zeros $z_n$ of $\Theta$ and the support of the singular measure associated with the singular factor in $\Theta$. In the case of meromorphic $\Theta$, $\sigma(\Theta)$ is just the zero set $\{z_n\}$. Given $\Theta$ and $\varepsilon\in (0,1)$, consider the sublevel sets $$\Omega_\varepsilon=\{z\in\mathbb{C^+}: |\Theta(z)|<\varepsilon\} \qquad\text{and} \qquad \Omega^c_\varepsilon = ({{\mathbb C}}^+\cup{{\mathbb R}}) \setminus\Omega_\varepsilon.$$ For $z\in{{\mathbb C}}^+ \cup \mathbb{R}$ put $$d_\varepsilon(z)={\rm dist} (z,\Omega_\varepsilon);\qquad\quad d_0(z)={\rm dist} (z,\sigma(\Theta)).$$ Then, for $0<\varepsilon<\delta<1$ there exist constants $A_1,A_2>0$ depending only on $\varepsilon, \delta$ (but not on $\Theta$!) such that $$\label{lev} A_1\min (d_0(z), \|k_z\|_2^{-2}) \le d_\varepsilon(z) \le A_2 \min (d_0(z), \|k_z\|_2^{-2}), \qquad z\in \Omega_\delta^c.$$ This inequality was established in [@bar1 Theorem 4.9] for the case $z=x\in{{\mathbb R}}$. The proof of is very similar to the proof in [@bar1]; we give its sketch in Section \[level\]. Any de Branges space can be realized as a Fock-type space with the representing weight depending on the geometry of level sets. By $\chi_e$ we denote the usual characteristic function of a set $e$. \[main\] Let ${{\mathcal H}}(E)$ be a de Branges space, let $W_0$ be defined by , and let $\Theta$ be the associated inner function. For $0<\varepsilon<\delta<1$ put $$\label{mw} W(z) = W_0(z)\big(1+ (d_{\varepsilon}(z))^{-1/2} \chi_{\Omega_\delta^c} (z) \big), \qquad z\in {{\mathbb C}}^+ \cup{{\mathbb R}},$$ and $W(z) = W(\overline z)$, $z\in {{\mathbb C}}^-$. Then ${{\mathcal H}}(E) = {{\mathcal F}}_W$. An important subclass of inner functions is formed by $\Theta$ such that the set $\Omega_{\varepsilon}$ is connected for some ${\varepsilon}\in(0,1)$. They are usually called [*one-component*]{} inner functions. This class was introduced by W. Cohn [@cohn] and studied extensively by A. Aleksandrov (see, e.g., [@al1] and references therein). Some results about de Branges spaces whose $\Theta$ is one-component can be found in [@MNO]. In particular this paper contains an observation (communicated to authors of [@MNO] by the first author) that a meromorphic $\Theta$ is one-component if and only if the measure ${\varphi}'(x) dx$ on ${{\mathbb R}}$ is locally doubling. It follows from the results of [@bar1] (and will be explained below) that for one-component inner function one has $d_{\varepsilon}(z) \asymp \|k_z\|_2^{-2}$, $z\in \Omega_\delta^c$. Thus, we have the following corollary: \[one1\] If $\Theta = E^\sharp/E$ is a one-component inner function, then for $0<\varepsilon<\delta<1$, we have ${{\mathcal H}}(E)={{\mathcal F}}_W$ where $W(z) = W_0(z)\big(1+ \|k_z\|_2\, \chi_{\Omega_\delta^c} (z) \big)$, $z\in {{\mathbb C}}^+ \cup{{\mathbb R}}$, extended symmetrically to ${{\mathbb C}}^-$. Weights depending on imaginary part ----------------------------------- We have seen that for the Paley–Wiener space $PW_a$, one can choose a representing Fock weight $W_0(z) = (|{{\rm Im}\,}z|+1)^{-1}e^{-a|{{\rm Im}\,}z|}$ which depends only on ${{\rm Im}\,}z$. Are there other de Branges spaces with the same property? The answer turns out to be negative: \[depy\] Let ${{\mathcal H}}(E)$ be a de Branges space. Then the following are equivalent: 1. There exists a representing weight $W$ for ${{\mathcal H}}(E)$ such that $W(z) = W({{\rm Im}\,}z)$, $z\in{{\mathbb C}}$; 2. ${{\mathcal H}}(E) = PW_a$ for some $a>0$. Note that there exist functions $E$ such that $|E|$ is not equivalent to 1 on ${{\mathbb R}}$, but ${{\mathcal H}}(E) = PW_a$. We discuss such construction (from paper [@ls]) in Section \[th3\]. We conclude with one question which we were not able to answer. Note that the property (i) implies, in particular, that all real translations $U_t: F\mapsto F(\cdot+t)$, $t\in\mathbb{R}$, are isomorphisms of ${{\mathcal H}}(E)$ onto itself and $\sup_{t\in{{\mathbb R}}}\|U_t:{{\mathcal H}}(E)\to{{\mathcal H}}(E)\|<\infty$. It would be interesting to know whether the converse is true, that is, whether there exist de Branges spaces where all real translations are isomorphisms with uniformly bounded norms which are different from the Paley–Wiener spaces. Organization of the paper ------------------------- In Section \[level\] we prove some auxiliary estimates for the distances to the level sets. In Section \[th2\] we prove the main result of the paper, Theorem \[main\]. In Section \[th1\] we prove necessity of the condition ${\varphi}'\in L^\infty$ in Theorem \[F Wb\] (its sufficiency follows almost immediately from Theorem \[main\]). In Section \[spect\] we give another construction of a representing weight for a general de Branges space which uses geometry of the spectral data for ${{\mathcal H}}(E)$. Finally, in Section \[th3\] we prove Theorem \[depy\] on the weights depending on the imaginary part only. Distances to the level sets {#level} =========================== In this section we prove inequality . \[lev1\] Let $0<\varepsilon<\delta<1$. Then there exist constants $A_1, A_2>0$ depending on ${\varepsilon},\delta$ such that for any inner function $\Theta$ in ${{\mathbb C}}^+$, $$\label{lev2} A_1\min (d_0(z), \|k_z\|_2^{-2}) \le d_\varepsilon(z) \le A_2 \min (d_0(z), \|k_z\|_2^{-2}), \qquad z\in \Omega_\delta^c.$$ Without loss of generality, we can assume that $\Theta$ is a Blaschke product. Indeed, by the Frostman theorem (see, e.g., [@ga Chapter II]), Blaschke products are uniformly dense in the set of all inner functions. So in what follows $\Theta=B$ is a Blaschke product with zeros $z_n=x_n+i y_n$, $n\in{{\mathbb N}}$, with $y_n>0$, and for each integer $n$, we set $$b_n(z)=\frac{z-z_n}{z-\overline z_n},\qquad B_n=\prod^{n-1}_{j=1} b_j \qquad (B_0\equiv 1).$$ Suppose that $z\in \Omega_{\delta}^c$ and $\left|w-z\right|<\alpha\min\left(d_{0}(z), \left\|k_z\right\|_2^{-2}\right)$. We will show that if $\alpha>0$ is small enough, then $\left|B(w)\right|>{\varepsilon}$. This will prove the estimate from below in . In what follows the constants in estimates $\lesssim$ and $\asymp$ depend only on $\delta$ and ${\varepsilon}$, but do not depend on $B$.\ [**Estimate of**]{} $d_{\varepsilon}$ [**from below.**]{} Since $z\in \Omega_{\delta}^c$, we have $$\label{esti bn} \delta\leq \left|B(z)\right|\leq \left|B_{n+1}(z)\right| \leq\left|\frac{z-z_n}{z-\overline z_n}\right|,\qquad n\in{{\mathbb N}}.$$ Using now the identities $$\begin{gathered} \label{ert} 2\pi\left\|k_z\right\|_2^{2}=\frac{1-\left|B(z)\right|^2}{{{\rm Im}\,}z}=\sum_{n}\left|B_{n}(z)\right|^2\frac{1-\left|b_n(z)\right|^2}{{{\rm Im}\,}z},\\ \label{ert0} \frac{1-\left|b_n(z)\right|^2}{{{\rm Im}\,}z}=\frac{4 y_n}{\left|z-\overline z_n\right|^2},\end{gathered}$$ we get that $$\label{esti1 k_z} \delta^2\sum_{n}\frac{4 y_n}{\left|z-\overline z_n\right|^2}\leq 2\pi\left\|k_z\right\|_2^2\leq\sum_{n}\frac{4 y_n}{\left|z-\overline z_n\right|^2}.$$ Recall that $$|B'(x)| = \sum_{n}\frac{2 y_n}{|x-z_n|^2}, \qquad x\in {{\mathbb R}}$$ (of course, in general $|B'(x)|$ need not be finite). It follows from that $$\label{ert1} \delta|z-\overline z_n|\le |x-z_n| \le |z-\overline z_n|$$ and so, by , $$\delta^4\left|B'(x)\right|\leq 2\pi\left\|k_z\right\|_2^2, \qquad z=x+iy\in\Omega_{\delta}^c.$$ Thus, for $z=x+iy\in\Omega_{\delta}^c$, we have $$\label{esti3 k_z} \left\|k_z\right\|_2^2\asymp \left|B'(x)\right|\qquad \text{and}\qquad d_{0}(z) \asymp d_{0}(x).$$ Now, let $w=u+iv$, $v>0$. We have $$\log\left|\frac{B(w)}{B(z)}\right|^2 = \sum_{n}\log\left[\frac{\left(u-x_n\right)^2+\left(v-y_n\right)^2} {\left(u-x_n\right)^2+\left(v+y_n\right)^2} \cdot \frac{\left(x-x_n\right)^2+\left(y+y_n\right)^2}{\left(x-x_n\right)^2+\left(y-y_n\right)^2}\right] =\sum_{n} \log\left[1-Y_n\right]$$ where $$Y_n = \frac{4vy_n\left(x-x_n\right)^2-4yy_n\left(u-x_n\right)^2} {\left|z-z_n\right|^{2}\left|w-\overline z_n\right|^{2}}-\frac{\left(y-y_n\right)^2 \left(v+y_n\right)^2 + \left(y+y_n\right)^2\left(v-y_n\right)^2}{\left|z-z_n\right|^{2}\left|w-\overline z_n\right|^{2}}.$$ Since $\left|w-z\right|<\alpha d_0(z)$, we observe that $$\label{esti w-zn bar} \left(1-\alpha\right)\left|z-\overline z_n\right|\leq \left|w-\overline z_n\right|\leq \left(1+\alpha\right)\left|z-\overline z_n\right|.$$ Using (\[esti bn\]) and (\[esti w-zn bar\]), it is easily seen that $$\begin{aligned} \sum_n\left|Y_n\right|&\leq C(\delta)\sum_n\frac{\left|w-z\right|y_n}{\left|z-\overline z_n\right|^2}\\ &\leq \tilde C(\delta)\left|w-z\right|\left\|k_z\right\|_2^2\leq \tilde C(\delta)\alpha.\end{aligned}$$ Hence, if $\alpha$ is sufficiently small, we conclude that $$\log\left|\frac{B(w)}{B(z)}\right|^2> 2 \log\frac{{\varepsilon}}{\delta},$$ whence $\left|B(w)\right|>{\varepsilon}.$ Thus, there exists a positive constant $C=C(\delta,{\varepsilon})$ such that $$d_{{\varepsilon}}(z)\geq C\min\left(d_{0}(z),\left\|k_z\right\|_2^{-2}\right).$$\ [**Estimate of**]{} $d_{\varepsilon}$ [**from above.**]{} Let us prove the converse inequality. Let $z=x+iy\in\Omega_{\delta}^c$. Inequality implies that $|z-z_n| \ge \delta (2y -|z-z_n|)$ whence $y\lesssim d_{0}(z)$. Also, $y = (2\pi\|k_z\|_2^2)^{-1} (1-|B(z)|^2) \lesssim \|k_z\|_2^{-2}$. Hence, $$\label{ert2} y\lesssim \min\left(d_{0}(z),\left\|k_z\right\|_2^{-2}\right).$$ In view of , it remains to show that $$d_{{\varepsilon}}(x) \lesssim \min\left(d_{0}(x),\left|B'(x)\right|^{-1}\right) \asymp \min\left(d_{0}(z),\left\|k_z\right\|_2^{-2}\right).$$ The latter estimate is proved in [@bar1 Theorem 4.9]. We include its proof to make the exposition self-contained. Put $\beta=(1+4\log\frac{1}{\varepsilon})$. Clearly, $d_0(x) \ge d_\varepsilon(x)$. Thus, if $2\beta|B'(x)|^{-1}\ge d_0(x)$, then $d_\varepsilon(x) \le 2 \beta \min\left(d_{0}(x), |B'(x)|^{-1}\right)$. Assume that $2\beta|B'(x)|^{-1}< d_0(x)$ and put $w=x+i\beta|B'(x)|^{-1}$. Then $|x-w|<d_0(x)/2$ and so $$|x-z_n|/2 < |w-\overline z_n|< 2|x-z_n|$$ for any $n$. We have also $$\begin{aligned} \log|B(w)|^2 & = \sum_n \log |b_n(w)|^2 = \sum_n \log \bigg(1- \frac{4y_n{{\rm Im}\,}w}{|w-\overline z_n|^2}\bigg) \\ & < -\sum_n \frac{4\delta|B'(x)|^{-1}y_n}{|w-\overline z_n|^2} \le -\sum_n \frac{\beta|B'(x)|^{-1}y_n}{|x-\overline z_n|^2} = -\beta/2<2\log\varepsilon. \end{aligned}$$ Hence, $|B(w)|<\varepsilon$, and so $d_\varepsilon(x)<|x-w|=\beta |B'(x)|^{-1}$. In the case when $\Theta$ is a one-component inner function, we have a simpler formula: \[onecom\] Let $\Theta$ be a one-component inner function in ${{\mathbb C}}^+$ and let $0<{\varepsilon}<\delta<1$. Then $$d_{\varepsilon}(z) \asymp \|k_z\|_2^{-2} \asymp |\Theta'({{\rm Re}\,}z)|^{-1}, \qquad z\in \Omega_\delta^c,$$ where the constants in the asymptotic equality depend on $\Theta$, ${\varepsilon}$ and $\delta$, but not on $z\in \Omega_\delta^c$. By a result of Aleksandrov [@al1 Theorem 1.2], $\Theta$ is one-component if and only if there exists a constant $A>0$ such that $\|k_z\|_\infty \le A \|k_z\|_2^2$, $z\in {{\mathbb C}}^+$. Note that this estimate implies that $k_x \in L^2({{\mathbb R}})$ for $x\in {{\mathbb R}}$ only if $x$ is separated from the spectrum of $\Theta$. Then $$|z_n- \overline{z}|^{-1} = 2\pi |k_z(z_n)| \le 2\pi A \|k_z\|_2^2,$$ whence $|z_n-\overline z| \gtrsim \|k_z\|_2^{-2}$, $z\in {{\mathbb C}}^+$. In particular, $|x-z_n| \gtrsim |\Theta'(x)|$, $x\in {{\mathbb R}}$, and so $d_0(x) \gtrsim |\Theta'(x)|$. It follows that $d_{\varepsilon}(x) \asymp |\Theta'(x)|$, $x\in {{\mathbb R}}$. Now let $z = x+iy \in\Omega_\delta^c$. Then we know from that $\|k_z\|_2^2 \asymp |\Theta'(x)|$ and $d_{\varepsilon}(z) \asymp d_{\varepsilon}(x)$. This completes the proof. Proof of Theorem \[main\] {#th2} ========================= In what follows it will be convenient to use the following equivalent description of the Hardy space $H^2$. We will say that $f$ is a [*function of bounded type*]{} in the upper half-plane if $f$ is analytic in ${{\mathbb C}}^+$ and $f=g/h$ where $g, h \in H^\infty$ (here $H^\infty$ is the space of all bounded analytic functions in ${{\mathbb C}}^+$). If, moreover, $h$ is outer (see, e.g., [@ga Chapter II]) then we say that $f$ belongs to the Smirnov class $\mathcal{N}_+$. A classical theorem of V.I. Smirnov says that $f\in H^2$ if and only if $f\in \mathcal{N}_+$ and $f\in L^2({{\mathbb R}})$ in the sense of nontangential boundary values. If $f$ is analytic in ${{\mathbb C}}^+$ and continuous in ${{\mathbb C}}^+\cup {{\mathbb R}}$, then $f\in \mathcal{N}_+$ if and only if $f$ is of bounded type and $$\limsup\limits_{y\to+\infty}\frac{\log|f(iy)|}{y} \le 0$$ ($f$ is of nonpositive [*mean type*]{}). In the proof of Theorem \[main\] we will need the following \[phi\] Let $f$ be analytic in ${{\mathbb C}}^+$ and continuous on ${{\mathbb C}}^+\cup {{\mathbb R}}$. Assume that there exists a nonincreasing function $\psi:(0,+\infty)\rightarrow(0,+\infty)$ such that, for some $r>0$, $$\int^{r}_{0}y\left|\log \psi(y)\right|dy<\infty$$ and that $\left|f(z)\right|\leq\psi(y)$, $z=x+iy \in {{\mathbb C}}^+$. If, moreover, $f\in L^2({{\mathbb R}})$, then $f\in H^2$. This lemma is a consequence of a remarkable result of L. de Branges [@db Theorem 11] which gives sufficient condition for the inclusion to the Smirnov class $\mathcal{N}_+$. We will need Lemma only for $\psi(y)=1/y$, but prefer to give a proof in full generality (note that the critical growth is $\psi(y) = \exp(1/y^2)$). By Smirnov’s theorem, it is sufficient to show that $f$ is in the Smirnov class $\mathcal{N}_+$. By de Branges’ theorem [@db Theorem 11], a function $f$ (analytic in ${{\mathbb C}}^+$ and continuous in ${{\mathbb C}}^+\cup{{\mathbb R}}$) is in $\mathcal N_+$ if the following conditions hold: (a) : $\lim_{y\to \infty}\dfrac{\log\left|f(iy)\right|}{y}\leq 0$; (b) : $\int_{{{\mathbb R}}}\dfrac{\log^+\left|f(t)\right|}{1+t^2}dt<\infty$; (c) : $\lim_{R\to \infty} \frac{1}{R^2} \int^{\pi}_{0}\log^+\left|f(R e^{i\theta})\right|\sin \theta\, d\theta=0. $ Thus, apart from obvious necessary conditions (a) and (b) one needs to satisfy a very mild growth restriction (c). In our case conditions (a) and (b) are immediate. Condition (c) follows from the inequality $\left|f(z)\right|\leq\psi(y)$, for $z=x+iy$. Indeed, $$\int^{\pi - \text{arcsin}\frac{1}{R}}_{\text{arcsin}\frac{1}{R}} \log^+\left|f(R e^{i\theta})\right|d \theta\lesssim R,$$ since in this integral ${{\rm Im}\,}(R e^{i\theta})\ge 1$ and $f$ is bounded in $\{{{\rm Im}\,}z\ge 1\}$, while $$\int^{\text{arcsin}\frac{1}{R}}_{0}\log^+\left|f(R e^{i\theta})\right | \sin \theta d \theta\lesssim \frac{1}{R}\int^{1}_{0} s \log^+\psi(s)ds$$ (in the last integral, we have used the change of variable $s=R\sin \theta$). Thus $f\in H^2$. For the proof, it will be useful to consider a special system of intervals and Carleson squares. We may cover the real line as ${{\mathbb R}}=\cup_{n}I_n$, where $(I_n)_n$ are intervals with mutually disjoint interiors such that $$\left|I_n \right| \asymp\text{dist}\left(I_n, \Omega_{\delta}\right)\quad \text{ and } \quad S(I_n)\subset \Omega_{\delta}^c,$$ where $S(I)=I\times [0,\left|I\right|]$ is the Carleson square. Note that $\left|I_n \right|\asymp\left|I_{n+1} \right|$ for two contiguous intervals. We also have $$\begin{gathered} d_{\delta}(x)\asymp d_{\delta}(y),\qquad x,\ y\in I_n,\\ d_{\delta}(x)\asymp d_{{\varepsilon}}(x), \qquad x\in{{\mathbb R}},\end{gathered}$$ where the involved constants depend on ${\varepsilon}$ and $\delta$ only. Then it follows from that $$d_{{\varepsilon}}(z)\asymp\left|I_n \right|, \qquad z\in S(I_n).$$ We shall consider a modified weight, $$\label{modified W} \widetilde W(z)= W_0(z) \left(1+\sum_{n}\frac{1}{ \left|I_n \right|^{1/2}}\chi_{S(I_n) }(z)\right), \qquad z\in {{\mathbb C}}^+\cup {{\mathbb R}},$$ and $\widetilde W(z)=\widetilde W(\overline z),\ z\in {{\mathbb C}}^-$. It is clear that $\widetilde W \lesssim W$ and so ${{\mathcal F}}_W \subset {{\mathcal F}}_{\widetilde W}$.\ [**Inclusion**]{} ${{\mathcal F}}_{\widetilde W}\subset {{\mathcal H}}(E)$. We need to show that $F/E\in L^2({{\mathbb R}})$, and also that $F/E, F^{\sharp}/E$ are in $H^2({{\mathbb C}}^+)$. Note that by dividing the intervals $I_n$ in smaller parts, we can achieve that $S(2I_n)\subset \Omega_{\delta}^c$, where by $\gamma I_n$ we always denote the interval of length $\gamma \left|I_n\right|$ with the same center as $I_n$. Let $\alpha\in\left(0,\frac{1}{2}\right)$. For $x\in I_n$, put $D(x)=D(x,\alpha\left|I_n\right|)$ if $\left|I_n\right|\leq 1$, and $D(x)=D(x,\alpha)$ otherwise. Note that $\left|\Theta(z)\right|\geq\delta$, whenever $z\in S(2I_n)$. Since $\Theta(z)=\frac{1}{\overline{\Theta(\overline z)}}$ for $z\in{{\mathbb C}}^{-}$, we have $\left|\Theta(z)\right|\leq\frac{1}{\delta}$ when $z\in D(x)\cap {{\mathbb C}}^{-}$, since $D(x)\subset S(2I_n)\cup \overline{S(2I_n)}$ (where, for a set $A\subset \mathbb{C}$ we put $\overline{A} = \{\overline z: z\in A\}$). In particular, $\left|E(z)\right|\geq\delta\left|E(\overline z)\right|$, for $z\in D(x)\cap {{\mathbb C}}^{-}$. We first assume that $\left|I_n\right|\leq 1$. Using subharmonicity, we have $$\left|\frac{F(x)}{E(x)}\right|^{2}\leq\frac{1}{\pi\alpha^2\left|I_n\right|^2}\int_{D(x)}\left|\frac{F(z)}{E(z)}\right|^{2}dm(z).$$ Integrating over $I_n$, we get $$\begin{aligned} \int_{I_n}\left|\frac{F(x)}{E(x)}\right|^{2}dx&\leq \frac{1}{\pi\alpha^2\left|I_n\right|^2}\int_{I_n}\left(\int_{D(x)}\left|\frac{F(z)}{E(z)}\right|^{2}dm(z)\right)dx\\ &\lesssim \frac{1}{\left|I_n\right|}\int_{\left(\alpha+1\right)I_n\times [-\alpha\left|I_n\right|, \alpha\left|I_n\right|]}\left|\frac{F(z)}{E(z)}\right|^{2}dm(z)\\ &\lesssim \int_{\left(\alpha+1\right)I_n\times [-\alpha\left|I_n\right|, \alpha\left|I_n\right|]}\left|F(z)\widetilde W(z)\right|^2dm(z),\end{aligned}$$ since $\left|E(z)\right|\geq\delta\left|E(\overline z)\right|$ for $z\in {{\mathbb C}}^{-}$ and $\widetilde W(z)\asymp \frac{1}{\left|E(z)\right|} \frac{1}{\left|I_n\right|^{1/2}}$ in $S(2I_n) \cup \overline{S(2I_n)}$. In the case when $\left|I_n\right|>1$, we have $$\left| \frac{F(x)}{E(x)}\right|^{2} \leq\frac{1}{\pi\alpha^2} \int_{D(x)}\left|\frac{F(z)}{E(z)}\right|^{2}dm(z)$$ whence, integrating over $x\in I_n$, we get $$\begin{aligned} \int_{I_n}\left|\frac{F(x)}{E(x)}\right|^{2}dx \lesssim \int_{\left(\alpha+1\right)I_n\times [-\alpha, \alpha]}\left|F(z)\widetilde W(z)\right|^2dm(z),\end{aligned}$$ since now $\widetilde W\asymp \frac{1}{|E|}$ in $(\alpha+1)I_n\times [-\alpha, \alpha]$. Recall that $|I_n| \asymp |I_{n+1}|$ for two contiguous intervals. Therefore the sets $\left(\alpha+1\right)I_n\times [-\alpha |I_n|, \alpha |I_n|]$ for $|I_n|\le 1$ (or $(\alpha+1)I_n\times [-\alpha, \alpha]$ for $|I_n|>1$) may intersect but with finite multiplicity, which implies that $$\sum_{n}\int_{I_n}\left|\frac{F(x)}{E(x)}\right|^{2}dx\lesssim \int_{{{\mathbb C}}}\left|F\widetilde W\right|^2 dm,$$ and so $$\int_{{{\mathbb R}}}\left|\frac{F(x)}{E(x)}\right|^{2}dx\lesssim \int_{{{\mathbb C}}}\left|F\widetilde W\right|^2 dm.$$ To show that $F/E$ is in $H^2({{\mathbb C}}^+)$, we can use Lemma \[phi\] and the fact that $\widetilde W \ge W_0$. For $z\in{{\mathbb C}}^+$, we obtain $$\begin{aligned} \left|\frac{F(z)}{E(z)}\right|^{2}&\leq\frac{1}{\pi\left({{\rm Im}\,}z\right)^2} \int_{D(z,{{\rm Im}\,}z/2)}\left|\frac{F}{E}\left(\zeta\right)\right|^2dm\left(\zeta\right)\\ &\lesssim\frac{({{\rm Im}\,}z+1)^2}{\left({{\rm Im}\,}z\right)^2}\int_{D(z,{{\rm Im}\,}z/2)}\left|FW_0\right|^2dm \le \frac{({{\rm Im}\,}z+1)^2}{\left({{\rm Im}\,}z\right)^2}\left\|F\right\|^{2}_{{{\mathcal F}}_{\widetilde W}}.\end{aligned}$$ By Lemma \[phi\] applied to ${\varphi}(y) = (y+1)/y$, we have $F/E\in H^2$; the reasoning for $F^{\sharp}$ is analogous.\ [**Inclusion**]{} ${{\mathcal H}}(E)\subset {{\mathcal F}}_W$. We have already seen that if $F/E\in H^2$, then $$\int_{{{\mathbb C}}^+}\left|\frac{F(z)}{E(z)} \right|^{2}\frac{1}{\left(1+{{\rm Im}\,}z\right)^2}dm(z)<\infty.$$ We need to show that $$\int_{ \Omega_{\delta}^c }\left|\frac{F(z)}{E(z)}\right|^{2} \frac{1}{d_{{\varepsilon}}(z)}dm(z)\lesssim \left\|F\right\|^{2}_{{{\mathcal H}}(E)},$$ i.e., that $$d\mu(z):=\frac{1}{d_{{\varepsilon}}( z)}\chi_{\Omega_{\delta}^c}(z)dm(z)$$ is a Carleson measure for the model space $K_{\Theta}$. As before, consider the system of intervals $I_n$ such that $S(2I_n)\subset \Omega_{\delta}^c$ and $$\left|I_n\right|\asymp \text{dist}\left(I_n, \Omega_{\delta}\right)\asymp \text{dist}\left(I_n, \Omega_{{\varepsilon}}\right).$$ In order to show that $\mu$ is a Carleson measure for $K_{\Theta}$, we will verify the conditions of the following theorem of A.L. Volberg and S.R. Treil [@vt]: [*if for any $I$ with $S(I)\cap \Omega_{{\varepsilon}}\neq \varnothing$, we have $\mu\left(S(I)\right)\lesssim \left|I\right|$ (with constants independent of $I$), then $\mu$ is a Carleson measure for $K_{\Theta}$.*]{} Let $I = [a,b]$ be an interval such that $S(I)\cap \Omega_{{\varepsilon}}\neq \varnothing$. If $a\in I_n$, then $\left|I\right|\geq c \left|I_n\right|$. We can write $I\subset \cup_{n\in\mathcal J}I_n$, for some finite set $\mathcal J$ and $\left|I\right|\geq\sum_{n\in\mathcal J}\left|I_n\right|$. If $z=x+iy\in \Omega_{\delta}^c$ with $x\in I_n$, we know that $d_{{\varepsilon}}(z)\asymp d_{{\varepsilon}}(x) \asymp \left|I_n\right|$, with constants independent on $I_n$ and $z$. We also have $y\lesssim d_{{\varepsilon}}(z) \lesssim \left|I_n\right|$ by . Hence, $$\mu\big((I_n\times [0,+\infty))\cap \Omega_{\delta}^c \big)= \int_{I_n\times [0,+\infty)\cap \Omega_{\delta}^c}\frac{1}{d_{{\varepsilon}}(z)}dm(z) \lesssim \left|I_n\right|.$$ Therefore, $$\mu\left(S(I)\right) \leq\sum_{n\in\mathcal J}\mu\big((I_n\times [0,+\infty)) \cap \Omega_{\delta}^c\big)\lesssim \sum_{n\in\mathcal J}\left|I_n\right|\lesssim \left|I\right|.$$ Thus $\mu$ is a Carleson measure for $K_\Theta$. Corollary \[one1\] for one-component inner function follows immediately from Theorem \[main\] and Corollary \[onecom\] since in this case $d_{\varepsilon}(z) \asymp \|k_z\|_2^{-2}$, $z\in\Omega_\delta^c$. Using that $\|k_z\|_2^{2} \asymp|\Theta'(x)| =2{\varphi}'(x)$, $z=x+iy\in \Omega_\delta^c$, one can easily deduce another representation for the one-component case: \[one2\] Let $\Theta = E^\sharp/E$ be one-component. Put $$W_1(z) = W_0(z) \big(1+ ({\varphi}'({{\rm Re}\,}z))^{1/2}\chi_A(z) \big), \qquad z\in {{\mathbb C}}^+\cup{{\mathbb R}},$$ where $A= \{z=x+iy: 0\le y\le ({\varphi}'(x))^{-1}\}$, and $W_1(z) = W_1(\overline z)$, $z\in {{\mathbb C}}^-$. Then ${{\mathcal H}}(E) = {{\mathcal F}}_{W_1}$. We conclude this section with an example which shows that the appearance of the level set is intrinsic in our problem. Let $\Theta = E^\sharp/E$ be one-component. Put $$W_2(z) = W_0(z) (1+\|k_z\|_2), \qquad z\in {{\mathbb C}}^+\cup{{\mathbb R}},$$ and $W_2(z) = W_2(\overline z)$, $z\in {{\mathbb C}}^-$. Since $W_2\gtrsim W$, where $W$ is the weight from Corollary \[one1\], we have ${{\mathcal F}}_{W_2}\subset {{\mathcal H}}(E)$. However, the converse statement need not be true. Consider the de Branges space ${{\mathcal H}}(E)$, defined by the Hermite–Biehler function $E$ with zeros $\left\{\overline z_n \right\}_{n\in\mathbb{Z}}$, such that $z_n= |n|^{\alpha}{\rm sign}\, n +i$, $\frac{1}{2}<\alpha<1$, and put $\Theta=E^{\sharp}/E$. It is easy to show (see, e.g., [@bbh]) that $|\Theta'(x)|\asymp (|x|+1)^{\frac{1}{\alpha} -1}$, $x\in{{\mathbb R}}$. Therefore, ${\varphi}'(x)dx$ is a locally doubling measure and $\Theta$ is one-component. Moreover, there exists a constant $c>0$ such that for $z=x+iy\in {{\mathbb C}}^+$, $$\|k_z\|_2^2 = \frac{1-\left|\Theta(z)\right|^2}{4\pi y} \asymp\begin{cases}x^{\frac{1}{\alpha}-1},\ \left|x\right|\leq c y^{-\frac{\alpha}{1-\alpha}},\\ y^{-1},\ \left|x\right|> c y^{-\frac{\alpha}{1-\alpha}}. \end{cases}$$ We will show that ${{\mathcal H}}(E)$ is not contained in $ {{\mathcal F}}_{W_2}$. Indeed, it is easy to find a function $f$ in the model space $K_{\Theta}$ such that $\left|f(x+iy)\right|\asymp (|x|+1)^{-1/2} \log(|x|+2)^{-1}$, $x\in{{\mathbb R}}$, $0\le y\le 2$ (for much more general results of this type see [@belov]). Then $F=fE \in {{\mathcal H}}(E)$. However, setting $\gamma=\frac{\alpha}{1-\alpha}$ we get $$\int_0^1 \int_{{{\mathbb R}}} |f(z)|^2\|k_z\|^2_2 \gtrsim \int^{1}_{0}\frac{1}{y}\int^{+\infty}_{c/y^{\gamma}}\frac{dx}{x\log^{2}x}dx \asymp\int_{0}^1 \frac{dy}{y|\log y|}=\infty.$$ Thus, $F\notin {{\mathcal F}}_{W_2}$. Proof of Theorem \[F Wb\] {#th1} ========================= The implication (ii)$\Longrightarrow$(i) follows from Theorem \[main\]. Indeed, if $\overline z_n = x_n -iy_n$ is a zero of $E$, then by ${\varphi}'(x_n)\ge y_n^{-1}$. Thus, the inclusion $\varphi'\in L^{\infty}({{\mathbb R}})$ implies that $\inf_n y_n >0$ and it follows from that $\inf_{x\in {{\mathbb R}}} d_{\varepsilon}(x)>0$ for any ${\varepsilon}\in (0,1)$. Therefore, we can choose $0<{\varepsilon}<\delta<1$ such that for the weight $W$ defined by we have $W \asymp W_0$. Thus, ${{\mathcal F}}_{W_0} \subset {{\mathcal H}}(E)$. In the implication (i)$\Longrightarrow$(ii) we will use the following lemma. \[bab\] If ${{\mathcal F}}_{W_0} = {{\mathcal H}}(E)$ with equivalence of norms and $Z_E = \{\overline z_n\}$, then $\inf_{n} {{\rm Im}\,}z_n>0$. Recall that the zeros of $E$ are denoted by $\overline z_n=x_n-iy_n,$ with $y_n>0$ for each $n$. Let us consider the following test functions $$f_n(z)=E(z)\frac{\sqrt{y_n}}{z-\overline z_n},\qquad z\in{{\mathbb C}},\ n\in{{\mathbb N}}.$$ We shall compare ${{{\mathcal F}}_W}$ and ${{\mathcal H}}(E)$ norms of $f_n$. We have $$\begin{aligned} \left\|f_n\right\|^{2}_{{{\mathcal F}}_W}&=\int_{{{\mathbb C}}^+}\frac{y_n}{\left|z-\overline z_n\right|^2}\frac{dm(z)}{\left(1+\left|{{\rm Im}\,}z\right|\right)^2}+\int_{{{\mathbb C}}^-}\frac{y_n}{\left|z-\overline z_n\right|^2}\left|\frac{E(z)}{E(\overline z)}\right|^2\frac{dm(z)}{\left(1+\left|{{\rm Im}\,}z\right|\right)^2}\\ &\leq \int_{{{\mathbb C}}^+}\frac{y_n}{\left|z-\overline z_n\right|^2}\frac{dm(z)}{\left(1+\left|{{\rm Im}\,}z\right|\right)^2}+\int_{{{\mathbb C}}^-}\frac{y_n}{\left|z-\overline z_n\right|^2} \left|\frac{z-\overline z_n}{\overline z-\overline z_n}\right|^2 \frac{dm(z)}{\left(1+\left|{{\rm Im}\,}z\right|\right)^2}\\\end{aligned}$$ since for $z\in {{\mathbb C}}^-$, $z=x-iy$, $y>0$, we have, by , $$\left|\frac{E(z)}{E(\overline z)}\right|^2 =e^{-2\tau y} \prod_{k\neq n}\left|\frac{z-\overline z_k} {\overline z-\overline z_k}\right|^2\left|\frac{z-\overline z_n} {\overline z-\overline z_n}\right|^2 \le \left|\frac{z-\overline z_n} {\overline z-\overline z_n}\right|^2.$$ Using the change of variable $w=\overline z$ in the integral over ${{\mathbb C}}^-$, we get $$\begin{aligned} \left\|f_n\right\|^{2}_{{{\mathcal F}}_W}&\leq 2\int_{{{\mathbb C}}^+}\frac{y_n}{\left|z-\overline z_n\right|^2}\frac{dm(z)}{\left(1+\left|{{\rm Im}\,}z\right|\right)^2}\\ &=2 \int^{+\infty}_{0}\left(\int^{+\infty}_{-\infty}\frac{y_n}{(x-x_n)^2+(y+y_n)^2}dx\right)\frac{dy}{\left(1+y\right)^2}\\ &=2\pi \int^{+\infty}_{0}\frac{y_n}{y+y_n}\frac{dy}{\left(1+y\right)^2}\\ &\lesssim \int^{1}_{0}\frac{y_n}{y+y_n}dy+y_n \int^{+\infty}_{1} \frac{dy}{y\left(1+y\right)^2}\\ &\asymp y_n \log\left(\frac{1+y_n}{y_n}\right)+y_n.\end{aligned}$$ Since $\left\|f_n\right\|_E\asymp 1$, we conclude that $y_n \gtrsim 1$. Suppose that ${{\mathcal F}}_W= {{\mathcal H}}(E)$ with equivalent norms, and that $\varphi'$ is not bounded. Introducing the following test functions $$g_x(z)=\frac{K_x(z)}{\overline{E(x)}}=\frac{E(z)}{z-x} \left(1-\overline{\Theta(x)}\Theta(z)\right),\qquad x\in{{\mathbb R}},\ z\in{{\mathbb C}},$$ and assuming that ${\varphi}'(x)>1$, we will again compare their ${{{\mathcal F}}_W} $ and ${{\mathcal H}}(E)$ norms. As in Lemma \[bab\], we have $$\begin{aligned} \left\|g_x\right\|^{2}_{{{\mathcal F}}_W}&=\int_{{{\mathbb C}}^+}\left|\frac{1-\overline{\Theta(x)} \Theta(z)}{z-x}\right|^{2} \frac{1}{\left(1+\left|{{\rm Im}\,}z\right|\right)^2}dm(z) \\ & \quad +\int_{{{\mathbb C}}^-}\left|\frac{1-\overline{\Theta(x)}\Theta(z)}{z-x}\right|^{2} \left|\frac{E(z)}{E(\overline z)}\right|^2 \frac{1}{\left(1+\left|{{\rm Im}\,}z\right|\right)^2}dm(z)\\ & \quad \quad \asymp \int_{{{\mathbb C}}^+}\left|\frac{1-\overline{\Theta(x)}\Theta(z)}{z-x}\right|^{2} \frac{1}{\left(1+\left|{{\rm Im}\,}z\right|\right)^2}dm(z).\end{aligned}$$ For any positive number ${\varepsilon}<1$, let us denote by $S_{\varepsilon}$ the strip $\{0\le {{\rm Im}\,}z \le {\varepsilon}\}$ and by $R_{{\varepsilon}}$ the rectangle $\left\{z=\xi+i\eta\in{{\mathbb C}}^+,\ \left|\xi-x\right| \le {\varepsilon},\ 0 \le \eta \le {\varepsilon}\right\}$. We split the integral as follows: $$\begin{aligned} \int_{{{\mathbb C}}^+} & \left|\frac{1-\overline{\Theta(x)}\Theta(z)}{z-x}\right|^{2} \frac{1}{\left(1+\left|{{\rm Im}\,}z\right|\right)^2}dm(z) \\ & =\left(\int_{\left\{{{\rm Im}\,}z > {\varepsilon}\right\}} + \int_{S_{\varepsilon}\setminus R_{{\varepsilon}}} +\int_{R_{{\varepsilon}}}\right)\left|\frac{1-\overline{\Theta(x)}\Theta(z)}{z-x}\right|^{2} \frac{1}{\left(1+\left|{{\rm Im}\,}z\right|\right)^2}dm(z). \end{aligned}$$ We have $$\begin{aligned} \int_{\left\{{{\rm Im}\,}z\geq{\varepsilon}\right\}}\left|\frac{1-\overline{\Theta(x)}\Theta(z)}{z-x}\right|^{2} \frac{1}{\left(1+\left|{{\rm Im}\,}z\right|\right)^2}dm(z) &\lesssim \int^{\infty}_{{\varepsilon}}\int_{{{\mathbb R}}}\frac{d\xi}{\left(\xi-x\right)^2+\eta^2}\frac{d\eta}{(1+\eta)^2}\\ & = \pi \int^{\infty}_{{\varepsilon}}\frac{1}{\eta}\frac{d\eta}{(1+\eta)^2} \asymp 1+\log \frac{1}{{\varepsilon}}.\end{aligned}$$ Now $$\begin{aligned} \int_{S_{{\varepsilon}}\setminus R_{{\varepsilon}}}\left| \frac{1-\overline{\Theta(x)}\Theta(z)}{z-x}\right|^{2} \frac{1}{\left(1+\left|{{\rm Im}\,}z\right|\right)^2}dm(z)& \lesssim\int^{{\varepsilon}}_{0}\int_{\left\{\left|\xi-x\right|>{\varepsilon}\right\}} \frac{d\xi}{\left(\xi-x\right)^2}\frac{d\eta}{(1+\eta)^2}\\ &\asymp \frac{1}{{\varepsilon}} \int^{{\varepsilon}}_{0} \frac{d\eta}{(1+\eta)^2} \asymp 1.\end{aligned}$$ To estimate the integral over $R_{\varepsilon}$, note first that, for $z=x+iy\in{{\mathbb C}}^+$, $|\Theta'(z)|\le |\Theta'(x)| = 2{\varphi}'(x)$ (this follows immediately from ). Also, if $\inf_n y_n>0$, it follows from , that ${\varphi}'(x) \asymp{\varphi}'(t)$, $|x-t|\le{\varepsilon}$, with the constants depending on $\inf_n y_n$ and independent on ${\varepsilon}\in (0,1)$. We conclude that $\max_{u\in R_{{\varepsilon}}}\left|\Theta'(u)\right| \lesssim |{\varphi}'(x)|$. Since $\overline{\Theta(x)}=1/\Theta(x)$, we get $$\begin{aligned} \int_{R_{{\varepsilon}}}\left|\frac{1-\overline{\Theta(x)}\Theta(z)}{z-x}\right|^{2} \frac{1}{\left(1+\left|{{\rm Im}\,}z\right|\right)^2}dm(z) &\lesssim \int_{R_{{\varepsilon}}}\left|\frac{\Theta(x)-\Theta(z)}{z-x}\right|^{2}dm(z)\\ &\lesssim {\varepsilon}^2 \max_{u\in R_{{\varepsilon}}}\left|\Theta'(u)\right|^{2} \lesssim {\varepsilon}^2 \left[\varphi'(x)\right]^2.\end{aligned}$$ Choosing ${\varepsilon}\in(0,1)$ such that ${\varepsilon}^2 \left[\varphi'(x)\right]^2\asymp {\varepsilon}^{-1}$, we have $\left\|g_x\right\|^2_{{{\mathcal F}}_W}\lesssim {\varepsilon}^{-1} \asymp [{\varphi}'(x)]^{2/3}$ which, altogether with the fact that $\left\|g_x\right\|_{{{\mathcal H}}(E)}^2\asymp \varphi'(x)$, contradicts the equivalence of the norms if ${\varphi}'(x)$ takes arbitrarily large values. Thus, $\varphi' \in L^\infty({{\mathbb R}})$. Weight associated with the spectral data {#spect} ======================================== Let ${{\mathcal H}}(E)$ be a de Branges space and $\alpha\in [0, \pi)$. Then the function $E_\alpha(z) = e^{i\alpha}E - e^{-i\alpha}E^\sharp$ has only simple real zeros that we denote $t_{n, \alpha}$. It is one of the basic results of the de Branges theory (see [@db Theorem 22]) that the reproducing kernels $\{K_{t_{n,\alpha}}\}$ form an orthogonal basis in ${{\mathcal H}}(E)$ for all values of $\alpha$ except at most one. This exceptional $\alpha$ corresponds to the case when $E_\alpha\in {{\mathcal H}}(E)$ or, equivalently, $e^{2i\alpha}-\Theta \in H^2$. Also, $$\frac{E_\alpha(z)}{z-t_{\alpha, n}} = \frac{\pi}{i E(t_{\alpha, n})} K_{t_{\alpha, n}}(z) \quad\text{and}\quad \Big\| \frac{E_\alpha(z)}{z-t_{\alpha, n}} \Big\|_E^2 = \frac{\pi}{{\varphi}'(t_{\alpha,n})}.$$ Since ${{\mathcal H}}(E) = {{\mathcal H}}(e^{i\alpha} E)$ for any $\alpha$, we can assume in what follows that $A= (E+E^{\sharp})/2 \notin {{\mathcal H}}(E)$ (which is equivalent to $1+\Theta\notin H^2$). Put $T:=Z_A=\left\{t_n\right\}$ and $\mu_n = [{\varphi}'(t_n)]^{-1}$. Then any function in ${{\mathcal H}}(E)$ has the representation $$\label{sp} F(z)=A(z)\sum_{n}\frac{c_n\mu^{1/2}_{n}}{z-t_n},\qquad \left\{c_n\right\}\in \ell^{2},$$ where $$c_n=\frac{1}{\mu^{1/2}_{n}}\frac{F(t_n)}{A'(t_n)}$$ and $\|F\|_E^2 = \pi \|\{c_n\}\|_{\ell^2}^2$. We will call the pair $(T, \mu)$, where $\mu = \sum_n \mu_n \delta_{t_n}$, the [*spectral data*]{} for the de Branges space ${{\mathcal H}}(E)$. Note that we have $\sum_n (t_n^2+1)^{-1}\mu_n <\infty$. It is clear that $\|F\|_E^2 = \pi\int_{{\mathbb R}}|F/E|^2 d\mu$. We will replace the discrete measure $\mu$ by a weight concentrated near the points $t_n$. Let $r_n>0$ be such that $r_n<{\rm dist}\,(t_n, \{t_j\}_{j\ne n})/2$ and $$\label{rn} \sum_n \frac{r_n^2}{\mu_n} \sum_{j\ne n} \frac{\mu_j}{|t_n-t_j|^2}<\infty.$$ Put $W(z) = W_0(z) +W_T(z)$, where $$W_T(z) = \sum_n \frac{1}{\mu_n^{1/2} r_n}\cdot \frac{|z-t_n|}{|A(z)|}\chi_{D(t_n, r_n)}.$$ \[clark\] ${{\mathcal H}}(E) = {{\mathcal F}}_W$. [**Inclusion**]{} ${{\mathcal H}}(E) \subset {{\mathcal F}}_W$. We know that ${{\mathcal H}}(E) \subset {{\mathcal F}}_{W_0}$. Let $F\in {{\mathcal H}}(E)$ have representation . Then for each disc $D_n=D(t_n, r_n)$ we have $$\begin{aligned} \int_{D_n}\left|FW_T\right|^2 dm&=\int_{D_n}\bigg| \sum_{j\in{{\mathbb N}}}\frac{c_j\mu^{1/2}_{j}}{z-t_j} \bigg|^2 \frac{\left|z-t_n\right|^2}{\mu_n r^{2}_{n}}dm(z) \\ &\le 2 \int_{D_n}\left(\frac{\left|c_n\right|^2} {r^{2}_{n}}+\frac{\left|z-t_n\right|^2}{\mu_n r^{2}_{n}} \bigg|\sum_{j\neq n}\frac{c_j \mu^{1/2}_{j}}{z-t_j}\bigg|^2\right)dm(z) \\ &\le 2\pi |c_n|^2 + 2 \int_{D_n} \frac{\left|z-t_n\right|^2}{\mu_n r^{2}_{n}} \bigg(\sum_{j\neq n}\frac{\mu_{j}}{\left|z-t_j\right|^2}\bigg) \left\|c\right\|^{2}_{l^2} dm(z)\end{aligned}$$ Since for $z\in D_n$ and $j\neq n$, we have $\left|z-t_j\right| >\frac{\left|t_n-t_j\right|}{2}$, and $$\int_{D_n}\left|z-t_n\right|^2dm(z)\asymp r^{4}_{n},$$ we get $$\int_{D_n}\left|FW_T\right|^2 dm \lesssim\left|c_n\right|^2+\left\|c\right\|^{2}_{l^2} \bigg(\sum_{j\neq n}\frac{\mu_{j}}{\left|t_n-t_j\right|^2}\bigg) \frac{ r^{2}_{n}}{\mu_n}.$$ Summing the integrals over $D_n$ and using condition we conclude that $F\in {{\mathcal F}}_{W_T}$. [**Inclusion**]{} ${{\mathcal F}}_W \subset {{\mathcal H}}(E)$. Let $F\in {{\mathcal F}}_W$. Using analyticity of the function $\frac{F(z)(z-t_n)}{A(z)}$ in $D_n$, we have, for each $n$, $$\begin{aligned} \frac{1}{\mu_n}\left|\frac{F(z)(z-t_n)}{A(z)}\right|^2_{z=t_n}&=\frac{1}{\mu_n}\left|\frac{F(t_n)}{A'(t_n)}\right|^2\\ &\leq\frac{1}{\pi r^{2}_{n}\mu_n}\int_{D_n}\left|\frac{F(w)}{A(w)}\right|^{2}\left|w-t_n\right|^2 dm(w)\\ &=\frac{1}{\pi}\int_{D_n}\left|F(w)W_T(w)\right|^2dm(w).\end{aligned}$$ Then, for $$c_n:=\frac{1}{\mu^{1/2}_{n}}\frac{F(t_n)}{A'(t_n)},$$ we have $\{c_n\}\in\ell^2$. To show that $F\in {{\mathcal H}}(E)$ we need to prove that $$\frac{F(z)}{A(z)}=\sum_{n} \frac{c_n\mu^{1/2}_{n}}{z-t_n}, \qquad z\in {{\mathbb C}}.$$ To do this, it is enough to show that the entire function $$H(z)=\frac{F(z)}{A(z)}-\sum_{n} \frac{c_n\mu^{1/2}_{n}}{z-t_n}, \qquad z\in {{\mathbb C}},$$ is identically zero ($H$ is entire because $A$ has simple zeros and so the residues of $H$ at $t_n$ are zero). We know from the inclusion $F\in {{\mathcal F}}_{W_0}$ that $$\bigg|\frac{F(z)}{E(z)}\bigg| \lesssim 1 +\frac{1}{|{{\rm Im}\,}z|}, \qquad z\in {{\mathbb C}}^+$$ (see Section \[th2\]). Recall that $A=E(1+\Theta)/2$. It follows from formulas – that $|1+\Theta(z)|\ge 1-|\Theta(z)| \gtrsim \frac{{{\rm Im}\,}z}{|z|^2+1}$, whence $$\bigg| \frac{F(z)}{A(z)} \bigg| \lesssim (|z|^2+1)\bigg(\frac{1}{|{{\rm Im}\,}z|^2} +\frac{1}{|{{\rm Im}\,}z|}\bigg), \qquad z\in {{\mathbb C}}^+.$$ The same estimate holds for $z\in{{\mathbb C}}^-$ since we can apply the above reasoning to $F^\sharp$. Finally, $$\bigg|\sum_{n} \frac{c_n\mu^{1/2}_{n}}{z-t_n}\bigg|^2 \le \|\{c_n\}\|^2_{\ell^2} \sum_{n} \frac{\mu_n}{|z-t_n|^2}\lesssim \sum_n \frac{\mu_n}{t_n^2}\bigg|\frac{t_n}{z-t_n}\bigg|^2 \lesssim \frac{|z|^2}{|{{\rm Im}\,}z|^2}.$$ Thus, $|H(z)| \lesssim (|z|^2+1)|(|{{\rm Im}\,}z|^{-2}+|{{\rm Im}\,}z|^{-1})$. Since $$\log|H(z)| \le \frac{1}{2\pi} \int_{|\zeta-z|=1} \log|H(\zeta)|\,|d\zeta|,$$ we conclude that $|H(z)| \le C(|z|^2+1)$, $|{{\rm Im}\,}z| \le 1$. Thus, $H$ is a polynomial of degree at most $1$. We have $$F(z) = A(z)H(z) + A(z)\sum_{n}\frac{c_n\mu^{1/2}_{n}}{z-t_n}.$$ The latter sum is in ${{\mathcal H}}(E)$ and we already know that ${{\mathcal H}}(E) \subset {{\mathcal F}}_W$. Hence, $AH\in {{\mathcal F}}_W$ and in particular, $AH\in {{\mathcal F}}_{W_0}$. Since $A=(1+\Theta)E/2$, we conclude that, for any $y>0$, $$\int_{{{\mathbb R}}} |1+\Theta(x+iy)|^2 |H(x+iy)|^2 dx <\infty.$$ If $H$ is a nonzero polynomial we conclude that $1+\Theta(\cdot+iy) \in L^2({{\mathbb R}})$ for any $y>0$ and so $1+\Theta(\cdot +iy) \in H^2$. From this it is easy to deduce that $1+\Theta\in H^2$, a contradiction to the fact that $A\notin {{\mathcal H}}(E)$ by the choice of the spectral data. To see that $1+\Theta\in H^2$ one can use, e.g., the results of [@bar2]. By [@bar2 Theorem 2], for a nonconstant bounded analytic function $f$ in ${{\mathbb C}}^+$ with $\|f\|_\infty\le 1$, the inclusion $1-f\in H^2$ is equivalent to the asymptotics $$f(iy) = 1 - \frac{q}{y} +o\Big(\frac{1}{y}\Big), \qquad y\to+\infty,$$ with some $q>0$. Thus, $$\Theta(iy+iy_0) = -1 + \frac{q}{y} +o\Big(\frac{1}{y}\Big), \qquad y\to+\infty.$$ Then, by the same theorem, $1+\Theta \in H^2$ (since it has the same asymptotics as $y\to+\infty$). This contradiction shows that $H\equiv 0$. Weights depending on the imaginary part {#th3} ======================================= In this section we prove Theorem \[depy\]. We need to prove only implication (i)$\Longrightarrow$(ii). Therefore, in what follows we assume that $W(z) = \Phi(|y|)$, $y={{\rm Im}\,}z$, where $\Phi: [0, \infty)\to (0, \infty)$ and, in view of , $\inf_I \Phi>0$ for any bounded interval $I\subset [0, \infty)$. We write ${{\mathcal F}}_\Phi$ in place of ${{\mathcal F}}_W$. \[l2\] If ${{\mathcal H}}(E) = {{\mathcal F}}_\Phi$, then there exists $M>0$ such that $$\label{rep} M^{-1}\|K_{iy}\|_{E} \le \|K_{x+iy}\|_{E} \le M\|K_{iy}\|_{E}, \qquad x, y\in {{\mathbb R}}.$$ In particular, for any $F\in {{\mathcal H}}(E)$ and $y_1, y_2\in {{\mathbb R}}$, $F$ is bounded in $\{z: y_1\le {{\rm Im}\,}z\le y_2\}$. The proof of the lemma follows immediately from the fact that the translation operators $F \mapsto F(\cdot + t)$, $t\in\mathbb{R}$, are unitary on ${{\mathcal F}}_\Phi$ and from the formula $\|K_z\|_E = \sup_{F\in {{\mathcal H}}(E), \|F\|_E\le 1} |F(z)|$. \[expo\] If ${{\mathcal H}}(E) = {{\mathcal F}}_\Phi$, then $E$ is of finite exponential type. Recall that $\|K_z\|_E^2 = |E(z)|^2 \frac{1-|\Theta(z)|^2}{4\pi {{\rm Im}\,}z}$, $z\in {{\mathbb C}}^+$. Then it follows from that $$\label{trot} |E(iy)|^2 (1-|\Theta(iy)|^2) \asymp |E(x+iy)|^2 (1-|\Theta(x+iy)|^2), \qquad x, y\in {{\mathbb R}},$$ with the constants independent on $x, y$. Note that, by the formulas –, $1-|\Theta(z)|^2 \gtrsim |z|^{-2}{{\rm Im}\,}z \gtrsim |z|^{-2}$, ${{\rm Im}\,}z\ge 1$, and so $|E(x+iy)| \lesssim |E(i|y|)|\cdot|z|$, $|y| \ge 1$. Also, by Lemma \[l2\], $|E(z)|\lesssim 1+|z|$, $-1\le {{\rm Im}\,}z\le 1$, since $E/(z-z_0) \in {{\mathcal H}}(E)$ for any zero $z_0$ of $E$. Thus, if $\liminf_{y\to +\infty} |E(iy)| <\infty$, then $E$ is a polynomial of degree at most 1 by the Phragmén–Lindelöf principle applied to the strips. Now assume that $\lim_{y\to +\infty}|E(iy)|=+\infty$. We also have, by , $|E(x+iy)| \geq y^{-1}|E(iy)|$, $y\ge 1$. Hence, there exists $C>0$ such that $C|zE(z)|\ge 2$, ${{\rm Im}\,}z\ge 1$. Then $h(z) = \log C|zE(z)|$ is a positive harmonic function in $\{{{\rm Im}\,}z\ge 1\}$, and we can write its Riesz representation $$h(z) = p(y-1) + \frac{y-1}{\pi}\int_{{\mathbb R}}\frac{d\mu(t)}{(t-x)^2+(y-1)^2}, \qquad z=x+iy, \ y\ge 1,$$ with some measure $\mu$ such that $\int_{{\mathbb R}}(1+t^2)^{-1} d\mu(t) <\infty$. By obvious estimates of the Poisson integral, we get, for any ${\varepsilon}\in (0, \pi)$, $$\begin{aligned} |h(z)| & \lesssim |z|^2, \qquad {{\rm Im}\,}z\ge 2 \\ |h(z)| & \lesssim |z|, \qquad {\varepsilon}\le \arg z\le \pi-{\varepsilon}, \ {{\rm Im}\,}z\ge 2.\end{aligned}$$ Since $E$ is in the Hermite–Biehler class, analogous estimates hold in the lower half-plane. Since, moreover, $|E(z)| \lesssim |z|$ in $\{-2\le {{\rm Im}\,}z \le 2\}$, we conclude by the Phragmén–Lindelöf principle applied to the angles $\{|\arg z|\le {\varepsilon}\}$ and $\{|\arg z - \pi|\le {\varepsilon}\}$ that $E$ is of finite exponential type. Let ${{\mathcal H}}(E) = {{\mathcal F}}_\Phi$. First we will show that $F(\cdot+ iy_0) \in L^2({{\mathbb R}})$ for any $F\in {{\mathcal H}}(E)$ and $y_0\in{{\mathbb R}}$. Indeed, $$|F(x+iy_0)|^2 \le \int_{D(x+iy_0, 1)} |F(\zeta)|^2 dm(\zeta),$$ whence $$\begin{gathered} \int_{{\mathbb R}}|F(x+iy_0)|^2 dx \lesssim \int_{y_0-1\le{{\rm Im}\,}\zeta\le y_0+1} |F(\zeta)|^2 dm(\zeta) \qquad \qquad\\ \le \int_{y_0-1\le{{\rm Im}\,}\zeta\le y_0+1} |F(\zeta)\Phi(\zeta)|^2 dm(\zeta) \cdot \sup\limits_{y_0-1\le{{\rm Im}\,}\zeta\le y_0+1} [\Phi(y)]^{-2} \le M^{-2} \|F\|^2_{{{\mathcal F}}_\Phi},\end{gathered}$$ where $M = \inf_{y_0-1\le y \le y_0+1} \Phi(y)>0$. By Lemma \[expo\], $E$ is of finite exponential type and so the number $a$ defined by $$a = \limsup\limits_{y\to+\infty} \frac{\log|E(iy)|}{y}$$ is finite. Also, for any zero $z_0$ of $E$, the function $G(z) = E(z)/(z-z_0)$ is in ${{\mathcal H}}(E)$ and so in $L^2({{\mathbb R}})$. Since $G$ is of finite exponential type, we conclude that $G$ belongs to some Paley–Wiener space. A function in the Paley–Wiener space has maximal growth along the imaginary axis and so $G\in PW_a$. In particular, $e^{iaz} G \in H^2$. The function $G$ does not vanish in ${{\mathbb C}}^+$ and also $\limsup_{y\to+\infty} y^{-1} \log |e^{-ay}G(iy)| = 0$. We conclude that $e^{iaz} G$ is an outer function in ${{\mathbb C}}^+$ (since its canonical Smirnov–Nevanlinna factorization contains no factors of the form $e^{i\tau z}$, $\tau >0$), and it follows from standard estimates of the Poisson integral that $$\label{esio} a = \lim\limits_{y\to+\infty} \frac{\log|E(x+iy)|}{y}$$ and the limit is uniform for $x\in [-1,1]$. Now let $F\in{{\mathcal H}}(E)$. Then $F\in L^2({{\mathbb R}})$ and $F/E$ is in $H^2$. Hence, $$\limsup\limits_{y\to+\infty} \frac{\log|e^{-ay}F(iy)|}{y} = \limsup\limits_{y\to+\infty} \frac{\log|e^{-ay}E(iy)| + \log|E^{-1}(iy)F(iy)|}{y} \le 0.$$ Thus, $F\in PW_a$ and we conclude that ${{\mathcal H}}(E) \subset PW_a$. It remains to prove the converse inclusion $PW_a\subset {{\mathcal H}}(E)$. Put $$b= \sup\bigg\{b': \int_0^\infty e^{2b'y}\Phi^2(y)dy<\infty \bigg\}.$$ Since $$\int_{-1}^1 \int_0^\infty \frac{|E(x+iy)|^2}{|x+iy-z_0|^2}\Phi^2(y)dy dx <\infty$$ and $|E(x+iy)|^2 \gtrsim e^{2a'y}$, $y\ge 0$, for any $a'<a$ by , we conclude that $a\le b$. Now let $b'<b$. Then $PW_{b'} \subset {{\mathcal F}}_\Phi$. Indeed, $\int_{{\mathbb R}}|f(x+iy)|^2 dx \le\|f\|^2_2 e^{2b'|y|}$ for any $f\in PW_{b'}$, whence $$\int_{{\mathbb C}}|f(x+iy)|^2 \Phi^2(|y|)dxdy \le 2\|f\|^2_2 \int_{0}^\infty e^{2b'y} \Phi^2(y) dy<\infty.$$ Thus, $PW_{b'} \subset {{\mathcal F}}_\Phi \subset PW_a$ for any $b'<b$ and so $b\le a$. We have seen that $PW_{b'} \subset {{\mathcal F}}_\Phi \subset PW_a$ for any $b'<a$. To complete the proof note that for any $g\in PW_{b'}$ we have $g/E \in L^2({{\mathbb R}})$. Now let $f\in PW_a$. Then $f = e^{iaz/2} f_1 + e^{-iaz/2} f_2$ where $f_1, f_2 \in PW_{a/2}$ and so $f/E \in L^2({{\mathbb R}})$. By the discussion in the beginning of Section \[th2\], it remains to show that $f/E$, $f^\sharp/E$ are in the Smirnov class $\mathcal{N}_+$. Since both $f$ and $E/(z-z_0)$ are in $PW_a$, the function $f/E$ is of bounded type in ${{\mathbb C}}^+$. Finally, $$\limsup\limits_{y\to+\infty} \frac{1}{y}\log\bigg|\frac{f(iy)}{E(iy)}\bigg| = \limsup\limits_{y\to+\infty} \bigg(\frac{1}{y}\log |e^{-ay} f(iy)| + \frac{1}{y}\log \frac{e^{-ay}}{|E(iy)|}\bigg) \le 0.$$ Thus, $f/E$ (and similarly, $f^\sharp/E$) are in $H^2$, whence $f\in {{\mathcal H}}(E)$. Final remarks ------------- Consider the following three statements: 1. ${{\mathcal H}}(E) = {{\mathcal F}}_{\Phi}$ ($\Longleftrightarrow {{\mathcal H}}(E) = PW_a$); 2. $\sup_{t\in{{\mathbb R}}}\|U_t:{{\mathcal H}}(E)\to{{\mathcal H}}(E)\|<\infty$; 3. $\|K_x\|_E \asymp \|K_0\|_E$, $x\in{{\mathbb R}}$, i.e., ${\varphi}'(x)|E(x)|^2 \asymp 1$. We have seen that (i)$\Longrightarrow$(ii)$\Longrightarrow$(iii). One can ask whether converse to any of these implications is true. We do not know whether (iii)$\Longrightarrow$(ii) or (ii)$\Longrightarrow$(i), but we can show that (iii) does not imply (i). Namely, we will construct a de Branges space that satisfies (iii), but does not coincide as a set with the Paley–Wiener space. Given $\delta>0$, consider the sequence $(z_n)_{n\in{{\mathbb Z}}}$ defined by $$z_n=\begin{cases} n-\delta+in^{-4\delta},\ n>0,\\ n+\delta+i\left|n\right|^{-4\delta},\ n<0,\\ i,\ n=0, \end{cases}$$ and let $E(z)=\lim_{R\to\infty} \prod_{|z_n|\le R}(1-z/\overline z_n)$ be the corresponding Hermite–Biehler function. This product converges uniformly on compact sets and $E$ is a function of Cartwright class. This example was given in [@ls] where it was shown that ${{\mathcal H}}(E) = PW_\pi$ if and only if $\delta<1/4$. However, we will show that for $1/4 \le\delta\le 1/2$ the function $E$ satisfies (iii). In what follows let $\delta\in (0,1)$. We will estimate $E(x)$, $x\in {{\mathbb R}}$, by comparing it to the function $\sin \pi(x+i)$ whose modulus is comparable to a constant. These are well-known calculations which we include for the sake of completeness. Let $|x-k|\le 1/2$ and $k>0$. Then it is easy to show that $$\left|\frac{E(x)}{\sin \pi(x+i)}\right| \asymp \prod_{n\in{{\mathbb Z}}}\left|\frac{1-\frac{x}{\overline z_n}}{1-\frac{x}{n-i}}\right|^{2} \asymp A^{2}\left|\frac{1-\frac{x}{\overline z_k}}{1- \frac{x}{k-i}}\right|^{2},$$ where $$A=\prod_{n\geq 1, n\neq k}\left| \frac{\left(1-\frac{k}{n-\delta}\right)\left(1-\frac{k}{n+\delta}\right)}{\left(1-\frac{k}{n}\right)\left(1+\frac{k}{n}\right)} \right|.$$ Note that $$A= \prod_{n\geq 1, n\neq k}\left|1+\frac{k^2\left(-2n\delta+\delta^2\right)} {\left(n-\delta\right)^{2}\left(n^2-k^2\right)} \right| \asymp \prod_{n\geq 1, n\neq k}\left| 1 + \frac{2\delta nk^2} {\left(n-\delta\right)^{2}\left(k^2-n^2\right)} \right|.$$ We have also $$\sum_{n\geq 1, n\neq k}\frac{n k^2}{\left(n-\delta\right)^{2} \left(k^2-n^2\right)}\\ = \sum_{n\geq 1, n\neq k}\frac{k^2}{n\left(k^2-n^2\right)}+O(1) = \sum_{1\leq n\leq 2k-1, n\neq k} \frac{k^2} {n\left(n^2-k^2\right)}+O(1)$$ (where $O(1)$ denotes a quantity uniformly bounded in $k$) and $$\begin{aligned} \sum_{1\leq n\leq 2k-1} \frac{k^2}{n\left(k^2-n^2\right)} & =\sum_{1\leq n\leq 2k-1}\left[\left(\frac{1}{n}-\frac{1}{k+n}\right)+\frac{1}{2} \left(\frac{1}{k-n}+\frac{1}{k+n}\right)\right]\\ &=\sum_{1\leq n\leq 2k-1}\left[\frac{1}{n}- \frac{1}{2(k+n)} \right] = \log k+O(1).\end{aligned}$$ Thus $$\begin{aligned} \log A & = \sum_{n\geq 1, n\neq k} \log\left|1 + \frac{2\delta nk^2} {\left(n-\delta\right)^{2}\left(k^2-n^2\right)} \right| \\ & = 2\delta \sum_{n\geq 1, n\neq k}\frac{k^2}{n\left(k^2-n^2\right)}+O(1) = 2\delta\log k+O(1)\end{aligned}$$ and so $$\left|E(x)\right|^2\asymp \left|k\right|^{4\delta}\left|x-z_k\right|^2,\qquad x\in{{\mathbb R}}, \ |x-k|\le 1/2.$$ Now we estimate the derivative of the phase function: $$\varphi'(x)=\sum_{n\geq 1, n\ne k} \frac{y_n}{\left(x-n+\delta\right)^2+y_n^2} +\sum_{n< 0} \frac{y_n}{\left(x-n-\delta\right)^2+y_n^2} + \frac{1}{x^2+1} +\frac{k^{-4\delta}}{\left(x-k+\delta\right)^2+k^{-8\delta}}.$$ By simple estimates of sums, $$\begin{aligned} \sum_{|n|\ge 3k/2}\frac{n^{-4\delta}}{(x-n)^2} \lesssim k^{-4\delta-1}, & \qquad \sum_{k/2\le n \le 3k/2, n\ne k} \frac{n^{-4\delta}}{(x-n)^2} \lesssim k^{-4\delta}, \\ \sum_{-3k/2\le n \le k/2, n\ne 0} \frac{|n|^{-4\delta}}{(x-n)^2} & \lesssim\frac{1}{k^2} \sum_{-3k/2\le n \le k/2, n\ne 0} |n|^{-4\delta}.\end{aligned}$$ If $0<4\delta<1$, then the last expression is $\asymp k^{-4\delta-1}$, while for $1< 4\delta \le 2$ it is comparable with $k^{-2}\lesssim k^{-4\delta}$. Finally, note that $$\frac{k^{-4\delta}}{\left(x-k+\delta\right)^2+k^{-8\delta}} \gtrsim k^{-4\delta}, \qquad |x-k|\le 1/2.$$ We conclude that ${\varphi}'(x) \asymp k^{-4\delta} |x-z_k|^{-2}$, $|x-k|\le 1/2$, when $0<\delta\le 1/2$. Thus, ${\varphi}'(x)|E(x)|^2\asymp 1$ for $0<\delta\le 1/2$ (and it is easy to see from the above estimates that it is no longer true for $\delta>1/2$). However, for $1/4 \le\delta\le 1/2$, ${{\mathcal H}}(E) \ne PW_\pi$.\ [**Problem.**]{} It would be interesting to describe the de Branges spaces ${{\mathcal H}}(E)$ satisfying (iii) or, especially, (ii). [99]{} A.B. Aleksandrov, Embedding theorems for coinvariant subspaces of the shift operator. II, [*Zap. Nauchn. Sem. Inst. Steklov. (POMI)*]{} [**262**]{} (1999), 5–48; English transl.: [*J. Math. Sci.*]{} [**110**]{} (2002), 2907–2929. A.D. Baranov, Isometric embeddings of the spaces $K_\Theta$ in the upper half-plane, [*J. Math. Sci.*]{} [**105**]{} (2001), 5, 2319–2329. A.D. Baranov, Bernstein-type inequalities for shift-coinvariant subspaces and their applications to Carleson embeddings, [*J. Funct. Anal.*]{} [**223**]{} (2005), 116–146. A.D. Baranov, A.A. Borichev, V.P. Havin, Admissible majorants for meromorphic functions with fixed poles, [*Indiana Univ. Math. J.*]{} [**56**]{} (2007), 4, 1595–1628. A. Baranov, Yu. Belov, A. Borichev, Spectral synthesis in de Branges spaces, [*Geom. Funct. Anal.*]{} ([*GAFA*]{}) [**25**]{} (2015), no. 2, 417–452. A. Baranov, Yu. Belov, A. Borichev, Riesz bases of reproducing kernels in Fock type spaces and de Branges spaces, [*Studia Math.*]{} [**236**]{} (2017), 2, 127–142. Yu.S. Belov, Model functions with nearly prescribed modulus, [*Algebra i Analiz*]{} [**20**]{} (2008), 2, 3–18; English transl.: [*St. Petersburg Math. J.*]{} [**20**]{} (2009), 2, 163–174. H. Bommier-Hato, M. Englis, E.-H. Youssfi, Bergman-type projections on generalized Fock spaces, [*J. Math. Anal. Appl.*]{} [**389**]{} (2012), 1086–1104. A. Borichev, R. Dhuez, K. Kellay, Sampling and interpolation in large Bergman and Fock spaces, [*J. Funct. Anal.*]{} [**242**]{} (2007), 563–606. L. de Branges, [*Hilbert Spaces of Entire Functions*]{}, Prentice Hall, Englewood Cliffs (NJ), 1968. W.S. Cohn, Carleson measures for functions orthogonal to invariant subspaces, [*Pacific J. Math.*]{} [**103**]{} (1982), 347–364. K.M. Dyakonov, Entire functions of exponential type and model subspaces in $H^p$, [*Zap. Nauchn. Sem. Leningrad. Otdel. Mat. Inst. Steklov. (LOMI)*]{} [**190**]{} (1991), 81–100; English transl.: [*J. Math. Sci.*]{} [**71**]{} (1994), 2222–2233. J. B. Garnett, [*Bounded Analytic Functions*]{}, Academic Press, New York, 1981. V.P. Havin, J. Mashreghi, Admissible majorants for model subspaces of $H^2$. Part I: slow winding of the generating inner function, [*Can. J. Math.*]{} [**55**]{}, 6 (2003), 1231–1263. V.P. Havin, J. Mashreghi, Admissible majorants for model subspaces of $H^2$. Part II: fast winding of the generating inner function, [*Can. J. Math.*]{} [**55**]{}, 6 (2003), 1264–1301. M. Kaltenbäck, H. Woracek, DeBranges spaces of exponential type: general theory of growth, [*Acta Sci. Math. (Szeged)*]{} [**71**]{} (2005), 1-2, 231–284. Yu. I. Lyubarskii, K. Seip, Weighted Paley–Wiener spaces, [*J. Amer. Math. Soc.*]{} [**15**]{} (2002), 4, 979–1006. J. Marzo, S. Nitzan, J.F. Olsen, Sampling and interpolation in de Branges spaces with doubling phase, [*J. Anal. Math.*]{} [**117**]{} (2012), 1, 365–395. J. Ortega-Cerdà, K. Seip, Fourier frames, [*Ann. of Math.*]{} (2), [**155**]{} (2002), 3, 789–806. K. Seip, Density theorems for sampling and interpolation in the Bargmann–Fock space. I, [*J. Reine Angew. Math.*]{} [**429**]{} (1992), 91–106. A.L. Volberg, S.R. Treil, Embedding theorems for invariant subspaces of the inverse shift operator, [*Zap. Nauchn. Sem. Inst. Steklov. (LOMI)*]{} [**149**]{} (1986), 38–51; English transl.: [*J. Soviet Math.*]{} [**42**]{} (1988), 1562–1572. [^1]: The work is supported by Russian Science Foundation grant 14-41-00010.
{ "pile_set_name": "ArXiv" }
--- abstract: 'When a binary fluid demixes under a slow temperature ramp, nucleation, coarsening and sedimentation of droplets lead to an oscillatory evolution of the phase separating system. The advection of the sedimenting droplets is found to be chaotic. The flow is driven by density differences between the two phases. Here, we show how image processing can be combined with particle tracking to resolve droplet size and velocity simultaneously. Droplets are used as tracer particles, and the sedimentation velocity is determined. Taking these effects into account, droplets with radii in the range of 4 – 40$\mu$m are detected and tracked. Based on this data we resolve the oscillations in the droplet size distribution which are coupled to the convective flow.' author: - 'T. Lapp' - 'M. Rohloff' - 'J. Vollmer' - 'B. Hof' date: 'Received: date / Accepted: date' title: Particle tracking for polydisperse sedimenting droplets in phase separation --- Introduction {#intro} ============ The characterization of particle distributions in fluids is important to control and optimize processes in food, pharmaceutical, oil and chemical industry, as [@Heffels1998] point out. To determine the particle mass flux, particle sizes and velocities have to be measured simultaneously, which was achieved by [@Petrak2002]. Here, we present a particle tracking algorithm, which uses droplets as marker particles to measure the flow field. They are created naturally in the phase separation process of demixing binary systems. Droplet positions and radii are detected simultaneously. The radius can therefore be used as a criterion to identify droplets in subsequent images. Assuming Stokes law the sedimentation velocity can be calculated from the droplet radius. The droplet velocity is decomposed into sedimentation velocity and advection by the flow. By subtracting the sedimentation velocity from the droplet velocity, the advection of *all* droplets can be used to measure the flow field. With the advective flow field and the sedimentation velocity of each droplet, its position in the next frame can be predicted and compared to the image. Previous studies have extensively used particle tracking velocimetry as a measurement technique to investigate turbulent flows carrying small particles ([@Maas1993; @Malik1993; @Ouellette2006; @Kreizer2010]). To that end monodisperse tracer particles are added to the fluid whose position is detected by image processing of high speed camera data. While this procedure is followed successfully for monodisperse particles in single-phase flows, it is not easily applicable for measuring flow patterns in demixing binary systems, which have been studied by [@Vollmer1997b; @Vollmer2002] and [@Emmanuel2006]. Tracer particles cannot be added as they would act as nucleation centers, and therefore affect the droplet number density. Further more, colloidal particles aggregate on the interfaces and change the growth and coalescence rate, as [@Thijssen2010] have shown. For a review of the stabilizing effect of colloidal particles in emulsions see [@Binks2002] and [@Aveyard2003]. There also is a broad range of acoustic / electro-acoustic and optical techniques for the measurement of droplet size distributions, an overview given by [@Maass2009]. Several laser based techniques measure chord lengths that have to be transfered into droplet size distributions ([@Hu2006]), e.g. focus beam reflectance measurement (FBRM, see e.g. [@Ruf2000]) or optical reflectance measurement (ORM, see e.g. [@Lovick2005]). However, [@Maass2011] and [@Simmons2000] have shown that these techniques give inaccurate results for liquid/liquid dispersions in comparison to image analysis of in situ microscope imaging. To investigate the evolution of demixing systems, it is therefore preferential to use particle tracking of droplets present in the demixing system, as described in the following. The article is organized as follows: We first characterize the chemical system. Then we introduce the experimental setup with a focus on the illumination technique, and explain the experimental procedure (sec. 2). In section 3, we present the particle tracking algorithm. It is based on an image processing method to detect the droplets. In section 4, we measure the sedimentation velocity of droplets and give an indication on the uncertainties in the radius detection. We then use the velocity information to filter out falsely detected droplets. We show how the droplet size distribution and the characteristics of the flow field evolve in time. Finally, in section 5, we discuss the applicability and the limitations of the droplet tracking procedure for the investigation of demixing binary fluids. Experimental Method =================== Model System ------------ As a model system a mixture of water and isobutoxyethanol (i-BE) is chosen. It has a lower critical point at 25.5°C and demixes under heating. A phase diagram is shown in Fig. \[fig:phase\_diagram\]. The sample can easily be prepared at room temperature and demixes at temperatures between 25.5°C and 80°C, as shown by [@Nakayama2001]. The i-BE (purity $\geq 97 \%$) is purchased from Wako Chemicals GmbH and used without purification. The data points in the phase diagram (Fig. \[fig:phase\_diagram\]) were determined by turbidity measurements. The solid lines show fits with sixth order (left branch) and fourth order (right branch) polynomials respectively. The density and the viscosity of the mixture depend on composition as well as on temperature. For the density, the data of [@Doi2000] is used, taking thermal expansion and molar excess volume explicitly into account. The determination of the viscosity is described in the appendix. ![Phase diagram of water/isobutoxyethanol mixture between 25 and 50°C. The samples are filled with the composition $\Phi_{\rm g}$. The mixture separates above a critical temperature $T_{\rm c} =$ 25.5°C. For slow temperature ramps the compositions $\Phi_{\rm b}, \Phi_{\rm t}$ of the bottom and top phase follow the left and right branch of the binodal, respectively. $2\Phi_{\rm 0}=\Phi_{\rm t}-\Phi_{\rm b}$ denotes the width of the miscibility gap. The left and the right branch are fitted with sixth and fourth order polynomials, respectively.[]{data-label="fig:phase_diagram"}](Phase_Diagram){width="0.8\linewidth"} Droplet detection is improved by using Nile Red as a fluorescent dye. Nile Red solved in butoxyethanol absorbs in the wavelength range of 500nm to 590nm (green light) and emits between 580nm and 700nm (red). The dye solves preferentially in organic compounds and poorly in water. [@Fowler1985] point out that the fluorescence of Nile Red depends strongly on the polarity of the solvent. In particlar, its fluorescence is quenched in water . Therefore, the i-BE-rich phase appears bright and the water-rich phase stays relatively dark. Since each phase consists of a mixture of water and i-BE, the contrast between dark and bright parts in the images recorded during a measurement depends on the composition of the two phases. It was checked that the dye has no significant influence on the phase diagram. Experimental Setup ------------------ The probe is contained in a fluorescence cell ($10\times10\times33$mm) 117.100F-QS made by Hellma GmbH. The measurement cell is mounted in a water bath with controlled temperature (see Fig. \[fig:setup\]). An immersion cooler Haake EK20 is cooling with constant power, and a temperature control module Haake DC30 is heating the water bath to a preset temperature. Additionally the temperature of the water near the sample is measured with a PT100 temperature sensor. The temperature is controlled with an accuracy of 15 mK. To get enough signal from the fluorescently labeled i-BE-rich phase, a bright light source is required. A laser light sheet is not suitable for illumination since it is unidirectional and the two phases are not index-matched. Each droplet acts as a little lens which focuses and diffracts the parallel laser light sheet. After passing a short distance of the sample, the droplets have transformed the uniform light sheet into a stripe pattern, which illuminates the droplets very inhomogeneously. In contrast a mercury short arc lamp (LOT-Oriel, 100W) with a bright light emitting spot gives far better results. It provides a high light intensity, which is diffuse enough that almost no stripes appear (figure \[fig:raw\_image\]). A sketch of the optical components is given in figure \[fig:setup\]. The light is collected by the collimator lens (C) and a bandpass filter (GF) selects the green emission lines (546 and 577/579nm) for the excitation of the fluorescent dye. The spherical lens (L1, $f=200\rm mm$) forms a parallel light beam. The cylindrical lens (L2, $f=80\rm mm$) focuses the light normal to the focal plane of the camera. A spherical diaphragm (D) allows to adjust the amount of light manually. With a narrow slit of 300$\mu$m width directly in front of the measurement cell a vertical light sheet is formed. The measurement cell is covered by black apertures to shield stray light. The fluorescent light from the illuminated plane of the sample is projected by a $f=35$mm objective (L3) to the chip of a BM-500CL monochrome progressive scan CCD camera. It takes 2058 x 2448 pixel images with a maximum frame rate of 15Hz, covering 1.3mm$\times$1.5mm of the sample (1.6 pixel/$\mu$m). With a longpass filter (RF, edge 594nm) the excitation light is filtered. The camera, temperature control and magnetic stirrer are connected to a computer. Measurements are fully automated using a LABVIEW program. ![Sketch of the experimental setup, top view. The light of a short arc Hg vapor lamp is collected by a collimator lens (C) and parallelized by a spherical lens (L1). A cylindrical lens (L2) forms a light sheet which is cut by a diaphragm (D) and a line aperture (300$\mu$m width) directly in front of the sample cell. The light sheet illuminates a plane of the binary mixture, a green filter (GF) selects the wavelength range suitable for excitation of the dye. The emission of the fluorescently labeled phase is passes a red filter (RF) and is projected by an objective (L3) onto a 5 Mpixel CCD camera chip. The sample cell is mounted in a water bath with controlled temperature.[]{data-label="fig:setup"}](setup_opticsA){width="0.8\linewidth"} Experimental Procedure ---------------------- The temperature protocol for our experimental procedure is adapted from [@Auernhammer2005]. After mixing the sample at 24°C for one hour, the temperature is set to $T=25.8$°C (0.3K above critical point) and the system equilibrates for four hours. Then a temperature ramp is run from 25.8°C to 50°C, which is designed to keep the volume flux between the phases constant. Following [@Auernhammer2005] and [@Cates2003] a driving parameter $\xi$ is defined $$\xi = \frac{1}{\Phi_{\rm 0}(T(t))}\frac{\partial \Phi_{\rm 0}(T(t))}{\partial t} \label{eqn:xi}$$ with $2\Phi_{\rm 0}=\Phi_{\rm t}-\Phi_{\rm b}$, where $\Phi_{\rm t}$ and $\Phi_{\rm b}$ denote the compositions of the top and the bottom phase, respectively (compare Fig. \[fig:phase\_diagram\]). Hence, $2\Phi_{\rm 0}(T)$ is the width of the miscibility gap for a given temperature $T$. The driving parameter $\xi$ is equal to the rate of droplet mass production in the two phases. Equation (\[eqn:xi\]) is inverted to calculate a temperature ramp with fixed driving, $$T(t+\delta t) = T(t) + \frac{\xi \Phi_{\rm 0}(T(t))}{\partial_{T} \Phi_{\rm 0}(T(t))} \delta t. \label{eqn:tempramp}$$ Equation (\[eqn:tempramp\]) is integrated with uniformly distributed time steps in the range \[$t_{\rm 0},t_{\textrm{end}}$\] chosen such that $T(t_{\rm 0})=25.8$°C and $T(t_{\textrm{end}})=50.0$°C. To implement the temperature scan, the temperature of the thermostat is set with a rate of 1 Hz according to the calculated temperature ramp. Particle Detection and Tracking =============================== In this section we describe a particle detection and tracking algorithm for polydisperse sedimenting particles. Radii and positions are detected simultaneously. The first part describes an image processing algorithm which detects droplets in separate images. In the next step the flow field is estimated. Then the droplets are tracked through a series of images. Finally the velocity field is recalculated based on the droplet trajectories. This procedure allows us to sort out artifacts (caused e.g. by dirt or overlaps of droplet images) of the image processing, and to determine Lagrangian particle velocities. The image processing is done with the MATLAB Image Processing Toolbox. The data (typically images per measurement) is processed on a computer cluster. For simplicity we restrict our description to the detection of the fluorescently labeled droplets in the bottom phase. By inverting the images and the direction of gravity the same algorithm can be used for the dark droplets detected in the top phase. Processing of Single Images --------------------------- ### Preprocessing In a first step, dark spots caused by dirt on the camera chip are removed with a flat field correction. Next the image is Fourier filtered and the contrast is optimized. To reduce the horizontal stripes (illumination from the right) produced by the light dispersion of the droplets a high pass filter (Gaussian filter with width 5 in horizontal and 200 in vertical direction) is applied. For reducing the noise, a low pass filter (isotropic Gaussian filter, $\sigma = 100$) is applied. In figure \[fig:raw\_prepro\], a raw image and a preprocessed image of fluorescent i-BE-rich droplets in the denser water-rich bottom phase are shown. For high droplet densities, there is a significant number of overlaps of droplet images. Therefore we need special algorithms to take care of this for detection. ### Droplet Detection The droplets are detected in the preprocessed image using two strategies. The first relies on thresholding the image, the other uses marker-controlled watershed segmentation. The first strategy is visualized in figure \[fig:thresholding\] giving an example. All pixels darker than a threshold level (Fig. \[fig:thresholding0\]) are identified as background (Fig. \[fig:thresholding1\]). Isolated background regions are deleted. Then, for each pixel the distance to the background is calculated (Fig. \[fig:thresholding2\]). Local maxima in this distance map correspond to droplet centers, their distance to the background being their radius (Fig. \[fig:thresholding3\]). This procedure is carried out twice with two thresholds, one being better for finding big droplets, the other for small droplets. The second strategy is based on the marker-controlled watershed segmentation algorithm proposed by [@Gonzalez2004] pp. 422. It is illustrated in figure \[fig:watershed\]. Extended maxima in the image (Fig. \[fig:watershed1\]) are used as marker for the droplets (Fig. \[fig:watershed2\]). Lines between these maxima as well as extended minima are markers for the background. With a sobel filter the intensity gradient in the image is calculated (Fig. \[fig:watershed3\]). The markers are imposed as minima on the gradient image upon which a watershed transform is operated. The watershed lines of the intensity gradient correspond to droplet edges. The markers are needed to avoid over-segmentation. Areas and centroids of the regions enclosed by the watershed lines are taken for the droplet radii and positions (Fig. \[fig:watershed4\]). Using these two approaches, even small droplets on top of bigger ones and overlapping images of droplets can be resolved. On the other hand, some droplets are found multiple times and some imaging artifacts are taken for droplets. If two droplets with radii $r_{\rm 1},r_{\rm 2}$ are detected close to each other (centroid distance less than one radius) with almost the same radius (ratio $0.7<r_{\rm 1}/r_{\rm 2}<1.4$), they are considered as two detections of the same physical droplet. ### Matching to Image Not fully compensated dirt on the camera chip, overlaps of droplet images or unsharp droplet images can cause false detections. Hence, a second step is implemented to distinguish droplets from artifacts and the estimates of the droplet radius $r$ and position $(x,y)$ \[pixels\] are compared to the preprocessed image. As characteristics of droplets in the image, we consider them being brighter than their environment and having a circular intensity gradient at the edge (figure \[fig:matchsketch\]). A match function $m$ is defined as $$\begin{aligned} m(x,y,r) = A(r) \left[I_{\textrm{drop}}(x,y,r)-I_{\textrm{ring}}(x,y,r)\right] \\ - B \frac{(x-x_{\rm r})^2+(y-y_{\rm r})^2}{r^2} \end{aligned} \label{eqn:matchfuntion}$$ where the radius $r$ and the coordinates $x$, $x_r$, $y$ and $y_r$ are given in pixels. The difference between the mean intensity of the droplet $I_{\textrm{drop}}(x,y,r)$ and the mean intensity $I_{\textrm{ring}}(x,y,r)$ of a ring of 3 pixels width around it measures the intensity gradient at the droplet edge (figure \[fig:matchsketch\_intensity\]). An empirical radius dependent prefactor $A(r) = 1/\left(30 + r^{1.3}\right)$ ensures comparable matches for small and big droplets. Often small droplets are detected erroneously at the edge of big droplets. In this case, a little part of its halo is covering the dark background and the remaining part is covering the big droplet. The intensity weighted centroid of the halo (denoted as $x_{\rm r}$ and $y_{\rm r}$) is calculated and compared to the droplet position (figure \[fig:matchsketch\_centroid\]). The deviation accounts for this asymmetry and reduces the value of the match. It is weighted with the empirical prefactor $B=3$. For every droplet found in the droplet detection step, the match with the image is calculated. By varying radius and position of the droplet a local maximum of the match is found. All droplets with a match bigger than 0.08 are used for further analysis. A final step to increase the reliability of the measured size distribution takes advantage of the time information of image sequences. Tracking the droplets allows to sort out unphysical trajectories of artifacts and to find undetected droplets in the image using its past or future trajectory. To implement this information in section \[sec:filtering\] we must evaluate the flow field and the droplet trajectories in the flow. Calculation of Flow Field ------------------------- ### Identification of Droplets in Consecutive Images Typically droplets with radii $4\mu$m $<r<40\mu \rm m$ are detected. While small droplets ($r<10\mu$m) closely follow the flow, large droplets ($r>20\mu$m) are mainly driven by their buoyancy and do not qualify as tracers. They sediment towards the interface due to the density difference of the two phases. In a Lagrangian frame, co-moving with the surrounding fluid, the sedimentation velocity $u_{\textrm{sed}}$ amounts to the Stokes velocity for a sphere of radius $r$, which is [@Batchelor p. 234] $$u_{\textrm{sed}} = \frac{2}{9} \frac{\Delta \rho \: g r^2}{\eta}. \label{eqn:vstokes}$$ Here $\Delta \rho$ denotes the mass density difference between sphere and fluid, $g$ the gravitational acceleration, and $\eta$ the dynamic viscosity of the fluid. Hence, the droplet velocity $u_{\textrm{drop}}$ can be split into a sedimentation and an advection term $$u_{\textrm{drop}}(x,y,r,t) = u_{\textrm{sed}}(r) + u_{\textrm{flow}}(x,y,t). \label{eqn:dropletvelocity}$$ This ansatz will be checked in section \[sec:sedimentationvelocity\]. To track droplets the corresponding images of droplets in consecutive frames must be identified. To this end, the velocity field of the previous time step $u_{\textrm{flow}}(x,y,t-\delta t)$ is taken as an initial guess for $u_{\textrm{flow}}(x,y,t)$. For each droplet found in one image the position in the next image is predicted using equation (\[eqn:dropletvelocity\]). The prediction is compared to the droplets found in the image, and the droplet pair with the closest distance between prediction and actual position and with similar radii is identified as one physical droplet. In figure \[fig:displfield\] the droplets in the image are marked with red circles and the corresponding droplets in the next image with green circles. The procedure used to determine the flow field, which is indicated by the arrows, is lined out in section \[sec:flowfield\]. Subsequently in section \[sec:particletracking\] we describe details of the scheme adopted for finding particle trajectories. ![Fluorescently labeled droplets in the bottom phase (Fig. \[fig:preprocessed\_image\]). The arrows correspond to the flow field. Red circles mark the droplets found in the image and green circles their position and radius in the next image.[]{data-label="fig:displfield"}](displfield_3435_b){width="\linewidth"} ### Determine the Flow Field {#sec:flowfield} In order to calculate the Eulerian flow field $u_{\textrm{flow}}(x,y,t)$ the image is sampled by an equidistant mesh with $25\times20$ grid points. The grid distance is about $60\mu\rm m$, which is in the order of the droplet distance. The correlation length of the flow field is about $200\mu \rm m$, so the grid can resolve the flow structures. A flow correlation time of 30s allows for small frame rates (order of 1Hz). By subtracting the sedimentation velocity $u_{\textrm{sed}}$ from the droplet velocity $u_{\textrm{drop}}$ the flow contribution $u_{\textrm{flow}}$ can be calculated: The advection of the $N$ droplets in each mesh cell $(i)$ is averaged, incorporating with half weight the $M$ droplets of neighboring cells $(ij)$: $$\begin{aligned} u^{(i)}_{\textrm{flow}} = \frac{1}{N+M/2} [ \sum_{k=1}^{N}(^{k}u^{(i)}_{\textrm{drop}}- ^{k}u^{(i)}_{\textrm{sed}}) \\ + \frac{1}{2} \sum_{k=1}^{M} \sum_{j=1}^{4}(^{k}u^{(ij)}_{\textrm{drop}}- ^{k}u^{(ij)}_{\textrm{sed}}) ] \end{aligned}$$ ### Smoothing of Flow Field The advection field $u^{(i)}_{\textrm{flow}}$ is smoothed with a weighted average of next-nearest neighboring cells to eliminate unphysical discontinuities. $$% \begin{aligned} \bar{u}^{(i)}_{\textrm{flow}} = \frac{1}{4} u^{(i)}_{\textrm{flow}} + \underbrace{\frac{1}{8} \sum_{j=1}^{4} u^{(ij)}_{\textrm{flow}}}_{\textnormal{next neighbours}} %\\ + \underbrace{\frac{1}{16} \sum_{j'=1}^{4} u^{(ij')}_{\textrm{flow}}}_{\textnormal{next-nearest neighbours}} % \end{aligned}$$ In figure \[fig:displfield\] a smoothed advection field is shown together with the matched droplets. The Eulerian flow field is used to predict the droplet positions in the next image. Note that the spatial distribution of detected droplets is not homogeneous and some grid points can only be calculated by interpolation. For the subsequent analysis of the flow field only the grid points which are directly computed from droplet displacements are taken into account. Particle Tracking {#sec:particletracking} ----------------- ### Forward Tracking Now that the advection field and sedimentation velocities for each droplet found in the image at time $t$ have been determined, the position at time $t+\delta t$ can be predicted. This prediction is matched to the image of $t+\delta t$ and a local maximum of the match is searched for. This procedure is illustrated in figure \[fig:forward\]. Having determined the displacement from time $t$ to $t+\delta t$, the advection field can be recalculated. In the next step, the positions of droplets tracked from image $t$ to $t+\delta t$ plus all other droplets also detected in image $t+\delta t$ are predicted for the time step $t+2\delta t$. They are compared to the droplets in image $t+ 2\delta t$ and the procedure is repeated. The identified droplet pairs are sorted into trajectories, which contain the position and radius values for the different frames. Based on the trajectories, the flow field $u_{\textrm{flow}}(x,y,t)$ is recalculated according to the procedure described in section \[sec:flowfield\]. ### Backward Tracking The forward tracking result can be improved by subsequently tracking the droplets backwards in time. The positions of droplets at the beginning of the trajectories are predicted for the time step before and matched to the image (compare figure \[fig:backward\]). The droplet is compared to the entries in the other trajectories to merge interrupted trajectories, and to suppress branching of trajectories. As yet, our tracking algorithm does not account for the collision of droplets. ### Filtering of Trajectories {#sec:filtering} Trajectories shorter than three time steps are removed. Finally, all droplets which have moved by less than half the average advection velocity during the tracked time interval are deleted. Most likely, those spots are imaging artifacts. Eventually, the flow field is recalculated once more based on the enlarged and filtered trajectories in order to arrive at a consistent data set. Results ======= Sedimentation Velocities {#sec:sedimentationvelocity} ------------------------ We first aim to confirm the validity of the ansatz (\[eqn:vstokes\]) and (\[eqn:dropletvelocity\]). In figure \[fig:velo\_sed\] the vertical velocity calculated from the trajectories of the tracked droplets is shown as a function of the radius (red data points). The error bars display the uncertainty of the radius detection (see below) and the change of the material parameters in the time interval of averaging. It is compared to the sedimentation velocity (blue points) due to equation \[eqn:vstokes\] modified by the mean vertical velocity of the flow, which is calculated from the flow field. The Stokes velocity of a sphere (\[eqn:vstokes\]) appears to describe the average settling rate very well. We attribute the deviations for small droplet radii to the fact that both the droplet size distribution and the average flow velocity oscillate. Most of the small droplets can be found at the beginning of the oscillations, when the average downward flow velocity is relatively small (see below). ![Comparison of the measured vertical velocity of the droplets (red) and the Stokes velocity modified by the mean flow (blue), in the bottom phase at 26°C.[]{data-label="fig:velo_sed"}](velo_sed){width="0.8\linewidth"} Stokes Number ------------- Equation (\[eqn:dropletvelocity\]) holds for droplets with a Stokes number $\textrm{St}\!\ll1$. The Stokes number is defined as the ratio of two time scales $$\textrm{St} = \frac{\tau_{\textrm{drop}}}{\tau_{\textrm{flow}}}.$$ In a flow which changes on the time scale $\tau_{\textrm{flow}}$ the advected droplets respond to accelerations on the time scale $$\tau_{\textrm{drop}}\equiv \frac{2 \Delta \rho \: r^{2}}{9 \eta}.$$ The correlation time of the flow is found to be $\tau_{\textrm{flow}}\sim 30\rm s$. As an upper limit for the droplet response time we find $\tau_{\textrm{drop}}=36\mu \rm s$ at $T=50$°C for droplets with $a=40\mu\rm m$. Therefore, typical Stokes numbers in the experiment are smaller than $10^{-6}$. Even for a flow with high turbulence intensity with a Kolmogorov time scale of the order of microseconds, the Stokes number would not exceed $10^{-3}$. Due to the small density differences of fluids, kinematic particle-turbulence interactions (for a review see [@Vaillancourt2000]) can hence be ruled out for droplets, even in a turbulent flow. Trajectory Length ----------------- For a quantitative analysis trajectories of 2000 images are sorted into radius bins according to their average radius. In each bin the number of trajectories of a given length is color coded on a decade logarithmic scale (figure \[fig:traj\_duration\]). The red line indicates the time $t = H/u_{sed}$ needed by a droplet settling with Stokes velocity to pass through the image of height $H$. It puts an upper bound on the trajectory length and leads to a decrease of the trajectory length for big radii. For small radii, the duration of individual droplet detection is observed to increase roughly linearly with the radius (figure \[fig:traj\_duration\]). Hence we assume that the depth, in which a droplet can be detected, is proportional to its radius. Small droplets have to be exactly in the focal plane of the camera to be detectable, where as big droplets can still be detected when their center is slightly off the focal plane. Since the radius is detected by the maximum of the intensity gradient, it is rather robust to slight blurring, and once the droplets are too far from the focal plane, they are no longer detected because the intensity gradient at their edge is below the threshold of the match function. ![Radius dependence of the trajectory length: The number of trajectories with a given radius and length are color coded on a logarithmic scale. The black squares denote the average length for each radius bin and the solid red line is the sedimentation boundary.[]{data-label="fig:traj_duration"}](traj_duration){width="0.8\linewidth"} Uncertainty in Radius Detection ------------------------------- The fluctuations of the detected droplet radii in the trajectories are calculated to estimate the uncertainty of the attributed radius. For radii $r\leq19\mu$m a linear increase of the standard deviation is observed, yielding a constant relative uncertainty of about 20%. Above 19$\mu$m, the standard deviation is more or less constant with $3.8 \mu$m. ![Radius-dependent standard deviation of the radius of tracked droplets..[]{data-label="fig:radius_fluct"}](radius_fluct){width="0.8\linewidth"} Oscillating Size Distribution of Droplets ----------------------------------------- By slowly heating the system, more than twenty oscillations in the size distribution can be found (Fig. \[fig:sizedistribution\]). A complex interplay of nucleation ([@Krishnamurthy1980; @Sagui1999; @Vollmer2008]), coarsening ([@Aarts2005; @Vollmer2007]) and sedimentation ([@Hayase2008]) causes oscillations of the supersaturation and turbidity, modeled by [@Vollmer2007] and [@Benczik2010]. They can also clearly be identified in false color plots of the droplet size distribution displayed in Fig. \[fig:sizedistribution\]. Comparing the two size distribution plots in Fig. \[fig:sizedistribution\] reveals the effectiveness of droplet tracking. The noise in the processing of single images is reduced significantly and the evolution of the size distribution is unraveled for small droplets also. A detailed analysis of the oscillating size distribution and its implications for the understanding of precipitation in other systems (for example rain formation, for a review see [@Shaw2003]) will be given elsewhere. Oscillating Flow Properties --------------------------- To characterize the time evolution of the flow field (using only non-interpolated grid points), the mean vertical and horizontal flow velocity as well as the root mean squared velocity components are calculated. For averaging in time, four bins for each oscillation are used (figure \[fig:oscillatingflow\]). The mean horizontal velocity fluctuates around zero as expected for symmetry reasons, and the vertical flow velocity is typically negative. This can be explained by two effects. On the one hand, the upward sedimenting droplet volume has to be balanced by a downward fluid motion to ensure a zero net volume flux. On the other hand, there is a large scale convection since the system is heated from outside. Close to the walls the compositions of the phases are slightly more separated. Hence the density of the bulk is higher close to the wall than in the central region of the sample. This drives a convection pattern with downward motion close to the wall. A coupling of the oscillating droplet size distribution with the flow field produces oscillations in the average vertical flow component and the root mean squared velocities of both components. ![Mean fluid velocity in horizontal (red circles) and vertical (blue circles) direction, and the rms velocity in horizontal (red triangles) and vertical (blue triangles) direction. Data of twelve oscillations are shown, where four consecutive data points correspond to one oscillation. The measurement is performed in the bottom phase with $\xi=1.05\cdot 10^{-5} $s$^{-1}$.[]{data-label="fig:oscillatingflow"}](velocity_oscillation){width="0.8\linewidth"} Oscillation Periods ------------------- In each measurement, the driving rate $\xi$ is kept approximately constant. One may therefore expect the system to react with a constant time scale. However, we find, that during one experimental run, the oscillation period $\Delta t$ decreases with increasing temperature (compare figures \[fig:sizedistribution\] and \[fig:oscillatingflow\]). This trend has been observed in turbidity measurements with other binary mixtures by [@Auernhammer2005] and [@Vollmer2007] before. Secondly, the oscillation period depends on the driving rate. For faster temperature ramps, the oscillation period decreases. We have measured the oscillation periods of the droplet size distribution for two decades in $\xi$ for the top phase and almost three decades in $\xi$ for the bottom phase (see figure \[fig:oscillationperiods\]). The colors and symbols encode different heating rates $\xi$: black stars, $\xi < 6\times 10^{-6}\,{\rm s}^{-1}$; blue crosses, $\xi < 1.3\times 10^{-5}\,{\rm s}^{-1}$; cobalt circles, $\xi < 3\times 10^{-5}\,{\rm s}^{-1}$; green triangles, $\xi < 6\times 10^{-5}\,{\rm s}^{-1}$; red squares, $\xi < 3\times 10^{-4}\,{\rm s}^{-1}$; and magenta diamonds, $\xi > 3\times 10^{-4}\,{\rm s}^{-1}$. A trend of the period $\Delta t$ with the driving $\xi$ is visible, but it is masked by the temperature dependence of $\Delta t$. Discussion & Conclusions ======================== To investigate the physics of phase separation we study the water/isobutoxyethanol-system. It has very convenient physical properties, a critical point just above room temperature, an upper miscibility gap, and low vapor pressure. Therefore, sample preparation is easy and a broad range of parameters can be studied with a fully automatized setup. Using Nile Red as a dye and choosing a system with one polar and one organic phase we obtained images with a high fluorescence contrast. A short arc mercury vapor lamp is suitable for the illumination of the sample. To follow the evolution of the broad droplet size distribution for several hours, high resolution images (5 Mpixel) are taken at small frame rates (0.3 to 5 Hz). The resolution of the camera has to be large enough to detect droplet radii, whereas the frame rate must be adapted to the maximal droplet velocities. For measuring well resolved flow fields the droplet number density has to be high enough. This sets a lower bound to the rate of temperature change, which controls the nucleation of droplets. On the other hand, for high droplet number densities, measurements are only possible close to the walls. As the droplets are not index matched, clear images can only be obtained at a wall distance up to once or twice the mean distance of droplets. Therefore the measurements were conducted 0.3mm away from the wall, i.e. at a distance of about 1.5 times the correlation length calculated from the flow fields. We presented a particle tracking algorithm based on image processing with MATLAB. It detects droplets in a range of 4 to 40 $\mu$m radius (bigger ones can be detected but are very scarce) in a sample field of $1.3\times1.5$mm$^2$. The motion of the droplets is decomposed into a sedimentation and an advection term. The sedimentation velocity is given by the Stokes velocity of a sphere. The advection of droplets is sampled on an Eulerian grid. By taking the sedimentation velocity explicitly into account, big droplets can also be used to calculate the flow field, which reduces the noise. This is possible because the Stokes number is much smaller than one for all droplets in our experiments. It is not necessary to know in advance the radius at which sedimentation starts to become significant. This threshold depends on the specific flow conditions. There are two advantages by incorporating information on the droplet radius: First, the radius is a good criterion to identify droplets in subsequent images. Secondly, the prediction of the droplet position in the next image is improved by knowledge of the radius dependent sedimentation velocity. The experimental setup and the tracking algorithm are optimized for long measurement times (several hours) and low Reynolds number flow. Our experimental procedure can be applied to a broad variety of binary mixtures, provided an appropriate fluorescent dye for labeling (only) one of the two phases is found and the excitation and emission filters are adapted to it. It enables then detailed investigations of the evolution of the system: The droplet size distribution, Lagrangian particle velocities and Eulerian flow fields can be measured simultaneously for a broad range of heating rates, temperatures and sample geometries. To investigate other laminar or turbulent flows a high speed camera has to be used. Our technique might then be a promising approach to investigate turbulent boundary layers with reactions or phase separation. Furthermore, one may consider placing an endoscope into the sample, as [@Maass2009] have done for example, to measure the size distribution outside the boundary layer. A further advantage of the technique is that the simultaneous measurement of particle position and particle size allows us to determine the size evolution of individual droplets as they progress across the measurement area. This information can be used to obtain information on coalescence rates and collision efficiency of sedimenting droplets: A preliminary study shows that droplets with radii about $40\mu$m grow on average by $0.5\mu$m/s while they travel across the measurement area. This growth amounts to a collision efficiency (see [@Pruppacher1997] pp. 569 and [@Pinsky1999]) of order unity. Further studies are under way to determine how the collision efficiency depends on the radius of the sedimenting droplet. This methodology should be of interest to precipitation processes such as rain formation, where the growth of medium sized droplets is believed to be driven by coalescence. We are grateful to Wilhelm Hüttner, Konstantin Christou, Kristian Hantke, Alberto de Lozar and Eric Stellamanns for enlightening discussions and experimental tests of several illumination techniques. We thank Doris Vollmer and Günther Auernhammer for advice in designing the setup, choosing the system and developing the experimental protocol. We thank Markus Holzner and Mukund Vasudevan for comments on the manuscript. Tobias Lapp acknowledges financial support from Deutsche Forschungsgemeinschaft FOR 1182. Appendix {#appendix .unnumbered} ======== The viscosity of i-BE is measured with an Ubbelohde viscosimeter type 537 10/I made by Schott. The temperature dependence of the viscosity $\eta$ \[kg/ms\] is fitted by $$\eta(T) = A \cdot 10^{\frac{B\cdot (20 - T) - C\cdot (T-20)^2}{T+D}} \label{eqn:viscosity_temp}$$ with temperatures expressed in °C. The coefficients are given in table \[tab:coeffvisc\], where the values for water are taken from [@Weast1988]. $A [kg/ms]$ $B$ $C$ $D$ ------- ---------------------- -------- ---------- ----- water $1.002\cdot 10^{-3}$ 1.3272 0.001053 105 i-BE $3.36\cdot 10^{-3}$ 1.730 0.001 108 : Fit coefficients for the viscosity of water ([@Weast1988]) and i-BE, defined by equation (\[eqn:viscosity\_temp\]).[]{data-label="tab:coeffvisc"} To interpolate the viscosities for a mixed phase of given mass fraction $\Phi$ we use the composition-dependent viscosities at 25°C for a homogeneous mixture in the single phase regime given in [@Menzel2003]. The data is fitted with a fifth order polynomial $$\begin{aligned} \eta(\Phi,{\rm T}=25^{\circ}{\rm C}) = -40.66\Phi^5 +103.44\Phi^4 -100.32\Phi^3 \\ +39.35\Phi^2 +0.17\Phi +0.91. \label{eqn:viscfit} \end{aligned}$$ ![From top to bottom: Viscosity of the top phase, isobutoxyethanol, the bottom phase and water, using equations (\[eqn:viscosity\_temp\],\[eqn:etafinal\]). At 40°C the viscosity of the two mixed phases was measured with an Ubbelohde viscosimeter.[]{data-label="fig:viscosity"}](viscosity_temp){width="0.80\linewidth"} Assuming that the coefficients of interpolation are not changing substantially in the temperature range of our measurements, a rescaled viscosity $\tilde{\eta}(\Phi)$ is defined. It only depends on the composition $\Phi$ $$\eta(\Phi,T) = \tilde{\eta}(\Phi)\cdot \eta_{iBE}(T) + (1-\tilde{\eta}(\Phi))\cdot \eta_{H_2O}(T). \label{eqn:etafinal}$$ The viscosities of the two phases are shown in figure \[fig:viscosity\] as a function of temperature. To check the strong assumption entering this interpolation, we measured the viscosity of the two phases at $T=40$°C. For both phases it followed the prediction of equation (\[eqn:etafinal\]) to within 2%. This is sufficiently accurate for our means. [38]{} \[1\][\#1]{} \[1\][[\#1]{}]{} urlstyle \[1\][DOI \#1]{} \[2\]\[\][[\#2](#2)]{} Aarts DGAL, Dullens RPA, Lekkerkerker HNW (2005) Interfacial dynamics in demixing systems with ultralow interfacial tension. New J Phys 7:40 Auernhammer GK, Vollmer D, Vollmer J (2005) Oscillatory instabilities in phase separation of binary mixtures: Fixing the thermodynamic driving. J Chem Phys 123(13):134,511–8 Aveyard R, Binks BP, Clint JH (2003) Emulsions stabilised solely by colloidal particles. Adv Colloid Interface Sci 100-102:503–546 Batchelor G (2005) An Introduction to Fluid Dynamics. Cambridge University Press Benczik IJ, Vollmer J (2010) A reactive-flow model of phase separation in fluid binary mixtures with continuously ramped temperature. Europhys Lett 91(3):\ 36 003 Binks BP (2002) Particles as surfactants — similarities and differences. Curr Opin Colloid Interface Sci 7(1-2):21–41 Cates ME, Vollmer J, Wagner A, Vollmer D (2003) Phase separation in binary fluid mixtures with continuously ramped temperature. Philos Trans R Soc London, Ser A 361(1805):793–807 Doi H, Tamura K, Murakami S (2000) [Thermodynamic properties of aqueous solution of 2-isobutoxyethanol at T =  (293.15, 298.15, and 303.15) K, below and above LCST]{}. J Chem Thermodyn 32(6):729–741 Emmanuel S, Berkowitz B (2006) An experimental analogue for convection and phase separation in hydrothermal systems. J Geophys Res 111:B09,103 Fowler S, Greenspan P (1985) [Application of Nile red, a fluorescent hydrophobic probe, for the detection of neutral lipid deposits in tissue sections: comparison with oil red O.]{} J Histochem Cytochem 33(8):833–836 Gonzalez RC, Woods RE, Eddins Sl (2004) Digital Image Processing Using MATLAB. Pearson Education, Inc. Hayase Y, Kobayashi M, Vollmer D, Pleiner H, Auernhammer GK (2008) Asymmetric oscillations during phase separation under continuous cooling: A simple model. J Chem Phys 129(18):184,109–7 Heffels C, Polke R, Rädle M, Sachweh B, Schäfer M, Scholz N (1998) Control of particulate processes by optical measurement techniques. Part Part Syst Charact 15(5):211–218 Hu B, Angeli P, Matar OK, Lawrence CJ, Hewitt GF (2006) Evaluation of drop size distribution from chord length measurements. AlChE J 52(3):931–939 Kreizer M, Ratner D, Liberzon A (2010) Real-time image processing for particle tracking velocimetry. Exp Fluids 48(1):105–110–110 Krishnamurthy S, Goldburg WI (1980) Kinetics of nucleation in a binary liquid mixture. Phys Rev A 22:2147–2155 Lovick J, Mouza A, Paras S, Lye G, Angeli P (2005) Drop size distribution in highly concentrated liquid/liquid dispersions using a light back scattering method. J Chem Technol Biotechnol 80(5):545–552 Maaß S, Wollny S, Voigt A, Kraume M (2011) Experimental comparison of measurement techniques for drop size distributions in liquid/liquid dispersions. Exp Fluids 50(2):259–269 Maas HG, Gruen A, Papantoniou D (1993) Particle tracking velocimetry in three-dimensional flows. Exp Fluids 15(2):133–146 Maass S, Grünig J, Kraume M (2009) Measurement techniques for drop size distributions in stirred liquid-liquid systems. Chem Process Eng 30:635–651 Malik NA, Dracos T, Papantoniou DA (1993) Particle tracking velocimetry in three-dimensional flows. Exp Fluids 15(4):279–294 Menzel K, Mirzaev SZ, Kaatze U (2003) Crossover behavior in micellar solutions with lower critical demixing point: Broadband ultrasonic spectrometry of the isobutoxyethanol-water system. Phys Rev E 68(1):011,501 Nakayama H, Kanenaga H, Fujioka T (2001) The study of water structure in aqueous solutions of butoxyethanol by enthalpy of mixing measurements. J Therm Anal Calorim 64(1):193–199 Ouellette N, Xu H, Bodenschatz E (2006) [A quantitative study of three-dimensional Lagrangian particle tracking algorithms]{}. Exp Fluids 40(2):301–313 Petrak D (2002) Simultaneous measurement of particle size and particle velocity by the spatial filtering technique. Part Part Syst Charact 19(6):391–400 Pinsky M, Khain A, Shapiro M (1999) [Collisions of Small Drops in a Turbulent Flow. Part I: Collision Efficiency. Problem Formulation and Preliminary Results]{}. Journal of the Atmospheric Sciences 56(15):2585–2600 Pruppacher HR, Klett JD (1997) Microphysics of Clouds and Precipitation, 2nd edn. Kluwer Academic Publications Ruf A, Worlitschek J, Mazzotti M (2000) [Modeling and Experimental Analysis of PSD Measurements through FBRM]{}. Part Part Syst Charact 17(4):167–179 Sagui C, Grant M (1999) Theory of nucleation and growth during phase separation. Phys Rev E 59(4):4175–4187 Shaw RA (2003) Particle-turbulence interactions in atmospheric clouds. Annu Rev Fluid Mech 35(1):183–227 Simmons MJH, Zaidi SH, Azzopardi BJ (2000) Comparison of laser-based drop-size measurement techniques and their application to dispersed liquid-liquid pipe flow. Opt Eng 39:505 Thijssen JHJ, Clegg PS (2010) Emulsification in binary liquids containing colloidal particles: a structure-factor analysis. J Phys Condens Matter 22(45):455,102 Vaillancourt PA, Yau MK (2000) Review of particle-turbulence interactions and consequences for cloud physics. Bull Am Meteorol Soc 81(2):285–298 Vollmer D, Vollmer J, Strey R (1997) Oscillations in the dynamics of temperature-driven phase separation. Europhys Lett 39(3):245–250 Vollmer D, Vollmer J, Wagner AJ (2002) Oscillatory kinetics of phase separation in a binary mixture under constant heating. Phys Chem Chem Phys 4:1380–1385 Vollmer J (2008) Phase separation under ultraslow cooling: Onset of nucleation. J Chem Phys 129(16):164,502–8 Vollmer J, Auernhammer GK, Vollmer D (2007) Minimal model for phase separation under slow cooling. Phys Rev Lett 98(11):115,701 Weast RC (1988-1989) CRC Handbook of Chemistry and Physics, 69th edn. CRC Press, Inc. Boca Raton, Florida
{ "pile_set_name": "ArXiv" }
--- author: - 'Konstantina Christakou, *Member, IEEE*, Jean-Yves Le Boudec, *Fellow, IEEE*, Mario Paolone, *Senior Member, IEEE*, Dan-Cristian Tomozei, *Member, IEEE* [^1]' bibliography: - 'mybibfile.bib' title: Efficient Computation of Sensitivity Coefficients of Node Voltages and Line Currents in Unbalanced Radial Electrical Distribution Networks --- [^1]: Konstantina Christakou, Mario Paolone, Jean-Yves Le Boudec and Dan-Cristian Tomozei (email: [email protected], jean-yves.leboudec @epfl.ch, [email protected], dan-cristian.tomozei @epfl.ch) are with the École Polytechnique Fédérale de Lausanne, CH-1015 Lausanne, Switzerland.
{ "pile_set_name": "ArXiv" }
--- abstract: | The problem of description of superintegrable systems (i.e., systems with closed trajectories in a certain domain) in the class of rotationally symmetric natural mechanical systems goes back to Bertrand and Darboux. We describe all superintegrable (in a domain of slow motions) systems in the class of rotationally symmetric magnetic geodesic flows. We show that all sufficiently slow motions in a central magnetic field on a two-dimensional manifold of revolution are periodic if and only if the metric has a constant scalar curvature and the magnetic field is homogeneous, i.e. proportional to the area form. [**Key words:**]{} superintegrable system, surface of revolution, magnetic geodesic, magnetic Bertrand system. author: - 'E.A. Kudryavtseva, S.A. Podlipaev' title: | Superintegrable Bertrand\ magnetic geodesic flows --- UDK 514.853, 517.938.5 Introduction ============ We study magnetic geodesic flows invariant under rotations. Such a dynamical system describes the motion of a charged particle in a central [*magnetic field*]{} on a two-dimensional Riemannian manifold of revolution. We suppose that the magnetic field is not an identical zero. We study [*slow motions*]{}, i.e. motions with sufficiently small positive level of energy (or with a sufficiently small velocity). From a geometrical point of view, we study [*magnetic geodesics*]{} of sufficiently small radii of curvature, where the radius of curvature depends on a position on the surface and is proportional to the ratio of the magnetic field to the area form, with a sufficiently small positive proportionality factor. We are interested in the systems satisfying the following [*Bertrand condition for slow motions*]{}: any sufficiently slow motion of the charged particle is periodic. On a geometrical language, the Bertrand condition is formulated as follows: all magnetic geodesics of sufficiently small radii of curvature are closed curves. We will show (Theorem \[thm:Bert\]) that the Bertrand condition holds if and only if the metric has a constant scalar curvature and the magnetic field is homogeneous, i.e. proportional to the area form. Let us give a short historical overview. The problem of description of superintegrable systems (i.e., systems with closed trajectories in a certain domain) in the class of rotationally symmetric natural mechanical systems goes back to Bertrand and Darboux. Natural mechanical systems of Bertrand’s type were described under several restrictions by Bertrand [@Ber], Darboux [@Dar], Besse [@Bes], Perlick [@Per], Zagryadskii, Kudryavtseva and Fedoseev [@ZKF], Kudryavtseva and Fedoseev [@KF1; @KF2; @KF3] and others (cf. a survey in [@ZKF]). In this paper, we obtain a description (i.e., a rigorous classification) of all [*magnetic Bertrand systems*]{} — superintegrable magnetic geodesic flows on two-dimensional configurational manifolds of revolution satisfying the periodicity condition for slow motions. The problem of describing all [*electromagnetic Bertrand systems*]{} remains open, i.e. superintegrable systems that define motion of a charged particle under the influence of a potential as well as a magnetic force fields on a two-dimensional configurational manifold of revolution, and we plan to solve it in future. Let us indicate a difference of our approach from the classical approach of Bertrand [@Ber] and many of his followers including [@ZKF]. The classical approach is based on studying a [*family of reduced systems*]{} obtained from the initial system via a [*time change*]{} on the phase trajectories by the angle coordinate $\varphi=\varphi(t)$ (longitude) on the configuration manifold of revolution. Such a time change $t\to\varphi(t)$ makes sense (i.e. is regular and monotone) for [*nonsingular*]{} trajectories: when the longitude is monotone in time ($\dot\varphi(t)\ne0$) along the whole trajectory. The reduced system has one degree of freedom, depends on one parameter (the constant of kinetic momentum) and, for each parameter value, has a nondegenerate equilibrium point of the “centre” type (corresponding to a circular solution). Clearly, a superintegrability will take place in the case of a [*common isochronicity*]{} of the family of reduced systems, i.e., when the periods of their solutions near the indicated equilibria coincide, and their common period is commensurable with $\pi$. Since, in the classical case (without a magnetic field) almost all solutions are nonsingular, the superintegrability is equivalent to the common isochronicity. But, in the “magnetic” case, there are “many” singular solutions (e.g. all sufficiently slow motions), and the indicated time change (together with the beautiful condition of a common isochronicity) does not make sense for them. For such solutions, we write the reduced system without any time change, in contrast to the classical approach. Besides that, we follow the classical approach [@Ber]. Let us proceed with precise statements. Formulation of the main result {#sec:slow} ============================== Let $(Q,g)$ be a smooth two-dimensional Riemannian manifold of revolution, $B$ a differential 2-form on $Q$ invariant under rotations. The 2-form $B$ is, of course, closed (${{\rm d}}B=0$) and, thus, defines a [*magnetic field*]{} on $Q$. Its invariance under rotations means that the magnetic field is [*central*]{}. The manifold of revolution $Q$ is diffeomorphic to either a 2-sphere or a 2-disk or a 2-torus or an open cylinder: $Q\setminus\mathrm{Fix}(S^1)\approx I\times S^1$ with coordinates $(r,\varphi)$, where\ $r\in I$ is a [*latitude*]{}, i.e. a natural parameter on meridians $I\times\{\varphi_0\}$,\ $\varphi\in{{\mathbb R}}/2\pi{{\mathbb Z}}=S^1$ is a [*longitude*]{}, i.e. an angle coordinate on parallels $\{r_0\}\times S^1$.\ Here $I=(r_1,r_2)\subseteq{{\mathbb R}}$ (in the cases of a sphere, a 2-disk and a cylinder) or $I=S^1={{\mathbb R}}/L{{\mathbb Z}}$, $L>0$ (in the case of a 2-torus). Without loss of generality, we will assume that $Q$ is diffeomorphic to a cylinder: $$Q\approx(r_1,r_2)\times S^1$$ (the cases of a 2-sphere and a 2-disk reduce to this case by considering $Q$ punctured at the fixed points of the rotation, while the case of a 2-torus $Q\approx S^1\times S^1$ reduces to it by considering a covering cylinder $\tilde Q={{\mathbb R}}\times S^1$). The Riemannian metric of revolution on $Q$ has the form $$g= {{\rm d}}r^2 + f^2(r) {{\rm d}}\varphi^2, \qquad (r,\varphi) \in Q,$$ where $f(r)>0$ is a smooth function (the radius of a parallel $\{r\}\times S^1$). The rotationally symmetric 2-form $B$ on $Q$ has the form $$B=b(r){{\rm d}}r\wedge {{\rm d}}\varphi=a'(r){{\rm d}}r\wedge {{\rm d}}\varphi, \qquad (r,\varphi) \in Q,$$ where $a(r)$ is a function on $(r_1,r_2)$ defined (up to an additive constant) by the condition $a'(r)=b(r)$. The motion of the charged particle in the magnetic field $B$ on the surface $(Q,g)$ is described by the Hamiltonian system on $T^*Q$, with the Hamilton function and the complex structure $$H=\dfrac{p_r^2}{2} + \dfrac{p_\varphi^2}{2f^2(r)}, \quad \omega = {{\rm d}}p_r \wedge {{\rm d}}r + {{\rm d}}p_\varphi \wedge {{\rm d}}\varphi + B = {{\rm d}}p_r \wedge {{\rm d}}r + {{\rm d}}(p_\varphi + a(r)) \wedge {{\rm d}}\varphi .$$ Since the Hamilton function $H$ is independ of the angle variable $\varphi$, the system is integrable via an additional first integral $$K=\tilde p_\varphi := p_\varphi + a(r)$$ (the kinetic momentum). The first integral $K$ is $2\pi$-periodic, i.e., defines a free Hamiltonian action of the circle on $T^*Q$. If the magnetic field $B=a'(r){{\rm d}}r\wedge {{\rm d}}\varphi$ has no zeros, the change $r\to a=a(r)$ of the latitude $r\in(r_1,r_2)$ is monotone and regular. The corresponding momentum change is $p_r=a'(r)p_a$. We obtain a manifold of revolution $Q\simeq(a_1,a_2)\times S^1$ with Riemannian metric $$g = \dfrac{{{\rm d}}a^2}{R(a)} + \dfrac{{{\rm d}}\varphi^2}{F(a)},$$ where $R(a(r))=a'(r)^2>0,$ $F(a(r))=1/f^2(r)>0.$ The Hamilton function and the symplectic structure are $$\label {eq:Ham} H=R(a)\dfrac{p_a^2}{2} + F(a)\dfrac{p_\varphi^2}{2}, \qquad \omega = {{\rm d}}p_a \wedge {{\rm d}}a + {{\rm d}}(p_\varphi+a) \wedge {{\rm d}}\varphi.$$ In the canonical variables $(a,\varphi,p_a,\tilde p_\varphi=K)$, they take the form $$\label {eq:Ham:} H=R(a)\dfrac{p^2_a}{2} + F(a)\dfrac{(\tilde p_\varphi-a)^2}{2}, \qquad \omega={{\rm d}}p_a \wedge {{\rm d}}a + {{\rm d}}\tilde p_\varphi\wedge {{\rm d}}\varphi,$$ and the equations of motion of the charged particle have the form $$\label {eq:ODE} \left\{ \begin{array}{l} \dot a = \dfrac{\partial H}{\partial p_a} = R (a)p_a, \\ \dot{\varphi} = \dfrac{\partial H}{\partial \tilde p_\varphi} = F(a)(\tilde p_\varphi - a),\\ \dot{p}_a = - \dfrac{\partial H}{ \partial a} = - R'(a)\dfrac{p^2_a}{2} - F'(a)\dfrac{(\tilde p_\varphi-a)^2}{2} + F(a)(\tilde p_\varphi-a), \\ \dot {\tilde p}_\varphi = - \dfrac{\partial H}{\partial\varphi} = 0. \\ \end{array} \right.$$ Let us study an analogue of the [*Bertrand problem*]{} (cf. e.g. [@ZKF]) for [*slow*]{} (i.e. with small velocity) motions of the charged particle in the magnetic field under consideration. \[[[@Pod2015 Theorem 7.1]]{}\] \[thm:Bert\] Suppose that the central magnetic field $B$ on a two-dimensional Riemannian manifold of revolution $(Q,g)$ is different from the identical zero. Suppose that the “Bertrand condition for slow motions” holds: any motion with a sufficiently small nonzero velocity is periodic (in detail: there exists a continuous function $f>0$ on the manifold $Q$ such that any motion with initial conditions $(q,p)\in T^*Q$, $0<|p|<f(q)$, is periodic). Then: [(a)]{} the magnetic field $B$ is [*homogeneous*]{}, i.e., proportional to the area form ${{\rm d}}\sigma=f(r){{\rm d}}r\wedge {{\rm d}}\varphi$ (hence $B$ has no zeros, the motion is described by the Hamiltonian system (\[eq:Ham\]) and (\[eq:ODE\]), the 2-forms $B={{\rm d}}a\wedge {{\rm d}}\varphi$ and ${{\rm d}}\sigma=\dfrac{{{\rm d}}a\wedge {{\rm d}}\varphi}{\sqrt{RF}}$ are proportional and $RF\equiv \lambda^2=\const$); [(b)]{} the minimal positive period of the solution with the initial condition $(q,p)\in T^*Q$, $0<|p|<f(q)$, continuously depends on the initial condition $(q,p)$ and tends to the minimal positive period of the linearized system at the equilibrium $(q,0)$ as $|p|\to0$; [(c)]{} the scalar curvature $\Scal$ of the manifold $(Q,g)$ is constant and equals $\Scal=\dfrac{\lambda^2}{F^3}(F''F-2(F')^2)=-\left(\dfrac{\lambda^2} F\right)''=-R''$. In particular, $R(a)=\dfrac{\lambda^2}{F(a)}=\lambda_1+\lambda_2a-\Scal\dfrac{a^2}{2}$ for some constants $\lambda_1,\lambda_2\in{{\mathbb R}}$. The converse is also true: if the scalar curvature $\Scal$ of the manifold $(Q,g)$ is constant, and the magnetic field $B$ is homogeneous and different from the identical zero, then the Bertrand condition for slow motions holds. Derivation of homogeneity of the magnetic field from the Bertrand condition for slow motions ============================================================================================ In this section, we prove items (a) and (b) of Theorem \[thm:Bert\]. Suppose that the magnetic field $B$ has no zeros (the general case is studied on Step 6 below). Thus, the motion is given by the Hamiltonian system (\[eq:Ham\]) and (\[eq:ODE\]). We pass to more convenient canonical variables obtained from the initial variables by a [ *guiding-centre transformation*]{} [@Nei §3], $$\label {eq:hat:varphi} h:(a,\varphi,p_a,p_\varphi) \mapsto (\hat a, \hat\varphi, p_a,p_\varphi), \qquad \hat a=a+p_\varphi=K, \qquad \hat\varphi=\varphi-p_a.$$ In the new variables, the symplectic structure takes a canonical form: $$\omega ={{\rm d}}p_a \wedge {{\rm d}}(\hat a-p_\varphi) + {{\rm d}}\hat a \wedge {{\rm d}}(\hat\varphi+p_a) = - {{\rm d}}p_a \wedge {{\rm d}}p_\varphi + {{\rm d}}\hat a \wedge {{\rm d}}\hat\varphi.$$ Perform a [*scale change*]{} of momenta $h_\varepsilon:(\hat a, \hat\varphi, \hat p_a, \hat p_\varphi) \mapsto (\hat a, \hat\varphi, p_a, p_\varphi)$ by the formula $$p_a = {\varepsilon} \hat p_a, \qquad p_\varphi = {\varepsilon} \hat p_\varphi,$$ where $\varepsilon>0$ is a small parameter (e.g., the modulus of the velocity vector). Then $$H = \varepsilon^2\left(R(\hat a - \varepsilon \hat p_\varphi)\dfrac{\hat p_a^2}{2} + F(\hat a - \varepsilon \hat p_\varphi)\dfrac{\hat p_\varphi^2}{2} \right) , \qquad \omega = {{\rm d}}\hat a \wedge {{\rm d}}\hat\varphi - \varepsilon^2 {{\rm d}}\hat p_a \wedge {{\rm d}}\hat p_\varphi .$$ The equations of motion in new variables $(\hat a, \hat\varphi, \hat p_a, \hat p_\varphi)$ (sometimes called “slow-fast” variables [@Kud] for the given system) have the form $$\label {eq:ODE:eps} \left\{ \begin{array}{l} \dot {\hat a} = - \dfrac{\partial H}{\partial\hat \varphi} = 0,\\ \dot {\hat \varphi} = \dfrac{\partial H}{\partial \hat a} = \varepsilon^2\left(R'(\hat a - \varepsilon\hat p_\varphi)\dfrac{{\hat p_a}^2}{2} + F'(\hat a - \varepsilon\hat p_\varphi)\dfrac{\hat p_\varphi^2}{2}\right),\\ \dot{\hat p}_a = \dfrac{1}{\varepsilon^2} \dfrac{H}{ \partial \hat p_\varphi} = - \varepsilon R'(\hat a - \varepsilon\hat p_\varphi)\dfrac{{\hat p_a}^2}{2} - \varepsilon F'(\hat a - \varepsilon\hat p_\varphi)\dfrac{{\hat p_\varphi}^2}{2} + F(\hat a - \varepsilon\hat p_\varphi)\hat p_\varphi,\\ \dot{\hat p}_\varphi = - \dfrac{1}{\varepsilon^2} \dfrac{\partial H}{\partial \hat p_a} = - R (\hat a - \varepsilon\hat p_\varphi)\hat p_a. \end{array} \right.$$ \[rem:circ\] Let us find all [*relative equilibria*]{}, i.e., all phase points at which ${{\rm d}}H$ and ${{\rm d}}K$ are proportional (so, the corresponding integral curves $\gamma\subset T^*Q$ are [*circular orbits*]{} and [*equilibria*]{}, since they satisfy $a \equiv \const$). Due to the equations (\[eq:ODE\]), we have $p_a \equiv 0,$ $F'(a)\dfrac{(\tilde p_\varphi - a)^2}{2} + F(a)(a-\tilde p_\varphi) = 0.$ The latter equality implies $\tilde p_\varphi \equiv a$ (the case of an equilibrium) or $\dfrac{F'(a)}{F(a)}(\tilde p_\varphi - a) \equiv 2$ (the case of a circular orbit). In the new variables, for $\varepsilon>0$, we obtain $\hat p_a \equiv 0$ and either $\hat p_\varphi=0$ (in the case of an equilibrium), or $\dfrac{F'(\hat a - \varepsilon\hat p_\varphi)}{F(\hat a - \varepsilon\hat p_\varphi)}\varepsilon\hat p_\varphi \equiv 2$ (in the case of a circular orbit). As $\varepsilon\to0$, the system (\[eq:ODE:eps\]) tends to a well defined [*limit system*]{} — the family of harmonic oscillators on each fibre, i.e. the family of Hamiltonian systems $$\label {eq:harm} \left(T^*_{q}Q, \ \omega_{q}=-{{\rm d}}\hat p_a \wedge {{\rm d}}\hat p_\varphi, \ H_{q}=R(\hat a) \dfrac{\hat p_a^2}{2} + F(\hat a)\dfrac{\hat p_\varphi^2}{2}\right)$$ with parameters $q=(\hat a,\hat\varphi)\in(a_1,a_2)\times S^1=Q,$ where $\dot{\hat a}=\dot{\hat\varphi}=0.$ Under the inverse change $h^{-1}\circ h_0$, any solution of the limit system is transformed to the corresponding equilibrium $h^{-1}(\hat a,\hat\varphi,0,0)=(\hat a,\hat\varphi,0,0)$ of the initial system (\[eq:ODE\]) indicated in Remark \[rem:circ\]. Thus, solutions of the system (\[eq:ODE:eps\]) correspond to [*slow*]{} (i.e. having a small velocity) motions given by the system (\[eq:ODE\]) with (\[eq:Ham\]). Notice that the limit system has the form $\dot{\hat a}=\dot{\hat\varphi}=0,$ $\dot{\hat p}_a=F(\hat a)\hat p_\varphi,$ $\dot{\hat p}_\varphi = -R (\hat a)\hat p_a.$ Therefore all its solutions, apart from the equilibria $\hat p_a=\hat p_\varphi=0,$ satisfy the equation $\ddot{\hat p}_a = -R(\hat a)F(\hat a)\hat p_a$ and, hence, determine harmonic oscillations with angular frequency $\sqrt{R(\varkappa)F(\varkappa)}$ and minimal positive period $$\label {eq:T} T(\varkappa)=\dfrac{2\pi}{\sqrt{R(\varkappa)F(\varkappa)}},$$ where $\varkappa\in(a_1,a_2)$ is the value of the first integral $K=\hat a=\tilde p_\varphi$ on the given solution. Step 1. Let us first suppose that the magnetic field $B$ has no zeros. Fix a number $\varkappa\in(a_1,a_2),$ and consider the [*effective potential*]{} $$U_{\varkappa}=U_{\varkappa}(a):=F(a)\dfrac{(\varkappa - a)^2}{2}, \qquad a\in(a_1,a_2),$$ of the system (\[eq:Ham\]). The Hamilton function has the form $H=R(a)\dfrac{p_a^2}{2} + U_{K} (a)$. In the new variables, the effective potential is $$\label {eq:Ueff} \dfrac1{\varepsilon^2}U_{\varkappa}(a) =F(\varkappa - \varepsilon\hat p_\varphi)\dfrac{\hat p_\varphi^2}{2} =:\Hat U_{\varepsilon,\varkappa}(\hat p_\varphi), \qquad \hat p_\varphi\in{{\mathbb R}},$$ and the Hamilton function is $\dfrac1{\varepsilon^2}H=R(K - \varepsilon\hat p_\varphi)\dfrac{\hat p_a^2}{2} + \Hat U_{\varepsilon,K} (\hat p_\varphi).$ Let us study the [*reduced system*]{} corresponding to the $2\pi$-periodic first integral $K$, with the Hamilton function and the symplectic structure $$\label {eq:reduc} H_{\varepsilon,\varkappa} := \dfrac1{\varepsilon^2}H|_{\{K=\varkappa\}} = R(\varkappa - \varepsilon\hat p_\varphi)\dfrac{\hat p_a^2}{2} + \Hat U_{\varepsilon,\varkappa} (\hat p_\varphi), \qquad - {{\rm d}}\hat p_a \wedge {{\rm d}}\hat p_\varphi,$$ where $\varkappa\in(a_1,a_2)$ is the parameter of the reduced system, $0<\varepsilon\ll1$ is a small (“scale”) parameter. The reduced system is defined on a “reduced phase plane” $\{K=\varkappa\}/S^1\subset(T^*Q)/S^1$ with phase variables $(\hat p_a,\hat p_\varphi)$ and an induced symplectic structure, where one considers the Hamiltonian action of the circle $S^1$ on $T^*Q$ generated by the $2\pi$-periodic first integral $K$. The Hamilton function $H_{\varepsilon,\varkappa}$ of the reduced system (\[eq:reduc\]) equals the sum of the “reduced kinetic energy” $R(\varkappa - \varepsilon\hat p_\varphi)\dfrac{\hat p_a^2}{2}$ (quadratic in the “reduced momentum” $\hat p_a$) and the effective potential $\Hat U_{\varepsilon,\varkappa}(\hat p_\varphi)$ (depending on the “reduced coordinate” $\hat p_\varphi$ only). That is, the reduced system is a natural mechanical system with one degree of freedom, therefore we can explicitly solve it by standard techniques. Let us do this. For the value $\Hat E$ of the reduced Hamilton function $H_{\varepsilon,\varkappa}$, we have $R(\varkappa-\varepsilon\hat p_\varphi) \hat p_a^2 = 2\Hat E-2\Hat U_{\varepsilon,\varkappa}(\hat p_\varphi)\ge0.$ Therefore the value $\hat p_a$ can be expressed in terms of $\varepsilon,\Hat E,\varkappa$ and $\hat p_\varphi$ by the formula $$\label {eq:p} \hat p_a =\pm\sqrt{\dfrac{2\Hat E-2\Hat U_{\varepsilon,\varkappa}(\hat p_\varphi)}{R(\varkappa-\varepsilon\hat p_\varphi)}} \stackrel{(\ref {eq:Ueff})} = \pm\sqrt{\dfrac{2\Hat E- F(\varkappa - \varepsilon\hat p_\varphi) \hat p_\varphi^2}{R(\varkappa-\varepsilon\hat p_\varphi)}}.$$ Let us fix real numbers $\varkappa\in(a_1,a_2)$, $\Hat E\in(0,1)$ and $\hat p_\varphi\in{{\mathbb R}}$ such that $|\hat p_\varphi|^2<\dfrac{2\Hat E}{F(\varkappa)}$. In the limit $\varepsilon\to0$, we obtain $$\label {eq:p:lim} \Hat U_{0,\varkappa}(\hat p_\varphi)=F(\varkappa)\dfrac{\hat p_\varphi^2}{2}, \qquad \hat p_a=\pm\sqrt{\dfrac{2\Hat E-F(\varkappa)\hat p_\varphi^2}{R(\varkappa)}}.$$ Notice that the numerator of the radical expression in (\[eq:p:lim\]) has two simple roots $\hat p_\varphi = A_\pm(0,\Hat E,\varkappa):=\pm\sqrt{\dfrac{2\Hat E}{F(\varkappa)}},$ while the denominator is positive everywhere. It follows from the Inverse Function Theorem that, for sufficiently small “perturbation” $0<\varepsilon\ll1$, the numerator of the radical expression in (\[eq:p\]) also has two simple roots, denoted by $A_\pm(\varepsilon,\Hat E,\varkappa)$, that are $O(\varepsilon)$-close to the indicated roots. This implies (cf. e.g. [@ZKF §4, Proposition 2]) that, for any $0<\varepsilon\ll1$, there exists a (unique up to time shifts $t\mapsto t+t_0$) solution $$\label {eq:sol:reduc} \bar\gamma_{\varepsilon;\Hat E,\varkappa}(t) = ((\hat p_\varphi)_{\varepsilon;\Hat E,\varkappa}(t),(\hat p_a)_{\varepsilon;\Hat E,\varkappa}(t)), \qquad t\in{{\mathbb R}},$$ of the reduced system (\[eq:reduc\]) on the given level $\Hat E$ of the Hamiltonian function, on which the variable $\hat p_\varphi=(\hat p_\varphi)_{\varepsilon,\Hat E,\varkappa}(t)$ takes at least one value from the segment $[A_-(\varepsilon,\Hat E,\varkappa), A_+(\varepsilon,\Hat E,\varkappa)].$ Actually, on the solution (\[eq:sol:reduc\]), the variable $\hat p_\varphi$ takes all values from the indicated segment, i.e. $(\hat p_\varphi)_{\varepsilon,\Hat E,\varkappa}({{\mathbb R}}^1)$ coincides with this segment. This implies (cf. e.g. [@ZKF §4, предложение 3]) that the time-dependence of the variable $\hat p_\varphi=(\hat p_\varphi)_{\varepsilon,\Hat E,\varkappa}(t)$ and, hence, of the [*latitude*]{} $a=\varkappa-\varepsilon(\hat p_\varphi)_{\varepsilon,\Hat E,\varkappa}(t)$ is a periodic function, whose half-period equals the time $T(\varepsilon,\hat E,\varkappa)/2=t_+-t_-$ of motion between its adjacent minimum and maximum, where $(\hat p_\varphi)_{\varepsilon,\Hat E,\varkappa}(t_-)=A_-(\varepsilon,\Hat E,\varkappa),$ $(\hat p_\varphi)_{\varepsilon,\Hat E,\varkappa}(t_+)=A_+(\varepsilon,\Hat E,\varkappa).$ Step 2. Now, let us “lift” the solution (\[eq:sol:reduc\]) of the reduced system to the phase space $T^*Q$, i.e., consider the corresponding solution $$\gamma_{\varepsilon;\Hat E,\varkappa}(t) = (\varkappa, \hat\varphi_{\varepsilon;\Hat E,\varkappa}(t), (\hat p_a)_{\varepsilon;\Hat E,\varkappa}(t), (\hat p_\varphi)_{\varepsilon;\Hat E,\varkappa}(t)), \qquad t\in{{\mathbb R}},$$ of the initial system (\[eq:ODE:eps\]) (such a solution is unique up to shifts of the angle variable $\varphi\mapsto\varphi+\varphi_0$). Let us project the corresponding phase curve $\{\gamma_{\varepsilon;\Hat E,\varkappa}(t)\mid t\in{{\mathbb R}}\}$ to the configuration cylinder $Q$ with coordinates $(a,\hat\varphi)\in(a_1,a_2)\times S^1$. We obtain the corresponding “orbit” of the charged particle [^1]: $$\{(a,\hat\varphi) =(\varkappa-\varepsilon(\hat p_\varphi)_{\varepsilon,\Hat E,\varkappa}(t), \hat \varphi_{\varepsilon,\Hat E,\varkappa}(t)) \mid t\in{{\mathbb R}}\} \subset Q.$$ On this “orbit”, consider the points (called the [*pericentres*]{} and [*apocentres*]{} of the orbit), whose [*latitudes*]{} $a=\varkappa-\varepsilon(\hat p_\varphi)_{\varepsilon,\Hat E,\varkappa}(t)$ are the (left and right, respectively) endpoints of the corresponding segment $[\varkappa-\varepsilon A_+(\varepsilon,\Hat E,\varkappa),\ \varkappa-\varepsilon A_-(\varepsilon,\Hat E,\varkappa)].$ Step 3. As above, let us fix the numbers $\varkappa\in(a_1,a_2),$ $\Hat E\in(0,1)$ and $\hat p_\varphi\in{{\mathbb R}}$ such that $|\hat p_\varphi|^2<\dfrac{2\Hat E}{F(\varkappa)}.$ Following [@ZKF], let us compute the minimal positive period $T(\varepsilon, \hat{E}, \varkappa)=2(t_+-t_-)$ of the latitude function $a=\varkappa-\varepsilon(\hat p_\varphi)_{\varepsilon,\Hat E,\varkappa}(t)$, when $0<\varepsilon\ll1$: $$T(\varepsilon, \hat{E}, \varkappa) = 2 \int\limits_{A_-(\varepsilon,\Hat E,\varkappa)}^{A_+(\varepsilon,\Hat E,\varkappa)} \dfrac{{{\rm d}}\hat p_\varphi}{{(\dot {\hat p}_\varphi)}_{\varepsilon,\Hat E,\varkappa}} \stackrel{(\ref {eq:ODE:eps}),(\ref {eq:p})}=$$ $$= 2 \int\limits_{A_-(\varepsilon,\Hat E,\varkappa)}^{A_+(\varepsilon,\Hat E,\varkappa)} \dfrac{{{\rm d}}u}{\sqrt{R (\varkappa - \varepsilon u)} \sqrt{2\Hat E- F(\varkappa - \varepsilon u)u^2}}.$$ In particular, $T(0,\hat{E},\varkappa)=\dfrac{2\pi}{\sqrt{R(\varkappa)F(\varkappa)}}=T(\varkappa),$ cf. (\[eq:T\]). Step 4. Consider the real number $$\Phi(\varepsilon, \Hat{E}, \varkappa) :=\varphi_{\varepsilon,\Hat E,\varkappa}(t_+)-\varphi_{\varepsilon,\Hat E,\varkappa}(t_-) \stackrel{(\ref {eq:hat:varphi}),(\ref {eq:p})} =\hat\varphi_{\varepsilon,\Hat E,\varkappa}(t_+)-\hat\varphi_{\varepsilon,\Hat E,\varkappa}(t_-),$$ i.e. the difference of [*longitudes*]{} of the adjacent pericentre and apocentre of the “orbit” (cf. Step 2). Following [@ZKF], for each $\varepsilon\in{{\mathbb R}}$, $0<|\varepsilon|\ll1,$ we have $$\dfrac1{\varepsilon^2}\Phi(\varepsilon, \hat{E}, \varkappa) = 2 \int\limits_{A_-(\varepsilon,\Hat E,\varkappa)}^{A_+(\varepsilon,\Hat E,\varkappa)} \dfrac{\dot{\hat{\varphi}}_{\varepsilon,\Hat E,\varkappa} {{\rm d}}\hat p_\varphi} {(\dot{\hat p}_\varphi)_{\varepsilon,\Hat E,\varkappa}} \stackrel{(\ref {eq:ODE:eps}),(\ref {eq:p})}=$$ $$= \sqrt2 \int\limits_{A_-(\varepsilon,\Hat E,\varkappa)}^{A_+(\varepsilon,\Hat E,\varkappa)} \left( \dfrac{R'(\varkappa - \varepsilon u) \sqrt{\Hat E- F(\varkappa - \varepsilon u)\dfrac{u^2}2}} {R(\varkappa-\varepsilon u)^{3/2}} + \right.$$ $$\left. + \dfrac{F'(\varkappa - \varepsilon u)\dfrac{u^2}{2}} {\sqrt{R (\varkappa - \varepsilon u)} \sqrt{\Hat E - F(\varkappa - \varepsilon u)\dfrac{u^2}2}} \right) {{\rm d}}u = O(1).$$ Now suppose that the Bertrand condition for slow motions holds, i.e., the solution $\gamma_{\varepsilon,\Hat E,\varkappa}(t)$ from Step 2 is periodic, when $0<\varepsilon\ll1$. Therefore, its minimal positive period is a multiple of the minimal positive period $T(\varepsilon,\hat E,\varkappa)$ of the latitude function $a=\varkappa-\varepsilon(\hat p_\varphi)_{\varepsilon,\Hat E,K}(t)$, i.e. has the form $k T(\varepsilon,\hat E,\varkappa)$ for some $k\in{{\mathbb N}}.$ Therefore, the increment $k\Phi(\varepsilon, \Hat{E}, \varkappa)$ of the “longitude” $\hat\varphi_{\varepsilon,\Hat E,K}(t)$ in time $k T(\varepsilon,\hat E,\varkappa)$ is a multiple of $2\pi,$ i.e. has the form $k\Phi(\varepsilon, \hat{E}, \varkappa)=2\pi\ell$ for some $\ell\in{{\mathbb N}}.$ Here the integers $k,\ell$ depend, generally speaking, on $\varepsilon, \hat{E}, \varkappa$. Thus, the number $\dfrac{1}{2\pi}\Phi(\varepsilon, \hat{E}, \varkappa)=\frac\ell k$ is rational for any $\varepsilon, \hat{E}, \varkappa$ such that $\Hat E\in(0,1)$ is fixed and $|\varepsilon|>0$ is small enough. Since the function $\Phi(\varepsilon, \hat{E}, \varkappa)$ is continuous (due to the fact that $A_{\pm}(\varepsilon,\Hat E,\varkappa)$ are simple roots of the equation $\Hat U_{\varepsilon,\varkappa}(\hat p_\varphi)=\Hat E$ by Step 1) and takes values in a discrete set $\pi{{\mathbb Q}},$ it must be constant. Since it has order $O(\varepsilon^2)$, it tends to 0 as $\varepsilon\to0$, hence it is identically equal to 0. Since $\Phi(\varepsilon, \Hat{E}, \varkappa)\equiv0,$ we conclude that the “longitude” function $\hat\varphi_{\varepsilon,\Hat E,\varkappa}(t)$ itself (rather than just its time derivative) is $T(\varepsilon, \hat{E}, \varkappa)$-periodic, similarly to the latitude function $a=\varkappa-\varepsilon(\hat p_\varphi)_{\varepsilon,\Hat E,K}(t)$. Therefore the solution $\gamma_{\varepsilon,\Hat E,\varkappa}(t)$ is also $T(\varepsilon, \hat{E}, \varkappa)$-periodic. We obtain from Step 3 that, for any fixed $\Hat E\in(0,1)$ and $0<\varepsilon\ll1$, $$T(\varepsilon,\hat{E},\varkappa)-T(\varkappa)=T(\varepsilon,\hat{E},\varkappa)-T(0,\hat{E},\varkappa)=O(\varepsilon),$$ hence item (b) is proved. Step 5. Let us prove item (a). By Step 4, we have $\Phi(\varepsilon, \Hat{E}, \varkappa)=0$ for any $0<|\varepsilon|\ll1$, whence $\lim_{\varepsilon\to0}\dfrac{1}{\varepsilon^2}\Phi(\varepsilon, \Hat{E}, \varkappa)=0$. On the other hand, taking into account formulae from Step 4, it is not hard to compute $$\lim_{\varepsilon\to0}\dfrac{1}{\varepsilon^2}\Phi(\varepsilon, \Hat{E}, \varkappa) = 2 \int\limits_{A_-(0,\Hat E,\varkappa)}^{A_+(0,\Hat E,\varkappa)} \dfrac{\dot{\hat\varphi}_{0,\Hat E,\varkappa} {{\rm d}}\hat p_\varphi}{(\dot{\hat p}_\varphi)_{0,\Hat E,\varkappa}}$$ $$= \sqrt2 \int\limits_{A_-(0,\Hat E,\varkappa)}^{A_+(0,\Hat E,\varkappa)} \left( \dfrac{R'(\varkappa)} {R(\varkappa)^{3/2}} \sqrt{\Hat E - F(\varkappa)\dfrac{u^2}2} + \dfrac{F'(\varkappa)\dfrac{u^2}{2}} {\sqrt{R (\varkappa)} \sqrt{\Hat E - F(\varkappa)\dfrac{u^2}2}} \right) {{\rm d}}u$$ $$= \pi \hat{E} \dfrac{R'(\varkappa)F(\varkappa) + R(\varkappa)F'(\varkappa)}{(R(\varkappa)F(\varkappa))^\frac{3}{2}} = - 2\pi \Hat E \dfrac{{{\rm d}}}{{{\rm d}}\varkappa} \left( \dfrac{1}{\sqrt{R(\varkappa)F(\varkappa)}} \right).$$ Due to the latter formula and the identity $\Phi(\varepsilon, \hat{E}, \varkappa) \equiv 0$ proved for any $\Hat E\in(0,1)$, $\varkappa\in(a_1,a_2)$ and $0<\varepsilon\ll1$, we obtain that the function $\dfrac{1}{\sqrt{R(\varkappa)F(\varkappa)}}$ is constant, whence $RF=\const$ (i.e., the magnetic field $B$ is homogeneous). Step 6. Let us prove (a) in the general situation, i.e. without the assumption that the magnetic field $B$ has no zeros. Let $Z$ be the zero locus of the field $B$. By assumption, $Q\setminus Z\ne\varnothing$. We showed on Steps 1—5 that the magnetic field $B$ is homogeneous on each connected component $Q_i$ of the open set $Q\setminus Z$. This implies that $B$ has no zeros on $\overline{Q_i}$ (since the area form ${{\rm d}}\sigma$ has no zeros). On the other hand, $B=0$ on $Z$. Therefore, $\overline{Q_i}\cap Z=\varnothing$, whence $Q_i=\overline{Q_i}$, and hence $Q_i=Q$. Thus, $B$ is homogeneous on the whole configuration manifold $Q=Q_i$, as required. Derivation of constancy of scalar curvature from the Bertrand condition for slow motions ======================================================================================== Here we finish the proof of Theorem \[thm:Bert\] by proving its item (c). First, we will give a sketch of the proof of item (c) of Theorem \[thm:Bert\]. Following [@ZKF §4, proof of Proposition 4], let us expand the value $\Phi(E, K):=\Phi(1,E,K)$, which is the difference of longitudes at the adjacent pericentres of the orbit, into a power series in the small parameter $h>0$, and let the coefficients at the lower powers $h^0, h^2, h^4$ be equal to 0. Here the functions $E=E(c,h)$ and $K=K(c,h)$ are determined by the conditions $a_-(E, K)=c-h,$ $a_+(E, K)=c+h,$ where $c\in(a_1,a_2),$ $a_\pm(E,K):=K-A_\mp(1,E,K)$. Actually, the coefficients at $h^0$ and $h^2$ equal 0 due to Steps 4 and 5 (respectively) from the previous section. In more detail: taking into account (\[eq:ODE\]) and $RF\equiv \lambda^2=\const$ (by item (а) of Theorem \[thm:Bert\]), we have $$\Phi(E, K) = 2\int\limits_{a_-(E, K)}^{a_+(E, K)} \dfrac{\dot{\varphi}}{\dot{a}}{{\rm d}}a = 2 \int\limits_{a_-(E, K)}^{a_+(E, K)} \dfrac{F(a)(K-a)}{\sqrt{R(a)}\sqrt{2E-F(a)(K-a)^2}}{{\rm d}}a$$ $$= 2 \int\limits_{a_-(E, K)}^{a_+(E, K)} \dfrac{F^\frac{3}{2}(a)(K-a)}{|\lambda|\sqrt{2E-F(a)(K-a)^2}}{{\rm d}}a .$$ Therefore, after the change $a = c + ht, a_- = c-h, a_+ = c+h, {{\rm d}}a = h{{\rm d}}t$, we obtain $$-\dfrac{|\lambda|}{2F(c)}\, \Phi(E(c,h), K(c,h)) =$$ $$= \int\limits_{-1}^1 \left(\dfrac{th}{\sqrt{1-t^2}} + \dfrac{(1-2t^2)(\hat F_1+\hat F_2th)}{2\sqrt{1-t^2}}h^2 + \dfrac{3\hat F_1^3 -6\hat F_1\hat F_2 + 4\hat F_3(1 + t^2 - 2t^4)}{8\sqrt{1-t^2}}h^4 \right){{\rm d}}t$$ $$+ \bar{\bar{o}}(h^4) = \dfrac{\pi}{8} \left(3\hat F_1^3 -6\hat F_1\hat F_2 + 4\hat F_3(1 + 1/2 - 3/4)\right) h^4 + \bar{\bar{o}}(h^4) =$$ $$= \dfrac{\pi}{8} \left(3\hat F_1^3 -6\hat F_1\hat F_2 + 3\hat F_3\right) h^4 + \bar{\bar{o}}(h^4),$$ where $\hat{F_i} := \dfrac{F_i(c)}{F(c)},$ and $F_i(c)$ is the coefficient at $(a-c)^i$ in the Taylor expansion of the function $F=F(a)$ at the point $c$. Since $\Phi(E, K)\equiv0$ by Step 4 of the proof of items (a,b) of Theorem \[thm:Bert\], we conclude that the coefficient at $h^4$ must be 0, i.e. $$\label {eq:2nd} 6F'^3 - 6FF'F'' + F^2F''' \equiv 0.$$ Let us show that the condition (\[eq:2nd\]) is equivalent to the constancy of the scalar curvature. On the manifold of revolution with the Riemannian metric $g = dr^2 + f^2(r)d\varphi^2$, the scalar curvature equals $\Scal=-2\dfrac{f''(r)}{f(r)},$ cf. [@ZKF §1.2, Remark 4]. Taking into account the change $a=a(r),$ $\dfrac{da}{dr}=\lambda f(r)=\dfrac{\lambda}{\sqrt{F(a)}},$ we obtain $$\Scal=\dfrac{\lambda^2}{F^3(a)}(F''(a)F(a)-2F'(a)^2)=-\left(\dfrac{\lambda^2}{F(a)}\right)''.$$ One easily checks that the condition $\Scal'\equiv0$ is equivalent to the condition (\[eq:2nd\]). Thus, the scalar curvature $\Scal\equiv\const$, and item (c) is proved. It remains to prove that the conditions (a,b) in Theorem \[thm:Bert\] are not only necessary, but also sufficient to fulfill the Bertrand condition for slow motions. Let $B=\lambda {{\rm d}}\sigma$, where ${{\rm d}}\sigma$ is the area form, $\lambda=\lambda(a,\varphi)$ is a smooth function (not necessarily constant). Then the solutions $\gamma(t)=(a(t),\varphi(t))$ of the corresponding Lagrange system of equations with energy level $\dfrac{\varepsilon^2}2>0$ are exactly smooth parametrized curves $\gamma=\gamma(t)$ with velocity $|\dot\gamma(t)|=\varepsilon$ and covariant acceleration $\lambda(\gamma(t))\varepsilon.$ Therefore $s\mapsto\gamma(s/\varepsilon)$ is a naturally parametrized curve (with a natural parameter $s$) with geodesic curvature $\dfrac{\lambda(\gamma(s/\varepsilon))}{\varepsilon}.$ Therefore, if the conditions (a,b) of Theorem \[thm:Bert\] hold, then the orbits of all sufficiently slow motions of the charged particle will be circles and, in particular, will be closed. Hence, the Bertrand condition for slow motions will be fulfilled. Theorem \[thm:Bert\] is completely proved. The authors are thankful to A. I. Neishtadt for indicating a construction of slow-fast variables in the problem on the motion of a charged particle in a slow magnetic field by means of a guiding-centre transformation [@Nei], to A. A. Oshemkov for the idea of a possible relation of the condition (\[eq:2nd\]) with the constancy of the scalar curvature, to A. Albouy for discussing the method of common isochronicity of a family of systems for solving the Bertrand problem. [99]{} J. Bertrand, Théorème relatif au mouvement d’un point attiré vers un centre fixe // C.R. Acad. Sci. Paris [**77**]{} (1873), 849–853. G. Darboux, Étude d’une question relative au mouvement d’un point sur une surface de révolution // Bulletin de la S. M. F. [**5**]{} (1877), 100–113. A. Besse. Manifolds all of whose geodesics are closed. Berlin Heidelberg New York: Springer-Verlag, 1978. V. Perlick, Bertrand spacetimes // Class. Quantum Grav. [**9**]{} (1992), 1009–1021. O.A. Zagryadskii, E.A. Kudryavtseva, and D.A. Fedoseev, A generalization of Bertrand’s theorem to surfaces of revolution // Sbornik Math. [**203**]{}:8 (2012), 1112–1150. arXiv:1109.0745. E.A. Kudryavtseva and D.A. Fedoseev, Mechanical systems with closed orbits on manifolds of revolution // Sbornik Math. [**206**]{}:5 (2015), 718–737. E.A. Kudryavtseva and D.A. Fedoseev, On Bertrand’s manifolds with equators // Mosc. Univ. Math. Bull. [**71**]{}:1 (2016), 23–26. E.A. Kudryavtseva and D.A. Fedoseev, Superintegrable natural mechanical Bertrand systems // Itogi Nauki i Tekhn. Ser. Modern Math. Appl. Themat. Surv. [**148**]{} (2018), 37–57 (in Russian). Engl. transl.: J. Math. Sci. [**238**]{} (2019). M.A. Podlipaev. Geometric properties of natural mechanical systems on surfaces of revolution. Course work of the 3rd course. Faculty of Mathematics and Mechanics of Moscow State University, 2015. http://dfgm.math.msu.su/files/0students/2015-kr3-Podlipaev.pdf A.I. Neishtadt, Averaging, adiabatic invariants and periodic trajectories of motion in multi-dimensional magnetic fields / Manuscript, 1998, 14 pp. (in Russian). E.A. Kudryavtseva, Periodic solutions of planetary systems with satellites and the averaging method in systems with slow and fast variables // arXiv:1201.6356. E.A. Kudryavtseva\ Lomonosov Moscow State University\ [*E-mail:*]{} eakudr at mech.math.msu.su S.A. Podlipaev\ Lomonosov Moscow State University\ [*E-mail:*]{} podlipaev.sergey at gmail.com [^1]: When we use here the term “orbit”, we abuse notation. Strictly speaking, on the configuration cylinder $Q$, we introduced the coordinates $(a,\varphi)$ rather than $(a,\hat\varphi)=(a,\varphi-p_a)=(a,\varphi-\varepsilon\hat p_a)$. Thus, the true orbit of the charged particle is obtained via a projection onto the configuration cylinder $Q$ with coordinates $(a,\varphi)$, hence it is only $O(\varepsilon)$-close to the “orbit” mentioned above (and does not coincide with it, generally speaking). We remark that, for the true orbit, the pericentres, apocentres and the time of motion between adjacent pericentre and apocentre are exactly the same as for our “orbit”. This similarity between our “orbit” and the true orbit will suffice for us to prove Theorem \[thm:Bert\].
{ "pile_set_name": "ArXiv" }
--- abstract: 'Hybrid quantum gates hold great promise for quantum information processing since they preserve the advantages of different quantum systems. Here we present compact quantum circuits to deterministically implement controlled-NOT, Toffoli, and Fredkin gates between a flying photon qubit and diamond nitrogen-vacancy (NV) centers assisted by microcavities. The target qubits of these universal quantum gates are encoded on the spins of the electrons associated with the diamond NV centers and they have long coherence time for storing information, and the control qubit is encoded on the polarizations of the flying photon and can be easily manipulated. Our quantum circuits are compact, economic, and simple. Moreover, they do not require additional qubits. The complexity of our schemes for universal three-qubit gates is much reduced, compared to the synthesis with two-qubit entangling gates. These schemes have high fidelities and efficiencies, and they are feasible in experiment.' address: 'State Key Laboratory of Low-Dimensional Quantum Physics and Department of Physics, Tsinghua University, Beijing 100084, China' author: - 'Hai-Rui Wei & Gui Lu Long[^1]' title: 'Hybrid quantum gates between flying photon and diamond nitrogen-vacancy centers assisted by optical microcavities' --- Introduction ============ A quantum computer[@book] is more powerful than a classical computer in solving certain computationally demanding tasks. Quantum logic gates are the fundamental building blocks of a quantum computer, and a quantum computing task can be completed using a sequence of quantum gates as described in a quantum circuit. It is well known that any quantum computing can be decomposed into a sequence of single-qubit gates and two-qubit entangling gates[@uni], and analytical expressions[@liuy] for an arbitrary $n$-qubit unitary gate have been explicitly derived using the methods provided in Ref. [@uni]. One of the most popular universal quantum gates is the controlled-NOT (CNOT) gate. Quantum circuit received great attention over the years, in particular the CNOT gate (or the controlled phase gate)[@uni; @liuy; @3CNOT2; @Theor; @Bull0; @Bull; @longprl; @xug1; @xug2] and the hyperparallel CNOT gate[@renCNOT; @renCNOT2; @RenCNOTpra]. The theoretical lower bound of an unstructured $n$-qubit quantum computation is $(4^n - 3n - 1)/4$ CNOT gates[@Theor]. In multi-qubit systems, the fundamental three-qubit Toffoli gate[@Toffoli] or Fredkin gate[@Fredkin] form a family of universal quantum gates with the help of Hadamard operations, and they are valuable in fault-tolerant quantum circuits and some quantum algorithms. The realization of a Toffoli gate or a Fredkin gate in terms of two-qubit entangling gates is troublesome as the optimal cost is six CNOT gates[@Toffolicost] for a Toffoli gate and five two-qubit entangling gates for a Fredkin gate[@Fredkincost]. It is desirable to seek efficient schemes for directly implementing the Toffoli and Fredkin gates so as to speedup the quantum computation. A single photon is a perfect information carrier and it has a flexible controllability. However, it seems unsuitable for quantum computing as the direct interaction between individual photons is very weak. Different to photonic qubit[@Bull1; @Bull3], matter qubits, such as atoms, quantum dots (QDs), superconduction junctions, and diamond nitrogen-vacancy (NV) defect centers, are widely utilized in quantum computing because of their long-lived coherence time and their good scalability. Compared with other candidates, a diamond NV center is a particularly promising one for a qubit as it has an ultralong coherence time (1.8 ms)[@coherence1] even at the room temperature. In a diamond NV center, the electron spin can be exactly populated by the optical pumping with 532 nm light[@population], and it can be manipulated[@population; @manipulate2; @manipulate4; @manipulate5] and readout[@readout2; @readout3] by using the microwave excitation. The techniques to transfer the information from electron spins to nuclear spins were developed well[@register1; @register3; @register4]. Besides, some important tasks in quantum computation have been investigated and even been realized in experiment on diamond NV centers. For example, in 2004, Jelezko *et al.*[@CROT] carried out the experiments for implementing the hybrid controlled-ROT gate on an electron-nuclear system. In 2012, Sar *et al.*[@decoherence-protected] realized the decoherence-protected conditional rotation gates on hybrid electron-nuclear systems. In 2010, Yang *et al.*[@CCPF] proposed a conditional phase gate on three diamond NV centers. In 2013, Wei and Deng[@our] proposed some compact schemes for implementing universal gates on diamond NV centers, and Wang *et al.*[@Wangchuan] designed a quantum circuit for the photonic controlled phase gate via a diamond NV center. In 2015, Ren, Wang and Deng[@RenCNOTpra] presented the dipole induced transparency of a diamond NV center embedded in a photonic crystal cavity coupled to two waveguides, and proposed two universal hyperparallel hybrid photonic quantum logic gates, including a hybrid hyper-controlled-NOT gate and a hybrid hyper-Toffoli gate, on photon systems in both the polarization and the spatial-mode degrees of freedom, which can be used to perform more quantum operations with less resources and depress the resources consumed and the photonic dissipation. Recently, some interesting works for quantum information processing have been achieved on diamond NV centers, such as entanglement generation[@photon-NV; @nuclear-nuclear; @NV-NV2; @NV-NV3; @NV-NV5; @threemeters; @entanglement-by-measurement], quantum manipulation[@cpb1; @apl1; @light], quantum teleportation between solid-state qubits separated by three meters[@teleportation], and hyperentanglement[@ren] and entanglement[@wangchuanscpma; @shengybcsb] purification and concentration. A light-matter system[@Hu1; @Hu2; @Bonato; @atom1] coupled to a cavity provides an important platform to study quantum information processing. For example, some important schemes for the conventional parallel quantum computation[@atom1; @weioe; @HuaMPRA; @HuaSR] or the hyperparallel photonic quantum computation[@renCNOT; @renCNOT2; @RenCNOTpra] were proposed with the light-matter platform coupled to optical microcavities. By using a flying photon as a bus, schemes for universal gates on atoms[@atom2] and QDs[@weioe20014] have been proposed. Hybrid quantum gates on two or more physical systems inherit all the advantages of the different systems. In this paper, we focus on designing compact quantum circuits to implement CNOT, Toffoli, and Fredkin gates between a flying photon and solid-state diamond NV centers coupled to cavities. These quantum circuits are constructed by utilizing the input-output process of the single photon as a result of cavity quantum electrodynamics and optical spin selection rules. The schemes well work at the degeneracy of the spin 1 system and the gate’s mechanism is deterministic in principle. The target qubits are encoded on the ground states of the electrons $|m_s=\pm1\rangle$ associated with the diamond NV centers. The control qubit is encoded on the polarizations of the flying single photon. Our schemes have some advantages. First, our quantum circuits for these universal quantum gates are compact and economic. Second, they do not require additional qubits. Third, the control qubit is the flying photon which has the flexible controllability. Fourth, the target qubits are encoded on the spins of the electrons associated with NV centers which have the relatively long coherence time even at room temperature and are perfect for the storage of quantum information. Fifth, the complexity of our schemes for three-qubit quantum gates beats their synthesis procedures largely. The high fidelities and efficiencies of our schemes show that they may be feasible with current technology. Results {#sec2} ======= [**A diamond nitrogen-vacancy center confined in an optical resonant microcavity.**]{} A diamond NV center consists of a vacancy adjacent to a substitutional nitrogen atom (typically $^{14}$N). In a diamond NV center, both the nuclear spins (typically $^{13}$C with $I$=1/2 or $^{14}$N with $I$=1) and the electron spins are promising for quantum information processing. The ground states of the electron, $|0\rangle\equiv|m_s=0\rangle$ and the two-fold degenerate states $|\pm\rangle\equiv|m_s=\pm1\rangle$, is split by $D\approx 2.87$ GHz in a zero external field due to the spin-spin interaction[@split]. The six excited states[@photon-NV] $|A_1\rangle=(|E_-\rangle|+\rangle-|E_+\rangle|-\rangle)/\sqrt{2}$, $|A_2\rangle=(|E_-\rangle|+\rangle+|E_+\rangle|-\rangle)/\sqrt{2}$, $|E_x\rangle=|X\rangle|0\rangle$, $|E_y\rangle=|Y\rangle|0\rangle$, $|E_1\rangle=(|E_-\rangle|-\rangle-|E_+\rangle|+\rangle)/\sqrt{2}$, and $|E_2\rangle=(|E_-\rangle|-\rangle+|E_+\rangle|+\rangle)/\sqrt{2}$ are dominated by the NV center’s C$_{3v}$ symmetry and the spin-spin, spin-orbit interactions without external strain and electric or magnetic fields. Here $|E_\pm\rangle$, $|X\rangle$, and $|Y\rangle$ are the orbital states of an NV center. The spin-orbit interaction (5.5 GHz)[@split1; @split2] splits the excited states into three two-fold degeneracy pairs ($A_1$, $A_2$) (to be shifted up), ($E_x$, $E_y$), and ($E_1$, $E_2$) (to be shifted down). The spin-spin interaction (1.42 GHz) shifts up states ($A_1$, $A_2$, $E_1$, and $E_2$) by 1.42/3 GHz and shifts down states ($E_x$, $E_y$) by $2*1.42/3$ GHz[@split1; @split2]. Besides, it splits $A_2$ and $A_1$ by $\pm$ 1.55 GHz[@split1; @split2]. The local non-axial high strain (10 GHz, larger than the spin-orbit splitting in the presence of the zero field) splits the excited states into two branches, ($A_2$, $A_1$, and $E_x$) and ($E_y$, $E_1$, and $E_2$). The state $|A_2\rangle$ is robust against the relatively small strain and magnetic fields with the stable symmetry properties, preserving the polarization properties of its optical transitions. The frequency of the spin-selective optical resonant transition can be tuned via an application of a controlled external electric field[@split2; @electricfield; @electricfield1; @electricfield2; @electricfield3]. In 2011, Bassett *et al.*[@electricfield3] experimentally demonstrated an exceeding 10 GHz optical transition frequency. The transitions between the ground states are in the microwave frequency regime, and the transitions between the ground states and the excited states are in the optical regime. With microwave and laser, one can prepare, store, and read out the states of the solid-state electron spins[@PRL]. Here we encode the qubit on the sublevels $|\pm\rangle$, and take $|A_2\rangle$ as an auxiliary state. $|A_2\rangle$ decays into $|\pm\rangle$ with the right-circularly-polarized ($R$) and left-circularly-polarized ($L$) photons \[see Fig. 1b\], respectively, owning to total angular momentum conservation. They take place with the equal probability. ![ (a) Schematic diagram of an NV-cavity system. (b) The energy-level diagram of an NV-cavity system. The triple ground states $|\pm\rangle\equiv|m_s=\pm1\rangle$ are chosen to act as the two levels for an electron-spin qubit. The excited state $|A_2\rangle=(|E_-\rangle|+1\rangle+|E_+\rangle|-1\rangle)/\sqrt{2}$ is an auxiliary state. $|\pm\rangle\rightarrow|A_2\rangle$ are driven by the right- and left- circularly-polarized photons, respectively. []{data-label="level"}](fig1_level.eps){width="8.2"} In 2011, Chen *et al.*[@NV-NV3] discussed a composite unit, that is, a diamond NV center confined inside a single-sided resonator \[see Fig. 1a\]. Combing the Heisenberg equations of motion[@QObook] $$\begin{aligned} \label{eq1} %\begin{cases} \begin{split} &\frac{d\hat{a}}{dt} = -\left[i(\omega_c-\omega_p)+\frac{\kappa}{2}\right]\hat{a}(t)-\emph{\text{g}}\sigma_{-}(t) - \sqrt{\kappa}\hat{a}_{in}, \\ &\frac{d\sigma_-}{dt} = -\left[i(\omega_{0}-\omega_p)+\frac{\gamma}{2}\right]\sigma_{-}(t)-\emph{\text{g}}\sigma_z(t)\hat{a}(t) + \sqrt{\gamma}\sigma_z(t)\hat{b}_{in}(t), %\nonumber\\ \end{split} %\end{cases}\end{aligned}$$ and the standard input-output relation for the cavity $$\begin{aligned} \label{eq2} \hat{a}_{out} = \hat{a}_{in}+ \sqrt{\kappa}\hat{a}(t),\end{aligned}$$ the explicit expression of the reflection coefficient for the NV-cavity unit in the weak excitation limit $\langle\sigma_z\rangle=-1$ can be written as[@NV-NV3; @Hersenberg] $$\begin{aligned} \label{eq3} r(\omega_p)=\frac{[i(\omega_{c}-\omega_p)-\frac{\kappa}{2}][i(\omega_{0}-\omega_p)+\frac{\gamma}{2}]+\emph{\text{g}}^2} {[i(\omega_{c}-\omega_p)+\frac{\kappa}{2}][i(\omega_{0}-\omega_p)+\frac{\gamma}{2}]+\emph{\text{g}}^2}.\end{aligned}$$ Here $\hat{a}$ and $\sigma_-$ are the annihilation operator of the cavity mode and the transition operator of the diamond NV center with the frequencies $\omega_c$ and $\omega_0$, respectively. $\omega_p$ is the frequency of the input single photon. $\sigma_z(t)$ presents the inversion operator of the NV center. $\gamma$ and $\kappa$ are the NV decay rate and the cavity damping rate, respectively. $\emph{\text{g}}$ is the coupling strength between an NV center and a cavity. The vacuum input field $b_{in}(t)$ has the commutation relation $[\hat{b}_{in}(t),\hat{b}_{in}^\dag(t')]=\delta(t-t')$. When the diamond NV center confined inside a resonant cavity interacts with a resonant single photon, i.e., $\omega_0=\omega_c=\omega_p$, the reflection coefficients for the hot cavity ($\emph{\text{g}}\neq0$) and the cold cavity ($\emph{\text{g}}=0$) can be written as $$\begin{aligned} \label{eq4} r=\frac{-\kappa\gamma+4\emph{\text{g}}^2}{\kappa\gamma+4\emph{\text{g}}^2},\qquad\qquad\qquad r_0=-1.\end{aligned}$$ That is, the change of the incident photon can be summarized as follows[@NV-NV3]: $$\begin{aligned} \label{eq5} \begin{split} |R\rangle|+\rangle \;\xrightarrow{\text{NV}}\;|r||R\rangle|+\rangle,\;\;\;\; |L\rangle|-\rangle\;\xrightarrow{\text{NV}}\; |r||L\rangle|-\rangle,\;\;\; |R\rangle|-\rangle\;\xrightarrow{\text{NV}}\;-|R\rangle|-\rangle,\;\;\;\; |L\rangle|+\rangle\;\xrightarrow{\text{NV}}\;-|L\rangle|+\rangle. \end{split}\end{aligned}$$ When $\emph{\text{g}}\geq5\sqrt{\gamma\kappa}$ and $\omega_0=\omega_c=\omega_p$, $r(\omega_p)\simeq1$ and $r_0(\omega_p) =-1$[@NV-NV3], and Eq. (\[eq5\]) becomes $$\begin{aligned} \label{eq6} \begin{split} |R\rangle|+\rangle \;\xrightarrow{\text{NV}}\;|R\rangle|+\rangle, \;\;\;\; |L\rangle|-\rangle \;\xrightarrow{\text{NV}}\;|L\rangle|-\rangle,\;\;\;\; |R\rangle|-\rangle\;\xrightarrow{\text{NV}}\;&-|R\rangle|-\rangle,\;\;\; |L\rangle|+\rangle\;\xrightarrow{\text{NV}}\;-|L\rangle|+\rangle. \end{split}\end{aligned}$$ [**Compact quantum circuit for implementing a hybrid CNOT gate.**]{} The framework of our proposal for implementing a CNOT gate is shown in Fig. 2. It performs a not operation on the diamond NV center when the flying single photon is in state $|L\rangle$. Let us describe its principle in detail as follows. ![ Compact quantum circuit for implementing the CNOT gate on a hybrid photon-NV system with the flying single photon polarization as the control qubit and the electron spin in the diamond NV center as the target qubit. PBS$_i$ ($i=1,2$) represents a polarizing beam splitter (PBS) which transmits the $R$-polarized photon and reflects the $L$-polarized photon, respectively. HWP represents a half-wave plate oriented at 0$^\circ$ and it is used to complete the unitary transformation $\sigma_z=|R\rangle\langle R|-|L\rangle\langle L|$ on a photon. []{data-label="CNOT"}](fig2_CNOT.eps){width="8"} Suppose that the input state of the composite system composed of the flying single photon and the diamond NV center is $$\begin{aligned} \label{eq7} |\psi_{\text{in}}\rangle=(\cos\alpha|R\rangle_c+\sin\alpha|L\rangle_c)\otimes(\cos\beta|+\rangle_t+\sin\beta|-\rangle_t).\end{aligned}$$ Here subscripts $c$ and $t$ represent the control qubit (the flying single photon) and the target qubit (the diamond NV center), respectively. First, the input single photon is split into two wave-packets by a polarizing beam splitter (PBS), say PBS$_1$. Second, the $R$-polarized component does not interact with the diamond NV center, whereas the $L$-polarized component interacts with the diamond NV center and then arrives at PBS$_2$ simultaneously with the $R$-polarized component. Third, before and after the photon interacts with the diamond NV center, a Hadamard operation $H_e$ is performed on the diamond NV center, respectively. Here $H_e$ completes the following transformations $$\begin{aligned} \label{eq8} |+\rangle \xrightarrow{H_e}|\rightarrow\rangle\equiv\frac{1}{\sqrt{2}}(|+\rangle+|-\rangle), \;\;\;\;\;\;\;\;\;\;\;\; |-\rangle\xrightarrow{H_e}|\leftarrow\rangle\equiv\frac{1}{\sqrt{2}}(|+\rangle-|-\rangle).\end{aligned}$$ Finally, a single-qubit operation $\sigma_z=|R\rangle\langle R|-|L\rangle\langle L|$ is performed on the output photon with a half-wave plate HWP oriented at $0^\circ$. With these operations, the state of the composite system evolves as follows: $$\begin{aligned} \label{eq9} \begin{split} &|\psi_{\text{in}}\rangle\xrightarrow{H_e,\;\text{PBS}_1}(\cos\alpha|R\rangle_c+\sin\alpha|L\rangle_c)(\cos\beta|\rightarrow\rangle_{t}+\sin\beta|\leftarrow\rangle_{t}) \\& \;\;\;\;\;\;\;\;\;\;\xrightarrow{\text{NV}}\cos\alpha|R\rangle_c(\cos\beta|\rightarrow\rangle_{t}+\sin\beta|\leftarrow\rangle_{t}) -\sin\alpha|L\rangle_c(\cos\beta|\leftarrow\rangle_{t}+\sin\beta|\rightarrow\rangle_{t}) \\& \;\;\;\;\;\;\;\;\;\;\xrightarrow{H_e,\;\text{PBS}_2}\cos\alpha|R\rangle_c(\cos\beta|+\rangle_{t}+\sin\beta|-\rangle_{t}) -\sin\alpha|L\rangle_c(\cos\beta|-\rangle_{t}+\sin\beta|+\rangle_{t}) \\& \;\;\;\;\;\;\;\;\;\;\xrightarrow{\text{HWP}}|\psi_{\text{out}}\rangle=\cos\alpha|R\rangle_c(\cos\beta|+\rangle_{t}+\sin\beta|-\rangle_{t})+\sin\alpha|L\rangle_c(\cos\beta|-\rangle_{t}+\sin\beta|+\rangle_{t}). \end{split}\end{aligned}$$ The quantum circuit shown in Fig. 2 completes the transformation $|\psi_{\text{in}}\rangle\xrightarrow{\text{CNOT}}|\psi_{\text{out}}\rangle$. That is, it implements a CNOT gate on a hybrid photon-NV system. If the flying single photon is in state $|L\rangle$, the spins of the electron associated with the diamond NV center are flipped; otherwise, the spins of the electron remain unchanged. [**Compact quantum circuit for implementing a Toffoli gate on a hybrid system.**]{} The principle of our hybrid Toffoli gate is shown in Fig. 3. This gate performs a CNOT operation on the two diamond NV centers, $\text{NV}_{c_2}$ and $\text{NV}_{t}$, when the flying single photon $c_1$ is in state $|L\rangle$. Suppose that the system composed of $c_1$, NV$_{c_2}$, and NV$_{t}$ is prepared in the state $$\begin{aligned} \label{eq10} |\Phi_{\text{in}}\rangle=(\cos\alpha|R\rangle_{c_1}+\sin\alpha|L\rangle_{c_1})\otimes(\cos\beta|+\rangle_{c_2}+\sin\beta|-\rangle_{c_2})\otimes(\cos\delta|+\rangle_{t}+\sin\delta|-\rangle_{t}).\label{Toffoliin}\end{aligned}$$ Our hybrid Toffoli gate works with the following steps. ![ Compact quantum circuit for implementing a Toffoli gate which performs a CNOT operation on the two diamond NV centers if the flying single photon is in state $\vert L\rangle$. []{data-label="Toffoli1"}](fig3_Toffoli.eps){width="11.50"} First, the $R$-polarized component of the input single photon $c_1$ is transmitted to spatial mode 1 by PBS$_1$ and then arrives at PBS$_8$ directly, whereas the $L$-polarized component is reflected to spatial mode 2 for interacting with the diamond NV centers. When the photon emits from spatial mode 2, it passes through the block composed of PBS$_2$, NV$_{c_2}$, and PBS$_3$, and a Hadamard operation $H_p$ is performed on it with a half-wave plate (HWP) oriented at 22.5$^\circ$ before and after it passes through the block, respectively. We can obtain the following transformation induced by the above operations ($\text{PBS}_1\rightarrow\text{HWP}_1\rightarrow \text{PBS}_2\rightarrow\text{NV}_{c_2}\rightarrow \text{PBS}_3\rightarrow\text{HWP}_2$) $$\begin{aligned} \label{eq11} \begin{split} |\Phi_{\text{in}}\rangle\rightarrow|\Phi\rangle_{1}=&\big[\cos\alpha|R_1\rangle_{c_1}(\cos\beta|+\rangle_{c_2}+\sin\beta|-\rangle_{c_2}) +\sin\alpha(\cos\beta|L_8\rangle_{c_1}|+\rangle_{c_2}\\&-\sin\beta|R_8\rangle_{c_1}|-\rangle_{c_2})\big](\cos\delta|+\rangle_{t}+\sin\delta|-\rangle_{t}). \end{split}\end{aligned}$$ Here and below, we use $|R_i\rangle$ ($|L_i\rangle$) denotes the $R$- ($L$-) polarized photon emitted from spatial mode $i$ ($i=1,2,\cdots, 19$). Second, the photon passes through the block composed of PBS$_4$, NV$_t$, and PBS$_5$, and before and after the photon interacts with NV$_{t}$, an $H_e$ is performed on NV$_{t}$, respectively. These operations ($H_e\rightarrow\text{PBS}_4\rightarrow\text{NV}_t\rightarrow\text{PBS}_5\rightarrow H_e$) transform $|\Phi\rangle_{1}$ into $|\Phi\rangle_{2}$. Here $$\begin{aligned} \label{eq12} \begin{split} |\Phi\rangle_{2} =& \big[\cos\alpha|R_{1}\rangle_{c_1}(\cos\beta|+\rangle_{c_2}+\sin\beta|-\rangle_{c_2}) +\sin\alpha\cos\beta|L_{12}\rangle_{c_1}|+\rangle_{c_2}\big]\\&\times(\cos\delta|+\rangle_t+\sin\delta|-\rangle_t) -\sin\alpha\sin\beta|R_{12}\rangle_{c_1}|-\rangle_{c_2}(\cos\delta|-\rangle_t+\sin\delta|+\rangle_t). \end{split}\end{aligned}$$ Third, the photon emitting from spatial mode 12 passes through the block composed of PBS$_6$, NV$_{c_2}$, and PBS$_7$ . Before and after the photon passes through the block, an $H_p$ is performed on it with HWP$_3$ and HWP$_4$, respectively. After the wave-packet emitting from spatial mode 18 arrives at PBS$_8$ simultaneously with the wave-packet emitting from spatial mode 1, the state of the system becomes $$\begin{aligned} \label{eq13} \begin{split} |\Phi_{\text{out}}\rangle=&\big[\cos\alpha|R_{19}\rangle_{c_1}(\cos\beta|+\rangle_{c_2}+\sin\beta|-\rangle_{c_2}) +\sin\alpha\cos\beta|L_{19}\rangle_{c_1}|+\rangle_{c_2}\big]\\&\times(\cos\delta|+\rangle_{t}+\sin\delta|-\rangle_{t}) +\sin\alpha\sin\beta|L_{19}\rangle_{c_1}|-\rangle_{c_2}(\cos\delta|-\rangle_{t}+\sin\delta|+\rangle_{t}).\label{Toffoliout} \end{split}\end{aligned}$$ From Eqs. (\[Toffoliin\])–(\[Toffoliout\]), one can see that the quantum circuit in Fig. 3 completes the transformation $|\Phi_{\text{in}}\rangle\xrightarrow{\text{controlled-CNOT}}|\Phi_{\text{out}}\rangle$. That is, it implements a Toffoli gate (it is also named a controlled-CNOT gate) which performs a CNOT operation on the two diamond NV centers when the control photon is in state $|L\rangle$; otherwise, the states of the two NV centers keep unchanged. [**Quantum circuit for implementing a deterministic Fredkin gate on a hybrid system.**]{} Our Fredkin gate is used to exchange the states of the two target diamond-NV-center-spin qubits, $\text{NV}_{t_1}$ and $\text{NV}_{t_2}$, when the flying single photon $c$ is in state $|L\rangle$; otherwise, the states of the two target qubits remain unchanged. The quantum circuit for implementing our Fredkin gate is shown in Fig. 4 and its principle can be explained as follows. ![ Quantum circuit for implementing a hybrid Fredkin gate with a flying single photon as the control qubit and the two diamond NV centers as the target qubits. []{data-label="Fredkin"}](fig4_fredkin.eps){width="11"} Let us consider an input state of the three-qubit hybrid system composed of the control photon $c$ and the two target diamond NV centers $\text{NV}_{t_1}$ and $\text{NV}_{t_2}$, $$\begin{aligned} \label{eq14} \begin{split} |\Xi_{\text{in}}\rangle=(\cos\alpha|R\rangle_{c}+\sin\alpha|L\rangle_{c})\otimes(\cos\beta|+\rangle_{t_1}+\sin\beta|-\rangle_{t_1})\otimes(\cos\delta|+\rangle_{t_2}+\sin\delta|-\rangle_{t_2}). \end{split}\end{aligned}$$ When the injecting control photon $c$ arrives at PBS$_1$, the state of the hybrid system is transformed from $|\Xi_{\text{in}}\rangle$ to $|\Xi_1\rangle$. Here $$\begin{aligned} \label{eq15} \begin{split} |\Xi_1\rangle =&(\cos\alpha|R_1\rangle_{c}+\sin\alpha|L_2\rangle_{c})(\cos\beta|+\rangle_{t_1}+\sin\beta|-\rangle_{t_1})(\cos\delta|+\rangle_{t_2}+\sin\delta|-\rangle_{t_2}). \end{split}\end{aligned}$$ The wave-packet emitting from spatial mode 1 arrives at PBS$_4$ directly and the optical switch $S_1$ leads the wave-packet emitting from spatial mode 2 to spatial mode 3. After an $H_p$ is performed on the photon with HWP$_1$, it first passes through the block composed of PBS$_2$, NV$_{t_1}$, NV$_{t_2}$, and PBS$_3$, and then arrives at $S_2$. $S_2$ leads the photon to spatial mode 10, followed with $S_1$ which leads the photon to spatial mode 3 for passing through HWP$_1$. These operations ($S_1\rightarrow\text{HWP}_1\rightarrow\text{PBS}_2\rightarrow\text{NV}_{t_1}\rightarrow\text{NV}_{t_2}\rightarrow\text{PBS}_3\rightarrow S_2\rightarrow S_1\rightarrow\text{HWP}_1$) transform the state of the hybrid system into $$\begin{aligned} \label{eq16} \begin{split} |\Xi_2\rangle =&\cos\alpha|R_1\rangle_{c}(\cos\beta|+\rangle_{t_1}+\sin\beta|-\rangle_{t_1})(\cos\delta|+\rangle_{t_2}+\sin\delta|-\rangle_{t_2})\\& +\sin\alpha(\cos\beta\cos\delta|L_4\rangle_{c}|+\rangle_{t_1}|+\rangle_{t_2} -\cos\beta\sin\delta|R_4\rangle_{c}|+\rangle_{t_1}|-\rangle_{t_2}\\& -\sin\beta\cos\delta|R_4\rangle_{c}|-\rangle_{t_1}|+\rangle_{t_2} +\sin\beta\sin\delta|L_4\rangle_{c}|-\rangle_{t_1}|-\rangle_{t_2}). \end{split}\end{aligned}$$ Before and after the second round, an $H_e$ is performed on each of NV$_{t_1}$ and NV$_{t_2}$. These operations ($H_{e_2},H_{e_3}\rightarrow\text{PBS}_2\rightarrow\text{NV}_{t_1}\rightarrow\text{NV}_{t_2}\rightarrow\text{PBS}_3\rightarrow H_{e_2},H_{e_3}\rightarrow S_2\rightarrow S_1$) transform $|\Xi_2\rangle$ into $$\begin{aligned} \label{eq17} \begin{split} |\Xi_3\rangle =& \cos\alpha|R_1\rangle_{c}(\cos\beta|+\rangle_{t_1}+\sin\beta|-\rangle_{t_1})(\cos\delta|+\rangle_{t_2}+\sin\delta|-\rangle_{t_2})\\& +\sin\alpha (\cos\beta\cos\delta|L_3\rangle_{c}|+\rangle_{t_1}|+\rangle_{t_2} -\cos\beta\sin\delta|R_3\rangle_{c}|-\rangle_{t_1}|+\rangle_{t_2}\\& -\sin\beta\cos\delta|R_3\rangle_{c}|+\rangle_{t_1}|-\rangle_{t_2} +\sin\beta\sin\delta|L_3\rangle_{c}|-\rangle_{t_1}|-\rangle_{t_2}). \end{split}\end{aligned}$$ Next, the photon passes through HWP$_1$ and the block composed of PBS$_2$, NV$_{t_1}$, NV$_{t_2}$, and PBS$_2$ in succession, and then $S_2$ leads it to spatial mode 11, followed with an $H_p$ (i.e., let it passes through HWP$_2$). Finally, the wave-packet emitting from spatial mode 12 arrives at PBS$_4$ simultaneously with the wave-packet emitting from spatial mode 1. That is, these operations ($\text{HWP}_{1}\rightarrow\text{PBS}_2\rightarrow\text{NV}_{t_1}\rightarrow\text{NV}_{t_2}\rightarrow\text{PBS}_3\rightarrow S_{2}\rightarrow \text{HWP}_2\rightarrow\text{PBS}_4$) transform $|\Xi_3\rangle$ into $$\begin{aligned} \label{eq18} \begin{split} |\Xi_{\text{out}}\rangle =& \cos\alpha|R_{13}\rangle_{c}(\cos\beta|+\rangle_{t_1}+\sin\beta|-\rangle_{t_1})(\cos\delta|+\rangle_{t_2}+\sin\delta|-\rangle_{t_2})\\& +\sin\alpha|L_{13}\rangle_{c}(\cos\beta|+\rangle_{t_2}+\sin\beta|-\rangle_{t_2})(\cos\delta|+\rangle_{t_1}+\sin\delta|-\rangle_{t_1}). \end{split}\end{aligned}$$ Putting all the pieces together, one can see that the quantum circuit shown in Fig. 4 completes the transformation $|\Xi_{\text{in}}\rangle\xrightarrow{\text{Fredkin}}|\Xi_{\text{out}}\rangle$. That is, the quantum circuit shown in Fig. 4 implements a Fredkin gate which exchanges the spins of the two electrons associated with the diamond NV centers NV$_{t_1}$ and NV$_{t_2}$ when the flying single photon is in state $|L\rangle$; otherwise, the states of the two target qubits remain unchanged. Discussion {#sec3} ========== By far, several groups have experimentally demonstrated the coupling between a diamond NV center and a microcavity, such as microspheres[@microsphere1; @microsphere2; @microsphere3; @microsphere4], microdisks[@microdisk], photonic crystals[@crystal1; @crystal2; @crystal3], microtoroidal resonators[@toridal1; @toridal2], and fiber-based microcavity[@fiber-based]. It is a challenge to achieve the strong coupling between the NV and the cavity in experiments with current technology. Fortunately, the strong coupling between NV centers in diamond nanocrystals and a whispering gallery mode (WGM) in a silica microsphere has been achieved[@microsphere1]. Larsson *et al.*[@microsphere2] showed that it is possible to achieve the strong coupling between NV centers in a diamond nanopillar coupled to a WGM in a silica microsphere. In 2013, Teissier *et al.*[@mechanical] realized an exceeding 10 MHz coupling strength between an NV center and a diamond mechanical oscillator. In 2006, Park *et al.*[@microsphere1] observed the strong coupling ($\emph{\text{g}}/2\pi=55$ MHz, $\gamma/2\pi=25$ MHz, $\kappa/2\pi=50$ MHz) in a diamond NV center coupled to a WGM in a silica microsphere. Barclay *et al.*[@Barclay] showed that the strong coupling with the parameters $[\emph{\text{g}},\kappa,\gamma_{\rm{tot}}]/2\pi=[2.25,0.16,0.013]$ GHz is possible in an NV nanocavity. In 2009, Barclay *et al.*[@microdisk] showed that the parameters $[\emph{\text{g}},\kappa,\gamma,\gamma_{\rm{ZPL}}]/2\pi=[0.30,26,0.013,0.0004]$ GHz can be achieved in experiment for coupling the NV centers in single crystal diamond to an chip-based microcavity. Here $\gamma_{\rm{ZPL}}$ is the spontaneous emission rate of a diamond NV center into the zero phonon line (ZPL). For NV-microtoridal resonators, $|r(\omega_p)|\sim1$ can be achieved when $\emph{\text{g}}=2\pi\times500$ MHz with $\kappa=2\pi\times10$ GHz or $\kappa=2\pi\times1$ GHz[@NV-NV3]. Our schemes work for the degenerate cavity modes, and it can be achieved by employing microtoroidal resonators[@toridal1; @toridal2; @degenerate2; @Lei], H1 photonic crystals[@unpolarized-photon1; @unpolarized-photon2], micropillars[@unpolarized-pillar1; @unpolarized-pillar2; @unpolarized-pillar3], or fiber-based[@fiber-based] cavities. Our schemes are deterministic in principle. Our schemes have high fidelities and efficiencies if the photon loss caused by the linear optics are not taken into account. Certainly, we should take the photon loss into account in the practical applications[@Bull1] as there are the cavity absorption and scattering, and the absorption from linear optical elements (such as the fibers, PBS, and HWP). Different to the protocol for generating entanglement between two NV centers[@threemeters], our gates cannot be heralded by the destructive detection of a single photon. Our schemes can be inferred by the successful instances in postselection in practical applications of our gates. For example, when our hybrid gate is used for quantum information transfer, the successful transfer of the information from the NV electron spin to the single photon polarization indicates the success of our CNOT gate. In principle, the photon loss can be reduced by improving experiment techniques and fabrication processing. The ZPL emission of an NV center is only 3%-4% of the total emission. In 2011, Barclay *et al.*[@PRX] enhanced the ZPL emission of an NV center in a WGM nanocavity from $\sim$3% to $\sim$16%. Subsequently, they[@toridal1] enhanced the ZPL emission of an NV center coupled to a microresonator from 3/100 to 36/133. In 2012, Faraon *et al.*[@Faraon] enhanced the ZPL emission by a factor of $\sim$70 in photonic crystal cavities. Fluctuations in the frequency of the optical transition of NV centers, due to the fluctuation in the charge environment, is a hurdle for our schemes. This spectral diffusion in the nanocavity devices results in an overall line width which can be much larger than the NV transition line width (13-16 MHz). Therefore, as that done by Delft’s group[@threemeters], we should first check the transition frequency of the NV centers before our schemes. Spectral diffusion can be reduced by active stabilization technique, preselection of the transition frequency technique, or combination of high temperature annealing and subsequent surface treatment technique[@Diffusion1; @Diffusion2; @Diffusion3]. The optical transition frequencies of the two NV centers in our schemes for Toffoli and Fredkin gates can be tuned into resonance with each other by applying an external electric field[@electricfield]. Our schemes work not only for the two-fold sublevels encoded for the electron-spin qubits but also for the non-degenerate spin sublevels lifted by a small external magnetic field. Dréau *et al.*[@anticrossing] demonstrated that the excited states occur sublevels anticrossing when $B\approx510$ G and the one for the ground states when $B\approx1020$ G. The state $A_2$ is robust against a relatively small magnetic field. For the non-degenerate one, if only the $R$-polarized photon matches the resonance transition, our schemes can implement the CNOT, Toffoli, and Fredkin gates only with a little modification on the quantum circuit in Fig. 2. Compared with the parity-measurement approach in [@PCG1; @PCG2] and the one based on control path and merging gates[@Kerr1], the auxiliary qubits are not required in our schemes, and the number of the nonlinear interactions required for our CNOT gate is fewer than that in [@PCG1; @PCG2; @Kerr1]. The complexity of our Toffoli and Fredkin gates beat their synthesis procedures in terms of two-qubit entangling gates largely as the well known cost of the Toffoli and Fredkin gates[@Toffolicost; @Kerr1; @Fredkincost] are six CNOT gates and five two-qubit entangling gates, respectively. In summary, we have presented compact quantum circuits for the hybrid universal quantum gates assisted by the input-output process of a single photon. Our CNOT, Toffoli, and Fredkin gates work with the single-photon polarizations as the control qubits and the electron spins associated with the diamond NV centers as the target qubits. Our schemes take the advantages of the theoretical and experimental progress in the fast electron-spin manipulation, the long-lived electron-spin coherence time, and the flexible controllability of the single photon. All our schemes are compact, economic, and simple. They have high fidelities and efficiencies with current technology. Methods {#sec4} ======= [**Average fidelities and efficiencies of the gates.**]{} We use the fidelity and the efficiency to characterize the performance of our universal quantum gates. In order to characterize the construction of these gates, we specify the evolutions of the hybrid systems from the initial sates $|\psi_{\text{in}}\rangle$ to the output states $|\psi_{\text{out}}\rangle$ in the ideal case. The fidelity of a quantum gate is defined as $F=|\langle\psi_{\text{out}}|\psi_{\text{out}}'\rangle|^2$, and it is the probability that the normalized output state of the whole system in the ideal case $|\psi_{\text{out}}\rangle$ overlaps with the realistic state $|\psi_{\text{out}}'\rangle$. Taking the CNOT gate as an example, in the ideal case (i.e., $r(\omega_p)\simeq1$ and $r_0(\omega_p) =-1$), the normalized output state of our scheme is given by Eq. (\[eq9\]), that is, $$\begin{aligned} \label{eq19} |\psi_{\text{out}}\rangle_{CT}\;=\;\cos\alpha|R\rangle_{c}(\cos\beta|+\rangle_{t}+\sin\beta|-\rangle_{t}) +\sin\alpha|L\rangle_{c}(\cos\beta|-\rangle_{t}+\sin\beta|+\rangle_{t}).\end{aligned}$$ By substituting Eq. (\[eq5\]) for Eq. (\[eq6\]) and combing the evolutions of the state for the CNOT gate, the non-normalized output state in the realistic case becomes $$\begin{aligned} \label{eq20} \begin{split} |\psi_{\text{out}}'\rangle_{CT} &= \cos\alpha|R\rangle_{c}(\cos\beta|+\rangle_{t}+\sin\beta|-\rangle_{t}) +\frac{\sin\alpha}{2}|L\rangle_{c}\big[\cos\beta(|r|+1)-\sin\beta(|r|-1)\big]|-\rangle_{t}\\&\;\;\;\;\; +\frac{\sin\alpha}{2}|L\rangle_{c}\big[\sin\beta(|r|+1)-\cos\beta(|r|-1)\big]|+\rangle_{t}. \end{split}\end{aligned}$$ That is, the average fidelity of our CNOT gate can be expressed as $$\begin{aligned} \label{eq21} \overline{F}_{CT}=\frac{1}{4\pi^2}\int_0^{2\pi}d\alpha\int_0^{2\pi}d\beta|\langle\psi_{\text{out}}|\psi_{\text{out}}'\rangle|^2.\end{aligned}$$ Using the same arguments for the CNOT gate, one can obtain the average fidelities of the Toffoli gate $\overline{F}_T$ and the Fredkin gate $\overline{F}_F$, shown in Fig. 5a. ![The average fidelities ($\overline{F}$) and the average efficiencies ($\overline{\eta}$) of our universal quantum gates on photon-NV hybrid systems vs $\emph{\text{g}}/\sqrt{\kappa\gamma}$. Here the red solid line, the green dashed line, and the blue dash-dotted line correspond to those of our CNOT, Toffoli, and Fredkin gates, respectively. $\emph{\text{g}}/\sqrt{\kappa\gamma}\geq 0.5$. []{data-label="Fidelity"}](fig5_fide_effi.eps){width="6.2"} Since the flying single photon may be lost during the operation for a gate, we can use $\eta=n_{\text{output}}/n_{\text{input}}$ to characterize the efficiency of a gate. Here $n_{\text{input}}$ and $n_{\text{output}}$ are the numbers of the input photons and the output photons, respectively. Combing the spin-selection rules in the realistic case described by Eq. (\[eq5\]) and the evolutions of the system from the input states to the output states, the average efficiencies of our gates, averaged over $\alpha,\beta,\delta\in[0,2\pi]$, can be obtained as follows: $$\begin{aligned} \label{eq22} \overline{\eta}_{CT}=\frac{3 + |r|^2}{4},\end{aligned}$$ $$\begin{aligned} \label{eq23} \overline{\eta}_{T}=\frac{(3 + |r|^2) (27 + 2 |r| + 4 |r| ^2 - 2 |r| ^3 + |r|^4)}{128},\end{aligned}$$ $$\begin{aligned} \label{eq24} \begin{split} \overline{\eta}_{F}=&\frac{1361 - 156 |r| + 286 |r| ^2 + 28 |r| ^3 + 239 |r| ^4 + 152 |r| ^5+148 |r| ^6 - 24| r| ^7 - | r| ^8 + 4 |r| ^9 + 14 |r| ^{10} - 4 |r| ^{11} + |r| ^{12}}{2048}. \end{split}\end{aligned}$$ The average efficiencies of our gates vary with $\emph{\text{g}}/\sqrt{\kappa\gamma}$, shown in Fig. 5b. [**The feasibility of the gates.**]{} The fidelities of our gates can be reduced by the few percent by the experimental operation imperfection, such as electronic spin preparation with a low limit fidelity of $99.7\pm 0.1\%$ to $m_s=0$ and $99.2\pm 0.1\%$ to $m_s=\pm 1$[@readout3]. Bernien *et al.*[@threemeters] showed that the fidelity of their setup can be reduced by the microwave pulse errors ($\sim$3.5%), off-resonant excitation errors ($\sim$1%), spin decoherence ($<$1%), the charge fluctuation due to the optical frequencies, and spin-flip errors in the excited states during the optical excitation ($\sim$1%). Togan *et al.*[@photon-NV] pointed out that the fidelity can be reduced by the imperfect optical transitions due to the moderate and high strain, the path length fluctuation ($\sim$4%), and the signal to noise ratio in the ZPL channel ($\sim$11%). The charge fluctuation and the imperfect electron-spin population can be decreased by exploiting a repeated-until-success (the negative charge state and on resonance) fashion[@threemeters] before performing our gates. Acknowledgments {#acknowledgments .unnumbered} =============== This work was supported by the National Natural Science Foundation of China under Grant Nos. 11175094 and 91221205, the National Basic Research Program of China under Grants No. 2009CB929402 and No. 2011CB9216002, and the China Postdoctoral Science Foundation under Grant No. 2014M550703. GLL is a member of the Center of Atomic and Molecular Nanosciences, Tsinghua University. [99]{} Nielsen, M. A. & Chuang, I. L. *Quantum Computation and Quantum Information* (Cambridge University, Cambridge, 2000). Barenco, A. *et al.* Elementary gates for quantum computation. [*Phys. Rev. A*]{} [**52**]{}, 3457–3467 (1995). Liu, Y., Long, G. L. & Sun, Y. Analytic one-bit and CNOT gate constructions of general n-qubit controlled gates. [*Int. J. Quant. Inf.*]{} [**06**]{}, 447-462 (2008). Vatan, F. & Williams, C. Optimal quantum circuits for general two-qubit gates. [*Phys. Rev. A*]{} [**69**]{}, 032315 (2004). Shende, V. V., Markov, I. L. & Bullock, S. S. Minimal universal two-qubit controlled-NOT-based circuits. [*Phys. Rev. A*]{} [**69**]{}, 062321 (2004). Feng, G. R., Xu, G. F. & Long, G. L. Experimental realization of nonadiabatic holonomic quantum computation. [*Phys. Rev. Lett.*]{} [**110**]{}, 190501 (2013). Xu, G. & Long, G. Universal nonadiabatic geometric gates in two-qubit decoherence-free subspaces. [*Sci. Rep.*]{} [**4**]{}, 6814 (2014). Xu, G. & Long, G. Protecting geometric gates by dynamical decoupling. [*Phys. Rev. A*]{} [**90**]{}, 022323 (2014). Heilmann, H., Gräfe, M., Nolte, S. & Szameit, A. A novel integrated quantum circuit for high-order W-state generation and its highly precise characterization. [*Sci. Bull.*]{} [**60**]{}, 96-100 (2015). Xu, J. S. & Li, C. F. Quantum integrated circuit: Classical characterization. [*Sci. Bull.*]{} [**60**]{}, 141–141 (2015). Ren, B. C., Wei, H. R. & Deng, F. G. Deterministic photonic spatial-polarization hyper-controlled-not gate assisted by a quantum dot inside a one-side optical microcavity. [*Laser Phys. Lett.*]{} [**10**]{}, 095202 (2013). Ren, B. C. & Deng, F. G. Hyper-parallel photonic quantum computation with coupled quantum dots. [*Sci. Rep.*]{} [**4**]{}, 4623 (2014). Ren, B. C., Wang, G. Y. & Deng, F. G. Universal hyperparallel hybrid photonic quantum gates with dipole-induced transparency in the weak-coupling regime. [*Phys. Rev. A*]{} [**91**]{}, 032328 (2015). Shi, Y. Y. Both Toffoli and controlled-not need little help to do universal quantum computation. [*Quant. Inf. Comput.*]{} [**3**]{}, 084–092 (2003). Fredkin, E. & Toffoli, T. Conservative logic. [*Int. J. Theor. Phys.*]{} [**21**]{}, 219–253 (1982). Shende, V. V. & Markov, I. L. On the CNOT-cost of Toffoli gate. [*Quant. Inf. Comput.*]{} [**9**]{}, 461–468 (2009). Smolin, J. A. & DiVincenzo, D. P. Five two-bit quantum gates are sufficient to implement the quantum Fredkin gate. [*Phys. Rev. A*]{} [**53**]{}, 2855–2856 (1996). Zhang, C., Li, C. F. & Guo, G. C. Experimental demonstration of photonic quantum ratchet. [*Sci. Bull.*]{} [**60**]{}, 249–255 (2015). Long, G. L. & Zhang, T. C. Quantum ratchet with photons. [*Sci. Bull.*]{} [**60**]{}, 278 (2015). Balasubramanian, G. *et al.* Ultralong spin coherence time in isotopically engineered diamond. [*Nat. Mater.*]{} [**8**]{}, 383–387 (2009). Jelezko, F., Gaebel, T., Popa, I., Gruber, A. & Wrachtrup, J. Observation of coherent oscillations in a single electron spin. [*Phys. Rev. Lett.*]{} [**92**]{}, 076401 (2004). Gaebel, T. *et al.* Room-temperature coherent coupling of single spins in diamond. [*Nat. Phys.*]{} [**2**]{}, 408–413 (2006). Fuchs, G. D., Dobrovitski, V. V., Toyli, D. M., Heremans, F. J. & Awschalom, D. D. Gigahertz dynamics of a strongly driven single quantum spin. [*Science*]{} [**326**]{}, 1520–1522 (2009). Buckley, B. B., Fuchs, G. D., Bassett, L. C. & Awschalom, D. D. Spin-light coherence for single-spin measurement and control in diamond. [*Science*]{} [**330**]{}, 1212–1215 (2010). Jiang, L. *et al.* Repetitive readout of a single electronic spin via quantum logic with nuclear spin ancillae. [*Science*]{} [**326**]{}, 267–272 (2009). Robledo, L. *et al.* High-fidelity projective read-out of a solid-state spin quantum register. [*Nature*]{} [**477**]{}, 574–578 (2011). Dutt, M. V. G. *et al.* Quantum register based on individual electronic and nuclear spin qubits in diamond. [*Science*]{} [**316**]{}, 1312–1316 (2007). Childress, L. *et al.* Coherent dynamics of coupled electron and nuclear spin qubits in diamond. [*Science*]{} [**314**]{}, 281–285 (2006). Fuchs, G. D., Burkard, G., Klimov, P. V. & Awschalom, D. D. A quantum memory intrinsic to single nitrogen-vacancy centers in diamond. [*Nat. Phys.* ]{} [**7**]{}, 789–793 (2011). Jelezko, F. *et al.* Observation of coherent oscillation of a single nuclear spin and realization of a two-qubit conditional quantum gate. [*Phys. Rev. Lett.*]{} [**93**]{}, 130501 (2004). van der Sar, T. *et al.* Decoherence-protected quantum gates for a hybrid solid-state spin register. [*Nature*]{} [**484**]{}, 82–86 (2012). Yang, W. L., Yin, Z. Q., Xu, Z. Y., Feng, M. & Du, J. F. One-step implementation of multi-qubit conditional phase gating with nitrogen-vacancy centers coupled to a high-$Q$ silica microsphere cavity. [*Appl. Phys. Lett.*]{} [**96**]{}, 241113 (2010). Wei, H. R. & Deng, F. G. Compact quantum gates on electron-spin qubits assisted by diamond nitrogen-vacancy centers inside cavities. [*Phys. Rev. A*]{} [**88**]{}, 042323 (2013). Wang, C., Zhang, Y., Jiao, R. Z. & Jin, G. S. Universal quantum controlled phase gates on photonic qubits based on nitrogen vacancy centers and microcavity resonators. [*Opt. Express*]{} [**21**]{}, 19252–19260 (2013). Neumann, P. *et al.* Multipartite entanglement among single spins in diamond. [*Science*]{} [**320**]{}, 1326–1329 (2008). Xu, Z. Y., Hu, Y. M., Yang, W. L., Feng, M. & Du, J. F. Deterministically entangling distant nitrogen-vacancy centers by a nanomechanical cantilever. [*Phys. Rev. A*]{} [**80**]{}, 022335 (2009). Bernien, H. *et al.* Heralded entanglement between solid-state qubits separated by three metres. [*Nature*]{} [**497**]{}, 86–90 (2013). Pfaff, W. *et al.* Demonstration of entanglement-by-measurement of solid state qubits. [*Nat. Phys.*]{} [**9**]{}, 29–33 (2013). Togan, E. *et al.* Quantum entanglement between an optical photon and a solid-state spin qubit. [*Nature*]{} [**466**]{}, 730–734 (2010). Yang, W. L., Xu, Z. Y., Feng, M. & Du, J. F. Entanglement of separate nitrogen-vacancy centers coupled to a whispering-gallery mode cavity. [*New J. Phys.*]{} [**12**]{}, 113039 (2010). Chen, Q., Yang, W. L., Feng, M. & Du, J. F. Entangling separate nitrogen-vacancy centers in a scalable fashion via coupling to microtoroidal resonators. [*Phys. Rev. A*]{} [**83**]{}, 054305 (2011). Zhang, D., Li, J. H. & Yang, X. X. Laser-polarization-dependent spontaneous emission of the zero phonon line from single nitrogen-vacancy center in diamond. [*Chin. Phys. B*]{} [**23**]{}, 044204 (2014). Chang, Y. C. [*et al.*]{} Band-selective shaped pulse for high fidelity quantum control in diamond. [*Appl. Phys. Lett.*]{} [**104**]{}, 262403 (2014). Chen, X. [*et. al.*]{} Subdiffraction optical manipulation of the charge state of nitrogen vacancy center in diamond. [*Light: Sci. & Appl.*]{} [**4**]{}, e230; doi:10.1038/lsa.2015.3 (2015). Pfaff, W. *et al.* Unconditional quantum teleportation between distant solid-state quantum bits. [*Science*]{} [**345**]{}, 532–535 (2014). Ren, B. C. & Deng, F. G. Hyperentanglement purification and concentration assisted by diamond NV centers inside photonic crystal cavities. [*Laser Phys. Lett.*]{} [**10**]{}, 115201 (2013). Wang, C., He, L., Zhang, L. Y., Zhang, Y., Ma, H. & Zhang, R. Complete entanglement analysis on electron spins using quantum dot and microcavity coupled system. [*Sci. China-Phys. Mech. Astron.*]{} [**56**]{}, 2054-2058 (2013). Sheng, Y. B., Liu, J., Zhao, S. Y. & Zhou, L. Multipartite entanglement concentration for nitrogen-vacancy center and microtoroidal resonator system. [*Chin. Sci. Bull.*]{} [**59**]{}, 3507-3513 (2013). Hu, C. Y., Young, A., O’Brien, J. L., Munro, W. J. & Rarity, J. G. Giant optical Faraday rotation induced by a single-electron spin in a quantum dot: Applications to entangling remote spins via a single photon. [*Phys. Rev. B*]{} [**78**]{}, 085307 (2008). Hu, C. Y., Munro, W. J., O’Brien, J. L. & Rarity, J. G. Proposed entanglement beam splitter using a quantum-dot spin in a double-sided optical microcavity. [*Phys. Rev. B*]{} [**80**]{}, 205326 (2009). Bonato, C. *et al.* CNOT and Bell-state analysis in the weak-coupling cavity QED regime. [*Phys. Rev. Lett.*]{} [**104**]{}, 160503 (2010). Duan, L. M. & Kimble, H. J. Scalable photonic quantum computation through cavity-assisted interactions. [*Phys. Rev. Lett.*]{} [**92**]{}, 127902 (2004). Wei, H. R. & Deng, F. G. Scalable photonic quantum computing assisted by quantum-dot spin in double-sided optical microcavity. [*Opt. Express*]{} [**21**]{}, 17671–17685 (2013). Hua, M., Tao, M. J. & Deng, F. G. Universal quantum gates on microwave photons assisted by circuit quantum electrodynamics. [*Phys. Rev. A*]{} [**90**]{}, 012328 (2014). Hua, M., Tao, M. J. & Deng, F. G. Fast universal quantum gates on microwave photons with all-resonance operations in circuit QED. [*Sci. Rep.*]{} [**5**]{}, 9274 (2015). Chen, Q. & Feng, M. Quantum gating on neutral atoms in low-$Q$ cavities by a single-photon input-output process. [*Phys. Rev. A*]{} [**79**]{}, 064304 (2009). Wei, H. R. & Deng, F. G. Universal quantum gates on electron-spin qubits with quantum dots inside single-side optical microcavities. [*Opt. Express*]{} [**22**]{}, 593–607 (2014). Manson, N. B., Harrison, J. P. & Sellars, M. J. Nitrogen-vacancy center in diamond: Model of the electronic structure and associated dynamics. [*Phys. Rev. B*]{} [**74**]{}, 104303 (2006). Batalov, A. *et al.* Low temperature studies of the excited-state structure of negatively charged nitrogen-vacancy color centers in diamond. [*Phys. Rev. Lett.*]{} [**102**]{}, 195506 (2009). Tamarat, P. *et al.* Spin-flip and spin-conserving optical transitions of the nitrogen-vacancy centre in diamond. [*New J. Phys.*]{} [**10**]{}, 045004 (2008). Sipahigil, A. *et al.* Quantum interference of single photons from remote nitrogen-vacancy centers in diamond. [*Phys. Rev. Lett.*]{} [**108**]{}, 143601 (2012). Tamarat, P. *et al.* Stark shift control of single optical centers in diamond. [*Phys. Rev. Lett.*]{} [**97**]{}, 083002 (2006). Maze, J. R. *et al.* Properties of nitrogen-vacancy centers in diamond: The group theoretic approach. [*New J. Phys.*]{} [**13**]{}, 025025 (2011). Bassett, L. C., Heremans, F. J., Yale, C. G., Buckley, B. B. & Awschalom, D. D. Electrical tuning of single nitrogen-vacancy center optical transitions enhanced by photoinduced fields. [*Phys. Rev. Lett.*]{} [**107**]{}, 266403 (2011). Kosaka, H. & Niikura, N. Entangled absorption of a single photon with a single spin in diamond. [*Phys. Rev. Lett.*]{} [**114**]{}, 053603 (2015). Walls, D. F. & Milburn, G. J. *Quantum Optics* (Springer-Verlag, Berlin, 1994). An, J. H., Feng, M. & Oh, C. H. Quantum-information processing with a single photon by an input-output process with respect to low-$Q$ cavities. [*Phys. Rev. A*]{} [**79**]{}, 032303 (2009). Park, Y. S., Cook, A. K. & Wang, H. Cavity QED with diamond nanocrystals and silica microspheres. [*Nano Lett.*]{} [**6**]{}, 2075–2079 (2006). Larsson, M., Dinyari, K. N. & Wang, H. Composite optical microcavity of diamond nanopillar and silica microsphere. [*Nano Lett.*]{} [**9**]{}, 1447–1450 (2009). Barbour, R. J., Dinyari, K. N. & Wang, H. A composite microcavity of diamond nanopillar and deformed silica microsphere with enhanced evanescent decay length. [*Opt. Express*]{} [**18**]{}, 18968–18974 (2010). Schietinger, S., Schröder, T. & Benson, O. One-by-one coupling of single defect centers in nanodiamonds to high-$Q$ modes of an optical microresonator. [*Nano Lett.*]{} [**8**]{}, 3911–3915 (2008). Barclay, P. E., Fu, K. M. C., Santori, C. & Beausoleil, R. G. Chip-based microcavities coupled to nitrogen-vacancy centers in single crystal diamond. [*Appl. Phys. Lett.*]{} [**95**]{}, 191115 (2009). McCutcheon, M. W. & Lončcar, M. Design of a silicon nitride photonic crystal nanocavity with a quality factor of one million for coupling to a diamond nanocrystal. [*Opt. Express*]{} [**16**]{}, 19136–19145 (2008). Wolters, J. *et al.* Enhancement of the zero phonon line emission from a single nitrogen vacancy center in a nanodiamond via coupling to a photonic crystal cavity. [*Appl. Phys. Lett.* ]{} [**97**]{}, 141108 (2010). Englund, D. *et al.* Deterministic coupling of a single nitrogen vacancy center to a photonic crystal cavity. [*Nano Lett.* ]{} [**10**]{}, 3922–3926 (2010). Faraon, A., Barclay, P. E., Santori, C., Fu, K. M. C. & Beausoleil, R. G. Resonant enhancement of the zero-phonon emission from a colour centre in a diamond cavity. [*Nat. Photon.*]{} [**5**]{}, 301–305 (2011). Gregor, M., Henze, R., Schröder, T. & Benson, O. On-demand positioning of a preselected quantum emitter on a fiber-coupled toroidal microresonator. [*Appl. Phys. Lett.*]{} [**95**]{}, 153110 (2009). Albrecht, R., Bommer, A., Deutsch, C., Reichel, J. & Becher, C. Coupling of a single nitrogen-vacancy center in diamond to a fiber-based microcavity. [*Phys. Rev. Lett.*]{} [**110**]{}, 243602 (2013). Teissier, J., Barfuss, A., Appel, P., Neu, E. & Maletinsky, P. Strain coupling of a nitrogen-vacancy center spin to diamond mechanical oscillator. [*Phys. Rev. Lett.*]{} [**113**]{}, 020503 (2014). Barclay, P. E., Fu, K. M., Santori, C. & Beausoleil, R. G. Hybrid photonic crystal cavity and waveguide for coupling to diamond NV-centers. [*Opt. Express*]{} [**17**]{}, 9588–9601 (2009). Shen, J. T. & Fan, S. Quantum critical coupling conditions for zero single-photon transmission through a coupled atom-resonator-waveguide system. [*Phys. Rev. A*]{} [**82**]{}, 021802 (2010). Peng, B., *et al.* Parity-time-symmetric whispering-gallery microcavities. [*Nat. Phys.*]{} [**10**]{}, 394–398 (2014). Luxmoore, I. J. *et al.* Restoring mode degeneracy in H1 photonic crystal cavities by uniaxial strain tuning. [*Appl. Phys. Lett.*]{} [**100**]{}, 121116 (2012). Hagemeier, J. *et al.* H1 photonic crystal cavities for hybrid quantum information protocols. [*Opt. Express*]{} [**20**]{}, 24714–24726 (2012). Bonato, C. *et al.* Tuning micropillar cavity birefringence by laser induced surface defects. [*Appl. Phys. Lett.*]{} [**95**]{}, 251104 (2009). Gudat, J. *et al.* Permanent tuning of quantum dot transitions to degenerate microcavity resonances. [*Appl. Phys. Lett.*]{} [**98**]{}, 121111 (2011). Bonato, C. *et al.* Strain tuning of quantum dot optical transitions via laser-induced surface defects. [*Phys. Rev. B*]{} [**84**]{}, 075306 (2011). Barclay, P. E., Fu, K. M., Santori, C., Faraon, A. & Beausoleil, R. G. Hybrid nanocavity resonant enhancement of color center emission in diamond. [*Phys. Rev. X*]{} [**1**]{}, 011007 (2011). Faraon, A., Santori, C., Huang, Z., Acosta, V. M. & Beausoleil, R. G. Coupling of nitrogen-vacancy centers to photonic crystal cavities in monocrystalline diamond. [*Phys. Rev. Lett.*]{} [**109**]{}, 033604 (2012). Bernien, H. *et al.* Two-photon quantum interference from separate nitrogen vacancy centers in diamond. [*Phys. Rev. Lett.*]{} [**108**]{}, 043604 (2012). Acosta, V. M. *et al.* Dynamic stabilization of the optical resonances of single nitrogen-vacancy centers in diamond. [*Phys. Rev. Lett.*]{} [**108**]{}, 206401 (2012). Chu, Y. *et al.* Coherent optical transitions in implanted nitrogen vacancy centers. [*Nano Lett.*]{} [**14**]{}, 1982–1986 (2014). Dréau, A., Spinicelli, P., Maze, J. R., Roch, J. F. & Jacques, V. Single-shot readout of multiple nuclear spin qubits in diamond under ambient conditions. [*Phys. Rev. Lett.*]{} [**110**]{}, 060502 (2013). Beenakker, C. W. J., DiVincenzo, D. P., Emary, C. & Kindermann, M. Charge detection enables free-electron quantum computation. [*Phys. Rev. Lett.* ]{} [**93**]{}, 020501 (2004). Nemoto, K. & Munro, W. J. Nearly deterministic linear optical controlled-NOT gate. [*Phys. Rev. Lett.*]{} [**93**]{}, 250502 (2004). Lin, Q. & He, B. Single-photon logic gates using minimal resources. [*Phys. Rev. A*]{} [**80**]{}, 042310 (2009). [^1]: Correspondence author: [email protected])
{ "pile_set_name": "ArXiv" }
--- abstract: | We report a Raman study of the so-called buffer layer with periodicity which forms the intrinsic interface structure between epitaxial graphene and SiC(0001). We show that this interface structure leads to a non-vanishing signal in the Raman spectrum at frequencies in the range of the D- and G-band of graphene and discuss its shape and intensity. *Ab-initio* phonon calculations reveal that these features can be attributed to the vibrational density of states of the buffer-layer.\ address: - '$^1$ Lehrstuhl für Technische Physik, Universität Erlangen-Nürnberg, Erwin-Rommel-Str. 1, 91058 Erlangen, Germany' - '$^2$ Paul-Drude-Institut für Festkörperelektronik, Hausvogteiplatz 5-7, 10117 Berlin, Germany' - '$^3$ Physics and Material Sciences Research Unit, University of Luxembourg, Campus Limpertsberg, L-1511 Luxembourg' - '$^4$ Institute for Electronics, Microelectronics, and Nanotechnology (IEMN), CNRS UMR 8520, Dept. ISEN, 59652 Villeneuve d’Ascq Cedex, France' - '$^5$ Institut für Physik, Technische Universität Chemnitz, Reichenhainer Str. 70, 09126 Chemnitz, Germany' author: - 'F Fromm$^{1}$, M H Oliveira Jr.$^{2}$, A Molina-Sánchez$^{3,4}$, M Hundhausen$^{1}$, J M J Lopes$^{2}$, H Riechert$^{2}$, L Wirtz$^{3,4}$ and T Seyller$^{5}$' title: 'Contribution of the buffer layer to the Raman spectrum of epitaxial graphene on SiC(0001)' --- [*Keywords*]{}: Graphene, silicon carbide, buffer layer, phonons, Raman spectroscopy, *ab-initio* calculations, vibrational density of states Introduction ============ Raman spectroscopy is a very powerful tool for investigating carbon materials and is intensively used for the characterization of graphene obtained by diffent methods.[@jorio2011a] For example, Raman spectroscopy has been shown to be extremely useful in order to discern monolayer graphene from bilayers and multilayers.[@ferrari2006a; @graf2007a] Furthermore, this technique provides information about the carrier concentration in graphene,[@das2008a] the edges of graphene flakes,[@pimenta2007a; @gupta2009a; @zhang2011a] and about the properties of graphene nano ribbons.[@gillen2010a; @gillen2010b; @ryu2011a] Hence it is no surprise that it is also used to investigate epitaxial graphene grown on silicon carbide. [@faugeras2008a; @roehrl2008a; @ni2008a; @ferralis2008a; @lee2008a; @robinson2009a; @robinson2009b; @domke2009a; @speck2011a; @oliveira2011a; @emtsev2009a; @tiberj2011a; @ferralis2011a] ![\[fig:raman\] Raman spectrum of 6H-SiC measured with a laser wavelength of 532 nm. The signals due to the longitudinal acoustic (LA), transverse optical (TO), and longitudinal optical (LO) phonon modes are indicated. The region between 1000 and 2000  is also shown after multiplication by a factor of 50. In this region, the signals due to two-phonon processes are visible. Three-phonon processes are very weak and can be neglected. The positions of the D, G, and 2D line of graphene are indicated by the vertical dashed lines.](fig1){width="12cm"} The Raman spectrum of graphene usually shows three main features: the D-band at around 1350 , the G-band at about 1580 , and the 2D-band at approximately 2680 . However, Raman spectroscopy is not a surface sensitive method and usually the probed sample volume is much larger, i.e. deeper, than the graphene sheet itself. This leads to the presence of substrate-related features in the spectrum as well. For many substrates such as SiO$_2$/Si this is not a problem because these features do not overlap with the graphene signals. This is different for epitaxial graphene grown on SiC where the Raman spectrum in the D- and G-range is dominated by the two-phonon modes of the SiC substrate.[@roehrl2008a] As an example we show in figure \[fig:raman\] a Raman spectrum of 6H-SiC. The region between around 1000 and 2000  is dominated by two-phonon processes. Therefore, it has become common to correct the spectra of epitaxial graphene on SiC by subtracting the spectrum of the bare substrate. This procedure, however, assumes that the spectrum contains only contributions from the epitaxial graphene and from the SiC bulk. In the case of epitaxial graphene on SiC(0001) this assumption, however, may not be correct because it is known that the graphene sheet resides on the so-called buffer layer[@emtsev2007a; @emtsev2008a]. It is nowadays widely accepted that the buffer layer itself is a graphene-like honeycomb lattice of carbon atoms ontop of an otherwise unreconstructed Si(0001) surface. [@emtsev2007a; @emtsev2008a] The buffer layer shows the undistorted -states of graphene but a distorted -band. The distortions are caused by the hybridization of the $\pi$-states with the states of the SiC(0001) surface and by the formation of covalent bonds between some of the graphene-carbon atoms and underlying silicon atoms. Several theoretical studies have investigated various aspects of this structure, all in good agreement with experimental results.[@varchon2007a; @mattausch2007a; @mattausch2007b; @mattausch2008a; @kim2008a] It is natural to ask about the contribution of the buffer layer to the total Raman spectrum measured from epitaxial graphene on SiC. In this paper we will show that the buffer layer leads to a non-negligible contribution in the Raman spectrum, and we will discuss the origin of the signal. Experimental details ==================== In order to identify the contribution of the buffer layer to the Raman signal of epitaxial graphene we have studied different samples. The structures of the samples are depicted schematically in fig. \[fig:structures\]. All samples were prepared on chips cut from nitrogen-doped, on-axis oriented 6H-SiC(0001) wafer purchased from SiCrystal AG. Despite the fact that the wafer had an epi-ready chemo-mechanical polish (CMP), the surfaces were treated with an additional hydrogen etch in 1 bar H$_2$ at 1500 .[@ostler2010a] Samples covered with the buffer layer (R3 for short, see fig. \[fig:structures\](a)) were prepared by annealing the SiC(0001) sample in 1 bar Ar at $T=1450$ .[@ostler2010a] Monolayer graphene on the buffer layer (termed MLG, see fig. \[fig:structures\](b)) was obtained by annealing the SiC substrate in 1 bar Ar at 1650 . [@ostler2010a; @emtsev2009a] From previous studies it is known that such samples may contain inclusions of bilayer graphene at positions close to the step edges.[@emtsev2009a] Therefore, our micro-Raman spectroscopy measurements (see below) also allowed us to obtain Raman spectra from bilayer graphene on the buffer layer (BLG for short, see fig. \[fig:structures\](c)). Finally, quasi-free standing graphene on hydrogen-saturated SiC(0001)[@riedl2010b; @speck2010a; @speck2011a; @forti2011a] (QFMLG, see fig. \[fig:structures\](d)) was obtained by annealing samples covered by the buffer layer in 1 bar hydrogen.[@speck2011a] Reference spectra of 6H-SiC were obtained from a hydrogen etched sample which is free of any carbonaceous surface layer. ![\[fig:structures\] Schematic structures (side view) of the samples used in the present study. (a) The buffer layer (R3) with periodicity. (b) Monolayer graphene (MLG) situated on the buffer layer. (c) Bilayer graphene (BLG) situated on top of the buffer layer. (c) Quasi-free standing monolayer graphene (QFMLG) on top of the hydrogen-saturated SiC surface. The carbon atoms of the buffer layer are plotted in blue.](fig2){width="12cm"} The samples prepared in the above mentioned ways were thorougly characterized by X-ray photoelectron spectroscopy (XPS) and atomic force microscopy. Micro-Raman spectroscopy measurements were performed using a Jobin Yvon T64000 triple spectrometer combined with a liquid nitrogen cooled CCD detector. A frequency doubled Nd:YVO$_4$ laser with a wavelength of 532 nm was employed. Additional spectra were measured using an Ar ion laser providing wavelengths of 476 nm and 514 nm. The laser beam was focused onto the sample by a 100x objective with numerical aperture $\mbox{NA}=0.9$ and the scattered light was detected in backscattering geometry. The laser spot size was 1 $\mathrm{~\mu m}$. Unless otherwise stated, the Raman spectra were measured under the same conditions. The raw data was normalized to the maximum of the TO phonon mode of 6H-SiC at about 780$\;$cm$^{-1}$. Results and Discussion ====================== Figure \[fig:spectra\](a) compiles typical Raman spectra of the different samples described above. The spectra were collected at a laser wavelength of 532 nm. The lowest spectrum (labelled R3) was measured on the sample covered by the buffer layer. At low energies that spectrum contains two rather broad features, one centered at around 1350  and one at 1580 . The latter is accompanied by a smaller peak at the low-energy side situated at around 1485-1490 . No 2D line is observed for the buffer layer. ![\[fig:spectra\] (a) Raman spectra of (from bottom to top) the buffer layer (R3), monolayer graphene on the buffer layer (MLG), bilayer graphene on the buffer layer (BLG), and quasi-free standing monolayer graphene (QFMLG) on H-terminated SiC(0001). The spectra were measured with a laser wavelength of 532 nm. (b) Raman spectra of the buffer layer measured with three different laser energies.](fig3){width="12cm"} The spectra of MLG and BLG are also displayed in figure \[fig:spectra\](a). Both spectra exhibit a G line and a 2D line. The 2D line of MLG at 2706  is very well described by a single Lorentzian with a full width of half maximum of 35 , wich agrees with the notion that the sample is covered mainly by monolayer epitaxial graphene. The shape of the 2D band of BLG in fig. \[fig:spectra\](a) is consistent with what has been observed previously on both exfolitated graphene[@ferrari2006a; @graf2007a] and epitaxial graphene.[@roehrl2008a] Note, that in this work we are not interested in the exact positions of the G and 2D bands, which might be influenced by strain and charge. For the discussion of this topic we refer the reader to previous published work. [@roehrl2008a; @ni2008a; @ferralis2008a; @lee2008a; @robinson2009a; @robinson2009b; @speck2011a] However, what is important for the present work is the observation that the spectra of MLG and BLG contain the same broad features between 1200 and 1665  that are observed for the R3 sample. Finally, figure \[fig:spectra\](a) also shows the spectrum of a sample of QFMLG, i.e. a layer of graphene on SiC(0001) without the buffer layer at the interface. The spectrum consists of three narrow lines: the D line, the G line, and the 2D line, as discussed in previous work.[@speck2011a] In contrast to the spectra of MLG and BLG, the broad features between 1200 and 1665  are absent and the spectrum is basically flat between D and G line. This fact provides important input. As mentioned above, the spectra shown are difference spectra where the spectrum of a clean SiC sample is subtracted from that of the graphene covered one. One could therefore think that the broad features described above are the result of an insufficient background correction. This is clearly ruled out by the fact that the spectrum of QFMLG, which was obtained in exactly the same way as those of the [R3]{}, MLG, and BLG samples, does not show this features. The only effect of the background subtraction is the increase of noise on both sides of the G line which can be seen in all spectra. This can be understood by considering that at those frequencies the intensity in both data sets, the one of the sample with graphene and the one used for background subtraction, is particularily large due to the contribution of the SiC substrate (see fig. \[fig:raman\]). The larger intensity at these frequencies leads to a larger statistical noise ($\sqrt{n}$ with $n$ being the count rate) which of course is not removed by the subtraction of the spectra. Therefore we can savely state that the spectrum labelled [R3]{} in fig. \[fig:spectra\] is the true Raman spectrum of the buffer layer which exists at the interface between SiC(0001) and epitaxial graphene. Figure \[fig:spectra\](b) shows three Raman spectra of the buffer layer measured with three different laser wavelengths of 476 nm, 514 nm, and 532 nm, which correspond to excitation energies of 2.33 eV, 2.41 eV, and 2.60 eV, respectively. Since the scattering intensity is zero for wavenumbers larger than approximately 1665 , we show only the low-energy part of the spectrum between 1100 and 1800 . As can be seen from figure \[fig:spectra\](b), the observed Raman spectrum is virtually independent of the laser wavelength, i.e. the broad peaks at 1350 , 1485-1490 , and 1580  show no dispersion and there is hardly any change in the shape of the signals. ![\[fig:model\] Unit cell used in the simulation of the buffer layer. The carbon atoms of the buffer layer are plotted in blue. For clarity, the top view shows only atoms of the topmost SiC bilayer. Hydrogen atoms passivate the dangling bonds of the carbon atoms of SiC at the bottom. In this configuration two carbon atoms of buffer layer are on the top of silicon atoms. Drawing not to scale.](fig4){width="12cm"} The Raman spectrum shown in Fig. \[fig:spectra\](b) does not seem to be composed of discrete peaks but rather resembles a vibrational density of states (vDOS). This is plausible since the unit-cell of the reconstruction is quite large and the corresponding reciprocal unit cell is small. Thus, a large part of the phonon-dispersion is folded back onto $\Gamma$ and becomes potentially Raman active. To a good approximation, one may therefore assume that the Raman spectrum of Fig. \[fig:spectra\](b) corresponds to the vDOS of the buffer layer. We have verified this hypothesis by an explicit calculation of the phonon dispersion and vDOS of the buffer layer using *ab-initio* methods as presented in the following. Since the unit-cell of the reconstruction (and even of the recently proposed $(5\times5)$ superstructure[@pankratov2010a]) is prohibitively large for *ab-initio* calculations of phonons, we have chosen to work with the [3]{} reconstruction which was also used in the electronic-structure calculations of Refs. [@varchon2007a] and [@mattausch2007a]. This unit-cell corresponds to a $(2\times2)$ supercell of graphene. The unit cell for our simulation of the buffer layer on SiC is shown in Fig. \[fig:model\]. With respect to free-standing graphene, it corresponds to a $(2 \times 2)$ unit cell, containing 8 carbon atoms. With the aim of obtaining reliable results for the phonons of the buffer layer, the commensurability between the buffer layer and SiC is obtained by squeezing the substrate by 8 %, adopting the experimental lattice constant of graphene (2.46 Å). (This is different from the procedure in Ref. [@varchon2007a] where the lattice constant of graphene was increased by 8 % in order to match the experimental lattice constant of SiC.) The SiC substrate is simulated by four atomic layers (2 Si layers and 2 C layers), passivated with hydrogen atoms at the bottom. Note that in this configuration two of the eight carbon atoms of the buffer layer are on top of the silicon atoms, forming a covalent bond. The atomic positions inside the unit cell have been calculated with density functional theory (DFT)[@kohn1965a; @parr1989a], in the local density approximation (LDA).[^1] The calculations were performed with the Quantum-Espresso code[@QE-2009] using ultra-soft Vanderbilt pseudopotentials, a $12 \times 12 \times 1$ k-point grid, and and energy cutoff of 35 Ry. Due to the formation of covalent bonds, the carbon atoms on top of surface Si atoms display an inward buckling by $\Delta$z = 0.39 Å(see Fig. \[fig:model\]). The distance between the other carbon atoms and the surface-Si plane is $d$ = 2.02 Å. ![\[fig:phonons\] Left panel: Phonon modes of the buffer layer (blue lines) and free-standing graphene (black dashed line) in a $(2 \times 2)$ unit cell. The red circles mark the Kohn anomalies at $\Gamma$ and $K$ of free-standing graphene. Right panel: vibrational density of states of the buffer layer (blue line) and of free-standing graphene (black-dashed line) in comparison with the experimental Raman data (red dots).](fig5){width="12cm"} Starting from the optimized geometry, we use density functional perturbation theory (DFPT),[@gonze1997a; @baroni2001a] to calculate the phonon dispersion of the buffer layer. The result is shown in Fig. \[fig:phonons\], where we have concentrated on the frequency range above 1200 cm$^{-1}$ which is important for the interpretation of the spectra in Fig. \[fig:spectra\](b). The broad blue lines correspond to the modes of the buffer layer (The SiC bulk modes have frequencies below 1200 cm$^{-1}$). For comparison we have included the phonon bands of isolated graphene in a $2\times2$ unit cell (containing 8 atoms, leading to 24 phonon branches marked by the black dashed lines).[^2] The right panel of Fig. \[fig:phonons\] shows the vibrational density of states (vDOS) of the buffer layer and free-standing graphene, together with the experimental Raman spectra (red dots). The phonon dispersion of the buffer-layer is considerably changed compared to the one of isolated graphene. The changes in the electronic structure (lifting of the linear crossing at K and separation of the $\pi$ and $\pi^*$ bands by more than 2 eV [@varchon2007a]) lead to the elimination of the Kohn anomalies at K and $\Gamma$ (marked by red circles in Fig. \[fig:phonons\]), similar to the findings for graphene on a Ni(111) surface.[@allard2010a] A consequence of this is the absence of the 2D line (around 2680 cm$^{-1}$) in the Raman spectrum of the buffer layer, as seen in Fig. \[fig:spectra\]. Furthermore, the buffer layer displays flatter bands than pure graphene. In particular, the overbending of the highest optical branch around $\Gamma$ almost disappears. This brings the frequency of the highest vDOS peak at 1630 cm$^{-1}$ down to 1620 cm$^{-1}$, in agreement with the highest peak in the Raman spectrum. Additionally, some degeneracies are broken at the $\Gamma$ and M points. These modifications lead to noticeable changes in the vDOS of the buffer layer. For instance, a clear gap emerges between 1500 and 1550 , in agreement with the minimum at the same frequency range in the Raman spectrum. The broad feature in the Raman spectrum around 1360 cm$^{-1}$ can also be associated with peaks in the vDOS of the buffer layer that are due to flat phonon bands. The phonon bands of pure graphene are very dispersive in this range and the corresponding vDOS is flat. ![\[fig:projection\] Left panel: Phonon bands weighted according to the contribution of two different subspaces (see main text). Right panel: same as in Fig. \[fig:phonons\].](fig6){width="12cm"} For a better understanding of the gap opening between 1500 and 1550 , we analyze the phonon eigenvectors of the buffer layer by projecting them onto the eigenvectors of the isolated (undisturbed) graphene. In Fig. \[fig:projection\] we have projected every eigenvector of the buffer layer onto two subspaces of eigenvectors of isolated graphene. The color of the phonon branch indicates which subspace dominates the character of the vibration. The first subspace (orange color) is composed by the two eigenvectors of highest frequency at the respective phonon-wave vector $\bm{q}$. This definition becomes ambiguous at the crossing point of the 2nd and 3rd highest mode in between the high-symmetry points. But it is well defined at the high-symmetry points $\Gamma$, M, and K, where double-degeneracies are observed. The second subspace (green color) includes the next four eigenvectors in energy order, and it has also two 2-fold degeneracies at $\Gamma$, M, and K. We focus on the dispersion around M. We can assign the first and third phonon branches as being predominantly due to the orange subspace. The frequency difference is 100 . All the other eigenvectors belong to the green subspace. The perturbation of the buffer-layer vibrations by the partially covalent bonding to the SiC substrate lifts the degeneracies. The splitting is so strong that it even changes the order of the phonon modes: The lower frequency mode due to the first subspace falls below the highest frequency mode of the second subspace. In the phonon dispersion, between $\Gamma$ and M (and also between $\Gamma$ and K), this leads to an avoided crossing between the 2nd and 3rd highest phonon mode and thus to the opening of a gap from 1500 to 1550 . For the lower frequency modes, similar analysis can be made. But the analysis becomes more complicated due to a large number of participating modes. In the right panel of Fig. \[fig:phonons\], one can observe an approximate agreement between dips in the Raman spectrum and gaps in the vDOS. The same holds for the peaks in the Raman spectrum and in the vDOS. Since we used a simplified supercell for the buffer-layer geometry, we do not expect perfect agreement here. But we consider the present calculation as a qualitative argument that the observed features in the Raman spectra of the buffer-layer can indeed be associated with the vibrational density of states. Conclusions =========== In conclusion we have been able to unambiguously identify the Raman spectrum of the buffer layer (R3) which exists at the interface between epitaxial graphene and SiC(0001). We have shown that it constitutes a non-negligible contribution underlying the graphene spectrum especially at frequencies around the D- and G-line. This implies that proper Raman analysis of graphene on SiC(0001) requires that the spectrum is also corrected for the buffer layer contributions. Neglecting the buffer layer will lead to errors in the interpretation of Raman spectra of epitaxial graphene on SiC(0001). Furthermore, we have discussed the Raman spectrum of the buffer layer in terms of the vibrational density of states. To that end, *ab initio* calculations on a 3 superstructure have been performed which revealed a complete extinction of the Kohn anomally, in agreement with the lack of a Dirac cone in the electronic structure [@emtsev2007a; @emtsev2008a] and with the absence of a 2D line in the Raman spectrum. As a consequence, phonon bands become flatter than in free-standing graphene. In addition, the carbon-silicon covalent bonds modify substantially the frequencies and lead to a mixing of the phonon branches of isolated graphene. This leads to a breaking of degeneracies in the phonon dispersion and the vDOS of the buffer layer is richer in structure than that of isolated graphene. In particular a clear gap between 1500 and 1550  emerges which agrees fairly well with the Raman spectrum. Acknowledgement =============== This work was supported by the European Science Foundation (ESF) in the framework of the EuroGRAPHENE project Graphic-RF (grant number 09-EuroGRAPHENE-FP-018) and by the European Union (EU) in the framework of the project ConceptGraphene (grant number 257829). A.M.-S. and L.W. acknowledge funding by the French National Research Agency (ANR) through project ANR-09-BLAN-0421-01. Calculations were done at the IDRIS supercomputing center, Orsay (Proj. No. 091827), and at the Tirant Supercomputer of the University of Valencia (group vlc44). The authors thank L. Ley, J. Ristein, and R. J. Koch for stimulating discussions. References ========== [52]{} A. Jorio, R. Saito, and M. S. Dresselhaus, Raman spectroscopy in graphene related systems (Wiley, 2011). A. C. Ferrari, J. C. Meyer, V. Scardaci, C. Casiraghi, M. Lazzeri, F. Mauri, S. Piscanec, D. Jiang, K. S. Novoselov, S. Roth, and A. K. Geim, Phys. Rev. Lett. 97, 187401 (2006). D. Graf, F. Molitor, K. Ensslin, C. Stampfer, A. Jungen, C. Hierold, and L. Wirtz, Nano Lett. 7, 238 (2007). A. Das, S. Pisana, B. Chakraborty, S. Piscanec, S. K. Saha, U. V. Waghmare, K. S. Novoselov, H. R. Krishnamurty, A. K. Geim, A. C. Ferrari, and A. K. Sood, Nat. Nano. 3, 210 (2008). M. A. Pimenta, G. Dresselhaus, M. S. Dresselhaus, L. G. Cancado, A. Jorio, and R. Saito, Phys Chem. Chem. Phys. 9, 1276 (2007). A. K. Gupta, T. J. Russin, H. R. Gutierrez, and P. C. Eklund, ACS Nano 3, 45 (2009). W. Zhang and L.-J. Li, ACS Nano 5, 3347 (2011). R. Gillen, M. Mohr, and J. Maultzsch, Phys. Rev. B 81, 205426 (2010). R. Gillen, M. Mohr, and J. Maultzsch, Phys. Stat. Sol. B 247, 2941 (2010). S. Ryu, J. Maultzsch, M. Y. Han, P. Kim, and L. E. Brus, ACS Nano 5, 4123 (2011). C. Faugeras, A. Nerriere, M. Potemski, A. Mahmood, E. Dujardin, C. Berger, and W. A. de Heer, Appl. Phys. Lett. 92, 011914 (2008). J. Röhrl, M. Hundhausen, K. Emtsev, T. Seyller, R. Graupner, , and L. Ley, Appl. Phys. Lett. 92, 201918 (2008). Z. H. Ni, W. Chen, X. F. Fan, J. L. Kuo, T. Yu, A. T. S. Wee, and Z. X. Shen, Phys. Rev. B 77, 115416 (2008). N. Ferralis, R. Maboudian , and C. Carraro, Phys. Rev. Lett. 101, 156801 (2008). D. S. Lee, C. Riedl, B. Krauss, K. von Klitzing, U. Starke, and J. H. Smet, Nano Lett. 8, 4320 (2008). J. Robinson, C. Puls, N. Staley, J. Stitt, M. Fanton, K. Emtsev, T. Seyller, and Y. Liu, Nano Lett. 9, 964 (2009). J. Robinson, M. Wetherington, J. Tedesco, P. Campbell, X.Weng, J. Stitt, M.Fanton, E. Frantz, D. Snyder, B. Van-Mil, G. Jernigan, R. Myers-Ward, C.R. Eddy Jr., and D. Gaskill, Nano Lett. 9, 2873 (2009). K. F. Domke and B. Pettinger, J. Raman Spectrosc. 40, 1427 (2009). F. Speck, J. Jobst, F. Fromm, M. Ostler, D. Waldmann, M. Hundhausen, H. B. Weber, and T. Seyller, Appl. Phys. Lett. 99, 122106 (2011). M. H. Oliveira Jr., T. Schumann, M. Ramsteiner, J. M. J. Lopes, and H. Riechert, Appl. Phys. Lett. 99, 111901 (2011). K. V. Emtsev, A. Bostwick, K. Horn, J. Jobst, G. L. Kellogg, L. Ley, J. L. McChesney, T. Ohta, S. A. Reshanov, J. Roehrl, E. Rotenberg, A. K. Schmid, D. Waldmann, H. B. Weber, and T. Seyller, Nat. Mater. 8, 203 (2009). A. Tiberj, N. Camara, P. Godignon, and J. Camassel, Nanoscale Res. Lett. 6, 478 (2011). N. Ferralis, R. Maboudian, and C. Carraro, Phys. Rev. B 83, 081410 (2011). K. V. Emtsev, T. Seyller, F. Speck, L. Ley, P. Stojanov, J. Riley, and R. Leckey, Mater. Sci. Forum 556-557, 525 (2007). K.V.Emtsev, F. Speck, T. Seyller, J.D.Riley, and L. Ley, Phys. Rev. B 77, 155303 (2008). F. Varchon, R. Feng, J. Hass, X. Li, B. N. Nguyen, C. Naud, P. Mallet, J.-Y. Veuillen, C. Berger, E. Conrad, and L. Magaud, Phys. Rev. Lett 99, 126805 (2007). A. Mattausch and O. Pankratov, Phys. Rev. Lett. 99, 076802 (2007). A. Mattausch and O. Pankratov, Mater. Sci. Forum 556-557, 693 (2007). A. Mattausch and O. Pankratov, Phys. Stat. Sol. B 245, 1425 (2008). S. Kim, J. Ihm, H. J. Choi, and Y.-W. Son, Phys. Rev. Lett. 100, 176802 (2008). M. Ostler, F. Speck, M. Gick, and T. Seyller, Phys. Stat. Sol. B 247, 2924 (2010). C. Riedl, C. Coletti, T. Iwasaki, A. A. Zakharov, and U. Starke, Phys. Rev. Lett. 103, 246804 (2009). F. Speck, M. Ostler, J. R¨ohrl, J. Jobst, D. Waldmann, M. Hundhausen, H. Weber, and T. Seyller, Mater. Sci. Forum 645-648, 629 (2010). S. Forti, K. V. Emtsev, C. Coletti, A. A. Zakharov, C. Riedl, and U. Starke, Phys. Rev. B 84, 125449 (2011). O. Pankratov, S. Hensel, and M. Bockstedte, Phys. Rev. B 82, 121416(R) (2010). W. Kohn and L. J. Sham, Phys. Rev. 140, A1133 (1965). R. G. Parr and W. Yang, Density-Functional Theory of Atoms and Molecules (Oxford University Press, 1989). P. Giannozzi, S. Baroni, N. Bonini, M. Calandra, R. Car, C. Cavazzoni, D. Ceresoli, G. L. Chiarotti, M. Cococcioni, I. Dabo, A. Dal Corso, S. de Gironcoli, S. Fabris, G. Fratesi, R. Gebauer, U. Gerstmann, C. Gougoussis, A. Kokalj, M. Lazzeri, L. Martin-Samos, N. Marzari, F. Mauri, R. Mazzarello, S. Paolini, A. Pasquarello, L. Paulatto, C. Sbraccia, S. Scandolo, G. Sclauzero, A. P. Seitsonen, A. Smogunov, P. Umari, and R. M. Wentzcovitch, J. Phys. Condens. Matter 21, 395502 (2009). X. Gonze and C. Lee, Phys. Rev. B 55, 10355 (1997). S. Baroni, S. de Gironcoli, A. Dal Corso, and P. Giannozzi, Rev. Mod. Phys. 73, 515 (2001). A. Allard and L. Wirtz, Nano Lett. 10, 4335 (2010). G. Kresse, J. Furthmüller, and J. Hafner, Europhys. Lett. 32, 729 (1995). L. Wirtz and A. Rubio, Solid State Communications 131, 141 (2004). G. Kern, G. Kresse, and J. Hafner, Phys. Rev. B 59, 8551 (1999). A. Molina-Sánchez and L. Wirtz, Phys. Rev. B 84, 155413 (2011). A. Marini, P. Garc?a-González, and A. Rubio, Phys. Rev. Lett. 96, 136404 (2012). F. Mittendorfer, A. Garhofer, J. Redinger, J. Klimes, J. Harl, and G. Kresse, Phys. Rev. B 84, 201401 (2011). H.-J. Kim, A. Tkatchenko, J.-H. Cho, and M. Scheffler, Phys. Rev. B 85, 041403(R) (2012). S. Berciaud, S. Ryu, L. E. Brus, and T. F. Heinz, Nano Lett. 9, 346 (2009). R. J. Koch, Th. Seyller, et al., manuscript in preparation. [^1]: We note that the use of DFT with purely (semi)local functionals is questionable for the use in layered systems where Van der Waals forces are expected to play an important role. Nevertheless, the LDA seems to work fine for the calculation of geometries and even of phonon frequencies (however, not for the binding energies) of several layered systems such as graphite[@kresse95; @wirtz2004a], hexagonal boron nitride[@kern1999a], graphene on a Nickel(111) surface[@allard2010a], and MoS$_{2}$.[@molina2011a] This seemingly good performance is probably due to a fortuitious error cancellation: the small (but non-negligible) covalent part of the inter-layer binding is overestimated while the Van der Waals part of the binding energy is completely neglected. For a more precise treatment of Van der Waals forces, calculating electron correlation in the random-phase approximation, we refer to the work of Marini et al. on hBN[@marini2006a], Mittendorfer et al. on graphene bound to metallic substrates[@mittendorfer2011a] or Kim et al. for the binding of benzene molecules on a Si surface [@kim2012a]. [^2]: There is the eternal question if one should use (*ab-initio*) optimized lattice constants or experimental lattice constants for the phonon calculations. Since the local-density approximation tends to overbind, the optimized lattice constant is smaller than the experimental one. The calculated phonon frequencies are in general a little bit higher than the experimental values and need to be scaled down by about 1%[@wirtz2004a]. We use here the experimental lattice constant of pure graphene (for both the isolated graphene and the buffer layer). In this case, the phonon frequencies are a little bit lower than the experimental ones. We thus rescale the calculated phonon dispersions (of both isolated and buffer graphene) by the respective ratio of the experimental and theoretical values of the E$_{2g2}$ (highest optical mode at $\Gamma$) phonon frequencies. For isolated graphene the Raman G-line has the value of 1580 cm$^{-1}$ (according to recent measurements on suspended graphene[@berciaud2009a]) and our calculated value is 1568 cm$^{-1}$. For the buffer layer on SiC(0001), the experimental value (measured by electron energy loss spectroscopy [@koch2013]) is 1595 cm$^{-1}$ and our calculated value is 1558 cm$^{-1}$. These differences are related to the unknown strain state of the buffer layer.
{ "pile_set_name": "ArXiv" }
--- abstract: | We consider a transmission problem in which the interior domain has infinitely ramified structures. Transmission between the interior and exterior domains occurs only at the fractal component of the interface between the interior and exterior domains. We also consider the sequence of the transmission problems in which the interior domain is obtained by stopping the self-similar construction after a finite number of steps; the transmission condition is then posed on a prefractal approximation of the fractal interface. We prove the convergence in the sense of Mosco of the energy forms associated with these problems to the energy form of the limit problem. In particular, this implies the convergence of the solutions of the approximated problems to the solution of the problem with fractal interface. The proof relies in particular on an extension property.\ Emphasis is put on the geometry of the ramified domain. The convergence result is obtained when the fractal interface has no self-contact, and in a particular geometry with self-contacts, for which an extension result is proved. author: - '[^1], [^2]' bibliography: - 'biblio.bib' nocite: '[@MR3108826]' --- Introduction {#Introduction} ============ Transmission problems naturally arise in various fields of physics and have been extensively studied. An introduction to this class of problems can be found in [@MR0400916] and several applications are detailed in [@MR969367].\ Such problems were more recently studied in the case when the interface is irregular, Lipschitz continuous or even fractal. These problems find many applications, such as the study of rough electrodes in electrochemistry or diffusion across irregular membranes in physiological processes, etc. (see [@PhysRevLett.73.3314; @PhysRevLett.84.5776]). Several transmission problems with fractal interfaces have been studied in the case of the Koch flake or the Sierpiński gasket in 2D and 3D (see *e.g.* [@MR1916407; @MR2056421; @MR2323380; @MR2679587; @MR2645992; @MR2557700; @MR3100114; @MR3100076]). This paper deals with transmission problems between two domains ${\Omega_\mathrm{int}}$ and ${\Omega_\mathrm{ext}}$ where ${\Omega_\mathrm{int}}$ is a ramified bounded domain as defined in Section \[domains\] (see Figure \[fig-domram\]). The domain ${\Omega_\mathrm{int}}$ presents infinitely many ramifications, and its boundary contains a fractal self-similar set $\Gamma$ which plays the role of the interface.\ The domain ${\Omega_\mathrm{int}}$ can be seen as a bidimensional idealization of the bronchial tree, for example. Since the exchanges between the lungs and the circulatory system take place only in the last generations of the bronchial tree (the smallest structures), it is reasonable to consider transmission problems with specific transmission conditions accross the fractal boundary $\Gamma$ of the tree. We will however limit ourselves to simple transmission conditions.\ The fractal boundary $\Gamma$ belongs to a family of self-similar sets introduced by Mandelbrot *et al.* in [@MR1684366]. Elliptic boundary value problems in the domain ${\Omega_\mathrm{int}}$ have been studied in [@ach-tch-2007-1], and traces and extension results for these domains have been proved in [@ADT; @SEP].\ The considered problem can be formally stated as $$\label{pb-transmission} \tag{$P$} \left\{ \begin{array}{ll} -\Delta u = f & \mathrm{in}~ {\Omega_\mathrm{int}}\cup{\Omega_\mathrm{ext}},\\[1mm] [u]=0 &\text{on}~ \Gamma,\\[1mm] [\partial_n u]=\alpha u &\text{on}~ \Gamma,\\[1mm] \partial_n {u_\mathrm{int}}=\partial_n {u_\mathrm{ext}}= 0 & \mathrm{on}~ \partial {\Omega_\mathrm{int}}\setminus \Gamma,\\[1mm] \partial_n {u_\mathrm{ext}}= 0,~~ {u_\mathrm{int}}= u_0 & \mathrm{on}~ \Gamma^0,\\[1mm] \partial_n {u_\mathrm{ext}}= 0 & \mathrm{on}~ \partial D, \end{array} \right.$$ where $\alpha$ is a positive real number, $D$ is a regular bounded open domain in the plane with $\overline{D} = \overline{{\Omega_\mathrm{int}}}\cup \overline{{\Omega_\mathrm{ext}}}$ such that $\Gamma \Subset D$, and $\Gamma^0$ is a line segment included in the boundary of ${\Omega_\mathrm{int}}$. The sets ${\Omega_\mathrm{int}}$ and ${\Omega_\mathrm{ext}}$ are disjoint subdomains of $D$, and $[u]$ (*resp.* $[\partial_n u]$) denotes the jump of $u$ (*resp.* of the “normal derivative" $\partial_n u$ of $u$) accross the fractal set $\Gamma$. Since the interface $\Gamma$ is fractal, the normal derivative on $\Gamma$ has to be understood in a suitable weak sense, which will be made precise later.\ Problem is a model problem. Its study is the first step in the modelling of physical transmission problems in ramified structures.\ The goal is to study approximations $(P_n)$ of problem , obtained by stopping the construction of the ramified domain ${\Omega_\mathrm{int}}$ at step $n$. The interfaces in the problems $(P_n)$ are called prefractal approximations of the fractal set $\Gamma$. They consist of disjoint finite unions of line segments, which makes problems $(P_n)$ much simpler than problem $(P)$. A natural question is to understand the asymptotic behavior of the problems $(P_n)$ as $n\to \infty$, and in particular to investigate the convergence of the solutions $u_n$ of the problems $(P_n)$ to the solution $u$ of problem $(P)$. \[sec:introduction-1\] In contrast with the references [@MR1916407; @MR2056421; @MR2323380; @MR2679587; @MR2645992; @MR2557700; @MR3100114; @MR3100076], the boundary value problem does not only involve transmission conditions at the interface between ${\Omega_\mathrm{int}}$ and ${\Omega_\mathrm{ext}}$: there are also homogeneous Neumann conditions on the polygonal part of $\partial {\Omega_\mathrm{int}}$; as a consequence, the traces of ${u_\mathrm{int}}$ and ${u_\mathrm{ext}}$ on this set do not match a priori. Coping with these discontinuities will be a difficulty in studying the convergence of the solutions $u_n$ of the problems $(P_n)$ to the solution $u$ of problem $(P)$. \[sec:introduction\] We have chosen that the source term in ($P$) appear both in the Dirichlet boundary condition on $\Gamma_0$ for ${u_\mathrm{int}}$ and in the Poisson equations in ${\Omega_\mathrm{int}}$ and ${\Omega_\mathrm{ext}}$; this is of course completely arbitrary. A crucial step in the study of the asymptotic behavior of $(P_n)$ is the question of extending functions defined in the domain ${\Omega_\mathrm{int}}$. More precisely, it is of particular importance that ${\Omega_\mathrm{int}}$ should satisfy a $W^{1,p}$-extension property for some $p\in [1,\infty]$, *i.e.* there should exist a bounded linear operator $${{\mathcal E}}: W^{1,p}({\Omega_\mathrm{int}})\to W^{1,p}({{\mathbb R}}^2)$$ such that ${{{\mathcal E}}(u)}_{|{\Omega_\mathrm{int}}} = u$ for all $u\in W^{1,p}({\Omega_\mathrm{int}})$. The domains satisfying this property for all $p\in [1,\infty]$ will often be referred to as [*Sobolev extension domains*]{}.\ It is well known that every Lipschitz domain in ${{\mathbb R}}^n$, that is every domain whose boundary is locally the graph of a Lipschitz function, is a Sobolev extension domain. Calder[ó]{}n proved the $W^{1,{p}}$-extension property for ${p}\in (1,\infty)$ (see [@MR0143037]), and Stein extended the result to the cases ${p}=1$ and ${p}=\infty$ (see [@MR0290095]).\ In [@MR631089], Jones generalized this result to the class of $({\varepsilon},\delta)$-domains, also referred to as Jones domain, or locally uniform domains (see [@MR565886]). In dimension two, the definition of $({\varepsilon},\delta)$-domains is equivalent to that of quasi-disks, see [@MR817985]. This extension result is almost optimal in the plane, in the sense that every plane finitely connected Sobolev extension domain is an $({\varepsilon},\delta)$-domain, see [@MR631089; @MR817985]. The case of an unbounded domain in ${{\mathbb R}}^2$ has been studied in [@MR530508]. Extension properties for domains which do not have the $({\varepsilon},\delta)$-property have been studied *e.g.* in [@MR1643072], where the authors examine in particular the case of domains with cusps.\ When the fractal boundary $\Gamma$ of the domains ${\Omega_\mathrm{int}}$ studied in this paper has no self-contact, ${\Omega_\mathrm{int}}$ can be proved to be an $({\varepsilon},\delta)$-domain, and and is therefore a Sobolev extension domain. However, in the case when the boundary self-intersects, ${\Omega_\mathrm{int}}$ does not have the $W^{1,p}$-extension property for all $p{\geqslant}1$. In particular, the extension property does not hold for $p=2$, which is the relevant case here, since the variational formulations of (\[pb-transmission\]) naturally involve the spaces $H^1({\Omega_\mathrm{int}})$. In the particular geometry considered in Section \[r=rstar\], it will be proved that the transmission condition imposed on $\Gamma$ yields a better regularity of the trace on $\Gamma$, for functions belonging to the function space arising in the variational formulation. It is then possible to deduce an extension result in this case (see Theorem \[ext-thm\]).\ The main question investigated in this paper is the question of the convergence in the sense of Mosco of the energy forms associated with problem $(P_n)$ to the energy form of the problem $(P)$. The notion of Mosco-convergence, or $M$-convergence, was introduced in [@MR0298508], see also [@MR1283033]. It is a stronger convergence in the space of forms than $\Gamma$-convergence. In particular, it also implies the convergence of minimizers of the energy forms to the minimizer of the limit form. The $M$-convergence of forms is equivalent to the convergence of the resolvent operators associated with the relaxed forms in the strong operator topology (see [@MR1283033]). The main results of this paper are Theorems \[M-cv\] and \[ext-thm\]. Theorem \[M-cv\] is about the convergence of the energy forms associated with $(P_n)$ in the sense of Mosco to the energy form associated with , in the case when the fractal interface has no self-contact. The proof uses the extension operator from $H^1({\Omega_\mathrm{int}})$ to $H^1({{\mathbb R}}^2)$ as a main ingredient. The existence of a continuous extension operator from $H^1({\Omega_\mathrm{int}})$ to $H^1({{\mathbb R}}^2)$, for a particular geometry where the fractal interface self-intersects, is stated in Theorem \[ext-thm\]. As a consequence, the proof of Theorem \[M-cv\] can be reproduced in this case, to show the $M$-convergence of the energy forms. The paper is organized as follows: the geometry of the interior and exterior domains is detailed in Section \[sec-geometry\], as well as the prefractal geometry. Section \[sec-function-spaces\] is devoted to the study of the function spaces involved in the paper, and emphasis is put on trace and extension results for the domains under study. The considered transmission problem is described in Section \[sec-transmission-pb\]. Section \[sec-convergence\] is devoted to the $M$-convergence of the energy forms associated with the problem with prefractal interface to the energy form of the problem with fractal interface, in the case when the boundary of ${\Omega_\mathrm{int}}$ has no self-contact. In Section \[r=rstar\], a particular geometry in which the fractal part of the boundary of ${\Omega_\mathrm{int}}$ self-intersects is considered; an extension result is proved in this particular case and the $M$-convergence of the energy forms follows. The geometry {#sec-geometry} ============ The fractal interface {#fractal-interface} --------------------- ### Definitions {#def} Consider four real numbers ${r}, \beta_1,\beta_2, \theta$ such that $1/2{\leqslant}{r}< 1/\sqrt{2}$, $\beta_1>0$, $\beta_2>0$ and $0 {\leqslant}\theta<\pi/2$. Let $f_i$, $i=1,2$ be the two similitudes in ${{\mathbb R}}^2$ given by $$ f_1 \begin{pmatrix} x_1\\x_2 \end{pmatrix} ~=~ \begin{pmatrix} -\beta_1 \\ \beta_2 \end{pmatrix} +{r}\begin{pmatrix} x_1 \cos \theta- x_2 \sin \theta\\x_1 \sin \theta+ x_2 \cos \theta \end{pmatrix},$$ $$f_2 \begin{pmatrix} x_1\\ x_2 \end{pmatrix} ~=~ \begin{pmatrix} \beta_1\\ \beta_2 \end{pmatrix} +{r}\begin{pmatrix} x_1 \cos \theta + x_2 \sin \theta\\-x_1 \sin \theta+ x_2 \cos \theta \end{pmatrix}.$$ The two similitudes have the same dilation ratio ${r}$ and opposite angles $\pm \theta$. One can obtain $f_2$ by composing $f_1$ with the symmetry with respect to the vertical axis $\{x_1=0\}$.\ Let $\Gamma$ denote the self-similar set associated with the similitudes $f_1$ and $f_2$, *i.e.* $\Gamma$ is the unique compact subset of ${{\mathbb R}}^2$ such that $$\Gamma = f_1(\Gamma)\cup f_2(\Gamma).$$ It was stated in [@MR1684366] (see [@deheuvelsphd] for a complete proof) that for any $\theta$, $0{\leqslant}\theta<\pi/2$, there exists a unique positive number ${r}^\star_\theta\in [1/2, 1 /{\sqrt 2}[$ which only depends on the angle $\theta$ such that $$\label{SEP:eq:10} \begin{array}[c]{ll} \diamond~~ \hbox{if}~ 0<{r}<{r}^\star_\theta, &\hbox{then}~ \Gamma ~\hbox{is totally disconnected,} \\ \diamond~~\hbox{if}~ {r}={r}^\star_\theta, &\hbox{then}~ \Gamma \hbox{ is connected}. \end{array}$$ In the following, we will always assume that ${r}{\leqslant}{r}^\star_\theta$. #### Notations For every integer $n>0$, we note ${{\mathcal A}}_n=\{1,2\}^n$. For $\sigma\in {{\mathcal A}}_n$, we note $f_\sigma$ the similitude $f_{\sigma_1}\circ\ldots \circ f_{\sigma_n}$. We agree to extend the notation to the case $n=0$: $f_\sigma=\text{Id}$ if $\sigma\in {{\mathcal A}}_0$. We also introduce the notation ${{\mathcal A}}:= \bigcup_{n{\geqslant}0} {{\mathcal A}}_n$.\ For $\sigma\in {{\mathcal A}}$, we note $\Gamma^\sigma = f_\sigma(\Gamma^0)$, and for every integer $n{\geqslant}0$, $\displaystyle \Gamma^n = \bigcup_{\sigma\in {{{\mathcal A}}_n}} f_\sigma(\Gamma^0)$. ### Hausdorff dimension of $\Gamma$ {#dimension-gamma} If ${r}{\leqslant}{r}^\star_\theta$, then it can be seen that the open set condition (or Moran condition) holds, see [@MR0014397] or [@MR1840042] for a definition. The open set condition is satisfied *e.g.* for the domain ${\Omega_\mathrm{int}}$ defined in , if Assumption \[assumption1\] below is satisfied (Theorem \[exist-ram-dom\] proves the existence of such a domain).\ The open set condition implies that the Hausdorff dimension of $\Gamma$ is $$d:= \dim_H \Gamma = -\frac{\log 2}{\log {r}}$$ see [@MR0014397; @MR1840042]. If $0{\leqslant}\theta<\pi/2$, then $1/2 {\leqslant}{r}{\leqslant}{r}^\star_\theta<1/\sqrt{2}$ and thus $1{\leqslant}d <2$.\ In the case when ${r}={r}^\star_\theta$, introduce the set $$\label{Xi} \Xi = f_1(\Gamma)\cap f_2(\Gamma).$$ In this case, the fractal set $\Gamma$ self-intersects, and union of the images of $\Xi$ by the similitudes $f_{\sigma_1}\circ \ldots \circ f_{\sigma_n}$, $\sigma_1,\ldots,\sigma_n \in \{1,2\}$ is the set of the multiple points of $\Gamma$.\ Two situations can occur, depending on the angle $\theta$ (see [@MR1684366]): [–]{}[ ]{} if $\theta$ is not of the form $\frac\pi{2k}$ for any integer $k>0$, then $\Xi$ is reduced to a single point, and $\Gamma$ has countably many multiple points, if $\theta$ is of the form $\frac\pi{2k}$ where $k>0$ is an integer, then $\Xi$ is a Cantor set, and the Hausdorff dimension of $\Xi$, noted $\dim_H \Xi$, is $\frac{\dim_H \Gamma}2$. ### The self-similar measure $\mu$ Recall the classical result on self-similar measures, see [@MR1449135; @MR625600] and [@MR1840042] page 26. There exists a unique Borel regular probability measure $\mu$ on $\Gamma$ such that for any Borel set $A\subset\Gamma$, $$\label{SEP:eq:18} \mu(A)=\frac 1 2 \mu\left(f_{1}^{-1}(A)\right)+\frac 1 2 \mu \left(f_{2}^{-1}(A)\right).$$ The measure $\mu$ is called the [*self-similar measure defined in the self-similar triplet $ \left(\Gamma, f_1,f_2\right)$*]{}.\ Let $L^{p}_\mu(\Gamma)$, ${p}\in [1,+\infty)$ be the space of the measurable functions $v$ on $\Gamma$ such that $\int_{\Gamma}|v|^{p}{\;\text{d}}\mu<\infty$, endowed with the norm $\Vert v\Vert_{L^{p}(\Gamma)}= \left(\int_{\Gamma}|v|^{p}{\;\text{d}}\mu\right)^{1/p}$. A Hilbert basis of $L^2_\mu(\Gamma)$ can be constructed *e.g.* with Haar wavelets.\ The space $W^{s,{p}}(\Gamma)$ for $s\in (0,1)$ and ${p}\in [1,\infty)$ is defined as the space of functions $v\in L^{p}(\Gamma)$ such that ${\vert v \vert}_{W^{s,{p}}(\Gamma)} < \infty$, where $${\vert v \vert}_{W^{s,{p}}(\Gamma)} ~=~ {\left( \int_{\Gamma} \int_{\Gamma} \frac{{\vert v(x)-v(y) \vert}^{p}}{{\vert x-y\vert}^{d+{p}s}} {\;\text{d}}\mu(x) {\;\text{d}}\mu(y) \right)}^\frac 1 {p}.$$ Endowed with the norm ${\Vert v \Vert}_{W^{s,{p}}(\Gamma)} = {\Vert v \Vert}_{L^{p}(\Gamma)} + {\vert v \vert}_{W^{s,{p}}(\Gamma)} $, the spaces $W^{s,p}(\Gamma)$ are Banach spaces. In the special case $p=2$, the space $W^{s,p}(\Gamma)$ is a Hilbert space, and is noted $H^s(\Gamma)$. \[norme-Lip\] In the special case when $\theta=0$ and $r=r^\star_\theta=1/2$, the set $\Gamma$ is in fact a line segment. This geometry will be discussed in Section \[r=rstar\] (see Figure \[trapeze\]). In this case, it can be proved that if $s\in (0,1)$, then an equivalent norm of the space $W^{s,p}(\Gamma)$ is given by $$\label{H-Lip} {\Vert v \Vert}_{\mathrm{Lip}_s^{p,p}(\Gamma)}^p ~:= ~ \int_\Gamma {|v|}^p {\;\text{d}}\mu + \sum_{k{\geqslant}0} 2^{skp} \sum_{\sigma\in {{\mathcal A}}_k} \int_{2f_\sigma (\Gamma)} {|v-{\langle v \rangle}_{2f_\sigma(\Gamma)}|}^p{\;\text{d}}x$$ where $2f_\sigma(\Gamma)$ is the intersection with $\Gamma$ of the segment obtained by expanding the line segment $f_\sigma(\Gamma)$ with a factor 2 around its center (see [@MR2032227; @MR820626]). As in the rest of the paper, if $v$ is a measurable function in a measured space $(X,m)$, the notation $\langle v \rangle_X$ refers to the mean value $\frac 1{m(X)} \int_X v {\;\text{d}}m$. The domains ${\Omega_\mathrm{int}}$ and ${\Omega_\mathrm{ext}}$ {#domains} --------------------------------------------------------------- Call $P_1=(-1,0)$ and $P_2=(1,0)$ and $\Gamma^0$ the line segment $\Gamma^0=[P_1 P_2]$. Let us assume that $f_2(P_1)$ and $f_2(P_2)$ have positive coordinates, *i.e.* that $$\label{cond1} {r}\cos\theta <\beta_1 ~ \text{ and } ~ {r}\sin\theta < \beta_2.$$ Let us also assume that the open domain $Y^0$ inside the closed polygonal line joining the points $ P_1$, $P_2$, $f_2(P_2)$, $f_2(P_1)$, $f_1(P_2)$, $f_1(P_1)$, $P_1$ in this order must be convex and hexagonal, except if $\theta=0$, in which case it is trapezoidal. With (\[cond1\]), this is true if and only if $$\label{cond2} (\beta_1-1)\sin\theta +\beta_2 \cos\theta > 0.$$ Under assumptions (\[cond1\]) and (\[cond2\]), the domain $Y^0$ is contained in the half-plane $x_2>0$ and symmetric with respect to the vertical axis $x_1=0$.\ Call $K^0= \overline{ Y^0}$. It is possible to glue together $K^0$, $f_1(K^0)$ and $f_2(K^0)$ and obtain a new polygonal domain. The assumptions and imply that $Y^0\cap f_1( Y^0)=\emptyset$ and $ Y^0\cap f_2( Y^0)=\emptyset$. Let the open domain ${\Omega_\mathrm{int}}$ (see Figure \[fig-domram\]) be defined as follows: $$\label{def-dom-ram} {\Omega_\mathrm{int}}= \hbox{Interior} \left(\bigcup_{\sigma\in {{\mathcal A}}} f_\sigma (K^0)\right),$$ with the notations of §\[def\].\ For a given $\theta$, with ${r}^\star_\theta$ defined as above, the following assumption on $(\alpha,\beta)$ will be made: \[assumption1\] For $0{\leqslant}\theta<\pi/2$, the parameters $\alpha$ and $\beta$ satisfy and for ${r}={r}^\star_\theta$, and are such that $$\left\{ \begin{array}[c]{l} \hbox{\textit{i}. for all ${r}$, $0<{r}{\leqslant}{r}^\star_\theta$, the sets $Y^0$, $f_\sigma ({ Y^0})$, $\sigma\in {{\mathcal A}}$, are pairwise disjoint,}\\ \hbox{\textit{ii}. for all ${r}$, $0<{r}< {r}^\star_\theta$, $f_1(\overline {\Omega}_\text{int})\cap f_2(\overline {\Omega}_\text{int})=\emptyset$,}\\ \hbox{\textit{iii}. for ${r}={r}^\star_\theta$, $f_1(\overline{\Omega}_\text{int}) \cap f_2(\overline{\Omega}_\text{int})\not=\emptyset$.} \end{array} \right.$$ Assumption \[assumption1\] implies that if $r=r^\star_\theta$, then $f_1( {{\Omega_\mathrm{int}}})\cap f_2( {{\Omega_\mathrm{int}}})=\emptyset$. In the case $\theta=0$, Assumption \[assumption1\] is satisfied by any $\alpha>{r}^\star_\theta=1/2$ and $\beta>0$. The following theorem, proved in [@YANT2010], asserts that for all $\theta\in (0,\pi/2)$, there exists $(\alpha,\beta)$ satisfying Assumption 1. \[exist-ram-dom\] If $\theta\in (0, \pi/2)$, then for every $\alpha>{r}^\star_\theta \cos \theta$, there exists $\bar \beta>0$ such that for all $\beta{\geqslant}\bar \beta$, $(\alpha,\beta)$ satisfies Assumption \[assumption1\]. Let $D$ be an open bounded domain with a Lipschitz boundary, containing the closure of ${\Omega_\mathrm{int}}$. The exterior domain ${\Omega_\mathrm{ext}}$ is defined by $${\Omega_\mathrm{ext}}:= \text{Interior} (D\setminus {\Omega_\mathrm{int}}).$$ The assumption that $ {\Omega_\mathrm{int}}\Subset D$ may be relaxed: in fact, it would be enough to assume that $\Gamma \Subset D$. Displayed on Figure \[fig-domram\] are examples of the domains ${\Omega_\mathrm{int}}$ and ${\Omega_\mathrm{ext}}$, for the parameters $\theta=\pi/5$ in the left-hand side and $\theta=\pi/4$ in the right-hand side. $$\scalebox{0.45}{\PandocStartInclude{fig1.pstex_t}\PandocEndInclude{input}{507}{37}} ~~~ \scalebox{0.4663}{\PandocStartInclude{fig2.pstex_t}\PandocEndInclude{input}{509}{39}}$$ The truncated domain ${\Omega_\mathrm{int}}^n$ and the prefractal interface --------------------------------------------------------------------------- For every integer $n{\geqslant}0$, the truncated domain ${\Omega_\mathrm{int}}^n$ is defined by $${\Omega_\mathrm{int}}^n = \text{Interior}\left(\bigcup_{0{\leqslant}k{\leqslant}n} \bigcup_{\sigma \in {{{\mathcal A}}_k}} f_\sigma(K^0)\right),$$ with the notations of §\[def\]. As above, the exterior domain associated to ${\Omega_\mathrm{int}}^n$ is $${\Omega_\mathrm{ext}}^n = \text{Interior}(D\setminus {\Omega_\mathrm{int}}^n).$$ Note that the set $\Gamma^n$ defined in §\[def\] is a part of the boundary of ${\Omega_\mathrm{int}}^n$. The sets $\Gamma^n$, $n{\geqslant}0$, will be referred to as *prefractal approximations* of the fractal set $\Gamma$. Function spaces {#sec-function-spaces} =============== Hereafter, we consider a domain ${\Omega_\mathrm{int}}$ as defined in \[domains\], with $\theta$ in $[0,\pi/2)$ and ${r}{\leqslant}{r}^\star_\theta$, and we assume that the parameters $\alpha$, $\beta$ are such that Assumption \[assumption1\] is satisfied.\ We define $W^{1,p}(\Omega)$, $p=[1,\infty]$, $\Omega = {\Omega_\mathrm{int}}$ or $\Omega={\Omega_\mathrm{ext}}$, to be the space of functions in $L^p(\Omega)$ with first order partial derivatives in $L^p(\Omega)$.\ The space $W^{1,p}(\Omega)$ is a Banach space with the norm $ \left(\|u\|_{L^p(\Omega)}^p +\| \frac {\partial u}{\partial x_1 }\|_{L^p(\Omega)}^p +\| \frac {\partial u}{\partial x_2 }\|_{L^p(\Omega)}^p \right)^{1/p}$, see for example [@MR2424078]. Elementary calculus shows that $ \|u\|_{W^{1,p}(\Omega)} := \left(\|u\|_{L^p(\Omega)}^p +\| \nabla u\|_{L^p(\Omega)}^p \right)^{1/p} $ is an equivalent norm, with $\| \nabla u\|^p_{L^p(\Omega)}:= \int_{\Omega} |\nabla u|^p $ and $|\nabla u|=\sqrt{|\frac {\partial u}{\partial x_1}|^2 + |\frac {\partial u}{\partial x_2}|^2}$.\ In the special case $p=2$, the space $W^{1,p}(\Omega)$ is a Hilbert space, and is noted $H^1(\Omega)$.\ The spaces $W^{1,p}({\Omega_\mathrm{int}})$ as well as elliptic boundary value problems in ${\Omega_\mathrm{int}}$ have been studied in [@ach-tch-2007-1], with, in particular Poincar[é]{} inequalities and a Rellich compactness theorem. The same results in a similar but different geometry were proved by Berger [@MR1800198] with other methods. Trace results ------------- ### The classical definition of traces We recall the classical definition of a trace operator on $\partial \omega$ when $\omega$ is an open subset of ${{\mathbb R}}^2$ (see for instance [@MR820626] p. 206). \[def-trace\] Consider an open set $\omega\subset {{\mathbb R}}^2$. The function $u\in L^1_{loc}(\omega)$ can be strictly defined at $x\in\overline \omega$ if the limit $$\label{eq:25} {\overline{u}}(x)=\lim_{r\rightarrow 0}\frac{1}{|B(x,r)\cap \omega|}\int_{B(x,r)\cap \omega}u(y) {\;\text{d}}y$$ exists, where $|B(x,r)\cap \omega|$ is the 2-dimensional Lebesgue measure of the set $B(x,r)\cap \omega$. In this case, $x$ is said to be a Lebesgue point of $u$.\ The trace $u_{|\partial \omega}$ is defined to be the function given by $u_{|\partial \omega}(x)= {\overline{u}}(x)$ at every point $x\in \partial \omega$ such that the limit ${\overline{u}}(x)$ exists. Recall that for any $p>1$, a function which belongs to $W^{1,p}({{\mathbb R}}^n)$ can be strictly defined except on a set with zero $p$-capacity, see for example [@MR0435361] and [@MR1404091]. ### A trace theorem on $\Gamma$ It has been shown in [@comparison] (see Theorem 11) that every function in $W^{1,p}({\Omega_\mathrm{int}})$ can be strictly defined on $\Gamma$ ${{\mathcal H}}^1$-almost everywhere, where ${{\mathcal H}}^1$ is the one-dimensional Hausdorff measure. Moreover, the following trace result holds. (see [@ADT])\[trace-thm\] [ ]{} Assume ${r}<{r}^\star_\theta$. For all $p\in ]1,\infty]$, if $u\in W^{1,p}({\Omega_\mathrm{int}})$, then $u_{|\Gamma}\in W^{1-\frac{2-d}p,p}(\Gamma)$, and there exists a constant $C>0$ independent of $u$ such that $${\|u_{|\Gamma}\|}_{W^{1-\frac{2-d}p,p}(\Gamma)} {\leqslant}C {\|u\|}_{W^{1,p}({\Omega_\mathrm{int}})}.$$ Assume ${r}={r}^\star_\theta$, then [$\diamond$]{}[ ]{} the previous result holds if $1<p<2-\dim_H \Xi$, if $p{\geqslant}2-\dim_H \Xi$, then $W^{1,p}({\Omega_\mathrm{int}})_{|\Gamma} \subset W^{s,p}(\Gamma) ~ \text{for all } s<\frac 1 p (d-\dim_H \Xi),$ and the embedding is continuous. Moreover, if $s>\frac 1 p (d-\dim_H \Xi)$, then ${W^{1,p}({\Omega_\mathrm{int}})}_{|\Gamma} \not\subset W^{s,p}(\Gamma)$. \[sec:trace-theorem-gamma\] The space of the traces on $\Gamma$ of functions in $W^{1,p}({\Omega_\mathrm{int}})$, $1<p<\infty$ was characterized in [@YANT2010], whether $r< r^\star_\theta$ or $r=r^\star_\theta$, as the space ${ \rm{JLip}}(1-\frac{2-d} p,p,p;0; \Gamma)$, which was first introduced in [@MR2032227]. Of course, if $r< r^\star_\theta$, then ${\rm{JLip}}(1-\frac{2-d} p,p,p;0; \Gamma)$ coincides with $W^{1-\frac{2-d} p,p}(\Gamma)$. An easy consequence of this characterization is that the space of the traces on $\Gamma$ of functions in $W^{1,p}({\Omega_\mathrm{int}})$ is relatively compact in $L^p_\mu (\Gamma)$. Proposition \[convergence-result\] below will be useful in the proofs of the main theorems of this paper. \[convergence-result\] For every $u\in H^1({\Omega_\mathrm{int}})$, $$\frac 1 {|\Gamma^n|} \int_{\Gamma^n} {u_{|\Gamma^n}}^2{\;\text{d}}x \underset{n\to \infty}{\longrightarrow} \int_\Gamma {u_{|\Gamma}}^2{\;\text{d}}\mu.$$ The present proof relies on Proposition 1 in [@ach-tch-2007-1], which states that for any $u\in H^1({\Omega_\mathrm{int}})$, the sequence of piecewise constant functions $(\tilde u_n)_{n\in {{\mathbb N}}}$ defined on $\Gamma$: $$\tilde u_n =\sum_{\sigma\in {{\mathcal A}}_n} \left\langle u \right\rangle_{\Gamma^\sigma} {\mathbbm{1}}_{f_\sigma( \Gamma) } ,$$ where $ \left\langle u \right\rangle_{\Gamma^\sigma}= \frac {1} {|\Gamma^\sigma|} \int_{\Gamma^\sigma} u_{|\Gamma^\sigma}(z) {\;\text{d}}z$, is such that $\lim_{n\to \infty} \|\tilde u_n -u_{|\Gamma} \|_{L_\mu ^2(\Gamma)}=0$. Note also that $ \int_{\Gamma} \tilde u_n ^2 d\mu= \sum_{\sigma\in {{\mathcal A}}_n} \frac {| \Gamma^\sigma|}{|\Gamma^n|} \left\langle u \right\rangle_{\Gamma^\sigma} ^2 $. Hence, in order to prove Proposition \[convergence-result\], it is enough to prove that\ $ \left| \frac 1 {|\Gamma^n|} \int_{\Gamma^n} {u_{|\Gamma^n}}^2{\;\text{d}}x - \int_{\Gamma} \tilde u_n ^2 d\mu\right| \underset{n\to \infty}{\longrightarrow}0$, or in an equivalent manner, that $ S_n \underset{n\to \infty}{\longrightarrow}0$, where $$S_n= \frac 1 {|\Gamma^n|} \sum_{\sigma\in {{\mathcal A}}_n} \left( \int_{\Gamma^\sigma } u^2_{|\Gamma^\sigma}(z) {\;\text{d}}z - | \Gamma^\sigma| \left\langle u \right\rangle_{\Gamma^\sigma} ^2 \right) = \frac 1 {|\Gamma^n|} \sum_{\sigma\in {{\mathcal A}}_n} \int_{\Gamma^\sigma} \left[ u_{|\Gamma^\sigma}(z)- \left\langle u \right\rangle_{\Gamma^\sigma} \right ]^2 {\;\text{d}}z.$$ From a standard trace result on $\Gamma^0$ and appropriate rescalings, we know that for a positive constant independent of $n$, $\sigma\in {{\mathcal A}}_n$ and $u\in H^1 ({\Omega_\mathrm{int}})$, $\int_{\Gamma^\sigma} \left[ u_{|\Gamma^\sigma}(z)- \left\langle u \right\rangle_{\Gamma^\sigma} \right ]^2 {\;\text{d}}z \le C |\Gamma^\sigma| \int_{f_\sigma( {\Omega_\mathrm{int}})} |\nabla u|^2$. Hence, $ S_n\le C \sum_{\sigma\in {{\mathcal A}}_n} \frac {|\Gamma^\sigma|}{|\Gamma^n|} \int_{f_\sigma( {\Omega_\mathrm{int}})} |\nabla u|^2= \frac C{2^n} \sum_{\sigma\in {{\mathcal A}}_n} \int_{f_\sigma( {\Omega_\mathrm{int}})} |\nabla u|^2$, which implies that $S_n \underset{n\to \infty}{\longrightarrow}0$. We also recall the following refined trace inequality, we refer to [@YANT2010] for the proof. \[SPI\] \[see [@YANT2010], Th. 11\] Assume that ${r}{\geqslant}1/2$, then for all real number $\kappa \in (2{r}^2,1)$, there exists a constant $C$ such that for all $v\in H^1({\Omega_\mathrm{int}})$ with $v_{|\Gamma^0}=0$, $${\Vert v_{|\Gamma}\Vert}_{L^2_\mu(\Gamma)}^2 ~{\leqslant}~ C \sum_{m{\geqslant}0} \kappa^m \sum_{\tau \in {{\mathcal A}}_n} {\Vert \nabla v\Vert}^2_{L^2(f_\tau(Y^0))}.$$ Extension results ----------------- ### The subcritical case ${r}<{r}^\star_\theta$ {#ext-subcritical} As seen in § \[Introduction\], it was proved in [@YANT2008] that if ${r}<{r}^\star_\theta$, then ${\Omega_\mathrm{int}}$ is an $({\varepsilon},\delta)$-domain (see [@MR631089] for a definition), or in an equivalent manner, a quasi-disk (see [@MR817985]). Hence, the extension result of Jones and Vodop’janov [*et al.*]{} applies, and ${\Omega_\mathrm{int}}$ is a Sobolev extension domain (see [@MR631089]), *i.e.* ${\Omega_\mathrm{int}}$ has the $W^{1,p}$-extension property for every $p\in [1,\infty]$: there exists a continuous linear operator ${{\mathcal E}}$ from $W^{1,p}({\Omega_\mathrm{int}})$ to $W^{1,p}({{\mathbb R}}^2)$ such that $$\label{ext-op} {{\mathcal E}}(u)_{|{\Omega_\mathrm{int}}}=u, \quad \forall u\in W^{1,p}({\Omega_\mathrm{int}}).$$ Similarly, the set ${\Omega_\mathrm{ext}}= D \setminus \overline \Omega_\text{int}$ is an $({\varepsilon},\delta)$-domain, and thus a Sobolev extension domain. ### The critical case ${r}={r}^\star_\theta$ When ${r}={r}^\star_\theta$, it is easily seen that ${\Omega_\mathrm{int}}$ is not an $({\varepsilon},\delta)$-domain, and the extension results of Jones and Vodop’janov [*et al.*]{} do not apply. In fact, if $p\in (1,\infty)$, it is easy to construct a function $u\in W^{1,p}({\Omega_\mathrm{int}})$ such that $u\equiv 1$ in $f_1({\Omega_\mathrm{int}})$ and $u\equiv -1$ in $f_2({\Omega_\mathrm{int}})$. If $p>2$, $u$ cannot be extended to a function belonging to $W^{1,p}({{\mathbb R}}^2)$ because the existence of such an extension would contradict the Sobolev imbedding of $W^{1,p}({{\mathbb R}}^2)$ in $C({{\mathbb R}}^2)$.\ In the case when ${r}={r}^\star_\theta$, the situation depends in fact on the Hausdorff dimension of the set $\Xi=f_1(\Gamma)\cap f_2(\Gamma)$. The following extension theorem holds. \[ext-theorem\] see [@SEP; @comparison] \[extension\] Set $p^\star= 2-\dim_H \Xi$ (recall that $\Xi$ is defined in ). 1. If $p\in (1,p^\star)$, then ${\Omega_\mathrm{int}}$ has the $W^{1,p}$-extension property. 2. If $p>p^\star$, then ${\Omega_\mathrm{int}}$ does not have the $W^{1,p}$-extension property. Point 1 in Theorem \[ext-theorem\] was obtained in [@SEP]. Point 2 is a consequence of [@ADT] and [@comparison]: by Theorem \[trace-thm\], if $p>p^\star$, then ${W^{1,p}({\Omega_\mathrm{int}})}_{|\Gamma}\not\subset W^{1-\frac{2-d}p,p}(\Gamma) = {W^{1,p}({{\mathbb R}}^2)}_{|\Gamma}$. This is in contradiction with the existence of a continuous extension operator from $W^{1,p}({\Omega_\mathrm{int}})$ to $W^{1,p}({{\mathbb R}}^2)$ (see [@ADT] for the proof that the notions of traces coincide $\mu$-almost everywhere on $\Gamma$). \[2-geometries\] As it was seen in §\[dimension-gamma\], only two situations can occur, depending on the geometry of ${\Omega_\mathrm{int}}$: [–]{}[ ]{} if $\theta$ is not of the form $\pi/(2k)$ for any integer $k$, then $\dim_H \Xi=0$ and $p^\star=2$, if $\theta$ is of the form $\pi/(2k)$ for an integer $k$, then $\dim_H \Xi=(\dim_H \Gamma)/2$, and $p^\star=2- (\dim_H \Gamma)/2$. \[p=pstar\] The special case $p=p^\star$ is not dealt with in Theorem \[extension\]. The latter is of particular importance in case 1 of Remark \[2-geometries\] above, since the case $p=p^\star=2$ corresponds to the question of the $H^1$-extension property.\ In fact, it was proved by Koskela in [@MR1658090] that if a domain in ${{\mathbb R}}^n$ has the $W^{1,n}$-extension property, then it must have the $W^{1,p}$-extension property for every $p{\geqslant}n$. Hence, a consequence of Theorem \[ext-theorem\] is that ${\Omega_\mathrm{int}}$ cannot have the $W^{1,p}$-extension property when $p=2$. In particular, the domains that we will consider in Section \[r=rstar\] fail to satisfy the $H^1$-extension property.\ To the best of our knowledge, the question of the extension property for $p=p^\star$ in case 2 of Remark \[2-geometries\] seems to be open. The transmission problem in the case $r<r^\star_\theta$ {#sec-transmission-pb} ======================================================= The transmission problem with fractal interface {#sec-pb-transmission} ----------------------------------------------- The transmission problem can be formally stated as $$\label{prob-transmission} \left\{ \begin{array}{ll} -\Delta u = f & \mathrm{in}~ {\Omega_\mathrm{int}}\cup{\Omega_\mathrm{ext}},\\[1mm] [u]=0 &\text{on}~ \Gamma,\\[1mm] [\partial_n u]=\alpha u &\text{on}~ \Gamma,\\[1mm] \partial_n {u_\mathrm{int}}=\partial_n {u_\mathrm{ext}}= 0 & \mathrm{on}~ \Sigma,\\[1mm] \partial_n {u_\mathrm{ext}}= 0,~~ {u_\mathrm{int}}= u_0 & \mathrm{on}~ \Gamma^0,\\[1mm] \partial_n {u_\mathrm{ext}}= 0 & \mathrm{on}~ \partial D, \end{array} \right.$$ where $\alpha>0$, $\Sigma = \partial {\Omega_\mathrm{int}}\setminus (\Gamma\cup \Gamma^0)$, $[u]$ (resp. $[\partial_n u]$) denotes the jump of $u$ (resp. of the normal derivative of $u$) across $\Gamma$, $f\in L^2(D)$ and $u_0\in H^{1/2}(\Gamma^0)$. We also use the notations ${u_\mathrm{int}}:= u_{|{\Omega_\mathrm{int}}}$ and ${u_\mathrm{ext}}:= u_{|{\Omega_\mathrm{ext}}}$.\ The transmission condition $[\partial_n u]=\alpha u$ on $\Gamma$ has no real meaning, since the normal is not defined on $\Gamma$. The rigorous meaning of is the following variational formulation: $$\tag{$P$}\label{prob-fv} \begin{array}{c} \text{find $u\in V$ such that for all $v\in V_0$,}\\[2mm] \displaystyle a(u,v) ~=~ \int_D f v {\;\text{d}}x, \end{array}$$ where $V$ is the affine space defined by $$\label{space-V} V ~=~ \{u\in L^2(D),~ {u_\mathrm{int}}\in H^1({\Omega_\mathrm{int}}),~ {u_\mathrm{ext}}\in H^1({\Omega_\mathrm{ext}}),~ {{u_\mathrm{int}}}_{|\Gamma^0}=u_0,~{{u_\mathrm{int}}}_{|\Gamma} = {{u_\mathrm{ext}}}_{|\Gamma}\}.$$ Recall that $v\in H^1({\Omega_\mathrm{int}}) \mapsto v_{|\Gamma} \in H^{1/2}(\Gamma)$ and $v\in H^1({\Omega_\mathrm{ext}}) \mapsto v_{|\Gamma} \in H^{1/2}(\Gamma)$ are continuous maps, hence $V$ is closed. Note that if $u\in H^1(D)$ and $u_{|\Gamma^0}=u_0$ then $u\in V$.\ The vector space $V_0$ is defined as $V$, except that the condition ${{u_\mathrm{int}}}_{|\Gamma^0}=u_0$ is replaced by ${{u_\mathrm{int}}}_{|\Gamma^0}=0$. Finally, $$\label{form} a(u,v) ~=~ \int_{{\Omega_\mathrm{int}}} \nabla {u_\mathrm{int}}\cdot \nabla {v_\mathrm{int}}{\;\text{d}}x + \int_{{\Omega_\mathrm{ext}}} \nabla {u_\mathrm{ext}}\cdot \nabla {v_\mathrm{ext}}{\;\text{d}}x + \alpha \int_\Gamma u_{|\Gamma} v_{|\Gamma} {\;\text{d}}\mu.$$ The traces in the condition ${{u_\mathrm{int}}}_{|\Gamma} = {{u_\mathrm{ext}}}_{|\Gamma}$ in are meant in the sense of Definition \[def-trace\]. The above definition of the space $V$ is suitable when $r<r^\star_\theta$ and in the special case when $r=r^\star_\theta$ and $\theta=0$ discussed in Section \[r=rstar\]. In the other cases, the transmission condition has to be considerably changed, see Remark \[rem-theta&gt;0\]. Note that the space $V_0$, equipped with the norm $a(u,u)^{1/2}$ is a Hilbert space. From the Lax-Milgram theorem, we see that for every function $f$ given in $L^2(D)$, there exists a unique weak solution $u\in V$ to ($P$). Moreover, $u$ minimizes the functional $$\label{min-u} v\in V\mapsto a(v,v) - 2\int_D f v {\;\text{d}}x.$$ The transmission problem with prefractal interface -------------------------------------------------- For any positive integer $n$, let us consider the similar transmission problem in which the interior domain has been truncated by stopping the construction at step $n$. This class of problems is much more standard since the interface $\Gamma^n$ consists of $2^n$ pairwise disjoint line segments. The boundary value problem reads: $$\label{prob-transmission-n} \left\{ \begin{array}{ll} -\Delta u = f & \mathrm{in}~ {\Omega_\mathrm{int}}^n\cup{\Omega_\mathrm{ext}}^n,\\[1mm] [u]=0 &\text{on } \Gamma^n,\\[1mm] [\partial_n u]= \frac \alpha {|\Gamma^n|} u &\text{on } \Gamma^n,\\[1mm] \partial_n {u_\mathrm{int}}^n =\partial_n {u_\mathrm{ext}}^n = 0 & \mathrm{on}~ \Sigma^n,\\[1mm] \partial_n {u_\mathrm{ext}}^n = 0,~~ {u_\mathrm{int}}^n = u_0 & \mathrm{on}~ \Gamma^0,\\[1mm] \partial_n {u_\mathrm{ext}}^n = 0 & \mathrm{on}~ \partial D, \end{array} \right.$$ where $\Sigma^n = \partial {\Omega_\mathrm{int}}^n\setminus (\Gamma^n\cup \Gamma^0)$, and $[u]$ (resp. $[\partial_n u]$) denotes the jump of $u$ (resp. of the normal derivative of $u$) across $\Gamma^n$. We also use the notations ${u_\mathrm{int}}^n = u_{|{\Omega_\mathrm{int}}^n}$ and ${u_\mathrm{ext}}^n = u_{|{\Omega_\mathrm{ext}}^n}$.\ The variational formulation of problem can be stated as follows: $$\tag{$P_n$}\label{prob-n-fv} \begin{array}{c} \text{find $u\in V^n$ such that for all $v\in V^n_0$,}\\[2mm] \displaystyle a_n(u,v) ~=~ \int_D f v {\;\text{d}}x, \end{array}$$ where $V^n$ is the affine space defined by $$\label{eq:7} V^n ~=~ \{u\in L^2(D),~ {u_\mathrm{int}}^n\in H^1({\Omega_\mathrm{int}}^n),~ {u_\mathrm{ext}}^n\in H^1({\Omega_\mathrm{ext}}^n),~ {{u_\mathrm{int}}^n}_{|\Gamma^0}=u_0,~ {{u_\mathrm{int}}^n}_{|\Gamma^n} = {{u_\mathrm{ext}}^n}_{|\Gamma^n}\}.$$ Let $G^n$ be the closure of the set $\partial {\Omega_\mathrm{int}}^n  \setminus (\Gamma^0\cup \Gamma^n)$, which is a finite union of polygonal lines. It is easy to see that $V^n$ is the set of the functions in $H^1(D\setminus G^n)$ such that ${{u_\mathrm{int}}}_{|\Gamma^0}=u_0$.\ Similarly, if we define $G$ as $\partial {\Omega_\mathrm{int}}\setminus (\Gamma^0 \cup \Gamma)$, we observe that $G$ is not closed, since its closure contains $\Gamma$. Observe that, in general, the functions $u\in H^1(D\setminus \overline G)$ do not satisfy ${{u_\mathrm{int}}}_{|\Gamma} = {{u_\mathrm{ext}}}_{|\Gamma}$, so $V$ cannot be identified with the set of the functions $u\in H^1(D\setminus \overline G)$ satisfying the Dirichlet boundary condition on $\Gamma^0$. On the other hand, since $D\setminus G$ is not an open set, dealing with $H^1(D\setminus G)$ is not very straightforward. Here also, $V^n_0$ is defined as $V^n$, except that the condition ${{u_\mathrm{int}}^n}_{|\Gamma^0}=u_0$ is replaced by ${{u_\mathrm{int}}^n}_{|\Gamma^0}=0$, and $a_n$ is defined by $$\label{form-n} a_n(u,v) ~=~ \int_{{\Omega_\mathrm{int}}^n} \nabla {u_\mathrm{int}}^n \cdot \nabla {v_\mathrm{int}}^n ~ {\;\text{d}}x + \int_{{\Omega_\mathrm{ext}}^n} \nabla {u_\mathrm{ext}}\cdot \nabla {v_\mathrm{ext}}^n {\;\text{d}}x + \frac \alpha {|\Gamma^n |}\int_{\Gamma^n} u_{|\Gamma^n} v_{|\Gamma^n} {\;\text{d}}x.$$ The space $V^n_0$, equipped with the norm $a_n(u,u)^{1/2}$ is a Hilbert space. We also remark that $V^n\subset V$ with a continuous imbedding. Again, the Lax-Milgram theorem implies that for every function $f$ given in $L^2(D)$, there exists a unique weak solution $u_n\in V^n$ to that problem, and $u_n$ minimizes the functional $$\label{min-u_n} v\in V^n \mapsto a_n(v,v) - 2\int_D f v {\;\text{d}}x.$$ \[sec:transm-probl-with\] The sequence $u_n$ is bounded in $V$. Let $\tilde u_0$ be a function in $H^1(D)$ such that $\tilde u_{0| \Gamma^0}=u_0$, and such that $\tilde u _0$ is supported in a compact set which does not intersect the sets $\Gamma^n$, $\forall n\ge 1$. It is clear that $\tilde u_0\in V$ and that $\tilde u_0\in V_n$ for all $n\ge 1$. Let us define $C_0=\int_D \left(|\nabla \tilde u_0|^2 -2 f \tilde u_0\right) {\;\text{d}}x$. Thus, for all $n\ge 1$, $$\label{eq:2} a_n(u_n,u_n) -2 \int_D f u_n {\;\text{d}}x ~{\leqslant}~ a_n(\tilde u_0,\tilde u_0) -2 \int_D f \tilde u_0 {\;\text{d}}x ~=~ C_0,$$ because $\tilde u_{0|\Gamma^n}=0$. On the other hand, since $V^n\subset V$, $$\label{eq:3} a_n(u_n,u_n) ~{\geqslant}~ \int_{{\Omega_\mathrm{int}}^n} |\nabla u_n|^2 {\;\text{d}}x +\int_{{\Omega_\mathrm{ext}}^n} |\nabla u_n|^2 {\;\text{d}}x ~=~ \int_{{\Omega_\mathrm{int}}} |\nabla u_n|^2 {\;\text{d}}x +\int_{{\Omega_\mathrm{ext}}} |\nabla u_n|^2 {\;\text{d}}x.$$ We shall also use the following Poincar[é]{} inequality: there exists a constant $C>0$ such that, for all $v\in V$, $$\label{eq:4} \|v\|^2 _{L^2 (D)} ~{\leqslant}~ C \left( \|v_{|\Gamma^0}\|^2_{L^2(\Gamma^0)}+ \int_{{\Omega_\mathrm{int}}} |\nabla v|^2 {\;\text{d}}x +\int_{{\Omega_\mathrm{ext}}} |\nabla v|^2 {\;\text{d}}x\right).$$ From (\[eq:2\]), (\[eq:3\]) and (\[eq:4\]), we deduce that $$\int_{{\Omega_\mathrm{int}}} |\nabla u_n|^2 {\;\text{d}}x +\int_{{\Omega_\mathrm{ext}}} |\nabla u_n|^2 {\;\text{d}}x- 2 \sqrt C \|f\|_{L^2 (D)} \left[ \|u_{0|\Gamma^0}\|^2_{L^2(\Gamma^0)}+ \int_{{\Omega_\mathrm{int}}} |\nabla u_n|^2 {\;\text{d}}x +\int_{{\Omega_\mathrm{ext}}} |\nabla u_n|^2 {\;\text{d}}x \right]^{\frac 1 2} {\leqslant}C_0,$$ which implies that the quantity $ \int_{{\Omega_\mathrm{int}}} |\nabla u_n|^2 {\;\text{d}}x +\int_{{\Omega_\mathrm{ext}}} |\nabla u_n|^2 {\;\text{d}}x$ is bounded by a constant independent of $n$. Using (\[eq:4\]) again, this implies that $ \|u_n\| _{L^2 (D)}$ is also bounded by a constant independent of $n$. Combining the previous two observations, we obtain that the sequence $u_n$ is bounded in $V$. M-convergence of the energy forms in the case ${r}< {r}^\star_\theta$ {#sec-convergence} --------------------------------------------------------------------- We start by extending the definition of the forms $a$ and $a_n$ to the whole space $L^2(D)$ by setting $$\begin{aligned} a(u,u) = \infty & \text{if } u\in L^2(D)\setminus V, \label{def-a}\\ a_n(u,u) = \infty & \text{if } u\in L^2(D)\setminus V^n. \label{def-a_n}\end{aligned}$$ We are interested in proving the convergence of the forms $a_n$ to $a$ in the following sense, introduced by Mosco (see [@MR1283033]). \[Mosco-convergence\] A sequence of forms ${(a_n)}_{n}$ is said to M-converge to a form $a$ in $L^2(D)$ if 1. \[point1\] for every sequence ${(u_n)}_{n}$ weakly converging to a function $u$ in $L^2(D)$, $$\label{M-conv1} {\underline{\lim}~}a_n(u_n,u_n) ~{\geqslant}~ a(u,u)~~\mathrm{as}~n\to \infty,$$ 2. \[point2\] for every $u\in L^2(D)$, there exists a sequence ${(u_n)}_{n}$ converging strongly in $L^2(D)$ such that $$\label{M-conv2} {\overline{\lim}~}a_n(u_n,u_n) ~{\leqslant}~ a(u,u)~~\mathrm{as}~n\to \infty.$$ \[M-cv\] Assume that ${r}< {r}^\star_\theta$, then the energy forms $a_n$ M-converge in $L^2(D)$ to the form $a$. The $M$-convegence of forms differs from the $\Gamma$-convergence only in that the sequence $(u_n)$ in point \[point1\] of Definition \[Mosco-convergence\] is assumed to converge weakly instead of strongly. In the following, only the $\Gamma$-convergence of the energy forms $a_n$ will be needed. We will prove separately points \[point1\] and \[point2\] in Definition \[Mosco-convergence\].\ *Proof of point \[point1\]*  Suppose that ${(u_n)}$ weakly converges to $u$ in $L^2(D)$. Without loss of generality, one can suppose ${\underline{\lim}~}a_n(u_n,u_n)$ is finite. We may further assume that there exists a subsequence, still called $(u_n)$, such that $a_n(u_n,u_n)$ converges to some real number as $n\to \infty$; as a consequence, there exists a constant $c$ independent of $n$ such that $$\label{u-bound} a_n(u_n,u_n) ~{\leqslant}~ c.$$ In particular, for all $n$, $u_n\in V^n$, which implies that $u_n\in V$. Then, implies that $({u_n}_{|{\Omega_\mathrm{int}}})$ is bounded in $H^1({\Omega_\mathrm{int}})$, and $({u_n}_{|{\Omega_\mathrm{ext}}})$ is bounded in $H^1({\Omega_\mathrm{ext}})$. Therefore, there is a subsequence that we still note $(u_n)$ such that $({u_n}_{|{\Omega_\mathrm{int}}})$ converges weakly in $H^1({\Omega_\mathrm{int}})$, and strongly in $L^2({\Omega_\mathrm{int}})$. Since $u_n \rightharpoonup u$ in $L^2(D)$, we see that ${u_n}_{|{\Omega_\mathrm{int}}} {\rightharpoonup}{u_\mathrm{int}}$ in $H^1({\Omega_\mathrm{int}})$. Similarly, up to a further extraction of a subsequence, ${u_n}_{|{\Omega_\mathrm{ext}}} {\rightharpoonup}{u_\mathrm{ext}}$ in $H^1({\Omega_\mathrm{ext}})$. Consequently, $${\underline{\lim}~}\left(\int_{{\Omega_\mathrm{int}}^n} {|\nabla u_n|}^2{\;\text{d}}x + \int_{{\Omega_\mathrm{ext}}^n} {|\nabla u_n|}^2{\;\text{d}}x \right) ~{\geqslant}~ \int_{{\Omega_\mathrm{int}}} {|\nabla u|}^2{\;\text{d}}x + \int_{{\Omega_\mathrm{ext}}} {|\nabla u|}^2{\;\text{d}}x.$$ We will now prove that $\frac 1{|\Gamma^n|}\int_{\Gamma^n} u_n^2{\;\text{d}}x \to \int_\Gamma u^2{\;\text{d}}\mu$ as $n\to \infty$, which will yield point \[point1\]\ The following inequality was proved in [@ach-tch-2007-1]: for every $v\in H^1({\Omega_\mathrm{int}})$, $\left\| v- \langle v \rangle_{\Gamma^0} \right \| _{L_\mu ^2 (\Gamma)} {\leqslant}C {\| \nabla v\|} _{L^2({\Omega_\mathrm{int}})}$. This implies that $$| \langle v \rangle_{\Gamma^0} | ~{\leqslant}~ \| v \| _{L_\mu ^2 (\Gamma)} + C {\| \nabla v\|} _{L^2({\Omega_\mathrm{int}})}.$$ Similarly, $ \left\| v- \langle v \rangle_{\Gamma^0} \right\| _{L ^2 (\Gamma^0 )} {\leqslant}C {\| \nabla v\|}_{L^2({\Omega_\mathrm{int}})}$ implies that $$\frac 1 {\sqrt { |\Gamma^0|} }\| v \| _{L^2(\Gamma^0)} ~{\leqslant}~ | \langle v \rangle_{\Gamma^0} | + C {\| \nabla v\|} _{L^2({\Omega_\mathrm{int}})}.$$ for some constant independent of $v$ that we still note $C$. Combining these two inequalities, we obtain that $$\frac 1 {\sqrt { |\Gamma^0|} }\| v \| _{L^2 (\Gamma^0)} ~ {\leqslant}~ \|v\|_{L_\mu ^2 (\Gamma)} + C {\| \nabla v\|}_{L^2({\Omega_\mathrm{int}})}.$$ Hence, for every $\sigma\in {{\mathcal A}}_n$, $$\begin{aligned} \frac 1 {\sqrt { |\Gamma^\sigma|} }\| v \| _{L ^2 (\Gamma^\sigma)} &{\leqslant}& \|v\circ f_\sigma \|_{L_\mu ^2 (\Gamma)} + C {\| \nabla (v \circ f_\sigma) \|}_{L^2({\Omega_\mathrm{int}})}\\ &=& 2^{\frac n 2 }\|v \|_{L_\mu ^2 (f_\sigma(\Gamma))}+C {\| \nabla v \|}_{L^2(f_\sigma({\Omega_\mathrm{int}}))}.\end{aligned}$$ This yields that $$\frac 1 { |\Gamma^\sigma| }\| v \|^2 _{L ^2 (\Gamma^\sigma)} {\leqslant}2^{n }\|v \|^2_{L_\mu ^2 (f_\sigma(\Gamma))}+ 2C 2^{\frac n 2 } \|v \|_{L_\mu ^2 (f_\sigma(\Gamma))} {\| \nabla v \|}_{L^2(f_\sigma({\Omega_\mathrm{int}}))} + C^2 {\| \nabla v \|}^2 _{L^2(f_\sigma({\Omega_\mathrm{int}}))}.$$ Therefore $$\begin{split} &\frac 1 {|\Gamma^n|} \int_{\Gamma^n} {u_n}^2 {\;\text{d}}x ~= ~ \frac 1{2^n} \sum_{\sigma\in {{\mathcal A}}_n} \frac 1 {|\Gamma^\sigma|} \int_{\Gamma^\sigma} {u_n}^2 {\;\text{d}}x\\ {\leqslant}~ & \|u_n \|^2 _{L_\mu ^2 (\Gamma)} + 2C 2^{-\frac n 2 } \sum_{\sigma\in {{\mathcal A}}_n} \|u_n \|_{L_\mu ^2 (f_\sigma(\Gamma))} {\| \nabla u_n \|}_{L^2(f_\sigma({\Omega_\mathrm{int}}))} + \frac {C^2} {2^n} \sum_{\sigma\in {{\mathcal A}}_n} {\| \nabla u_n \|}^2 _{L^2(f_\sigma({\Omega_\mathrm{int}}))}\\ {\leqslant}~ &\|u_n \|^2 _{L_\mu ^2 (\Gamma)} + 2C 2^{-\frac n 2 } \|u_n \| _{L_\mu ^2 (\Gamma)} \left( \sum_{\sigma\in {{\mathcal A}}_n} {\| \nabla u_n \|}^2 _{L^2(f_\sigma({\Omega_\mathrm{int}}))}\right)^{\frac 1 2} + \frac {C^2} {2^n} \sum_{\sigma\in {{\mathcal A}}_n} {\| \nabla u_n \|}^2 _{L^2(f_\sigma({\Omega_\mathrm{int}}))}. \end{split}$$ Since ${u_n}_{|{\Omega_\mathrm{int}}}$ is a bounded sequence in $H^1({\Omega_\mathrm{int}})$, there exists a constant $M$ such that $$\frac 1 {|\Gamma^n|} \int_{\Gamma^n} {u_n}^2 {\;\text{d}}x {\leqslant}\|u_n \|^2 _{L_\mu ^2 (\Gamma)} + 2C M 2^{-\frac n 2 } \|u_n \| _{L_\mu ^2 (\Gamma)} + \frac {C^2 M^2} {2^n} .$$ Moreover, since $u_n$ weakly converges to $u$ in $H^1({\Omega_\mathrm{int}})$, then up to the extraction of a subsequence, ${u_n}_{|\Gamma}$ strongly converges to $u_{|\Gamma}$ in $L^2_\mu(\Gamma)$, from Remark \[sec:trace-theorem-gamma\]. Hence, we obtain that $${\overline{\lim}~}\frac 1 {|\Gamma^n|} \int_{\Gamma^n} {u_n}^2 {\;\text{d}}x ~{\leqslant}~ \int_{\Gamma} u^2 {\;\text{d}}\mu.$$ Similarly, the following inequality holds for every $v\in H^1({\Omega_\mathrm{int}})$: $$\label{poincare-ineq2} {\| v\|}_{L^2_\mu(\Gamma)} ~{\leqslant}~ \frac 1 {\sqrt{|\Gamma^0|}} {\| v\|}_{L^2(\Gamma^0)} + C' {\| \nabla v\|}_{L^2({\Omega_\mathrm{int}})}$$ for some constant $C'$ independent of $v$. As above, we deduce that $${\underline{\lim}~}\frac 1 {|\Gamma^n|} \int_{\Gamma^n} {u_n}^2 {\;\text{d}}x ~{\geqslant}~ \int_{\Gamma} u^2 {\;\text{d}}\mu,$$ and we obtain the desired result.\ *Proof of point \[point2\]*  Take $u\in L^2(D)$. By , we may assume that $u\in V$. We must construct $(u_n)$ converging strongly in $L^2(D)$ such that holds. Note that the choice $u_n=u$ cannot be made, since $u\not\in V^n$ in general.\ Take $\delta>0$ and consider a neighborhood $\omega \subset D$ of ${\Omega_\mathrm{int}}$ such that ${\Omega_\mathrm{int}}\Subset \omega$ and $\sup_{x\in \omega}d(x,{\Omega_\mathrm{int}})<\delta$, where $d(x,{\Omega_\mathrm{int}}) = \inf_{y\in {\Omega_\mathrm{int}}} |x-y|$. We introduce the notations $\omega^\sigma = f_\sigma(\omega)$ for $\sigma \in {{\mathcal A}}$ and $\omega_n = \bigcup_{\sigma\in {{\mathcal A}}_n} \omega^\sigma$ for all integer $n$.\ For every $n$, introduce the cut-off function $\chi_n$ in $D$ defined by $$\chi_n(x) ~=~ {(1-\delta r^{-n} d(x,\omega_n))}^+,$$ where $\alpha^+$ stands for the positive part of a real number $\alpha$. Hence, $\chi_n \equiv 1$ in $\omega_n$ and $\chi_n \equiv 0$ outside $\tilde \omega_n := \{x\in D,~ d(x,\omega_n)<\delta r^n\}$. Note that if we set $\tilde \omega := \{x\in D,~ d(x,\omega)<\delta\}$ and $\tilde \omega^\sigma := f_\sigma(\tilde \omega) = \{x\in D,~ d(x,\omega^\sigma)<\delta r^n\}$ for $\sigma\in {{\mathcal A}}_n$, then $\tilde \omega_n = \bigcup_{\sigma\in {{\mathcal A}}_n} \tilde \omega^\sigma$.\ We can assume that $\delta$ is small enough so that $\tilde \omega^\sigma\cap \tilde \omega^\tau =\emptyset$ when $\sigma,\tau \in {{\mathcal A}}_n$ and $\sigma\not=\tau$, since $f_\sigma({\Omega_\mathrm{int}})\cap f_\tau({\Omega_\mathrm{int}})=\emptyset$.\ We now define a sequence of functions $u_n$ by $$\label{u_n} u_n = (1-\chi_n)u + \chi_n {{\mathcal E}}({u_\mathrm{int}}),$$ where ${{\mathcal E}}$ is the extension operator introduced in and as above, ${u_\mathrm{int}}= u_{|{\Omega_\mathrm{int}}}$. Obviously, $u_n$ belongs to the space $V^n$ and the sequence $(u_n)$ strongly converges to $u$ in $L^2(D)$. We will prove that $\lim a_n(u_n,u_n) = a(u,u)$ as $n\to \infty$. We start by showing that $$\label{an->a-1} I_n :=\int_{{\Omega_\mathrm{int}}} {|\nabla u|}^2 {\;\text{d}}x + \int_{{\Omega_\mathrm{ext}}} {|\nabla u|}^2 {\;\text{d}}x -\left( \int_{{\Omega_\mathrm{int}}^n} {|\nabla u_n|}^2 {\;\text{d}}x + \int_{{\Omega_\mathrm{ext}}^n} {|\nabla u_n|}^2 {\;\text{d}}x\right) ~\longrightarrow~ 0$$ as $n\to \infty$. First observe that $$I_n ~=~ \int_{\tilde \omega_n\setminus {\Omega_\mathrm{int}}} {|\nabla u|}^2 {\;\text{d}}x - \int_{\tilde \omega_n\setminus {\Omega_\mathrm{int}}} {|\nabla u_n|}^2 {\;\text{d}}x.$$ Hence, it is enough to show that $\int_{\tilde \omega_n\setminus {\Omega_\mathrm{int}}} {|\nabla (u-u_n)|}^2{\;\text{d}}x \to 0$ as $n\to \infty$. Note that $$\int_{\tilde \omega_n\setminus {\Omega_\mathrm{int}}} {|\nabla(u-u_n)|}^2 {\;\text{d}}x~=~ \int_{\tilde \omega_n\setminus {\Omega_\mathrm{int}}} {|\nabla(\chi_n({{\mathcal E}}({u_\mathrm{int}}) -u))|}^2 {\;\text{d}}x ~{\leqslant}~ 2(I_n^1 + I_n^2)$$ where $I_n^1 = \int_{\tilde \omega_n\setminus {\Omega_\mathrm{int}}} {|\nabla({{\mathcal E}}({u_\mathrm{int}}) -u)|}^2 {\;\text{d}}x$ and $I_n^2 = \int_{\tilde \omega_n\setminus {\Omega_\mathrm{int}}} {|(\nabla \chi_n)({{\mathcal E}}({u_\mathrm{int}}) -u)|}^2 {\;\text{d}}x$.\ First observe that $I_n^1 \to 0$ as $n\to \infty$ since $\nabla ({{\mathcal E}}({u_\mathrm{int}}) -u)\in L^2({\Omega_\mathrm{ext}})$. We are left with dealing with $I_n^2$. One has $$\begin{aligned} I_n^2 &{\leqslant}& \displaystyle c\; r^{-2n} \int_{\tilde \omega_n\setminus {\Omega_\mathrm{int}}} {|{{\mathcal E}}({u_\mathrm{int}}) -u|}^2{\;\text{d}}x\\ &{\leqslant}& \displaystyle c\; r^{-2n} \sum_{\sigma\in {{\mathcal A}}_n} \int_{\tilde \omega^\sigma \setminus {\Omega_\mathrm{int}}} {|{{\mathcal E}}({u_\mathrm{int}}) -u|}^2{\;\text{d}}x.\end{aligned}$$ where $c>0$ is a constant independent of $n$. Introduce the set $\hat \omega = {f_1}^{-1}(f_1(\tilde \omega)\setminus {\Omega_\mathrm{int}})$. We have the following Poincar[é]{} inequality: there exists a constant $C$ such that for every $v\in H^1(\hat \omega)$ such that $v_{|\Gamma}=0$, $$\label{PW} \int_{\hat \omega} {|v|}^2{\;\text{d}}x ~{\leqslant}~ C \int_{\hat \omega} {|\nabla v|}^2 {\;\text{d}}x.$$ Observe that if $\delta$ is small enough, then $\tilde \omega^\sigma\setminus {\Omega_\mathrm{int}}= f_\sigma(\hat \omega)$ for every $\sigma\in {{\mathcal A}}$. Therefore, applying a rescaled version of to the function ${{\mathcal E}}({u_\mathrm{int}})-u$, we obtain that there is a constant $c'>0$ indenpendant of $n$ such that $$\begin{aligned} I_n^2 &{\leqslant}& c'\sum_{\sigma\in {{\mathcal A}}_n} \int_{\tilde \omega^\sigma\setminus {\Omega_\mathrm{int}}} {|\nabla ({{\mathcal E}}({u_\mathrm{int}}) - u)|}^2{\;\text{d}}x\\ &=& c' \int_{\tilde \omega_n\setminus {\Omega_\mathrm{int}}} {|\nabla ({{\mathcal E}}({u_\mathrm{int}}) - u)|}^2{\;\text{d}}x \end{aligned}$$ since the sets $\tilde \omega^\sigma$, $\sigma \in {{\mathcal A}}_n$ are pairwise disjoint. We deduce that $I_n^2 \to 0$ as $n\to \infty$, since ${{\mathcal E}}({u_\mathrm{int}}) - u\in H^1({\Omega_\mathrm{ext}})$, which yields .\ We will now prove that $$\label{an->a-2} \frac 1 {|\Gamma^n|} \int_{\Gamma^n} {u_n}^2 {\;\text{d}}x \longrightarrow \int_\Gamma u^2 {\;\text{d}}\mu$$ as $n\to \infty$, which will conclude the proof of point \[point2\]. Observe that for every integer $n$, ${{\mathcal E}}({u_\mathrm{int}})_{|\Gamma^n} = u$, which implies that ${u_n}_{|\Gamma^n}=u$. We are left with proving that $\frac{1}{|\Gamma^n|} \int_{\Gamma^n} u^2{\;\text{d}}x \longrightarrow \int_{\Gamma} u^2{\;\text{d}}\mu $ as $n\to \infty$, which holds by Proposition \[convergence-result\]. A standard consequence of the Mosco-convergence of the energy forms proved in Theorem \[M-cv\] is the convergence of the solutions of the problems to the solution of problem , in $L^2(D)$ and in $V$ (recall that $u_n$ is bounded in $V$ from Lemma \[sec:transm-probl-with\]). Take $f\in L^2(D)$, and note $u_n$ (resp. $u$) the solution of problem (resp. ). The sequence $(u_n)$ converges to $u$ in the space $V$. A particular geometry with ${r}={r}^\star_\theta$ {#r=rstar} ================================================= As seen before, the proof of Theorem \[M-cv\] is based on the extension result of §\[ext-subcritical\]. In the case ${r}={r}^\star_\theta$, the $H^1$-extension property is no longer true for the domain ${\Omega_\mathrm{int}}$ (see Remark \[p=pstar\]). In what follows, we focus on the special case when $\theta=0$. We will see that in this case, the transmission condition imposed on $\Gamma$ yields an extension result (see Theorem \[ext-thm\]) which is the main ingredient for proving the $M$-convergence of the energy forms.\ In the case $\theta=0$, it can be seen that $r^\star_\theta =\frac 1 2$, and the ramified domain described in §\[domains\] is as in Figure \[trapeze\]. In this particular case, the set $f_1(\Gamma)\cap f_2(\Gamma)$ is reduced to a single point that we call $A$. Observe that the self-similar part $\Gamma$ of the boundary is a line segment, and the self-similar measure $\mu$ associated with $\Gamma$ is the normalized one-dimensional Hausdorff measure.\ Since ${r}={r}^\star_\theta$, the domain ${\Omega_\mathrm{ext}}$ has infinitely many connected components. Call $U$ the outer connected component of ${\Omega_\mathrm{ext}}$, *ie* the only connected component which has a nonempty intersection with $\partial D$ (see Figure \[trapeze\]). Observe that $\Gamma$ is a subset of $\partial U$, and that the intersection of $\Gamma$ with the boundary of every other connected component of ${\Omega_\mathrm{ext}}$ is reduced to a single point. Apart from $U$, each connected component of ${\Omega_\mathrm{ext}}$ is a triangle whose top vertex is at a dyadic point of $\Gamma$. The largest triangle is named $T$, see Figure \[trapeze\], and all the other triangles are the images of $T$ by $f_\sigma$, $\sigma\in {{\mathcal A}}_n$, $n\ge 1$.\ We consider the transmission problem $$\label{pb-transmission_2} \left\{ \begin{array}{ll} -\Delta u + \beta u= f & \mathrm{in}~ {\Omega_\mathrm{int}}\cup{\Omega_\mathrm{ext}},\\[1mm] {{u_\mathrm{int}}}_{|\Gamma} = {{u_\mathrm{ext}}}_{|\Gamma} & \\[1mm] [\partial_n u]=\alpha u &\text{on}~ \Gamma,\\[1mm] \partial_n {u_\mathrm{int}}=\partial_n {u_\mathrm{ext}}= 0 & \mathrm{on}~ \partial {\Omega_\mathrm{int}}\setminus \Gamma,\\[1mm] \partial_n {u_\mathrm{ext}}= 0,~~ {u_\mathrm{int}}= u_0 & \mathrm{on}~ \Gamma^0,\\[1mm] \partial_n {u_\mathrm{ext}}= 0 & \mathrm{on}~ \partial D, \end{array} \right.$$ where $\alpha$ and $\beta$ are positive numbers. The trace ${{u_\mathrm{ext}}}_{|\Gamma}$ in the transmission condition is meant as the trace of the function $u_{|U}$ on the set $\Gamma$. \[sec:part-geom-with\] The reason for considering the operator $-\Delta u+\beta u$ with $\beta>0$ instead of $-\Delta u$ as in the former case is that, in the present case, ${\Omega_\mathrm{ext}}$ is an infinite union of disjoint connected sets: ${\Omega_\mathrm{ext}}=U\cup \bigcup_{\sigma\in {{\mathcal A}}} f_\sigma(T)$. Therefore, (\[pb-transmission\_2\]) involves Neumann problems in $T$ and in $ f_\sigma(T)$, $\sigma\in {{\mathcal A}}_n$, $n\ge 1$, which are not well posed if $\beta=0$ and the average of $f$ in these sets is not zero. It would also be possible to consider the case $\beta=0$ under additional assumptions, on the support of $f$ for example, but this would imply further technical details, because the solutions of Neumann problems in the holes would then be defined up to the addition of constants. \[rem-theta&gt;0\] When $\theta>0$ and $r=r^*$, the situation is quite different: $\Gamma$ is not entirely contained in the boundary of any connected component of ${\Omega_\mathrm{ext}}$. It can be shown that there exists $\delta\in (1,d)$ such that the intersections of $\Gamma$ with the boundary of the connected components of ${\Omega_\mathrm{ext}}$ have Hausdorff dimension $\delta$. These sets are called the canopies of the domain ${\Omega_\mathrm{int}}$. In this case, the transmission condition has to be described more carefully. This is the topic of a work in progress. The meaning of (\[pb-transmission\_2\]) is the variational formulation (\[prob-fv\]) where $V$ is defined by (\[space-V\]) and $$\label{eq:6} a(u,v) ~=~ \int_{{\Omega_\mathrm{int}}} ( \nabla {u_\mathrm{int}}\cdot \nabla {v_\mathrm{int}}+\beta {u_\mathrm{int}}{v_\mathrm{int}}) ~ {\;\text{d}}x + \int_{{\Omega_\mathrm{ext}}} (\nabla {u_\mathrm{ext}}\cdot \nabla {v_\mathrm{ext}}+ \beta {u_\mathrm{ext}}{v_\mathrm{ext}})~{\;\text{d}}x + \alpha \int_\Gamma u_{|\Gamma} v_{|\Gamma} {\;\text{d}}\mu.$$ In order to set the transmission problems in the geometries with prefractal interfaces, we first need to define some trapezoidal subsets of the triangular holes as follows: let $H$ be the height of the triangle $T$. Choosing the coordinates in such a way that $\Gamma^0$ is a segment of the line $\{x_2=0\}$, we see that $\Gamma$ is a segment of the line $\{x_2=2H\}$. Then, we can also define $\widehat T^n$ and $ \widehat T_\sigma ^n$ by $\widehat T^n = T\cap \{ (2- 3 /2^{n+1}) H < x_2< (2- 2^{-n}) H \}$ and $ \widehat T_\sigma ^n= f_\sigma(T)\cap \{ (2- 3 /2^{n+1}) H < x_2< (2- 2^{-n}) H \}$ for $\sigma\in {{\mathcal A}}_m$ and $m<n$.\ Finally we define $$\label{eq:1} \widehat \omega^n_{\rm ext}= \bigcup_{m=0}^{n-1} \bigcup_{\sigma\in {{\mathcal A}}_m} \widehat T_\sigma ^n \;\; \subset \; {\Omega_\mathrm{ext}}^n \cap \bigg\{ 2H- \frac 3 {2^{n+1}} H < x_2< 2H- 2^{-n} H \bigg\}.$$ The transmission problem with interface $\Gamma^n$ is then $$\label{prob-transmission-2n} \left\{ \begin{array}{ll} - {\rm div } \left ( \nu_n \nabla u\right) +\beta u = f & \mathrm{in}~ {\Omega_\mathrm{int}}^n\cup{\Omega_\mathrm{ext}}^n,\\[1mm] [u]=0 &\text{on } \Gamma^n,\\[1mm] [\partial_n u]= \frac \alpha {|\Gamma^n|} u &\text{on } \Gamma^n,\\[1mm] \partial_n {u_\mathrm{int}}^n =\partial_n {u_\mathrm{ext}}^n = 0 & \mathrm{on}~ \Sigma^n,\\[1mm] \partial_n {u_\mathrm{ext}}^n = 0,~~ {u_\mathrm{int}}^n = u_0 & \mathrm{on}~ \Gamma^0,\\[1mm] \partial_n {u_\mathrm{ext}}^n = 0 & \mathrm{on}~ \partial D, \end{array} \right.$$ where $$\label{eq:5} \nu_n = 2^{-2n} {\mathbbm{1}}_{ \widehat \omega^n_{\rm ext} } + {\mathbbm{1}}_{D\backslash \widehat \omega^n_{\rm ext}}.$$ Note that $\nu_n =1$ in ${\Omega_\mathrm{int}}^n$ and that the Lebesgue measure of the set where $\nu_n = 2^{-2n}$ vanishes as $n\to \infty$. Hence $\nu_n$ tends to $1$ almost everywhere in $D$. The variational formulation of (\[prob-transmission-2n\]) is (\[prob-n-fv\]) with $V_n$ defined in (\[eq:7\]), and $a_n$ defined as follows: $$\label{eq:8} \begin{split} &a_n(u,v)\\ =& \int_{{\Omega_\mathrm{int}}^n} (\nabla {u_\mathrm{int}}^n \cdot \nabla {v_\mathrm{int}}^n +\beta {u_\mathrm{int}}^n {v_\mathrm{int}}^n) ~ {\;\text{d}}x + \int_{{\Omega_\mathrm{ext}}^n} (\nu_n \nabla {u_\mathrm{ext}}^n \cdot \nabla {v_\mathrm{ext}}^n + \beta {u_\mathrm{ext}}^n {v_\mathrm{ext}}^n ){\;\text{d}}x + \frac \alpha {|\Gamma^n |}\int_{\Gamma^n} u_{|\Gamma^n} v_{|\Gamma^n} {\;\text{d}}x. \end{split}$$ \[sec:part-geom-with-1\] The reason for modifying the partial differential equation in (\[prob-transmission-2n\]) by taking $-{\rm div } \left ( \nu_n \nabla u\right)$ instead of $-\Delta u$ in (\[prob-transmission-n\]) is that for a function $u\in V $, $u_{|{\Omega_\mathrm{int}}}$ is completely independent from $u_{|f_\sigma(T)}$. This explains why the construction of a sequence of functions $(u_n)$ such that $u_n\in V_n$, $u_n\to u$ in $L^2(D)$ and $a_n(u_n, u_n)\to a(u,u)$, is difficult without modifying the coefficients of the partial differential equation near the top of the triangles $T$ and $f_\sigma(T)$ in order to cope with the possibly strong gradients of $u_n$. Although we have not tried it, it may be possible to choose a parameter larger than $2^{-2n}$ in the definition of $\nu_n$. The main result of this paragraph is the following theorem \[M-cv-theta=0\] Assume that $\theta=0$ and ${r}={r}^\star_\theta=\frac 1 2$. Then the energy forms $a_n$ defined in (\[eq:8\]) M-converge in $L^2(D)$ to the form $a$ defined in (\[eq:6\]). Since $\partial U$ is Lipschitz-continuous, a standard trace result yields that for every $u\in V$, ${{u_\mathrm{ext}}}_{|\Gamma} \in H^{1/2}(\Gamma)$. Hence, the transmission condition in (\[pb-transmission\_2\]) implies that $$\label{trace-theta=0} {{u_\mathrm{int}}}_{|\Gamma}\in H^{1/2}(\Gamma).$$ Note that (\[trace-theta=0\]) is not only a consequence of the fact that ${u_\mathrm{int}}\in H^1({\Omega_\mathrm{int}})$, because the latter property only implies that ${{u_\mathrm{int}}}_{|\Gamma}$ in $H^s(\Gamma)$ for all $s<\frac 1 2$ (see Theorem \[trace-thm\]).\ For proving Theorem \[M-cv-theta=0\], we need the following extension result, which is not available in the literature: \[ext-thm\] There exist a linear extension operator ${{\mathcal F}}$ from $\{v\in H^1({\Omega_\mathrm{int}}),~ v_{|\Gamma}\in H^{1/2}(\Gamma)\}$ to $H^1(D)$ and a constant $C>0$ such that for every $v\in H^1({\Omega_\mathrm{int}})$ with $v_{|\Gamma}\in H^{1/2}(\Gamma)\}$, $$\label{continuity} {\Vert \tilde v \Vert}_{H^1(D)}^2 ~{\leqslant}~ C \left( {\Vert v \Vert}_{H^1({\Omega_\mathrm{int}})}^2 + {\Vert v_{|\Gamma} \Vert}_{H^{1/2}(\Gamma)}^2 \right).$$ Theorems \[auscher-badr\] and \[lem-PT\] below will play an important role in the proof of Theorem \[ext-thm\]. We start by recalling an extension result for multiple cones from [@MR2790816]. \[auscher-badr\]\[see [@MR2790816], Th. 5.1\] Call $C$ the double cone in ${{\mathbb R}}^2$ defined by $|x_1| <|x_2|$. Write $\rho(x)=\Vert x \Vert$ (the notation $\Vert \; . \; \Vert$ stands for the euclidean norm), and, for every $v\in H^1(C)$, introduce the antiradial part $v_a$ of $v$ in the cone $C$, defined by $$v_a(x) ~=~ v(x) - {\langle v \rangle}_{S_{\rho(x)}},$$ where for any $R>0$, $S_R=\{x\in C,~ \Vert x \Vert =R\}$, and ${\langle v \rangle}_{S_R}$ is the mean value of $v$ along the arc $S_R$.\ There exists a linear extension operator $$\Lambda : \left\{v\in H^1(C),~ \frac{v_a}\rho \in L^2(C)\right\} \to H^1({{\mathbb R}}^2),$$ such that for every $v\in \{v\in H^1(C),~ {v_a}/\rho \in L^2(C)\}$, $${\Vert \Lambda v \Vert}_{H^1({{\mathbb R}}^2)} ~{\leqslant}~ c\left( {\Vert v \Vert}_{H^1(C)} + {\left\Vert \frac {v_a} \rho \right\Vert}_{L^2(C)} \right)$$ where $c>0$ is a constant independent of $v$. \[rem-AB\] The construction in [@MR2790816] is such that if $v$ is radial (*resp.* constant) in $C\cap B(0,R)$, then $\Lambda v$ is radial (*resp.* constant) in $B(0,R)$. As mentioned in [@MR2790816], Theorem \[auscher-badr\] can be immediately extended in ${{\mathbb R}}^n$ to the case of a union of two half-cones sharing the same vertex, separated by a hyperplane passing through the vertex and not containing any direction of the boundaries. \[lem-PT\]\[see [@MR0221282; @Tartar]\] Let $E,E_1,E_2,F$ be Banach spaces, and let $A_i$, $i=1,2$, be continuous linear operators from $E$ to $E_i$, and suppose $A_1$ is compact. Further assume that there exists a constant $c_0>0$ such that for any $v\in E$, $$\label{cond-PT} {\Vert v\Vert}_E ~{\leqslant}~ c_0({\Vert A_1 v\Vert}_{E_1} + {\Vert A_2 v\Vert}_{E_2}).$$ If $L$ is a coninuous linear operator from $E$ to $F$ such that $L_{|\ker A_2} \equiv 0$, then there exists a constant $c_1>0$ such that for any $v\in E$, $$\label{PT} {\Vert Lv \Vert}_F ~{\leqslant}~ c_1 {\Vert A_2 v \Vert}_{E_2}.$$ #### Notations We start by introducing notations for the proof of Theorem \[ext-thm\].\ We first introduce a domain $C$ which is the union of two truncated half-cones included in ${\Omega_\mathrm{int}}$, whose common vertex is the point $A$. Recall that $T$ is the main hole of the domain ${\Omega_\mathrm{int}}$. Call ${\varphi}_0\in (0,\frac \pi 2)$ the upper half-angle of the triangular domain $T$ (see Figure \[cone\]), and take ${\varphi}_1>{\varphi}_0$. Call $\mathcal{C}$ the half-cone whose boundary is made of the two half-lines through $A$ with respective angles ${\varphi}_0$ and ${\varphi}_1$ with the vertical axis (see Figure \[cone\]). Call $C_2 = \mathcal{C}\cap ({\widetilde \Omega_\text{int}}\setminus \overline{Y^0})$. We can assume that ${\varphi}_1>{\varphi}_0$ is small enough so that $C_2\subset {\Omega_\mathrm{int}}$, in other words $C_2$ does not intersect any of the holes of ${\Omega_\mathrm{int}}$. We define $C_1$ to be the symmetric of $C_2$ with respect to the vertical axis $x_2=0$, and we write $C=C_1\cup C_2$.\ We also introduce the sets $Y^{k,1} := f_1\circ {f_2}^{k-1}(Y^0)$ and $Y^{k,2} := f_2\circ {f_1}^{k-1}(Y^0)$ for every $k{\geqslant}1$ (see Figure \[cone\]), and we write $Y^k := Y^{k,1}\cup Y^{k,2}$. We also note $\gamma := {f_1}^2(\Gamma)\cup {f_2}^2(\Gamma)$ (see Figure \[Yki\]). $$\scalebox{0.55}{\PandocStartInclude{cone-trapeze.pstex_t}\PandocEndInclude{input}{1459}{45}} $$ Call $\Omega^{1,1}:=f_1\big({\Omega_\mathrm{int}}\setminus \overline{f_2({\Omega_\mathrm{int}})}\big)$ and $\Omega^{1,2}:=f_2\big({\Omega_\mathrm{int}}\setminus \overline{f_1({\Omega_\mathrm{int}})}\big)$ (see Figure \[Yki\]). We introduce the sets $\Omega^{k,i}$ defined by $\Omega^{k,i}:=g^{k-1}(\Omega^i)$, $k{\geqslant}1$, $i=1,2$, where $g$ is the homothety centered at $A$ with ratio $1/2$, see Figure \[Yki\]. [6]{}(8.5,2) $\gamma = {f_1}^2(\Gamma)\cup {f_2}^2(\Gamma)$ $$\scalebox{0.5}{\PandocStartInclude{omega-k-i.pstex_t}\PandocEndInclude{input}{1474}{41}}$$ For every integer $k{\geqslant}1$, we introduce $\Sigma^k := \{\tau\in {{\mathcal A}},~ f_\tau(Y^0) \subset \Omega^{k,1}\cup\Omega^{k,2}\}$, with the notations of §\[def\].\ Take $\kappa\in (1/2,1)$. We introduce the space $G=\{v\in L^1_\text{loc}(\Omega^{1,1}\cup \Omega^{1,2}),~ {\Vert v \Vert}_G<\infty\}$, where $$\label{normG} {\Vert v \Vert}_G^2 = \sum_{m{\geqslant}1} \kappa^m \sum_{\tau \in \Sigma^1\cap {{\mathcal A}}_m} \int_{f_\tau(Y^0)} {|v |}^2 {\;\text{d}}x.$$ Endowed with the norm ${\Vert \cdot \Vert}_G$, the space $G$ is a Hilbert space.\ We also introduce the space $H=\{v\in L^2(\Omega^{1,1}\cup \Omega^{1,2}),~ \nabla v\in G\}$, which is a Hilbert space with the norm $$\left( {\|\nabla v\|}^2 _G + {\|v\|}^2 _{L^2(\Omega^{1,1}\cup \Omega^{1,2})} \right)^{\frac 1 2}.$$ Moreover, from Theorem \[SPI\], we see that $v\in H\mapsto v_{|\gamma}$ is a continuous operator from $H$ to $L^2_\mu (\gamma)$. Arguing by contradiction, we can show that $$\left( {\|\nabla v\|}^2_G + {\|v_{|\gamma}\|}^2 _{L^2 _\mu (\gamma)} \right)^{\frac 1 2}$$ is an equivalent norm on $H$.\ We first state and prove two lemmas which will be useful in the proof of Theorem \[ext-thm\]. \[lem1\] There exists a constant $c>0$ such that for every $v\in H$, $$\int_{Y^1} \left({|v(x)-{\langle v \rangle}_{Y^{1,1}}|}^2 + {|v(x)-{\langle v \rangle}_{Y^{1,2}}|}^2\right){\;\text{d}}x ~{\leqslant}~ c\left(\int_\gamma {|v-{\langle v \rangle}_\gamma|}^2{\;\text{d}}\mu + {\Vert \nabla v\Vert}^2_G \right). \label{eq-pt}$$ We introduce the Hilbert space $E = \{(v,w)\in H\times L^2_\mu(\gamma),~ v_{|\gamma} = w\}$, endowed with the norm ${\Vert \cdot \Vert}_E$ given by ${\Vert (v,w)\Vert}_E^2 ={\Vert \nabla v \Vert}_G^2 + {\Vert w \Vert}_{L^2_\mu(\gamma)}^2$.\ We now introduce the operators $$\begin{aligned} &&A_1 : (v,w)\in E \mapsto {\langle v \rangle}_{\gamma},\\[1mm] &&A_2 : (v,w)\in E \mapsto (\nabla v, w-{\langle w \rangle}_\gamma),\\[1mm] &&L : (v,w)\in E \mapsto (v-{\langle v \rangle}_{Y^{1,1}}, v-{\langle v \rangle}_{Y^{1,2}}).\end{aligned}$$ With the notations of Theorem \[lem-PT\], $E_1={{\mathbb R}}$, $E_2=G \times L^2_\mu(\gamma)$ and $F={L^2(Y^{1})}^2$. It is easily seen that $A_1$, $A_2$ and $L$ are continuous linear operators, and $A_1$ is compact. Moreover, is clearly satisfied.\ Observe that $(v,w)\in E$ lies in $\ker A_2$ if and only if $v$ is constant in $\Omega^{1,1}$ and in $\Omega^{1,2}$. Hence, it is obvious that $L_{|\ker A_2}\equiv 0$. From this, we deduce by Theorem \[lem-PT\] that there exists a constant $c>0$ such that ${\Vert L(v,w) \Vert}_F ~{\leqslant}~ c {\Vert A_2 (v,w) \Vert}_{E_2}$ for all $(v,w)\in E$, which yields . \[lem2\] Assume that $v\in H^1({\Omega_\mathrm{int}})$ and $v_{|\Gamma}\in H^{1/2}(\Gamma)$, then $\frac {v_a}\rho \in L^2(C)$, and $$\begin{aligned} \int_C {\left| \frac {v_a}\rho \right|}^2 {\;\text{d}}x &{\leqslant}& c \left( \sum_{k{\geqslant}1} 2^k \int_{\gamma_k} {|v-{\langle v \rangle}_{\gamma_k}|}^2{\;\text{d}}\mu + \sum_{k{\geqslant}1} \sum_{p{\geqslant}k} \kappa^{p-k+1} \sum_{\tau \in \Sigma^k\cap {{\mathcal A}}_p} \int_{f_\tau(Y^0)} {|\nabla v|}^2{\;\text{d}}x \right) \label{eq-lem2}\end{aligned}$$ for some constant $c>0$ independent of $v$, where $\gamma^k=g^{k-1}(\gamma)$ (recall that $g$ is the homothety centered at $A$, with ratio $1/2$). We first observe that, by self-similarity, there exists a constant $c_1>0$ such that for all $x\in Y^k$, $\rho(x){\geqslant}\frac {c_1}{2^k}$. Therefore, $$\int_{C}{\left|{\frac{v_a}{\rho}}\right|}^2 {\;\text{d}}x ~{\leqslant}~ c_1 \sum_{k{\geqslant}1} 2^{2k} \int_{Y^k} {|v_a|}^2 {\;\text{d}}x,$$ since ${C}\subset \bigcup_{k{\geqslant}1} Y^k$ by construction. Hence, there is a constant $c_2>0$ such that ${\displaystyle}\int_{C}{\left|\frac{v_a}\rho\right|}^2{\;\text{d}}x {\leqslant}c_2(I_1+I_2)$, where $$\begin{aligned} I_1 &=& \sum_{k{\geqslant}1} 2^{2k}\int_{Y^k} {|v(x)-{\langle v \rangle}_{Y^k}|}^2 {\;\text{d}}x,\\ I_2 &=& \sum_{k{\geqslant}1} 2^{2k}\int_{Y^k} {|{\langle v \rangle}_{Y^{k}} -{\langle v\rangle}_{S_{\rho(x)}}|}^2 {\;\text{d}}x.\end{aligned}$$ We start by dealing with $I_1$. We note that $$I_1 ~{\leqslant}~ \frac 1 2 \sum_{k{\geqslant}1} 2^{2k} \int_{Y^{k}} \left({|v-{\langle v \rangle}_{Y^{k,1}}|}^2 +{|v-{\langle v \rangle}_{Y^{k,2}}|}^2\right) {\;\text{d}}x.$$ For every $k{\geqslant}1$, we can apply Lemma \[lem1\] to the function $v\circ g^{k-1}$. Since $\gamma^k=g^{k-1}(\gamma)$, we obtain $$\begin{aligned} I_1 &{\leqslant}& c\left( \sum_{k{\geqslant}1} 2^k \int_{\gamma_k} {|v-{\langle v \rangle}_{\gamma_k}|}^2{\;\text{d}}\mu + \sum_{k{\geqslant}1} \sum_{m{\geqslant}1} \kappa^{m} \sum_{\tau \in \Sigma^1\cap {{\mathcal A}}_m} \int_{g^{k-1}(f_\tau(Y^0))} {|\nabla v|}^2{\;\text{d}}x \right) \notag \\ &=& c\left( \sum_{k{\geqslant}1} 2^k \int_{\gamma_k} {|v-{\langle v \rangle}_{\gamma_k}|}^2{\;\text{d}}\mu + \sum_{k{\geqslant}1} \sum_{p{\geqslant}k} \kappa^{p-k+1} \sum_{\tau \in \Sigma^k\cap {{\mathcal A}}_p} \int_{f_\tau(Y^0)} {|\nabla v|}^2{\;\text{d}}x \right), \label{ineq-I1}\end{aligned}$$ for some constant $c>0$ independent of $v$, since $\{g^{k-1}\circ f_\tau,~ \tau \in \Sigma^1\cap {{\mathcal A}}_m\} = \{f_\tau,~ \tau \in \Sigma^k\cap {{\mathcal A}}_{m+k-1}\}$.\ Let us now deal with $I_2$. For every $R>0$ and $i=1,2$, call $S_R^i = S_R \cap {C}_i$ and ${\langle v \rangle}_{S_R^i}$ the mean value of $v$ on the set $S_R^i$. We note that $$\label{maj-I2} I_2 ~{\leqslant}~ \frac 1 2 \sum_{i=1,2} \; \sum_{k{\geqslant}1} 2^{2k} \left( \int_{Y^{k}} {|{\langle v \rangle}_{Y^{k,i}} - {\langle v \rangle}_{S_{\rho(x)}^i}|}^2 \right).$$ Take $i\in \{1,2\}$ and $x_0\in f_i(Y^0)$, and, for every integer $k{\geqslant}1$, $\rho_k = \rho(x_0)/2^{k-1}$.\ We observe that $\sum_k 2^{2k} \int_{Y^{k}} {|{\langle v \rangle}_{Y^{k,i}} - {\langle v \rangle}_{S_{\rho(x)}^i}|}^2 {\leqslant}2(J_1+J_2)$, where $$\begin{aligned} J_1 &=& \sum_{k{\geqslant}1} 2^{2k+1} |Y^{k,i}| {({\langle v \rangle}_{Y^{k,i}} - {\langle v \rangle}_{S_{\rho_k}^i})}^2, \label{J1}\\ J_2 &=& \sum_{k{\geqslant}1} 2^{2k} \int_{Y^{k}} {|{\langle v \rangle}_{S_{\rho_k}^i} - {\langle v \rangle}_{S_{\rho(x)}^i}|}^2 {\;\text{d}}x.\end{aligned}$$ Let us first examine $J_1$. The following Poincar[é]{} inequality holds in $Y^{1,i}=f_i(Y^0)$: for every $v\in H^1(Y^{1,i})$, $$\label{poincare} \int_{Y^{1,i}} {|v(x)-{\langle v \rangle}_{S_{\rho_i}}|}^2 {\;\text{d}}x ~{\leqslant}~ M \int_{Y^{1,i}} {|\nabla v|}^2 {\;\text{d}}x,$$ where the constant $M>0$ is independent of $v$.\ Observe that for every integer $k{\geqslant}1$, $Y^{k,i} = g^{k-1}(Y^{1,i})$, and $S^i_{\rho_k} = g^{k-1}(S^i_{\rho_1})$. Then $$\begin{aligned} {|{\langle v \rangle}_{Y^{k,i}} - {\langle v \rangle}_{S_{\rho_k}^i}|}^2 &=& \frac{1}{{|Y^{k,i}|}^2} {\left|\int_{Y^{k,i}} (v(x)-{\langle v \rangle}_{S_{\rho_k}^i}){\;\text{d}}x \right|}^2\\ &{\leqslant}& \frac 1 {|Y^{k,i}|} \int_{Y^{k,i}} {|v(x)-{\langle v \rangle}_{S_{\rho_k}^i}|}^2 {\;\text{d}}x\\ &{\leqslant}& \frac{1}{2^{2(k-1)}|Y^{k,i}|} \int_{Y^{1,i}} {|v\circ g^{k-1} (x)-{\langle v\circ g^{k-1} \rangle}_{S_{\rho_1}^i}|}^2 {\;\text{d}}x\\ &{\leqslant}& \frac{M}{2^{2(k-1)}|Y^{k,i}|} \int_{Y^{1,i}} {|\nabla(v\circ g^{k-1})|}^2{\;\text{d}}x\\ &=& \frac{M}{2^{2(k-1)}|Y^{k,i}|} \int_{Y^{k,i}} {|\nabla v|}^2{\;\text{d}}x,\end{aligned}$$ where we used . From this and , we deduce that $$\label{ineq-J1} J_1~{\leqslant}~ 8M \sum_{k{\geqslant}1} \int_{Y^{k,i}} {|\nabla v |}^2.$$ To deal with $J_2$, we use polar coordinates $(\rho,{\varphi})$ centered at $A$. Introduce the positive constants $R_0,R_1$ such that for all $(\rho,{\varphi})\in Y^{1,i}$, $R_0{\leqslant}\rho{\leqslant}R_1$. Observe that if $x\in Y^{k}$, $$\begin{aligned} {|{\langle v \rangle}_{S_{\rho(x)}^i} -{\langle v \rangle}_{S_{\rho_k}^i} |}^2 &=& {\left|\frac{1}{{\varphi}_1-{\varphi}_0} \int_{{\varphi}_0}^{{\varphi}_1} (v(\rho(x),{\varphi})-v(\rho_k,{\varphi})) {\;\text{d}}{\varphi}\right|}^2\\ &=& {\left| \frac{1}{{\varphi}_1-{\varphi}_0} \int_{{\varphi}_0}^{{\varphi}_1} \int_{\rho_k}^{\rho(x)} \frac{\partial v}{\partial \rho}(s,{\varphi}){\;\text{d}}s {\;\text{d}}{\varphi}\right|}^2\\ &{\leqslant}& \frac{1}{({\varphi}_1-{\varphi}_0)^2} \left|\int_{{\varphi}_0}^{{\varphi}_1} \int_{\rho_k}^{\rho(x)} {\left|\frac{\partial v}{\partial \rho}(s,{\varphi})\right|}^2 s {\;\text{d}}s {\;\text{d}}{\varphi}\right| \times \left|\int_{{\varphi}_0}^{{\varphi}_1} \int_{\rho_k}^{\rho(x)} \frac{{\;\text{d}}s}s {\;\text{d}}{\varphi}\right|\\ &{\leqslant}& \frac{\log \frac {R_1}{R_0}}{{\varphi}_1-{\varphi}_0} \int_{{\varphi}_0}^{{\varphi}_1} \int_{R_0/2^{k-1}}^{R_1/2^{k-1}} {\left|\frac{\partial v}{\partial \rho}(s,{\varphi})\right|}^2 s {\;\text{d}}s {\;\text{d}}{\varphi}\\ &=& \frac{\log \frac {R_1}{R_0}}{{\varphi}_1-{\varphi}_0} \int_{C(k)} {|\nabla v(y)|}^2 {\;\text{d}}y,\end{aligned}$$ where $C(k) = \{(\rho,{\varphi}),~ \frac{R_0}{2^{k-1}} < \rho < \frac{R_1}{2^{k-1}},~ {\varphi}_1<|{\varphi}|<{\varphi}_0\}$. Hence, $$J_2 ~{\leqslant}~ \frac{\log \frac {R_1}{R_0}}{{\varphi}_1-{\varphi}_0} \sum_{k{\geqslant}1} 2^{2k} \int_{Y^{k}} \int_{C(k)} {|\nabla v(y)|}^2 {\;\text{d}}y {\;\text{d}}x,$$ and there is a constant $c_3>0$ independent of $v$ such that $$J_2 ~{\leqslant}~ c_3 \sum_{k{\geqslant}1} \int_{C(k)} {|\nabla v(y)|}^2 {\;\text{d}}y.$$ Note that every point $(\rho,{\varphi})$ in $C(k)$ lies in at most $\log_2 \frac {R_1}{R_0}$ sets $C(l)$, $l{\geqslant}1$. Therefore, $$\label{ineq-J2} J_2 ~{\leqslant}~ c_3 \log_2 \frac {R_1}{R_0} \int_{C} {|\nabla v(y)|}^2{\;\text{d}}y ~{\leqslant}~ c_3 \log_2 \frac {R_1}{R_0} \sum_{k{\geqslant}1} \int_{Y^k} {|\nabla v|}^2{\;\text{d}}x.$$ By the inequalities and , $I_2 {\leqslant}c_4 \sum_{k{\geqslant}1} \int_{Y^k} {|\nabla v|}^2{\;\text{d}}x$ for some constant $c_4>0$ independent of $v$. Hence, $$\label{ineq-I2} I_2 ~{\leqslant}~ \frac{c_4}\kappa \sum_{k{\geqslant}1} \sum_{p{\geqslant}k} \kappa^{p-k+1} \sum_{\tau \in \Sigma^k\cap {{\mathcal A}}_p} \int_{f_\tau(Y^0)} {|\nabla v|}^2{\;\text{d}}x.$$ Indeed, observe that the terms in the sum of for which $p=k$ are exactly $\kappa \int_{Y^k}{|\nabla v|}^2{\;\text{d}}x$.\ Therefore, and yield, which , which achieves the proof. In the proof we will write $\lesssim$ when there may arise in the inequality a constant that does not depend on the function $v\in H^1({\Omega_\mathrm{int}})$ we consider.\ By Lemma \[lem2\] and Theorem \[auscher-badr\], for every $v\in H^1({\Omega_\mathrm{int}})$ such that $v_{|\Gamma}\in H^{1/2}(\Gamma)$, there exists $\Lambda v\in H^1({{\mathbb R}}^2)$ such that $(\Lambda v)_{|C}=v$ and $${\Vert \Lambda v \Vert}_{H^1({{\mathbb R}}^2)}^2 ~\lesssim ~ {\Vert v \Vert}_{H^1(C)}^2 + {\left\Vert \frac{v_a}\rho\right\Vert}_{L^2(C)}^2.$$ Define $\widehat {C} = \text{Int}(\overline C \cup \overline Y^0)$ (see Figure \[cone\]). Introducing a cut-off function with support in the main hole $T$ and using the operator $\Lambda$ and Remark \[rem-AB\], we can construct a linear extension operator ${{\mathcal F}}_0$ from $H^1(\widehat C)$ to $H^1(\widehat C \cup \overline T)$ such that ${{\mathcal F}}1 = 1$ and for all $v\in H^1(\widehat C)$ $$\label{eq-th9} \int_T {|\nabla ({{\mathcal F}}_0 v)|}^2 {\;\text{d}}x ~\lesssim ~\int_{\widehat C} {|\nabla v|}^2{\;\text{d}}x + \int_C {\bigg| \frac{v_a}{\rho}\bigg|}^2{\;\text{d}}x,$$ which also implies $$\label{a-b} {\Vert {{\mathcal F}}_0 v \Vert}_{H^1(\widehat C\cup \overline T)}^2 ~\lesssim ~ {\Vert v \Vert}^2_{H^1(\widehat C)} + {\left\Vert \frac{v_a}\rho\right\Vert}_{L^2(C)}^2.$$ We will now define an extension $\tilde v\in H^1(\widetilde \Omega_\text{int})$ of a function $v$ as in Theorem \[ext-thm\], where $\widetilde \Omega_\text{int}$ is the convex hull of the domain ${\Omega_\mathrm{int}}$. Recall that $T$ is the main hole and $\{f_\sigma(T)$, $\sigma\in {{\mathcal A}}\}$ is the collection of the holes of the domain $\Omega_\text{int}$ (see Figure \[cone\]). Introduce the function $\tilde v$ defined in $\widetilde \Omega_\text{int}$ by $$\left\{ \begin{array}{ll} \tilde v := v &\text{in } {\Omega_\mathrm{int}},\\ \tilde v := {{\mathcal F}}_0(v\circ f_\sigma)\circ {f_\sigma}^{-1} & \text{in } f_\sigma(T),~ \sigma \in {{\mathcal A}}. \end{array}\right.$$ By Lemma \[lem2\] and , we get the estimate $$\label{estimation-T} \int_T {|\nabla ({{\mathcal F}}_0 v)|}^2 {\;\text{d}}x ~\lesssim ~ \sum_{k{\geqslant}1} 2^k \int_{\gamma_k} {|v-{\langle v \rangle}_{\gamma_k}|}^2{\;\text{d}}\mu + \sum_{k{\geqslant}0} \sum_{p{\geqslant}k} \kappa^{p-k+1} \sum_{\tau \in \Sigma^k\cap {{\mathcal A}}_p} \int_{f_\tau(Y^0)} {|\nabla v|}^2{\;\text{d}}x.$$ Indeed, $\kappa \int_{Y^0} {|\nabla v|}^2 {\;\text{d}}x$ (*resp.* $\kappa \int_{C} {|\nabla v|}^2{\;\text{d}}x$) is bounded from above by the terms for which $k=p=0$ (*resp.* $k{\geqslant}1$, $p=k$) in the second sum in .\ Observe that for every integer $k{\geqslant}1$, $\gamma^k \subset \widetilde \Gamma^{\sigma^k}$ where $\sigma^k = (1,2,\ldots,2)\in {{\mathcal A}}_{k-2}$, (recall that the sets $\widetilde \Gamma^{\sigma}$ have been introduced in Remark \[norme-Lip\]). Therefore, $$\label{ineq-lip} \int_{\gamma_k} {|v-{\langle v \rangle}_{\gamma_k}|}^2 {\;\text{d}}\mu ~\lesssim~ \int_{\widetilde \Gamma^{\sigma^k}} {|v-{\langle v \rangle}_{\widetilde \Gamma^{\sigma^{k}}}|}^2{\;\text{d}}x,$$ for $i=1,2$, where the constant in the inequality does not depend on $k$. Take $\sigma\in {{\mathcal A}}_n$, one has $$\begin{aligned} \int_{f_\sigma(T)} {|\nabla \tilde v|}^2 {\;\text{d}}x &=& \int_{T} {|\nabla({{\mathcal F}}_0(v\circ f_\sigma))|}^2{\;\text{d}}x\\ &\lesssim & \sum_{k{\geqslant}1} 2^{k+n} \int_{f_\sigma(\widetilde \Gamma^{\sigma^k})} {|v-{\langle v \rangle}_{f_\sigma(\widetilde \Gamma^{\sigma^k})}|}^2{\;\text{d}}\mu \notag \\ && ~~ + \sum_{k{\geqslant}0} \sum_{p{\geqslant}k} \kappa^{p-k+1} \sum_{\tau \in \Sigma^k\cap {{\mathcal A}}_p} \int_{f_{\sigma \tau}(Y^0)} {|\nabla v|}^2{\;\text{d}}x , \label{estimation-Tsigma}\end{aligned}$$ where we applied to the function $v\circ f_\sigma$, and we used . The constant in the inequality does not depend on $n$. The notation $\sigma \tau$ for $\tau \in {{\mathcal A}}_k$ stands for $(\sigma(1),\ldots,\sigma(n),\tau(1),\ldots,\tau(k))\in {{\mathcal A}}_{n+k}$.\ We can write $$\begin{aligned} \int_{{\widetilde \Omega_\text{int}}} {|\nabla \tilde v|}^2 {\;\text{d}}x &=& \int_{{\Omega_\mathrm{int}}} {|\nabla \tilde v|}^2 {\;\text{d}}x + \sum_{\sigma\in {{\mathcal A}}} \int_{f_\sigma(T)} {|\nabla \tilde v|}^2 {\;\text{d}}x\\ & \lesssim& \int_{{\Omega_\mathrm{int}}} {|\nabla \tilde v|}^2 {\;\text{d}}x + S_1+S_2,\end{aligned}$$ where $$\begin{aligned} S_1 &=& \sum_{n{\geqslant}0} \sum_{\sigma\in {{\mathcal A}}_n} \sum_{k{\geqslant}1} 2^{k+n} \int_{f_\sigma(\widetilde \Gamma^{\sigma^k})} {|v-{\langle v \rangle}_{f_\sigma(\widetilde \Gamma^{\sigma^k})}|}^2{\;\text{d}}\mu,\\ S_2 &=& \sum_{\sigma\in {{\mathcal A}}} \sum_{k{\geqslant}0} \sum_{p{\geqslant}k} \kappa^{p-k+1} \sum_{\tau \in \Sigma^k\cap {{\mathcal A}}_p} \int_{f_{\sigma \tau}(Y^0)} {|\nabla v|}^2{\;\text{d}}x.\end{aligned}$$ We first deal with $S_1$. Take $\sigma,\tau\in {{\mathcal A}}$ and $k,l{\geqslant}1$. Note that if $\sigma \sigma^k = \tau \sigma^l$, then $k=l$ and $\sigma=\tau$. Therefore, $$\label{S1} S_1 ~{\leqslant}~ \sum_{k{\geqslant}0} 2^k \sum_{\sigma\in {{\mathcal A}}_k} \int_{\widetilde \Gamma^\sigma} {|v-{\langle v \rangle}_{\widetilde \Gamma^\sigma}|}^2{\;\text{d}}\mu ~\lesssim ~ {\Vert v_{|\Gamma} \Vert}_{H^{1/2}(\Gamma)}^2$$ by .\ We are left with dealing with $S_2$. Assume that $\eta\in {{\mathcal A}}_N$ and $\eta=\sigma \tau$ with $\tau\in \Sigma^k\cap {{\mathcal A}}_p$, then $p{\leqslant}N$. Since the sets $\Sigma^k$, $k{\geqslant}1$, are pairwise disjoint, this means that the term $\int_{f_\eta(Y^0)} {|\nabla v |}^2{\;\text{d}}x$ appears at most $N$ times in the sum $S_2$.\ Moreover, we observe that $p-k+1 \in [1,N+1]$. It can be seen that there is at most one quadruplet $(\sigma',\tau',l,q)$ with $\sigma'\in {{\mathcal A}}$, $\tau'\in \Sigma^l\cap {{\mathcal A}}_{q}$, $l{\geqslant}1$ and $q{\geqslant}l$ distinct from $(\sigma,\tau,k,p)$ such that $\eta=\sigma'\tau'$ and $p-k+1=q-l+1$. As a consequence, $$\begin{aligned} S_2 &{\leqslant}& 2 \sum_{N{\geqslant}0} \sum_{\eta\in {{\mathcal A}}_N} \sum_{m=1}^{N+1} \kappa^m \int_{f_\eta(Y^0)} {|\nabla v|}^2{\;\text{d}}x \notag\\ &{\leqslant}& \frac 2{1-\kappa} \sum_{\eta\in {{\mathcal A}}} \int_{f_\eta(Y^0)} {|\nabla v|}^2{\;\text{d}}x ~=~ \frac 2{1-\kappa} \int_{{\Omega_\mathrm{int}}} {|\nabla v|}^2{\;\text{d}}x. \label{S2}\end{aligned}$$ Therefore, and give $${\Vert \tilde v \Vert}_{H^1({\widetilde \Omega_\text{int}})}^2 ~\lesssim~ {\Vert v \Vert}_{H^1({\Omega_\mathrm{int}})}^2 + {\Vert v_{|\Gamma} \Vert}_{H^{1/2}(\Gamma)}^2.$$ Since $\widetilde\Omega_\text{int}$ is a polygonal domain, we can further extend $\tilde v$ into a function ${{\mathcal F}}v$ in $H^1(D)$, where ${{\mathcal F}}$ is a linear operator satisfying . #### Proof of Theorem \[M-cv-theta=0\] We will prove separately points \[point1\] and \[point2\] in Definition \[Mosco-convergence\].\ *Proof of point \[point1\]*  Suppose that ${(u_n)}$ weakly converges to $u$ in $L^2(D)$. Without loss of generality, one can suppose ${\underline{\lim}~}a_n(u_n,u_n)$ is finite. We may further assume that there exists a subsequence, still called $(u_n)$, such that $a_n(u_n,u_n)$ converges to some real number as $n\to \infty$; as a consequence, there exists a constant $c$ independent of $n$ such that $$\label{u-bound} a_n(u_n,u_n) ~{\leqslant}~ c.$$ In particular, for all $n$, $u_n\in V^n$, which implies that $u_n\in V$. Then, implies that $({u_n}_{|{\Omega_\mathrm{int}}})$ is bounded in $H^1({\Omega_\mathrm{int}})$ and that $\sqrt {\nu_n} \nabla {u_n}_{|{\Omega_\mathrm{ext}}} $ is bounded in $L^2 ({\Omega_\mathrm{ext}})$. Therefore, there exists a subsequence that we still denote $(u_n)$ such that - $({u_n}_{|{\Omega_\mathrm{int}}})$ converges to $u_{|{\Omega_\mathrm{int}}}$ weakly in $H^1({\Omega_\mathrm{int}})$, and strongly in $L^2({\Omega_\mathrm{int}})$ - $\sqrt {\nu_n} \; \nabla {u_n}_{|{\Omega_\mathrm{ext}}} $ converges weakly in $L^2({\Omega_\mathrm{ext}})$ to $\nabla {u}_{|{\Omega_\mathrm{ext}}}$ (recall that $\nu_n$ converges to $1$ almost everywhere, so the weak limit of $\sqrt {\nu_n} \; \nabla {u_n}_{|{\Omega_\mathrm{ext}}} $ must be $\nabla {u}_{|{\Omega_\mathrm{ext}}}$.) Thus $$\begin{split} & {\underline{\lim}~}\left(\int_{{\Omega_\mathrm{int}}^n} \left( {|\nabla u_n|}^2 +\beta u_n^2 \right ) {\;\text{d}}x + \int_{{\Omega_\mathrm{ext}}^n} \left ( \nu_n {|\nabla u_n|}^2 + \beta u_n^2\right) {\;\text{d}}x \right) \\ {\geqslant}~ & \int_{{\Omega_\mathrm{int}}} \left({|\nabla u|}^2+\beta u^2\right) {\;\text{d}}x + \int_{{\Omega_\mathrm{ext}}} \left({|\nabla u|}^2+\beta u ^2\right) {\;\text{d}}x. \end{split}$$ Moreover, exactly as in the proof of Theorem \[M-cv\], we see that $\frac 1{|\Gamma^n|}\int_{\Gamma^n} u_n^2{\;\text{d}}x \to \int_\Gamma u^2{\;\text{d}}\mu$ as $n\to \infty$. We have proved point \[point1\]\ *Proof of point \[point2\]*  Take $u\in L^2(D)$. By , we may assume that $u\in V$. We must construct $(u_n)$ converging strongly in $L^2(D)$ such that holds.\ Recall that $H$ is the height of the triangle $T$ and that $\Gamma^0$ is a segment of the line $\{x_2=0\}$, $\Gamma$ is a segment of the line $\{x_2=2H\}$. We then introduce a sequence of smooth cut-off functions $\chi_n (x_2)$ such that $\chi_n(x_2)=0$ if $x_2\le 2H - \frac {3H} {2^{n+1}}$, $\chi_n(x_2)=1$ if $x_2\in [2H- \frac H {2^{n}} , 2H ]$ and that $ 2^{-n} \|\chi_n' \|_{L^\infty} + \|\chi_n \|_{L^\infty}$ is bounded by a constant independent of $n$.\ We now define the functions $u_n$ by $$\label{u_n} u_n(x)=\left\{ \begin{array}[c]{ll} {u_\mathrm{int}}(x)\quad & \forall x\in {\Omega_\mathrm{int}}, \\ {u_\mathrm{ext}}(x)\quad & \forall x\in U, \\ {{\mathcal E}}( {u_\mathrm{int}})(x) \quad &{\displaystyle}\forall x \in \bigcup_{m\ge n} \bigcup_{\sigma\in {{\mathcal A}}_m } f_\sigma(T),\\ \chi_n(x_2) {{\mathcal E}}( {u_\mathrm{int}})(x) + (1-\chi_n(x_2)) {u_\mathrm{ext}}(x) \quad &{\displaystyle}\forall x\in \bigcup_{m=0}^{n-1} \bigcup_{\sigma\in {{\mathcal A}}_m } f_\sigma(T), \end{array} \right.$$ where ${{\mathcal E}}$ is the extension operator introduced in Theorem \[ext-thm\], ${u_\mathrm{int}}= u_{|{\Omega_\mathrm{int}}}$ and ${u_\mathrm{ext}}= u_{|{\Omega_\mathrm{ext}}}$. It is easy to check that $u_n$ belongs to the space $V^n$ and that the sequence $(u_n)$ strongly converges to $u$ in $L^2(D)$. We claim that $$\begin{aligned} \label{eq:9} \sum_{m=0}^{n-1} \sum_{\sigma\in {{\mathcal A}}_m } \int_{f_\sigma(T)} \nu_n(x) \left| \nabla \left ( \chi_n ( {{\mathcal E}}( {u_\mathrm{int}}) - {u_\mathrm{ext}}) \right)\right|^2{\;\text{d}}x \longrightarrow 0, \\ \label{eq:10} \sum_{m\ge n} \sum_{\sigma\in {{\mathcal A}}_m } \int_{f_\sigma(T)} \left| \nabla \left ( {{\mathcal E}}( {u_\mathrm{int}}) - {u_\mathrm{ext}}\right)\right|^2 {\;\text{d}}x\longrightarrow 0.\end{aligned}$$ Indeed, we readily obtain (\[eq:10\]) from the fact that the measure of $ \bigcup_{m\ge n} \bigcup_{\sigma\in {{\mathcal A}}_m } f_\sigma(T)$ tends to zero and the fact that $\| {{\mathcal E}}({u_\mathrm{int}})\|_{H^1 (D)}$ and $\| {u_\mathrm{ext}}\|_{H^1 (D)}$ are finite. We obtain (\[eq:9\]) because $\| \sqrt {\nu_n} \nabla \chi_n\|_{L^\infty} $ is bounded uniformly with respect to $n$ and $\chi_n$ is supported in a region with vanishing measure, and because $\| {{\mathcal E}}({u_\mathrm{int}})\|_{H^1 (D)}$ and $\| {u_\mathrm{ext}}\|_{H^1 ({\Omega_\mathrm{ext}})}$ are finite.\ Therefore, $$\begin{split} & \lim_{n\to \infty} \left( \int_{{\Omega_\mathrm{int}}^n} ( |\nabla {u_\mathrm{int}}^n |^2 +\beta |{u_\mathrm{int}}^n |^2) {\;\text{d}}x + \int_{{\Omega_\mathrm{ext}}^n} (\nu_n |\nabla {u_\mathrm{ext}}^n|^2+ \beta |{u_\mathrm{ext}}^n |^2){\;\text{d}}x \right) \\ =& \int_{{\Omega_\mathrm{int}}} ( |\nabla {u_\mathrm{int}}|^2 +\beta {u_\mathrm{int}}^2) {\;\text{d}}x + \int_{{\Omega_\mathrm{ext}}^n} ( |\nabla {u_\mathrm{ext}}|^2 + \beta {u_\mathrm{ext}}^2 ){\;\text{d}}x, \end{split}$$ where ${u_\mathrm{int}}^n := {u_n}_{|{\Omega_\mathrm{int}}^n}$ and ${u_\mathrm{ext}}^n := {u_n}_{|{\Omega_\mathrm{ext}}^n}$. Finally, from Proposition \[convergence-result\], $\frac{1}{|\Gamma^n|} \int_{\Gamma^n} u^2{\;\text{d}}x \longrightarrow \int_{\Gamma} u^2{\;\text{d}}\mu $ as $n\to \infty$. Collecting all the above results, we obtain that $\lim a_n(u_n,u_n) = a(u,u)$ as $n\to \infty$, thus point [*ii*]{}. [^1]: Universit[é]{} Paris Diderot, UFR Math[é]{}matiques, 175 rue du Chevaleret, 75013 Paris, France, UMR 7598, Laboratoire Jacques-Louis Lions, F-75005, Paris, France, [email protected] [^2]: [É]{}cole normale sup[é]{}rieure de Rennes, avenue Robert Schuman, 35170 Bruz, France, [email protected]
{ "pile_set_name": "ArXiv" }
--- abstract: 'The *ab initio* anisotropic Migdal-Eliashberg formalism has been used to examine the pairing mechanism and the nature of the superconducting gap in the recently discovered lithium-decorated monolayer graphene superconductor. Our results provide evidence that the superconducting transition in Li-decorated monolayer graphene can be explained within a standard phonon-mediated mechanism. We predict a single anisotropic superconducting gap and a critical temperature $T_c=$ 5.1-7.6 K, in very good agreement with the experimental results.' author: - 'Jing-Jing Zheng' - 'E. R. Margine' title: 'First-principles calculations of the superconducting properties in Li-decorated monolayer graphene within the anisotropic Migdal-Eliashberg formalism' --- INTRODUCTION ============ During the past decade graphene has revolutionized many areas of nanotechnology from organic electronics to photovoltaics, plasmonics, photonics, and energy storage [@Novoselov_NAT12]. One notable application that was missing from this list was superconductivity, despite numerous theoretical predictions of either a conventional or an unconventional pairing mechanism [@einenkel; @profeta; @Margine2014; @uchoa; @nandkishore; @kiesel]. Very recently, a high-resolution angle-resolved photoemission spectroscopy (ARPES) study has presented evidence supporting the appearance of a superconducting phase in Li-decorated monolayer graphene (LiC$_6$) around 5.9 K [@Ludbrook], within the standard phonon-mediated coupling mechanism. This work has been followed by two more studies that reported the observation of superconductivity in Ca-intercalated bilayer graphene [@Hasegawa] and in Ca-intercalated graphene laminates [@Chapman]. In this article, we investigate from first principles the nature of the superconducting gap in LiC$_6$. To this end, we solve the fully anisotropic Migdal-Eliashberg equations [@allen_mitrovic; @Margine2013] to obtain the superconducting transition temperature ($T_c$) and the variation of the superconducting energy gap on the Fermi surface. While previous *ab initio* calculations have shown that the electron-phonon coupling is sufficient to yield a critical temperature in 6.7-10.3 K range using the Allen-Dynes formula [@profeta; @Guzman; @kaloni] or the isotropic Eliashberg formalism [@Szcz2014], the nature of the superconducting gap has not yet been addressed. We find that, similar to bulk CaC$_{6}$ [@Sanna_PRB07; @Margine2016; @Szcz2014; @Szcz2015], Li-decorated monolayer graphene exhibits a single anisotropic gap in agreement with the experimental work [@Ludbrook]. ![(Color online)(a) Top- and side-view of a ball-and- stick model of LiC$_6$, with C in gray and Li in red. (b) Band structure of LiC$_6$. The inner and outer $\pi^*$ bands (with respect to $\Gamma$ point) are labeled as $\alpha$ and $\beta$ (blue dots). The interlayer band is labeled as IL (red dots). The size of the blue and red symbols is proportional to the contribution of C-$\emph{p$_{z}$}$ and Li-$\emph{s}$ character. (c) The two-dimensional Fermi surface of LiC$_{6}$ with the same color code as in (b). The Brillouin zones of a graphene unit cell and a $\sqrt{3} \times \sqrt{3} R$30$^\circ$ graphene supercell are shown as black full and dashed lines, respectively. []{data-label="fig1"}](fig1){width="\linewidth"} ![image](fig2){width="150mm"} METHODOLOGY =========== The calculations are performed within the local density approximation to density-functional theory [@lda1; @lda2] using planewaves and norm-conserving pseudopotentials [@nc1; @nc2], as implemented in the [Quantum-ESPRESSO]{} package [@QE]. The planewaves kinetic energy cutoff is 100 Ry and the structural optimization is performed until the forces on atoms are less than 10 meV/Å. Li-decorated monolayer graphene is described in the $\sqrt{3} \times \sqrt{3} R$30$^\circ$ graphene supercell with one lithium atom per unit cell. The optimized lattice constant and the adatom-graphene distance are $a=4.24$ Å and $h=1.78$ Å. A Brillouin-zone (BZ) $\Gamma-$centered $\mathbf{k}$-point mesh of 24$\times$24 and a Methfessel-Paxton smearing [@mp] of 0.02 Ry are adopted for the electronic charge density calculations. The phonon modes are computed within density-functional perturbation theory [@baroni2001] on a 6$\times$6 $\mathbf{q}$-mesh. We employ the EPW code [@Giustino2007; @Giustino2016; @EPW; @Margine2013; @Ponce] to obtain the superconducting gap. The calculation of the electronic wavefunctions required for the Wannier-Fourier interpolation [@Marzari; @Mostofi] in EPW is performed on a uniform unshifted BZ $\mathbf{k}$-point grid of size 12$\times$12. For the anisotropic Eliashberg equations, we use 120$\times$120 and 60$\times$60 $\mathbf{k}$- and $\mathbf{q}$-point grids. The Matsubara frequency cutoff is set to five times the largest phonon frequency ($5\times 200$ meV), and the Dirac delta functions are replaced by Lorentzians of widths 100 meV and 0.5 meV for electrons and phonons, respectively. ELECTRONIC AND VIBRATIONAL PROPERTIES ===================================== In Figs. \[fig1\] (a)-(c), we show the crystal structure of Li-decorated monolayer graphene along with the corresponding decomposed electronic band structure and Fermi surface. Three bands cross the Fermi level around the $\Gamma$ point, in agreement with previous reports [@profeta; @Guzman]. The inner and outer C $\pi^*$ bands, labeled as $\alpha$ and $\beta$ (blue dots), are obtained by folding the $\pi^*$ states of graphene from $K$ to $\Gamma$, following the superstructure induced by Li adsorption. The Li-derived band, labeled as IL (red dots), displays a nearly-free electron like dispersion upwards from about 0.56 eV below the Fermi energy. Similar weakly bound free-electron states have been observed in other layered materials [@profeta; @Calandra; @Boeri; @Kolmogorov] and nanotubes [@Margine_NFE], and their rapid downshift under doping is due to the combined effects of quantum confinement and electrostatic response [@profeta; @Boeri; @Margine_NFE]. The corresponding Fermi surface of LiC$_6$ can be divided into two concentric regions centered around the $\Gamma$ point. The inner region is characterized by a snowflake-like electron pocket intersecting a hexagonal electron pocket, which arises from the mixing of the inner C $\pi^*$ states with the Li $s$ states. These Fermi sheets resemble the $\Gamma$-centered Fermi surface observed in Ca and Li intercalated bilayer graphene [@Margine2016; @Shimizu_PRL14]. The outer region also has a snowflake-like shape and originates on the outer C $\pi^*$ states and the Li $s$ states. We now focus on the vibrational properties and the electron-phonon coupling (EPC) in LiC$_{6}$. Similar to bulk CaC$_6$ [@Calandra] and bilayer C$_6$CaC$_6$ [@Margine2016], one can clearly identify in Fig. \[fig2\](a) three regions in the phonon dispersion associated to (i) the Li-related modes (up to 50 meV, where above 37 meV are Li$_z$ modes mixed with carbon out-of-plane C$_z$ modes), (ii) the carbon out-of-plane C$_z$ vibrations (50-100 meV), and (iii) the carbon in-plane C$_{xy}$ modes (above 100 meV). The size of the symbols in Fig. \[fig2\](a) is proportional to the atomic displacements corresponding to Li and C in-plane and out-of-plane contributions. The isotropic Eliashberg spectral function $\alpha^{2}F(\omega)$ $${{\alpha^{2} F(\omega)}}= \frac{1}{{{N_{\rm F}}}N_{\bf k} N_{\bf q}} \sum_{{{\bf k}},{{\bf {k'}}},\nu} |{{g_{\bf k \bf {k'}}^{\nu}}}|^2 {{\delta(\epsilon_{\bf k})}}{{\delta(\epsilon_{\bf {k'}})}}\delta(\omega-{{\omega_{\bf q \nu}}}),$$ and the cumulative electron-phonon coupling strength $\lambda(\omega)$ $$\lambda(\omega) = 2 \int_{0}^{\omega} d\omega' \alpha^2 F(\omega')/\omega', \vspace{-0.1cm}$$ are shown in Figs. \[fig2\](b)-(c). In these expressions ${{N_{\rm F}}}$ represents the density of electronic states per spin at the Fermi level, $N_{\bf k}$ and $N_{\bf q}$ are the total numbers of ${\bf k}$ and ${\bf q}$ points, ${{\epsilon_{\bf k}}}$ is the Kohn-Sham eigenvalue with respect to the Fermi level, and ${{g_{\bf k \bf {k'}}^{\nu}}}$ is the screened electron-phonon matrix element for the scattering between the electronic states ${{\bf k}}$ and ${{\bf {k'}}}$ through a phonon with wave vector ${{\bf q}}\!=\!{{\bf {k'}}}\!-{{\bf k}}$, frequency ${{\omega_{\bf q \nu}}}$ and branch index $\nu$. Here ${{\bf k}}$ and ${{\bf {k'}}}$ indicate both the electron wavevector and the band index. We find that the low-energy phonons are key to achieving a high electron-phonon coupling in LiC$_6$ as they account for 0.28 (51%) of the total EPC ($\lambda$ = 0.55). On the other hand, the electron-phonon coupling strengths associated with the out-of-plane C$_z$ and in-plane C$_{xy}$ modes are 0.12 (22%) and 0.15 (27%), respectively. Such behavior has also been found in bilayer C$_6$CaC$_6$, where the most significant contribution to the EPC comes from the low-energy phonon modes [@Margine2016]. Overall, our calculated EPC $\lambda=0.55$ is in good agreement with the experimental value $0.58\pm0.05$ observed at the highest Li coverage [@Ludbrook] and the values reported in previous theoretical studies [@profeta; @Guzman]. ![(Color online) Distribution of the electron-phonon coupling strength $\lambda_{\mathbf{k}}$ of LiC$_6$. Inset: Momentum-resolved electron-phonon coupling parameters $\lambda_{\mathbf{k}}$ on the Fermi surface (the data points correspond to electrons within $\pm 150$ meV from the Fermi energy).[]{data-label="fig3"}](fig3){width="75mm"} ![image](fig4){width="150mm"} To quantify the anisotropy in the electron-phonon coupling, we further evaluate the momentum-resolved EPC $\lambda_\mathbf{k}$ [@Margine2013], defined as: $$\lambda_{{{\bf k}}} = \sum_{{{\bf {k'}}},\nu} {{\delta(\epsilon_{\bf {k'}})}}|{{g_{\bf k \bf {k'}}^{\nu}}}|^2/{{\omega_{\bf q \nu}}}. \vspace{-0.1cm}$$ The calculated $\lambda_\mathbf{k}$ displays a significant anisotropy with a distribution in the 0.42-0.78 range as shown in Fig. \[fig3\]. This is in line with experimental ARPES measurements where a marked anisotropy in the electron-phonon coupling has been observed in the case of decorated graphene [@Ludbrook; @Fedorov_NatCom14], intercalated bilayer graphene [@Kleeman_JPSJ14], and intercalated graphite [@Gruneis_PRB09; @Valla_PRL09]. An alternative way to look at the EPC anisotropy is presented in the inset of Fig. \[fig3\], where the variation of $\lambda_\mathbf{k}$ on the Fermi surface is shown. When compared with the Fermi surface plot in Fig. \[fig1\](c), one can clearly see that the largest value of $\lambda_\mathbf{k}$ is attained on the portions of the Fermi surface dominated by the Li states. Notably, in bulk CaC$_6$, $\lambda_\mathbf{k}$ was also found to be larger for the states with Ca dominant orbital character on the Fermi surface [@Sanna_PRB07]. An important implication of this finding is that the IL state and its associated interaction play a critical role in the superconducting paring of LiC$_6$. A recent ARPES study has provided compelling evidence regarding the importance of the IL band in the pairing mechanism of bulk CaC$_6$ [@Yang]. Furthermore, the lack of any sign of superconductivity down to 3.5 K in few-layer graphene under large charge doping induced by electrochemical gating [@Gonnelli] provides additional proof of the vital role of dopant atoms. SUPERCONDUCTING PROPERTIES ========================== The superconducting properties of LiC$_6$ are obtained by solving self-consistently the fully anisotropic Migdal-Eliashberg equations along the imaginary axis at the fermion Matsubara frequencies ${{\omega_n}}=(2n+1)\pi T$ (with $n$ an integer) for each temperature $T$  [@Margine2013; @Margine2014; @allen_mitrovic; @Choi]: $$\begin{aligned} \hspace{-0.25cm} &&Z({{\bf k}},i{{\omega_n}}) = 1 + \frac{\pi T}{{{N_{\rm F}}}{{\omega_n}}} \sum_{{{\bf {k'}}}n'} \frac{ {{\omega_{n'}}}}{ \sqrt{{{\omega_{n'}}}^2+\Delta^2({{\bf {k'}}},i{{\omega_{n'}}})} } \nonumber \\ &&\qquad\qquad\qquad\times {{\delta(\epsilon_{\bf {k'}})}}\lambda({{\bf k}},{{\bf {k'}}},n\!-\!n'), \label{Znorm_surf} \\ &&Z({{\bf k}},i{{\omega_n}})\Delta({{\bf k}},i{{\omega_n}}) = \frac{\pi T}{{{N_{\rm F}}}} \sum_{{{\bf {k'}}}n'} \frac{ \Delta({{\bf {k'}}},i{{\omega_{n'}}}) }{ \sqrt{{{\omega_{n'}}}^2+\Delta^2({{\bf {k'}}},i{{\omega_{n'}}})} } \nonumber \\ &&\qquad\qquad\qquad\times {{\delta(\epsilon_{\bf {k'}})}}\left[ \lambda({{\bf k}},{{\bf {k'}}},\!n-\!n')-\mu_c^*\right].\nonumber\\ \label{Delta_surf}\end{aligned}$$ $Z({{\bf k}},i{{\omega_n}})$ is the mass renormalization function, $\Delta({{\bf k}},i{{\omega_n}})$ is the superconducting gap function, $\lambda({{\bf k}},{{\bf {k'}}},\!n-\!n')$ is the momentum- and energy-dependent EPC, and $\mu_c^*$ is the semiempirical Coulomb parameter. The anisotropic $\lambda({{\bf k}},{{\bf {k'}}},\!n-\!n')$ to be used in the Eliashberg equations is given by: $$\label{lambda} \lambda({{\bf k}},{{\bf {k'}}},n - n') = {{N_{\rm F}}}\sum_{\nu} \frac{2\omega_{{{\bf q}}\nu}}{({{\omega_n}}- {{\omega_{n'}}})^2+\omega_{{{\bf q}}\nu}^2} |{{g_{\bf k \bf {k'}}^{\nu}}}|^2.$$ Figure \[fig4\](a) shows the superconducting energy gap $\Delta_\mathbf{k}$ as a function of temperature, calculated for a screened Coulomb parameter $\mu^* = 0.14$, together with the average value of the gap (red squares). The superconducting gap $\Delta_\mathbf{k}$ on different parts of the Fermi surface at 0.5 K is shown in Fig. \[fig4\](b). We find that monolayer LiC$_{6}$ displays a single anisotropic gap with an average value $\Delta_0 = 0.89$ meV in the $T=0$ K limit, in very good agreement with the ARPES result of $0.9 \pm 0.2$ meV, measured at 3.5 K [@Ludbrook]. This situation is similar to bulk CaC$_{6}$ where the multiple-sheet Fermi surface gives rise to a single gap structure with a sizable anisotropy [@Sanna_PRB07; @Margine2016], but unlike bilayer C$_{6}$CaC$_{6}$ for which a two gap structure has been recently predicted [@Margine2016]. The superconducting $T_{c}$ is identified as the highest temperature at which the gap vanishes. From Fig. \[fig4\](a) we find $T_{c} = 5.9$ K and a ratio $2 \Delta_0 / k_B T_c = 3.50$, very close to the ideal BCS value of 3.53 [@BCS]. The predicted superconducting critical temperature is in excellent agreement with the experimental estimation of 5.9 K based on measurements of the size of the superconducting gap [@Ludbrook]. The temperature dependence of the superconducting gap can be well fitted with a BCS model, as obtained by solving numerically the BCS gap equation [@Johnston] with $\Delta_0$ and $T_c$ from our first-principles calculations. This is shown by the blue dashed line in Fig. \[fig4\](a). These results provide support for a conventional phonon-mediated mechanism as the superconducting origin in Li-decorated graphene. For completeness, we also explore the sensitivity of the calculated superconducting energy gap and critical temperature to the choice of the Coulomb parameter $\mu^*$, as shown in Figs. \[fig4\](c)-(d). For $\mu^*$ = 0.12 and 0.16, we obtain $\Delta_0$ = 1.10 meV and 0.69 meV and $T_{c}$ = 7.6 and 5.1 K, respectively. Finally, using the $\alpha$–model [@Johnston; @Padamsee], we obtain the temperature dependence of the reduced electronic specific heat in the superconducting state. Within this model, the ratio $\alpha$ = $\Delta_0 / k_B T_c$ is an adjustable parameter and the normalized superconducting state electronic entropy $S_{es}$ and heat capacity $C_{es}$ are expressed in terms of $\gamma_n T_c$ as: $$\label{entropy} \frac{S_{es}(t)}{\gamma_{n}T_{c}} =- \frac{6\alpha}{\pi^{2}}\int^{\infty}_{0}[f\ln(f)+(1-f)\ln(1-f)]d \tilde{\varepsilon},$$ $$\label{specificheat} \frac{C_{es}(t)}{\gamma_{n}T_{c}} = t\frac{d(S_{es}/\gamma_{n}T_{c})}{dt},$$ where $f= \left[\exp(\alpha \tilde{E}/t)+1\right]^{-1}$ is the Fermi-Dirac distribution function, $t=T/T_c$ is the reduced temperature, and $\gamma_n =(2/3) \pi^2 k_B^2 {{N_{\rm F}}}$ is the Sommerfeld coefficient. The reduced quasi-particle energy is defined as $\tilde{E}=\sqrt{\tilde{\varepsilon}^{2}+\delta^{2}(t)}$, where $\tilde{\varepsilon}= \varepsilon/\Delta_{0}$ is the reduced normal state single-particle energy relative to the Fermi level and $\delta(t) = \Delta(T)/\Delta_{0}$ is the reduced gap function. The upper limit in the integral in Eq. (\[entropy\]) is set to $500 \gg 1$. Figure \[fig4\](e) shows the calculated $C_{es}/\gamma_n T_c$ for $\alpha=1.75$. The temperature dependence of the normalized gap $\delta(t)$ is assumed to be the same as in the BCS theory [@Johnston; @Padamsee]. We checked the numerical results by comparing the data for $\alpha_{BCS}=1.764$ with Tables II–IV in Ref. \[\] and by verifying that the entropy at the critical temperature is equal to that of the normal state. The shape of the calculated specific heat curve is consistent with a one-gap BCS model and undergoes a discontinuous jump at the critical temperature. The specific heat jump at $T_c$ is found to be $\Delta C_e (T_c) / \gamma_n T_c = 1.385$ \[shown in the inset of Fig. \[fig4\](e)\], close to the weak limit BCS value of 1.426 [@BCS]. Furthermore, this result is comparable to the experimetal and theoretical values reported for the normalized specific heat jump in bulk CaC$_6$ [@Kim; @Sanna_PRB07]. CONCLUSIONS =========== In conclusion, we have studied the superconducting properties in Li-decorated monolayer graphene within the [*ab initio*]{} anisotropic Migdal-Eliashberg theory. Our results provide support for a standard phonon-mediated mechanism at the origin of the superconducting transition. Most of the electron-phonon coupling originates from the low-energy modes dominated by the motion of Li atoms similar to bilayer C$_6$CaC$_6$. We find a sizable anisotropy in the electron-phonon coupling which yields a single anisotropic gap over the Fermi surface. Further enhancement in the critical temperature of LiC$_6$ is expected in the presence of a substrate [@kaloni] or under applied strain [@Pesic2014]. ACKNOWLEDGMENTS =============== J-J Zheng acknowledges the support from the China Scholarship Council (Grant No. 201508140043). [99]{} K. S. Novoselov, V. I. Fal$'$ko, L. Colombo, P. R. Gellert, M. G. Schwab, and K. Kim, Nature [**490**]{}, 192 (2012). M. Einenkel and K. B. Efetov, Phys. Rev. B [**84**]{}, 214508 (2011). G. Profeta, M. Calandra, and F. Mauri, Nat. Phys. [**8**]{}, 131 (2012). E. R. Margine and F. Giustino, Phys. Rev. B [**90**]{}, 014518, (2014). B. Uchoa and A. H. Castro Neto, Phys. Rev. Lett. [**98**]{}, 146801 (2007). R. Nandkishore, L. S. Levitov, and A. V. Chubukov, Nat. Phys. [**8**]{}, 158 (2012). M. L. Kiesel, C. Platt, W. Hanke, D. A. Abanin, and R. Thomale, Phys. Rev. B [**86**]{}, 020507(R) (2012). B. M. Ludbrook, G. Levy, P. Nigge, M. Zonno, M. Schneider, D. J. Dvorak, C. N. Veenstra, S. Zhdanovich, D. Wong, P. Dosanjh, C. Stra[ß]{}er, A. Stöhr, S. Forti, C. R. Ast, U. Starke, and A. Damascelli, Proc. Natl. Acad. Sci. USA. [**112**]{}, 11795 (2015). S. Ichinokura, K. Sugawara, A. Takayama, T. Takahashi, and S. Hasegawa, ACS Nano [**10**]{}, 2761 (2016). J. Chapman, Y. Su, C. A. Howard, D. Kundys, A. Grigorenko, F. Guinea, A. K. Geim, I. V. Grigorieva, and R. R. Nair, Sci. Rep. [**6**]{}, 23254 (2016). P. B. Allen and B. Mitrović, Solid State Phys. [**37**]{}, 1 (1982). E. R. Margine and F. Giustino, Phys. Rev. B [**87**]{}, 024505 (2013). D. M. Guzman, H. M. Alyahyaei, and R. A. Jishi, 2D Mater. [**1**]{}, 021005 (2014). T. P. Kaloni, A. V. Balatsky, and U. Schwingenschlögl, Eur. Phys. Lett. [**104**]{}, 47013 (2013). D. Szczȩśniak, A. P. Durajski, and R. Szczȩśniak, J. Phys.: Condens. Matter [**26**]{}, 255701 (2014). A. Sanna, G. Profeta, A. Floris, A. Marini, E. K. U. Gross, and S. Massidda, Phys. Rev. B [**75**]{}, 020511(R) (2007). E. R. Margine, H. Lambert, and F. Giustino, Sci. Rep. [**6**]{}, 21414 (2016). R. Szczȩśniak, E. A. Drzazga, and D. Szczȩśniak, Eur. Phys. J. B [**88**]{}, 52 (2015). D. M. Ceperley and B. J. Alder, Phys. Rev. Lett. [**45**]{}, 566 (1980). J. P. Perdew and A. Zunger, Phys. Rev. B [**23**]{}, 5048 (1981). N. Troullier and J. L. Martins, Phys. Rev. B [**43**]{}, 1993 (1991). M. Fuchs and M. Scheffler, Comput. Phys. Commun. [**119**]{}, 67 (1999). P. Giannozzi [*et al.*]{}, J. Phys. Condens. Matter [**21**]{}, 395502 (2009). M. Methfessel and A. T. Paxton, Phys. Rev. B [**40**]{}, 3616 (1989). S. Baroni, S. de Gironcoli, A. Dal Corso, and P. Giannozzi, Rev. Mod. Phys. [**73**]{}, 515 (2001). F. Giustino, M. L. Cohen, and S. G. Louie, Phys. Rev. B [**76**]{}, 165108 (2007). F. Giustino, arXiv:1603.06965. J. Noffsinger, F. Giustino, B. D. Malone, C.-H. Park, S. G. Louie, and M. L. Cohen, Comput. Phys. Commun. [**181**]{}, 2140 (2010). S. Poncé, E. R. Margine, C. Verdi, and F. Giustino, Comput. Phys. Commun. (2016) doi:10.1016/j.cpc.2016.07.028. N. Marzari, A. A. Mostofi, J. R. Yates, I. Souza, and D. Vanderbilt, Rev. Mod. Phys. [**84**]{}, 1419 (2012). A. A. Mostofi, J. R. Yates, Y.-S. Lee, I. Souza, D. Vanderbilt, and N. Marzari, Comput. Phys. Comm. [**178**]{}, 685 (2008). M. Calandra and F. Mauri, Phys. Rev. Lett. [**95**]{}, 237002 (2005). L. Boeri, G. B. Bachelet, M. Giantomassi, and O. K. Andersen, Phys. Rev. B [**76**]{}, 064510 (2007). M. Calandra, A. N. Kolmogorov, and S. Curtarolo, Phys. Rev. B [**75**]{}, 144506 (2007). E. R. Margine and V. H. Crespi, Phys. Rev. Lett. [**96**]{}, 196803 (2006). R. Shimizu, K. Sugawara, K. Kanetani, K. Iwaya, T. Sato, T. Takahashi, and T. Hitosugi, Phys. Rev. Lett. [**114**]{}, 146103 (2015). A. V. Fedorov, N. I. Verbitskiy, D. Haberer, C. Struzzi, L. Petaccia, D. Usachov, O. Y. Vilkov, D. V. Vyalikh, J. Fink, M. Knupfer, B. Büchner, and A. Grüneis, Nat. Commun. [**5**]{}, 3257 (2014). J. Kleeman, K. Sugawara, T. Sato, and T. Takahashi, J. Phys. Soc. Jpn. [**83**]{}, 124715 (2014). A. Grüneis, C. Attaccalite, A. Rubio, D. V. Vyalikh, S. L. Molodtsov, J. Fink, R. Follath, W. Eberhardt, B. Büchner, and T. Pichler, Phys. Rev. B [**79**]{} 205106 (2009). T. Valla, J. Camacho, Z. H. Pan, A. V. Fedorov, A. C. Walters, C. A. Howard, and M. Ellerby, Phys. Rev. Lett. [**102**]{}, 107007 (2009). S.-L. Yang, J. A. Sobota, C. A. Howard, C. J. Pickard, M. Hashimoto, D. H. Lu, S.-K. Mo, P. S. Kirchmann, and Z.-X. Shen, Nat. Commun. [**5**]{}, 3493 (2014). R. S. Gonnelli, F. Paolucci, E. Piatti, K. Sharda, A. Sola, M. Tortello, J. R. Nair, C. Gerbaldi, M. Bruna, and S. Borini, Sci. Rep. [**5**]{}, 9554 (2015). H. J. Choi, M. L. Cohen, and S. G. Louie, Physica C [**385**]{}, 66 (2003). J. Bardeen, L. N. Cooper, and J. R. Schrieffer, Theory of superconductivity, Phys. Rev. [**108**]{}, 1175 (1957). D. C. Johnston, Supercond. Sci. Technol. [**26**]{}, 115011 (2013). H. Padamsee, J. E. Neighbor, and C. A. Shiffman, J. Low Temp. Phys. [**12**]{}, 387 (1973). J. S. Kim, R. K. Kremer, L. Boeri, and F. S. Razavi, Phys. Rev. Lett. [**96**]{}, 217002 (2006). J. Pesić, R. Gajić, K. Hingerl, and M. Belić, Europhys. Lett. [**108**]{}, 67005 (2014).
{ "pile_set_name": "ArXiv" }
--- abstract: 'Recent studies of novel four-body lepton number violating decays of $\tau$ leptons and neutral $B$ mesons are summarized and updated. These decays are assumed to be enhanced by the exchange of resonant Majorana neutrinos. It is shown that the $\tau^- \to \pi^+l^-l^-\nu_{\tau}$ decay channels, with $l=e$ or $\mu$, can provide stronger constraints on the mixing vs. mass parameter space of resonant Majorana neutrinos than analogous three-body decays of charged $B$ mesons.' address: 'Departamento de Física, Cinvestav del IPN, Apartado Postal 14-740, 07000 México, D.F. México' author: - 'G. López Castro' - 'N. Quintero' bibliography: - 'martin.bib' nocite: '[@*]' title: Lepton number violation in tau lepton decays --- Lepton number violation, Majorana neutrinos, heavy flavor decays, tau lepton Introduction ============ Total lepton number $L=L_e+L_{\mu}+L_{\tau}$ is an absolutely conserved quantum number in the Standard Model (SM). Some extensions of the SM include interactions that can induce $L$ non-conservation [@moha]. Minimal extensions of the SM aiming to include massive neutrinos can contain Majorana mass terms, like ${\cal L}_M= \overline{{\nu}_R^c}M_M \nu_R + {\rm h.c.}$, which provides an appealing mechanism that violates lepton number by two units ($\Delta L=2$) [@paul]. A clear signal of Majorana mass terms are $L$-number violating processes that involve the production of two equal-sign charged leptons, the most well known and widely studied example being neutrinoless double beta decay in nuclei [@rode]. In this contribution we consider the exchange of Majorana neutrinos as a source of $\Delta L=2$ lepton number violation (LNV) in decays of heavy flavors, and more specifically in four-body decays of the $\tau$ lepton. These Majorana neutrinos are assumed to be sterile, such that their coupling to the weak charged current are very suppressed by tiny mixings with active neutrinos. Typical neutrino-exchange diagrams contributing to LNV in decays of the $\tau$ lepton are shown in Figure 1. Under this scheme, the sensitivity of different heavy flavor LNV decays ($M$ denotes a vector or pseudoscalar meson and $l,l'=e, \mu, \tau$ whenever they are allowed by kinematics) $$\begin{aligned} D^+_{(s)}, B^+, B^+_c &\to& l^+l'^+M^- \nonumber \\ D^0, B^0, B_s &\to& l^-l'^- M_1^-M_2^-\nonumber \\ \tau^- &\to& l^+M_1^-M_2^- \nonumber \\ \tau^- &\to& \nu_{\tau}l^-l'^-M^+ \nonumber \end{aligned}$$ is determined by comparing the mass scale of the exchanged Majorana neutrinos with typical energies of the decay process. Thus, we distinguish three cases [@atre]: - If neutrinos are very light compared to their four-momenta in the propagator (actually, $m_{\nu}^2 << q^2$), the decay rates become sensitive to the [*effective*]{} Majorana mass defined by $\langle m_{ll'}\rangle\equiv \sum_i U_{li}U_{l'i}m_i$, where $U_{li}$ denote the mixings of light (active) neutrinos described by the PMNS matrix; - If neutrinos are heavy compared to the mass of the decaying state, the rate is sensitive to $\sum_N V_{lN}V_{l'N}/m_N$, where $V_{lN}$ are the mixings of light (active) and heavy (sterile) neutrinos of type $N$ (see definition in Section 2). - Finally, if heavy neutrinos are of the order of the heavy flavor mass scale such that they can be produced on their mass-shell ($q^2=m_N^2$), the rates are largely enhanced due to the resonant effect associated to their decay widths $\Gamma_N$, with their decay amplitudes proportional to $\sum_{N}V_{lN}V_{l'N}/\Gamma_N$. This is the so-called [*resonant enhancement mechanism*]{} [@atre] for LNV decays and can occur only for time-like neutrino momenta as in the case of mesons and $\tau$ lepton decays . Note that in the first two cases, the rates of heavy flavor decays turn out to be very suppressed, making uninteresting their searches at flavor factories [@ali; @atre1]. ![Neutrino-exchange diagrams induced by crossings of the $W^-W^-\to l^-l'^-$ $\Delta L=2$ kernel leading to LNV in $(a)$ three- and $(b)$ four-body tau decays.](taulnv.pdf){width="11.5cm"} Lepton number violation in three-body decays of $\tau$ leptons and charged ($D,\ D_s,\ B, \ B_c$) mesons have been widely investigated previously, both from the theoretical [@atre; @ali; @atre1; @chile1; @bao] and experimental [@babar1; @lhcb1; @belle2; @belle1; @pdg; @danielmstau2012; @kiyoshitau2012] points of view. The current best experimental upper bounds available on these decay channels are shown in Figure 2; in addition, very stringent bounds of the order of $10^{-9}$ have been obtained (see for example [@pdg]) for $K^+ \to \pi^-l^+l'^+$ decays, with $l,\ l'=e,\ \mu$. The measured upper limits allow to exclude a region in the $|V_{lN}|^2$ vs. $m_N$ plane of the parameter space, by assuming that a single resonant neutrino (usually denoted by the subindex $N$ or $4$) dominates the decay amplitude. Such sterile Majorana neutrinos, with masses in the range of 1$\sim$10 GeV, can appear in the framework of some minimal extensions of the SM; for instance, it has been suggested that they can play an important role to explain simultaneously the oscillations of neutrinos, the baryon asymmetry of the Universe and the dark matter problem [@shapo]. In this paper we present a summary and update of our recent proposals [@nos2011; @nos2012] which consider the four-body decays $\tau^- \to \nu_{\tau}l^-l'^-M^+$ and $B^0 \to D^-l^+l'^+M^-$, where $M$ is a pseudoscalar or vector meson that can be allowed by kinematics (the analogous decay $\pi^+\to e^+e^+\mu^-\nu$ was considered in [@chile3]). We illustrate our studies with results on di-muonic channels (results on di-electrons modes can be found in [@nos2011; @nos2012]). Searches for these decay channels have not been undertaken by experiments up to now. Here we emphasize that they can provide competitive or even stronger bounds on the parameter space of Majorana neutrinos as compared to three-body decays of heavy flavors. ![Experimental upper limits on branching ratios of 3-body decays of charged heavy mesons and $\tau$ lepton \[9-15\]. ](UL_3body.pdf){width="8.2cm"} Resonant three-body decays ========================== The addition of right-handed singlet neutrinos to the SM leads in a natural way to the appearance of Majorana and Dirac mass terms [@paul], with Majorana mass terms allowing $\Delta L=2$ lepton number violation. The heavier (sterile) neutrinos get involved into charged weak interactions, since after diagonalization of the full neutrino matrix, neutrinos of defined flavor becomes a mixture of light and heavy mass eigentstates, namely, $$\nu_l=\sum_{i=1}^3 U_{li}\nu_i+\sum_{N= 4}^{n+4}V_{lN}\nu_N\$$ if $n$ right-handed singlets are considered. Here, $U_{li}$ are essentially the entries of the PMNS matrix, and $V_{lN}$ are the tiny mixings of the active and sterile neutrinos. The charged current interaction Lagrangian in the flavor basis becomes: $$\!\!\!{\cal L}_{cc}=\frac{g}{2\sqrt{2}}\bar{\nu}_l\gamma^{\mu}(1-\gamma_5)l\cdot W^-_{\mu} + {\rm h.c.}$$ where $\nu_l$ is given above. Under the assumption that only one Majorana neutrino $N$ is resonant in three-body $\tau^- \to l'^+ M_1^- M_2^-$ and $M_1^- \to l^-l'^-M_2^+$ decays, the generic form of the decay amplitudes is (properly antisymmetrization under exchange of identical leptons in the final state must be understood) $$\!\!\!\!\!\! {\cal M}_{3}^{\tau,M_1^-}\sim G_F^2 V_{lN}V_{l'N}m_N{\cal F}(q^2)V_{M_1}^{CKM}V_{M_2}^{CKM}f_{M_1}f_{M_2}\ ,$$ where $V^{CKM}_{M_i}$ is the Cabibbo-Kobayashi-Maskawa (CKM) matrix element for the charged meson $M_i$ and $f_{M_i}$ its decay constant. The resonance factor is determined by the neutrino propagator ${\cal F}(q^2)\sim (q^2-m_N^2+im_N\Gamma_N)^{-1}$ where $q$ is the momenta of the exchanged neutrino. The neutrino decay width $\Gamma_N\leq 10^{-3}$ eV for the mass scales of interest in $\tau$ lepton and in other me son decays ($m_N \leq 5$ GeV) [@atre]. From Eq. (3) we note that the bigger is the CKM matrix element, the stronger is the constraint that can be set on the neutrino mixings from the measured upper limits on LNV branching fractions. Since the rates of $B^{\pm}$, $D^{\pm}$ meson decay vertices are Cabibbo-suppressed by $|V_{ub}|^2$ and $|V_{cd}|^2$ factors, respectively, the constraints that are derived from them will not be very strong . On another hand, $\tau$ lepton decays (taking $l=\tau$ in Eq. (3) above) allow to constrain only the product $|V_{\tau N}V_{l'N}|$ ($l'=e,\mu$) from the measured upper limits of $\tau^- \to l'^+ M_1^-M_2^-$ branching fractions. ![Constrains on mixing vs. mass of Majorana neutrino parameter space from di-muonic three-body decays of charged mesons. Constraints from four-body $B^0 \to D^-\pi^-\mu^+\mu^+$ and $\tau^- \to \nu_{\tau}\mu^-\mu^-\pi^+$ by assuming upper bounds of $10^{-7}$ and $10^{-8}$ on their branching ratios, respectively, are also shown for comparison.](VmuN2.pdf){width="8.5cm"} In Figure 3 we show the constraints in the $|V_{\mu N}|^2$ vs $m_N$ region that can be gotten from three-body LNV decays of charged $D_{(s)}$ and $B$ mesons by using the experimental upper limits on di-muonic channels. These plots were obtained by using the rates calculated in Ref. [@atre] and incorporate recently updated measurements obtained by $B$-factory experiments [@babar1] and LHCb [@lhcb1]. Note that despite the large improvement recently obtained by LHCb [@lhcb1; @danielmstau2012], $B(B^+ \to \pi^-\mu^+\mu^+) \leq 1.3 \times 10^{-8}$, the constraints on the $|V_{\mu N}|^2$ mixing are not largely improved owing to the Cabibbo suppresion factor described in the previous paragraph. Four-body $B^0$ decays ====================== In order to avoid the suppression due to CKM factors in the $B^+$ decay vertex, we have proposed to consider the four-body decays of neutral mesons, namely $B^0 \to D^-M^-l^+l'^+$ with $l,\ l'=e,\ \mu$ or $\tau$ and $M$ a pseudoscalar or vector meson [@nos2011]. The decay amplitude for this four-body decay is given by: $${\cal M}_4^{B^0} \sim G_F^2 V_{lN}V_{l'N}m_N{\cal F}(q^2)V_{cb}^{CKM}f_{M}F_+^{B\to D}(t)\ ,$$ where $F_{+}^{B\to D}(t)$ is the vector form factor for the $B \to D$ transition and $t=(p_B-p_D)^2$ is the square of the momentum transfer (the contribution of the scalar form factor is negligible in this case). In this updated contribution we use the $B \to D$ vector form factor obtained from Lattice QCD [@kronfeld], in order to avoid the model dependence of the vector form factor used in [@nos2011]. In the neutrino narrow width approximation, the generic expression for the branching ratios of three- and four-body decays can be written as: $$B_{ll'}\sim \frac{|V_{lN}V_{l'N}|^2}{\Gamma_N}G(m_N)\ ,$$ where $\Gamma_N=\sum_l f_l(m_N)|V_{lN}|^2$ is the neutrino decay width; the sum extends over the lepton flavors that are allowed by kinematics for a given neutrino mass $m_N$, and $f_l(m_N)$ depends on decay constants and masses of final states in neutrino decay channels [@atre]. The function $G(m_N)$ contains the product of fundamental constants, hadronic parameters as well as the integrated four-body phase space of the specific channels [@nos2011]. No upper limits have been reported so far for four-body LNV decays of neutral $B$ mesons. Upper limits for four-body LNV decays have been reported only in the case of $D^0\to M_1^-M_2^-l^+l'^+$ decays with $l,\ l'=e,\ \mu$ and $M_{1,2}=\pi,\ K$ [@e791]; in all cases, the upper limits obtained for the branching ratios are at the level of $10^{-5}\!\sim \!10^{-4}$ which loosely constrain the mixing angles. Very recently, the LHCb collaboration has reported the first upper limit on the charged $B$ decay channel, $B(B^-\to D^0\pi^+\mu^-\mu^-) \leq 1.5 \times 10^{-6}$ at the 95% c. l. [@lhcb1]. The constraints in the $|V_{\mu N}|^2$ vs. $m_N$ plane that are obtained by assuming $B(B^0 \to D^-\mu^+\mu^+\pi^-) \leq 10^{-7}$ are shown in Figure 3. We observe that the constraint of the $\mu N$ mixing angle that can be obtained from this decay channel is competitive or even better when compared to other three-body LNV decays of $B^+$ mesons, for a mass region $m_N \leq 1.5$ GeV. Four-body $\tau$ lepton decays ============================== In order to look for better constraints on mixing of resonant neutrinos, we consider the $\tau^- \to \nu_{\tau}l^-l'^-M^+$ decays, where $l,l'=e$ or $\mu$ and $M=\pi, K, \rho$ and $K^*$ mesons. The use of these novel decay channels to constraint the parameter space of Majorana neutrinos were proposed in Ref. [@nos2012]. A previous estimate of the branching ratio for the dimuonic channel $B(\tau^-\! \to \! \nu_{\tau}\mu^-\mu^-\pi^+) \leq 8.2 \times 10^{-5}$ was reported in [@chile2] using $|V_{\mu N}|^2 \leq 10^{-3}$ and $m_N= 400\sim 600$ MeV as typical values within a model where the sterile neutrino lifetime is dominated by the radiative $N\!\to\! \nu\gamma$ decay [@gninenko]. The four-body $\tau^- \to \nu_{\tau}l^-l'^-M^+$ decays can provide constrains directly on the $|V_{lN}|^2$ mixing angles contrary to their three-body decays which are sensitive only to the product $|V_{\tau N}V_{lN}|$. The decay amplitude in this case is given by $${\cal M}_4^{\tau} \sim G_F^2V_{lN}V_{l'N}m_N{\cal F}(q^2)V^{CKM}_{uq}f_M\ ,$$ where $q=d$ or $s$. ![Constraints on Majorana neutrino parameter space from four-body di-muonic channels of $\tau$ lepton decays.](VmuNtau.pdf){width="8.6cm"} In Figure 4 we plot the constraints in the $|V_{\mu N}|^2$ vs. $m_N$ plane that are obtained by assuming a common upper limit $B(\tau^- \to \nu_{\tau}\mu^-\mu^-M^+) \leq 10^{-7}$ for all channels. In Figure 3, the constraints obtained by assuming $B(\tau^- \to \nu_{\tau}\mu^-\mu^-\pi^+)\leq 10^{-8}$ are compared to those obtained from heavy meson decays. As we can observe, these constraints are stronger than the ones that can be currently obtained from three-body decays of $B^+$ and $D^+$ mesons. In order to compare with the results of Ref. [@chile2], we compute the following branching fraction (we use similar values of mixings and masses for the heavy sterile neutrino) $$B(\tau^- \to \nu_{\tau}\mu^-\mu^-\pi^+) \leq 1.4 \times 10^{-5}\ .$$ Our results turns out to be of similar size. The main difference comes from the models we have used to compute the heavy neutrino lifetime. In summary, LNV decays of heavy flavors can provide important constraints on tiny mixing angles of Majorana neutrinos with masses in the range $m_{\pi} \leq m_N \leq m_B$. This is possible if a single heavy neutrino resonantly enhances the decay amplitudes. In this contribution we have shown that the four-body decays $B^0 \to D^-\pi^-\mu^+\mu^+$ and $\tau^- \to \nu_{\tau}\mu^-\mu^-\pi^+$ can provide stronger constrains on the $|V_{\mu N}|^2$ mixing angles than the ones obtained from three-body decays of charged heavy mesons. [The authors are grateful to the organizing committee of Tau2012 for the opportunity to present this work. They acknowledge the financial support from Conacyt (Mexico) and the collaboration of D. Delepine at the early stage of this work.]{} [00]{} R. N. Mohapatra [*et al*]{}, Rep. Prog. Phys. [**70**]{}, 1757 (2007). P. G. Langacker, [*The Standard Model and Beyond*]{}, CRC Press Taylor and Francis Group, (2010). For for example: S. M. Bilenky, Phys. Part. Nucl. [**41**]{}, 690 (2010); W. Rodejohann, Int. J. Mod. Phys. E[**20**]{}, 1833 (2011). A. Atre, T. Han, S. Pascoli, and Zhang, JHEP 0905, 030 (2009). A. Ali, A. V. Borisov and N. B. Zamorin, Eur. Phys. J. C[**21**]{}, 123 (2001). A. Atre, V. Barger and T. Han, Phys. Rev. D[**79**]{}, 113014 (2005) J. C. Helo, S. Kovalenko and I. Schmidt, Nucl. Phys. B853, 80 (2011). M. A. Ivanov and S. G. Kovalenko, Phyd. Rev. D[**71**]{}, 053004 (2005); S. S. Bao et al, arXiv:1208.5136 \[hep-ph\]; J. M. Chang and G. L. Wang, Eur. Phys. C[**71**]{}, 1715 (2011). J. P. Lees et al \[BABAR Collab.\], Phys. Rev. D[**84**]{}, 072006 (2011); Phys. Rev. D[**85**]{}, 071103(R) (2012). R. Aaij et al \[LHCb Collab.\], Phys. Rev. Lett. [**108**]{}, 101601 (2012); Phys. Rev. D[**85**]{}, 112004 (2012). O. Seon [*et al*]{} \[Belle Collab.\], Phys. Rev. D[**84**]{}, 071106(R) (2011). Y. Miyazaki et al \[Belle Collab.\], Phys. Lett. B682, (2010). J. Beringer [*et al*]{}, Phys. Rev. D[**86**]{}, 010001 (2012). D. Martínez Santos, [*these proceedings*]{} K. Hasayaka, [*these proceedings*]{} L. Canneti [*et al*]{}, arXiv:1208.4607 \[hep-ph\]. D. Delepine, G. López Castro and N. Quintero, Phys. Rev. D[**84**]{}, 096011 (2011); [*ibid*]{} D[**86**]{}, 079905(E) (2012). G. López Castro and N. Quintero, Phys. Rev. D[**85**]{}, 076006 (2012); [*ibid*]{} D[**86**]{}, 079904(E) (2012). G. Cvetic, C. Dib, and C. S. Kim, JHEP1206, 149 (2012). E. M. Aitala [*et al*]{} \[E791 Collab.\], Phys. Rev. Lett. 86, (2001). J. A. Bailey [*et al*]{} \[Fermilab Lattice and MILC\], Phys. Rev. D[**85**]{}, 114502 (2012); [*ibid*]{} D[**86**]{}, 039904(E) (2012). C. Dib et al, Phys. Rev. D[*85*]{}, 011301(R) (2012). S. N. Gninenko, Phys. Rev. Lett. [**103**]{}, 241802 (2009).
{ "pile_set_name": "ArXiv" }
--- abstract: 'We show the explicit expression of the geometric phase for $n$-partite Gaussian states. In our analysis, the covariance matrix can be obtained as a boundary term of the geometric phase.' author: - 'Angel Garcia-Chung' title: On the geometric phase for Gaussian states --- Introduction ============ The covariance matrix of Gaussian states plays a prominent role detecting entanglement and squeezing in quantum states. The entanglement, which is one of the main features of quantum mechanics, not only due its philosophical implications but also for its applications in many different areas [@bennett2000quantum; @divincenzo1995quantum; @bouwmeester1997experimental; @pan2001entanglement; @axline2018demand; @adesso2007entanglement; @horodecki2009quantum; @richens2017entanglement], can be detected using several criteria [@peres1996separability; @horodecki1997separability; @simon2000peres; @duan2000inseparability; @werner2001bound; @giedke2001separability]. In the case of continuous variable (CV) systems, these criteria rely on some features of the covariance matrix [@simon2000peres; @duan2000inseparability; @werner2001bound; @giedke2001separability]. On the other hand, squeezed states can be used to improve the sensitivity of measurement devices beyond the usual quantum noise limits [@simon1988gaussian; @braunstein2005quantum; @walls2007quantum; @adesso2014continuous; @ma1990multimode; @schnabel2017squeezed; @pirandola2009correlation; @weedbrook2012gaussian]. These states are a particular type of Gaussian states, i.e., those states whose Wigner functions are Gaussian functions on the phase space [@simon1988gaussian; @braunstein2005quantum; @walls2007quantum; @adesso2014continuous]. Due to the covariance matrix contains all the information related with the uncertainty relations, it is used to determine whether a state is squeezed or not [@simon1988gaussian]. The squeezed states are generated by the squeeze operators which can be considered as elements of the unitary representation of the symplectic group $Sp(2n, \mathbb{R})$, specifically, those close to the group identity [@moshinsky1971linear; @Arvind:1995ab; @torre2005linear; @wolf2016development; @de2011symplectic]. When the squeeze operator acts on the vacuum state the resulting squeezed state is named vacuum squeezed state. A distinctive feature of the vacuum squeezed state is the expression for its covariance matrix $${\bf V}^{(2)} = \frac{1}{2} {\bf M} \; {\bf M}^T, \label{CVIntro}$$ where the matrix ${\bf M}$ is an element of the symplectic group $Sp(2n, \mathbb{R})$ [@pirandola2009correlation; @adesso2014continuous]. We can notice in (\[CVIntro\]) the connection between the covariance matrix and the symplectic group $Sp(2n, \mathbb{R})$. This connection paves the way to the following question: is there a direct relation between the covariance matrix and the geometric phase for vacuum squeezed states? Let us provide some context for this question. Geometric phases are obtained when a physical system evolve through a set of states [@hariharan2005geometric; @chruscinski2012geometric]. These phases are ubiquitous in many different research areas [@hariharan2005geometric; @sachdev2007quantum; @cabra2012modern; @chruscinski2012geometric], particularly, in quantum optics where they have been studied extensively at both levels, theoretically and experimentally [@hariharan2005geometric; @chaturvedi1987berry; @simon1993bargmann; @kwiat1991observation; @mukunda1993quantum; @brendel1995geometric; @strekalov1997two; @seshadri1997geometric; @mendas1997pancharatnam; @fuentes2000proposal; @de2002squeezing]. Among the manifestations of geometric phases in optics, we are interested in the phase resulting from a cycle of changes in squeezed states of light [@hariharan2005geometric; @chaturvedi1987berry; @Chiao; @seshadri1997geometric]. For simplicity, in this work we will refer to this phase as: [*squeeze phase*]{}. For a system with a single photon ($n=1$), the squeeze phase results from a cyclic variation of the parameter $\theta \in [0, 2 \pi]$. Recall that $\theta$ together with $r$, define the complex parameter $\zeta = r e^{i \theta}$ which in turn, define the squeeze operator $\widehat{S}(\zeta)$ [@walls2007quantum; @adesso2014continuous]. In the single photon scenario and for the squeezed state in the $j$-th energy level, the geometric phase $\gamma^{(1)}_j$, was reported in [@chaturvedi1987berry]. In this notation, the index ${}^{(1)}$ for the phase, labels the number of photons. In the case of general coherent squeezed states, the geometric phase $\gamma^{(n=1)}_j$, and the phase difference, $\Delta \gamma^{(n=1)}_j = \gamma^{(n=1)}_j - \gamma^{(n=1)}_{j-1}$, was treated in [@seshadri1997geometric]. As can be seen and to the best of the author’s knowledge, only the single photon cases have been considered. Hence, the analysis of the squeeze phase for a multi-photon system, i.e., $\gamma^{(n>1)}_0$, is still open. This becomes more relevant if we consider that the experimental results [@brendel1995geometric; @hariharan2005geometric] imply that the squeeze phase for pairs of photons with parallel polarizations acquire twice the geometric phase of single photons. On the other hand, due to the squeeze operator can be seen as an element of the unitary representation of $Sp(2n,\mathbb{R})$, a cyclic variation of $\theta$ yields a cyclic variation of ${\bf M}$. As a result, due to the connection in (\[CVIntro\]) between ${\bf V}^{(2)}$ and ${\bf M}$, we might expect also a cyclic variation of ${\bf V}^{(2)}$. Moreover, we can go beyond and generalize our analysis and consider an arbitrary Gaussian state. That being said, we consider that the motivation for our previous question seems now more natural. The purpose of this work is to determine the geometric phase of a $n$-partite Gaussian state and for an arbitrary symplectic matrix ${\bf M}$. Our results will be valid for both, systems described with quadrature operators and for the standard dimension-full operators $\widehat{x}$ and $\widehat{p}$ of the harmonic oscillator. We will show that the covariance matrix ${\bf V}^{(2)}$ for a general Gaussian state can be seen as a boundary term. To do so, we briefly describe in Section (\[preliminaries\]) the mathematical tools regarding the symplectic group analysis and its group action in the phase space $\mathbb{R}^{2n}$. Then, in Section (\[CovarianceMatrix\]) we summarize the main aspects related with its unitary representation. We also derive in this section the explicit form of the covariance matrix in terms of the symplectic group elements. In Section (\[BerryPhase\]) we calculate the explicit relation between the covariance matrix and the geometric phase and in Section (\[Discussion\]) we provide our conclusions. Preliminaries ============= The symplectic group $Sp(2n, \mathbb{R})$ is given by $2n \times 2n$ matrices ${\bf M}$ satisfying the equation $${\bf \Omega} = {\bf M} \; {\bf \Omega} \; {\bf M}^T, \label{SpCond}$$ where ${\bf M}^T$ is the transpose matrix of ${\bf M}$ and ${\bf \Omega}$ is defined as $${\bf \Omega} = \left( \begin{array}{cc} {\bf 0} & {\bf 1} \\ - {\bf 1} & {\bf 0} \end{array}\right).$$ Here ${\bf 0}$ and ${\bf 1}$ are the $n\times n$ zero and identity matrix respectively. Matrix ${\bf M}$ can be written in block form as follows $${\bf M} = \left( \begin{array}{cc} {\bf A} & {\bf B} \\ {\bf C} & {\bf D} \end{array}\right), \label{NoTildeMatrix}$$ where ${\bf A}$, ${\bf B}$, ${\bf C}$ and ${\bf D}$ are $n \times n$ real matrices. The condition (\[SpCond\]) now reads as $${\bf A} {\bf D}^T - {\bf B} {\bf C}^T = {\bf 1}, \quad {\bf A} {\bf B}^T = {\bf B} {\bf A}^T, \quad {\bf C}{ \bf D}^T = {\bf D} {\bf C}^T. \label{Sp1Coord}$$ Additionally, the group elements close to the identity can also be related with the elements of the Lie algebra $sp(2n, \mathbb{R})$ via the exponential map $${\bf M}({\bf L})= e^{ {\bf \Omega} {\bf L}}. \label{ExpMap}$$ where ${\bf L} \in sp(2n, \mathbb{R})$ is a real symmetric $2n \times 2n$ matrix. Here, ${\bf M}({\bf L})$ refers to the explicit the relation between the group element ${\bf M}$ and its associated Lie algebra matrix ${\bf L}$ via the exponential map. There are, of course, elements in $Sp(2n, \mathbb{R})$ that are not close to the identity and therefore, cannot be written as the exponential map of any symmetric matrix ${\bf L}$. In the appendix \[LAAnalysis\] we derive the relation ${\bf M}(\bf L)$ between the group $Sp(4,\mathbb{R})$ and its Lie algebra $sp(4, \mathbb{R})$. This result will be very useful in the analysis of the symplectic group representation for bi-partite systems. The symplectic group is used to provide the action of the linear canonical transformations on the phase space $(\mathbb{R}^{2n}, \{, \})$ where $\{, \}$ is the standard Poisson bracket. The group action on $\mathbb{R}^{2n}$ is then given as $$\vec{\bf X}'^T = {\bf M} \, \vec{\bf X}^T, \label{1Coord}$$ where $\vec{\bf X}^T = (\vec{q} \; \vec{p})^T$ and $\vec{q}=(q_1, q_2, \dots, q_n)$ and $\vec{p}=(p_1, p_2, \dots, p_n)$ are the coordinates on the space $\mathbb{R}^{2n}$. Using these coordinates, the Poisson bracket can be written as $$\{ \vec{\bf X}^T , \vec{\bf X} \} = {\bf \Omega}. \label{PoissonB}$$ If we now impose that (\[PoissonB\]) also holds for ${\vec{\bf X}}'$ then the relation in (\[1Coord\]) yields the condition (\[SpCond\]). This way of defining the symplectic group action on the phase space is very useful to obtain a unitary representation on a Hilbert space. However, to study entanglement conditions a more convenient form of the group action on the phase space $\mathbb{R}^{2n}$ is required. Consider the different array $\vec{Y}^T = ( q_1, \, p_1 , \, q_2 , \, p_2 , \, \hdots , \, q_n , \, p_n )$. Then, similarly to (\[1Coord\]) we consider the group action to be of the form $$\vec{\bf Y}'^T = \widetilde{\bf M}\; \vec{\bf Y}^T, \label{2Coord}$$ where $\widetilde{\bf M}$ is the new form of the symplectic matrix. We now insert the expression (\[2Coord\]), [*mutatis mutandis*]{}, in the Poisson bracket (\[PoissonB\]) to obtain the following symplectic group condition $$\left( \begin{array}{cccc} {\bf J} & {\bf 0} & \dots & {\bf 0} \\ {\bf 0} & {\bf J} & \dots & {\bf 0} \\ \vdots & \vdots & \ddots & \vdots \\ {\bf 0} & {\bf 0} & \dots & {\bf J} \end{array} \right) = \widetilde{\bf M} \left( \begin{array}{cccc} {\bf J} & {\bf 0} & \dots & {\bf 0} \\ {\bf 0} & {\bf J} & \dots & {\bf 0} \\ \vdots & \vdots & \ddots & \vdots \\ {\bf 0} & {\bf 0} & \dots & {\bf J} \end{array} \right) \widetilde{\bf M} ^T, \label{2SpCond}$$ where ${\bf J}$ is the $2\times 2$ matrix given by ${\bf J} = \left( \begin{array}{cc} 0 & 1 \\ -1 & 0 \end{array}\right)$. Analogously to the matrix ${\bf M}$, the matrix $\widetilde{\bf M}$ can also be written in block form as $$\widetilde{\bf M} = \left( \begin{array}{cccc} {\bf A}_{11} & {\bf A}_{12} & \dots & {\bf A}_{1n} \\ {\bf A}_{21} & {\bf A}_{22} & \dots & {\bf A}_{2n} \\ \vdots & \vdots & \ddots & \vdots \\ {\bf A}_{n1} & {\bf A}_{n2} & \dots & {\bf A}_{nn} \end{array}\right) , \label{TildeMatrix}$$ where ${\bf A}_{i j}$, with $i,j = 1,2, \dots, n$, are $n^2$ real $2\times 2$ matrices. The condition (\[2SpCond\]) now reads on these block matrices as $${\bf J} = \sum^{j=n}_{j=1} {\bf A}_{i\, j} {\bf J} {\bf A}^T_{i\,j}, \quad {\bf 0} = \sum^{k=n}_{k=1} {\bf A}_{i\, k} {\bf J} {\bf A}^T_{j\, k}, \label{Sp2Coord}$$ for all $i \neq j$ in the second condition. Expressions (\[Sp1Coord\]) and (\[Sp2Coord\]) define different conditions for the symplectic matrices although, they describe the same Lie group [@adesso2014continuous]. Notice in (\[TildeMatrix\]) that when the off-diagonal block matrices ${\bf A}_{i \neq j} = {\bf 0}$ the symplectic matrix becomes block-diagonal, i.e., $$\widetilde{\bf M}= \mbox{diag}({\bf A}_{11}, {\bf A}_{22} , \dots, {\bf A}_{nn}), \label{Entang}$$ then, notably, as a result of this form of the group action (\[Sp2Coord\]), each matrix ${\bf A}_{ii} $ is an element of the symplectic group $ Sp(2,\mathbb{R})$. Both group actions (\[NoTildeMatrix\]) and (\[TildeMatrix\]) are related via the transformation ${\bf \Gamma}$ [@adesso2014continuous] which is defined as $${\bf M} = {\bf \Gamma} \; \widetilde{\bf M} \; {\bf \Gamma}^{-1}, \label{RelationbetweenMs}$$ where ${\bf \Gamma}$ is given by $$\vec{\bf X}^T = {\bf \Gamma} \, \vec{\bf Y}^T , \label{CTransformation}$$ and is such that ${\bf \Gamma}^T = {\bf \Gamma}^{-1}$. For example, in the case of $n=2$ the matrix transformation, denoted by $\Gamma^{(2)}$, is $$\Gamma^{(2)} = \left( \begin{array}{cccc} 1 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 0 & 1 \end{array}\right). \label{Gamma2}$$ Let us now move to the next section, where we briefly show the analysis to calculate the covariance matrix using the unitary representation of the symplectic group. These results will be use in the calculation of the Berry phase in (\[BerryPhase\]). Covariance matrix {#CovarianceMatrix} ================= As we mentioned in the introduction, the calculation of the Berry phase requires the expression of some of the covariance matrix components. For completeness, let us proceed in this section to derive the general expression for the covariance matrix of Gaussian states. The symplectic group is a non-compact group which requires an infinite Hilbert space for its unitary representation. Let us consider the Hilbert space ${\cal H} = L^2(\mathbb{R}^n, d\vec{x})$. The unitary representation of $Sp(2n,\mathbb{R})$ is the map ${\bf M} \mapsto \widehat{C}_{\bf M}$, where $\widehat{C}_{\bf M}$ is a unitary operator $\widehat{C}_{\bf M} \in {\cal L}(\cal H)$. Here, ${\cal L}(\cal H)$ is the space of linear operators over the Hilbert space ${\cal H}$. The operator $\widehat{C}_{\bf M}$ is related with the fundamental operators $\widehat{x}_j$ and $\widehat{p}_j$ via the following condition $$\widehat{C}_{\bf M} \left( \begin{array}{c} \vec{\widehat{x}} \\ \vec{\widehat{p}}\end{array} \right) \widehat{C}^{-1}_{\bf M} = {\bf M}^{-1} \left( \begin{array}{c} \vec{\widehat{x}} \\ \vec{\widehat{p}}\end{array} \right) . \label{QASG}$$ This condition enables us to consider $\widehat{C}_{\bf M}$ as the (unitary) representation of $Sp(2n, \mathbb{R})$ on ${\cal H}$ [@moshinsky1971linear; @torre2005linear; @wolf2016development; @de2011symplectic]. The action of this operator on the Hilbert space is given by $$\widehat{C}_{\bf M} \Psi(\vec{x}) = \int d\vec{x}' C_{\bf M}(\vec{x} , \vec{x}') \Psi(\vec{x}') =: \Psi_{\bf M}(\vec{x}), \label{UnitaryRep}$$ for $\Psi(\vec{x}) \in {\cal H}$ and where the kernel of this integral operator is $$C_{\bf M}(\vec{x}, \vec{x}') = \frac{e^{ \frac{i}{2 \hbar} \left[ \vec{x}^T {\bf D} {\bf B}^{-1} \vec{x} - 2 \vec{x}'^T {\bf B}^{-1} \vec{x} + \vec{x}'^T {\bf B}^{-1} {\bf A} \vec{x}'\right]}}{\sqrt{ (2 \pi i \hbar)^n \det {\bf B}}} . \label{Kernel}$$ It can be checked that this representation (\[UnitaryRep\])-(\[Kernel\]) is unitary and also, the factor $\det {\bf B}$ in (\[Kernel\]) gives rise to a well define action even in the case where the matrix ${\bf B}$ is singular [@moshinsky1971linear; @wolf2016development]. Moreover, the representation is valid for the entire symplectic group and not just for those elements close to the group identity. Due to we are interested in group elements where ${\bf B}$ is non-singular we omit this analysis and provide in appendix \[SingularCase\] the analysis of the Berry phase for the very special case where ${\bf B}={\bf 0}$. Let us proceed with the calculation of the covariance matrix. Consider the state $$| \Psi_{\bf M} \rangle = \widehat{C}_{\bf M} | 0 \rangle = \int d\vec{x} \; \Psi_{\bf M}(\vec{x}) \, | \vec{x} \rangle, \label{InitialS}$$ where $| 0 \rangle = \int d\vec{x} \; \Psi_0(\vec{x}) \, | \vec{x} \rangle $ is the state $|0\rangle = |0\rangle_1 \otimes | 0\rangle_2 \dots | 0 \rangle_n$. The ket $| 0 \rangle_j$ is the vacuum state of the j-th quantum harmonic oscillator. These states are Gaussian states due to the kernel (\[Kernel\]) and moreover, they are eigenstates of the Hamiltonian $$\widehat{H}_{\bf M} = \widehat{C}_{\bf M} \widehat{H} \widehat{C}^{-1}_{\bf M},$$ where $\widehat{H} = \sum^n_j \widehat{H}_j$ is the Hamiltonian of $n-$decoupled quantum harmonic oscillators each of them characterized by the Hamiltonian $\widehat{H}_j = \frac{1}{2 m_j} \widehat{p}^2_j + \frac{m_j \omega^2_j}{2} \widehat{q}^2_j$. To calculate the covariance matrix we first calculate the following amplitude $$\langle \Psi_{\bf M} | \widehat{W}(\vec{a}, \vec{b}) | \Psi_{\bf M} \rangle = \langle \Psi_0 | \widehat{C}^\dagger_{\bf M} \widehat{W}(\vec{a}, \vec{b}) \widehat{C}_{\bf M} | \Psi_0 \rangle, \label{Amplitude}$$ where $\widehat{W}(\vec{a}, \vec{b})$ is a Weyl-algebra generator whose representation on ${\cal H}$ is $$\widehat{W}(\vec{a}, \vec{b}) \Psi(\vec{x}) = e^{\frac{i}{2 \hbar} \vec{a}^T \vec{b}} e^{\frac{i}{\hbar} \vec{a}^T \vec{x}} \Psi(\vec{x} + \vec{b}).$$ A tedious but standard calculation gives the amplitude in (\[Amplitude\]) the following form $$\langle \Psi_{\bf M} | \widehat{W}(\vec{a}, \vec{b}) | \Psi_{\bf M} \rangle = \exp\left\{ - \frac{1}{4} \left( \begin{array}{cc} \vec{a} & \vec{b} \end{array} \right)^T {\bf \Lambda} \left( \begin{array}{c} \vec{a} \\ \vec{b} \end{array}\right) \right\}, \label{Amplitude2}$$ where the matrix ${\bf \Lambda}$ is given by $${\bf \Lambda} := {\bf M} \left( \begin{array}{cc} \frac{1}{\hbar^2} {\bf L}^2 & {\bf 0} \\ {\bf 0} & {\bf L}^{-2} \end{array}\right) {\bf M}^T,$$ and $ {\bf L} = \mbox{diag}( l_1 , l_2 , \dots , l_n )$. Here, $l_j = \sqrt{\frac{\hbar}{m_j \, \omega_j}}$, with $j=1,2, \dots, n$, is the characteristic length of the quantum harmonic oscillator labelled by $j$. The covariance matrix ${\bf V}^{(2)}$, whose components can then be written as $${\bf V}^{(2)} = \left( \begin{array}{cc} \langle \Psi_{\bf M} | \widehat{x}_j \; \widehat{x}_k | \Psi_{\bf M} \rangle & \frac{1}{2} \langle \Psi_{\bf M} | \{ \widehat{x}_j , \widehat{p}_k \} | \Psi_{\bf M} \rangle \\ \frac{1}{2} \langle \Psi_{\bf M} | \{ \widehat{p}_j , \widehat{x}_k \} | \Psi_{\bf M} \rangle & \langle \Psi_{\bf M} | \widehat{p}_j \; \widehat{p}_k | \Psi_{\bf M} \rangle \end{array}\right), \label{CVMatrixDef}$$ can be obtained from (\[Amplitude2\]) using the following relations $$\begin{aligned} \langle \Psi_{\bf M} | \widehat{x}_j \; \widehat{x}_k | \Psi_{\bf M} \rangle = \frac{\hbar^2}{i^2} \partial^2_{ a_j a_k} \langle \Psi_{\bf M} | \widehat{W}(\vec{a}, \vec{b}) | \Psi_{\bf M} \rangle \vert_{\vec{a}, \vec{b}=0}, \label{EqCV1}\\ \frac{1}{2} \langle \Psi_{\bf M} | \{ \widehat{x}_j , \widehat{p}_k \} | \Psi_{\bf M} \rangle = \frac{\hbar^2}{i^2} \partial^2_{a_j b_k} \langle \Psi_{\bf M} | \widehat{W}(\vec{a}, \vec{b}) | \Psi_{\bf M} \rangle \vert_{\vec{a}, \vec{b}=0},\label{EqCV2} \\ \frac{1}{2}\langle \Psi_{\bf M} | \{ \widehat{p}_j , \widehat{x}_k \} | \Psi_{\bf M} \rangle = \frac{\hbar^2}{i^2} \partial^2_{b_j a_k} \langle \Psi_{\bf M} | \widehat{W}(\vec{a}, \vec{b}) | \Psi_{\bf M} \rangle \vert_{\vec{a}, \vec{b}=0}, \label{EqCV3} \\ \langle \Psi_{\bf M} | \widehat{p}_j \; \widehat{p}_k | \Psi_{\bf M} \rangle = \frac{\hbar^2}{i^2} \partial^2_{b_j b_k} \langle \Psi_{\bf M} | \widehat{W}(\vec{a}, \vec{b}) | \Psi_{\bf M} \rangle \vert_{\vec{a}, \vec{b}=0}, \label{EqCV4}\end{aligned}$$ an a straightforward calculation yields the following expression for ${\bf V}^{(2)} $ $${\bf V}^{(2)} = \frac{\hbar^2}{2}{\bf M} \left( \begin{array}{cc} \frac{{\bf L}^2}{\hbar^2} & {\bf 0} \\ {\bf 0} & {\bf L}^{-2} \end{array}\right) {\bf M}^T . \label{CVMatrix}$$ It is worth to mention that the matrix for the first order moments is zero due to the symmetry of the Gaussian function in (\[Amplitude2\]). In quantum optics we usually work with quadrature operators, which are dimensionless operators, and hence, can be derived by taking dimensionless parameters $l_j = 1$ and $\hbar=1$. As a result, the covariance matrix takes the simplified form $${\bf V}^{(2)}_q = \frac{1}{2} {\bf M} {\bf M}^T, \label{QCVMatrix}$$ where the sub-index $q$ in ${\bf V}^{(2)}_q$ means we are considering quadrature operators instead of ${\bf V}^{(2)}$ which stands for dimension-full parameters. Notice that ${\bf V}^{(2)}$ and ${\bf V}^{(2)}_q$ cannot be considered as symplectic matrices due to the $1/2$ factor. Additionally, it is worth to recall that this result is valid for systems with n-degrees of freedom and Gaussian states of the form given in (\[InitialS\]). In the case of $n=2$, the appendix \[LAAnalysis\] provides the general expression for the covariance matrices ${\bf V}^{(2)}$ and ${\bf V}^{(2)}_q$ using symplectic matrices ${\bf M}({\bf L})$. We are now ready to derive, in the next section, the Berry phase for general Gaussian states of the form (\[InitialS\]). Geometric phase for Gaussian states {#BerryPhase} =================================== To derive the geometric phase let us first consider the expression for the kernel $C_{\bf M}(\vec{x}, \vec{x'})$ and notice that it does not depend on the block matrix ${\bf C}$ defined in (\[NoTildeMatrix\]). Therefore, every variation will involve only those parameters in the kernel defined in (\[Kernel\]). For the very special case when ${\bf B} = {\bf 0}$, the representation is rather different to (\[Kernel\]). Due to we are interested in non-singular group elements and in order to simplify our presentation, we showed in the appendix (\[SingularCase\]) the analysis of the geometric phase when ${\bf B} = {\bf 0}$. Having said this, let us consider a variation in the parameters $A_{a b}$, $B_{a b}$ and $D_{a b}$, which are the components of the block matrices ${\bf A}$, ${\bf B}$ and ${\bf D}$ respectively and recall that, the parameters $m_j$ and $\omega_j$, defined in the Hamiltonians $\widehat{H}_j$, remain fixed. With all these considerations the geometric phase can be written as $$\gamma_{\bf M} = i \int_{\cal C} \left[ \langle 0 | \widehat{C}^\dagger_{\bf M} \frac{\partial \widehat{C}_{\bf M}}{\partial A_{a b}} | 0 \rangle \, dA_{a b} + \langle 0 | \widehat{C}^\dagger_{\bf M} \frac{\partial \widehat{C}_{\bf M}}{\partial B_{ ab}} | 0 \rangle \, dB_{a b} + \langle 0 | \widehat{C}^\dagger_{\bf M} \frac{\partial \widehat{C}_{\bf M}}{\partial D_{ij}} | 0 \rangle \, dD_{a b} \right], \label{BerryDef}$$ where each of the amplitudes in (\[BerryDef\]) have to be determined separately. Another tedious but standard calculation yields the following expressions for these amplitudes $$\begin{aligned} \langle 0 | \widehat{C}^\dagger_{\bf M} \frac{\partial \widehat{C}_{\bf M}}{\partial A_{ab}} | 0 \rangle &=&\frac{i}{2 \hbar} \langle 0 | \widehat{x}_b \; \widehat{x}_j | 0 \rangle B^{-1}_{j a}= \frac{i}{4 \hbar} l^2_b B^{-1}_{ba}, \label{Amp1}\\ \langle 0 | \widehat{C}^\dagger_{\bf M} \frac{\partial \widehat{C}_{\bf M}}{\partial B_{ab}} | 0 \rangle &=&\frac{1}{2 i \hbar} B^{-1}_{b l} \langle 0 | \widehat{C}^\dagger_{\bf M} \, \widehat{x} _l \; \widehat{x}_j \widehat{C}_{\bf M} | 0 \rangle D_{j k} B^{-1}_{k a} + \frac{i}{ \hbar} B^{-1}_{b l} \langle 0 | \widehat{C}^\dagger_{\bf M} \, \widehat{x} _l \; \widehat{C}_{\bf M} \; \widehat{x}_k | 0 \rangle B^{-1}_{k a} + \frac{1}{2 i \hbar} B^{-1}_{b l} A_{l j} \langle 0 | \widehat{x}_j \; \widehat{x}_k | 0 \rangle B^{-1}_{k a} + \nonumber \\ && - \frac{1}{2} B^{-1}_{b a} = - \frac{i}{4 \hbar} B^{-1}_{b c} A_{c d} l^2_d C_{a d} - \frac{i \hbar}{4} l^{-2}_b D_{ab}, \label{Amp2} \\ \langle 0 | \widehat{C}^\dagger_{\bf M} \frac{\partial \widehat{C}_{\bf M}}{\partial D_{ab}} | 0 \rangle &=& \frac{i}{2 \hbar} B^{-1}_{b j} \langle 0 | \widehat{C}^\dagger_{\bf M} \; \widehat{x}_j \; \widehat{x}_a \; \widehat{C}_{\bf M} | 0 \rangle = \frac{i}{4 \hbar} B^{-1}_{b c} A_{c d} l^2_d A_{a d} + \frac{i \hbar}{4} l^{-2}_b B_{a b}. \label{Amp3}\end{aligned}$$ Here, the covariance matrix expression (\[CVMatrix\]) is needed. The amplitudes $ \langle 0 | \widehat{x}_b \; \widehat{x}_j | 0 \rangle$ and $ \langle 0 | \widehat{C}^\dagger_{\bf M} \; \widehat{x}_b \; \widehat{x}_j \; \widehat{C}_{\bf M} | 0 \rangle$ are replaced by the expressions in (\[EqCV1\]) and (\[CVMatrix\]) of the covariance matrix components. For the amplitude $\langle 0 | \widehat{C}^\dagger_{\bf M} \, \widehat{x} _l \; \widehat{C}_{\bf M} \; \widehat{x}_k | 0 \rangle$ we used the expression (\[QASG\]) $$\langle 0 | \widehat{C}^\dagger_{\bf M} \, \widehat{x} _l \; \widehat{C}_{\bf M} \; \widehat{x}_k | 0 \rangle = \langle 0 | \widehat{C}^\dagger_{\bf M} \, \widehat{x} _l \; \left( \widehat{C}_{\bf M} \; \widehat{x}_k \widehat{C}^{-1}_{\bf M}\right) \widehat{C}_{\bf M} | 0 \rangle = \langle 0 | \widehat{C}^\dagger_{\bf M} \, \widehat{x} _l \; \sum_j \left( {\bf D}_{j k} \, \widehat{x}_j - {\bf B}_{j k} \,\widehat{p}_j \right) \widehat{C}_{\bf M} | 0 \rangle ,$$ and again, we inserted the expressions for the resulting amplitudes using the covariance matrix components. We now plug in (\[Amp1\]), (\[Amp2\]) and (\[Amp3\]) in the relation for the geometric connection (\[BerryDef\]), and after some simplifications using the block form of matrix ${\bf M}$ and the symplectic group conditions (\[Sp1Coord\]), we obtain the following form for the geometric phase $$\gamma_{\bf M} = - \frac{1}{4 \hbar} \int_{\cal C} \mbox{Tr} \left[ \left( \begin{array}{cc} {\bf L}^2 & {\bf 0} \\ {\bf 0} & \hbar^2 {\bf L}^{-2} \end{array}\right) {\bf M}^T \, {\bf \Omega} \, d {\bf M} \right] ,\label{BerryPhaseFormula}$$ where the symbol ‘Tr’ stands for the trace of the matrix. This expression exhibits an invariance under canonical transformation. To show this, consider a unitary operator $\widehat{C}_{\bf M'}$ related with a new symplectic matrix ${\bf M'}$. The Hamiltonian is given by $$\widehat{H}' = \widehat{H}_{{ \bf M'} \cdot {\bf M}} = \widehat{C}_{\bf M'} \, \widehat{H}_{\bf M} \widehat{C}^{-1}_{\bf M'} = \widehat{C}_{{\bf M'} \cdot {\bf M}} \, \widehat{H} \widehat{C}^{-1}_{ {\bf M'} \cdot {\bf M}} ,$$ where the group multiplication was used in the second line. The new symplectic matrix ${\bf M'} \cdot {\bf M}$, with ${\bf M'}$ fixed, can be inserted in the geometric phase (\[BerryPhaseFormula\]) and this yields $$\gamma_{\bf M} = - \frac{1}{4 \hbar} \int_{\cal C} \mbox{Tr} \left[ \left( \begin{array}{cc} {\bf L}^2 & {\bf 0} \\ {\bf 0} & \hbar^2 {\bf L}^{-2} \end{array}\right) ({\bf M'} \cdot {\bf M})^T \, {\bf \Omega} \, d ({\bf M'} \cdot {\bf M}) \right] = - \frac{1}{4 \hbar} \int_{\cal C} \mbox{Tr} \left[ \left( \begin{array}{cc} {\bf L}^2 & {\bf 0} \\ {\bf 0} & \hbar^2 {\bf L}^{-2} \end{array}\right) {\bf M}^T \cdot {\bf M'}^T \, {\bf \Omega} \, {\bf M'} \cdot d{\bf M} \right] =\gamma_{\bf M},$$ which shows the invariance of the geometric phase under canonical transformations. On the other hand, it can be notice that after an integration by parts in (\[BerryPhaseFormula\]), the boundary term is given by the covariance matrix ${\bf V}^{(2)}$ $$\gamma_{\bf M} = \frac{1}{4 \hbar} \int_{\cal C} \mbox{Tr} \left[ {\bf \Omega} \, {\bf M} \, \left( \begin{array}{cc} {\bf L}^2 & {\bf 0} \\ {\bf 0} & \hbar^2 {\bf L}^{-2} \end{array}\right) \, d {\bf M}^T \right] - \frac{1}{2\hbar} \int_{\cal C} d \mbox{Tr}\left[ {\Omega} \, {\bf V}^{(2)} \right] ,\label{BerryPhaseFormula2}$$ which is null, when a closed path in the parameters space is considered. Summarizing, the expression (\[BerryPhaseFormula\]) (or (\[BerryPhaseFormula2\])) yields the geometric phase for a variation on the parameters of the group element ${\bf M} \in Sp(2n, \mathbb{R})$. This expression is valid for an arbitrary group element ${\bf M}$ and not just for those close to the group identity. Also, from this expression we can read the Berry’s connection ${\cal A}_{\bf M}$ for the symplectic group acting on the Gaussian state $$\gamma_{\bf M} = \int_{\cal C} {\cal A}_{\bf M}.$$ Due to the group $Sp(2n,\mathbb{R})$ has dimension $n(2n+1)$, there will be $n (2n-1)$ spurious components which can be fixed with a gauge transformation of ${\cal A}_{\bf M}$. However, although a detailed analysis of this connection ${\cal A}_{\bf M}$ is required, it is beyond the scope of this paper and will be studied elsewhere. Instead, let us now show in the following subsections two examples of the geometric phase associated to the squeeze operator, for $n=1$ and $n=2$. Recall that the singular case when ${\bf B} = {\bf 0}$ is treated in the appendix \[SingularCase\]. Geometric phase for the $n=1$ squeeze operator. ----------------------------------------------- The squeeze operator for the $n=1$ case is given by $$\widehat{S}^{(1)}(\zeta) = e^{\frac{1}{2} ( \zeta^* \widehat{a}^2 - \zeta (\widehat{a}^\dagger)^2 )},$$ where $\widehat{a}$ and $\widehat{a}^\dagger $ are the ladder operators satisfying the canonical commutation relation $[\widehat{a}, \widehat{a}^\dagger] = 1$. The real parameter $r$, which is the absolute value of the parameter $\zeta = r e^{i \theta} \in \mathbb{C}$, labels the squeezing degree of the system. We now use the transformation $\widehat{a} = \frac{1}{\sqrt{2}} \frac{\widehat{x}}{l} + \frac{i}{\sqrt{2}} \frac{l \, \widehat{p}}{\hbar}$, to obtain the squeeze operator in the Schrödinger representation $$\widehat{S}^{(1)}(\zeta) = e^{- \frac{i}{2 \hbar} \left( \begin{array}{cc} \widehat{x} & \widehat{p} \end{array} \right) \left( \begin{array}{cc} \frac{\hbar r}{l^2} \sin \theta & - r \cos \theta \\ - r \cos \theta & - \frac{l^2 \, r}{\hbar} \sin \theta \end{array} \right) \left( \begin{array}{c} \widehat{x} \\ \widehat{p} \end{array}\right) }.\label{SqueezeOp}$$ Here, the Lie algebra isomorphism between $sp(2,\mathbb{R})$ and the Lie algebra of the second order polynomials [@de2011symplectic; @hall2018theory] allows us to take the matrix $${\bf L}^{(1)}_s = \left( \begin{array}{cc} \frac{\hbar}{l^2} r \sin \theta & - r \cos \theta \\ - r \cos \theta & - \frac{l^2}{\hbar} r \sin \theta \end{array} \right),$$ as the Lie algebra element of $sp(2,\mathbb{R})$. To make a difference between this description and the $n=2$ case further below, we introduce the index ${}^{(1)}$, which indicates we are working with $n=1$. In both cases, ${\bf L}^{(j)}_s$ with $j=1$ or $j=2$, corresponds to the Lie algebra element in $sp(2j,\mathbb{R})$ arising from the corresponding squeeze operator $\widehat{S}^{(j)}(\zeta)$. The symplectic matrix ${\bf M}_s({\bf L}^{(1)}_s)$ associated to this operator is given by the exponential map of the matrix ${\bf L}^{(1)}_s$ which yields $${\bf M}_s({\bf L}^{(1)}_s) = \left( \begin{array}{cc} \cosh r - \cos \theta \sinh r & - \frac{l^2}{\hbar}\sin \theta \sinh r \\ - \frac{\hbar}{l^2} \sin \theta \sinh r & \cosh r + \cos \theta \sinh r \end{array} \right).$$ We now insert this matrix in (\[BerryPhaseFormula\]) and consider the trajectory ${\cal C}$ given by a constant $r=R$ and $\theta = [0, 2 \pi)$. A quick calculation gives the following expression for the corresponding geometric phase $$\gamma^{(1)}_{\bf s} = - \pi \sinh^2 (R), \label{Berryn1}$$ which coincides with the same calculation reported in [@chaturvedi1987berry]. Remarkably, despite the factor ${\frac{1}{\hbar}}$ in (\[BerryPhaseFormula\]) and the factors $\hbar/l^2$ and $l^2/\hbar$ in the squeeze operator, the geometric phase is $\hbar$ and $l^2$ independent. It only depends on the squee parameter $r$. We might expect that this feature is intrinsic of the system with $n=1$, but, as we will soon notice, is also present in the $n=2$ case. Geometric phase for the $n=2$ squeeze operator. ----------------------------------------------- The squeeze operator $\widehat{S}^{(2)}(\zeta)$ for a bi-partite system is of the form $$\widehat{S}^{(2)}(\zeta) = e^{ \left( \zeta^* \widehat{a}_1 \widehat{a}_2 - \zeta \widehat{a}^\dagger_1 \widehat{a}^\dagger_2 \right) }, \label{SqOperator}$$ where as before, $\widehat{a}_1$ and $\widehat{a}_2$ are the annihilation operators for the sub-systems, say, 1 and 2, of the bi-partite system. $\widehat{a}^\dagger_1$ and $\widehat{a}^\dagger_2$ are their adjoint operators respectively and $\zeta = r e^{i \phi}$ is a complex number labelling the amount of squeezing. Again, the operator in (\[SqOperator\]) is given in the Fock representation and we write it in the Schrödinger representation using the transformation $$\begin{aligned} \widehat{a}_j = \frac{1}{\sqrt{2}} \frac{\widehat{x}_j}{l_j} + \frac{i }{\sqrt{2}} \frac{ l_j \widehat{p}_j}{\hbar} , \qquad \widehat{a}^\dagger_j = \frac{1}{\sqrt{2}} \frac{\widehat{x}_j}{l_j} - \frac{i }{\sqrt{2}} \frac{ l_j \widehat{p}_j}{\hbar} , \qquad j=1,2.\end{aligned}$$ Once we insert these expressions for $\widehat{a}_j$ and $\widehat{a}^\dagger_j$ in the expression for $\widehat{S}^{(2)}(\zeta)$, it takes the following form $$\begin{aligned} \widehat{S}^{(2)}(\zeta) &=& \exp \left\{ \frac{-i}{\hbar} \left[ \frac{\hbar \, \zeta_y}{l_1 l_2} \; \widehat{x}_1 \widehat{x}_2 - \frac{l_2 \, \zeta_x}{l_1} \; \widehat{x}_1 \widehat{p}_2 - \frac{l_1\, \zeta_x}{l_2} \; \widehat{p}_1 \widehat{x}_2 - \frac{l_1 l_2 \, \zeta_y}{\hbar} \widehat{p}_1 \widehat{p}_2 \right] \right\}, \label{SqGen}\end{aligned}$$ where $\zeta_x$ and $\zeta_y$ are the real and imaginary parts of $\zeta$. The operators in the argument of (\[SqGen\]) can be accommodated in a matrix form as follows $$\widehat{S}^{(2)}(\zeta) = \exp\left\{ - \frac{i}{2\hbar} (\vec{\widehat{R}}^T_1, \vec{\widehat{R}}^T_2) \left( \begin{array}{cc} {\bf 0} & {\bf b} \\ {\bf b}^T & {\bf 0} \end{array}\right) \left( \begin{array}{c} \vec{\widehat{R}}_1 \\ \vec{\widehat{R}}_2 \end{array}\right) \right\}. \label{SqueezedGen}$$ where ${\bf b}$ is given by $${\bf b} = \left( \begin{array}{cc} \frac{\hbar}{l_1 l_2} \zeta_y & -\frac{l_2}{l_1} \zeta_x \\ -\frac{l_1}{l_2} \zeta_x & -\frac{l_1 l_2}{\hbar} \zeta_y \end{array}\right). \label{DimensionFullb}$$ The Lie algebra isomorphism between $sp(4,\mathbb{R})$ and the Lie algebra of the second order polynomials [@de2011symplectic; @hall2018theory], allows us to take the $4 \times 4$ square matrix in (\[SqueezedGen\]) to be the Lie algebra element ${\bf L}^{(2)}_s \in sp(4,\mathbb{R})$. This matrix ${\bf L}^{(2)}_s$ has the block matrices ${\bf a} = {\bf c} = {\bf 0}$ (see the appendix \[LAAnalysis\] for more details) and ${\bf b}$ is given in (\[DimensionFullb\]). The matrix ${\bf L}^{(2)}_s$ is now inserted in the formula (\[FormulaForSQ\]) of Appendix \[LAAnalysis\] yielding the expression for $\widetilde{\bf M}$ which, after using the transformation $\Gamma$ in (\[RelationbetweenMs\]) gives the following form for the matrix ${\bf M}({\bf L}^{(2)}_s)$ $${\bf M}({\bf L}^{(2)}_s) = \left( \begin{array}{cccc} \cosh(r) & - \frac{l_1}{l_2} \sinh(r) \cos\phi & 0 & - \frac{l_1 l_2}{\hbar} \sinh(r) \sin\phi \\ - \frac{l_2}{l_1} \sinh(r) \cos\phi & \cosh(r) & - \frac{l_1 l_2}{\hbar} \sinh(r) \sin\phi & 0 \\ 0 & - \frac{\hbar}{l_1 l_2} \sinh(r) \sin\phi & \cosh(r) & \frac{l_2}{l_1} \sinh(r) \cos\phi \\ - \frac{\hbar}{l_1 l_2} \sinh(r) \sin\phi & 0 & \frac{l_1}{l_2} \sinh(r) \sin\phi & \cosh(r)\end{array}\right). \label{MforN2}$$ Let us now consider the same trajectory used for the $n=1$ case, that is to say, $r = R$ is constant and $\phi \in [0, 2 \pi)$ and insert this matrix in (\[BerryPhaseFormula\]). After some matrix calculations and the corresponding integration in $\phi$ it yields the following result $$\gamma^{(2)}_{\bf s} = - 2\pi \sinh^2 (R). \label{Berryn2}$$ As can be notice, this value for the geometric phase doubles the geometric phase for the $n=1$ case using the same trajectory in the parameters space. This implies that for the same amount of squeezing $r$ the observed geometric phase for $n=2$ must double that of the $n=1$ case. Remarkably, both geometric phases (\[Berryn1\]) and (\[Berryn2\]) are $\hbar$ and $l_j$ independent and depend only on the squeezing parameter $r$. Conclusions {#Discussion} =========== The main results of this work are the expressions (\[BerryPhaseFormula\]) and (\[BerryPhaseFormula2\]) for the geometric phases of a general Gaussian state of the form $\Psi_{\bf M}$ and an arbitrary path ${\cal C}$. This Gaussian state corresponds to a state generated by the action of the operator $\widehat{C}_{\bf M}$ on the vacuum state $\Psi_0(\vec{x})$ for a $n$-partite system. The operator $\widehat{C}_{\bf M}$ is the unitary representation of the symplectic group in the Hilbert space ${\cal H}$ given in section \[CovarianceMatrix\]. Our first observation is that, despite the ${\bf B}^{-1}$ term in the amplitudes (\[Amp1\]), (\[Amp2\]) and (\[Amp3\]), we managed to re-write the geometric phase in a presentable way and also, we showed how the covariance matrix ${\bf V}^{(2)}$ arises as a boundary term in (\[BerryPhaseFormula2\]). The special case for ${\bf B}= {\bf 0}$ was considered in the appendix (\[SingularCase\]). We also showed the invariance of the geometric phase under canonical transformations. Surprisingly, modulo other constant matrices, the Berry connection ${\cal A}_{\bf M}$ results to be a ‘bilinear’ term in the symplectic group matrices, that is to say, ${\cal A}_{\bf M} \sim {\bf M}^T \, {\Omega} \, d{\bf M}$ . Therefore, this result paves the way to explore the Berry curvature and its singular points. Additionally, we calculated the geometric phase for the squeeze operators $\widehat{S}^{(j)}(\zeta)$ with $j=1,2$ and showed, in the case of $n=2$, that the geometric phase is twice the geometric phase for $n=1$. Moreover, both geometric phases are $\hbar$ and $l_j$ independent. This feature might induce some sort of ‘classical nature’ to these phases. However, due to the squeezed states are quantum states with no classical analog, the quantum nature of the their corresponding geometric phase is granted. The absence of dimension-full parameters like $\hbar$ or $l_j$ in the squeeze phase suggest that the dynamical of the system does not play a relevant role in these phases, i.e., their nature is more like a kinematical feature of the symplectic group topology [@mukunda2014classical]. Finally, it is worth to mention that in order to derive the expression for the geometric phase, we first derived the covariance matrix for Gaussian states in (\[CVMatrix\]), and we also provided the full relation between $sp(4,\mathbb{R})$ and $Sp(4,\mathbb{R})$ in the appendix \[LAAnalysis\]. This result showed in the appendix \[LAAnalysis\], to the best of the author’s knowledge, have not been reported elsewhere. The full relation between the symplectic group $Sp(4,\mathbb{R})$ and its Lie algebra can be used to explore different geometric phases and not just those related with the squeeze operator. Acknowledgments =============== I thank Andrea Mari, Alessandro Bravetti and M. Bermúdez-Montaña for their comments and helpful discussions. Appendix: Relation between $Sp(4,\mathbb{R})$ and $sp(4,\mathbb{R})$ {#LAAnalysis} ==================================================================== In this appendix we derive the direct relation between the Lie algebra elements $sp(4,\mathbb{R})$ and the symplectic group matrices $Sp(4,\mathbb{R})$. Lie algebra and group analysis ------------------------------ Consider an arbitrary element ${\bf m} \in sp(4,\mathbb{R})$ of the form $${\bf m} = \left( \begin{array}{cc} {\bf J} & 0 \\ 0 & {\bf J} \end{array} \right) {\bf L}, \label{LieAlegebraElem}$$ where the matrix ${\bf L}$ is a real symmetric matrix written as $${\bf L}= \left( \begin{array}{cc} {\bf a} & {\bf b} \\ {\bf b}^T & {\bf c} \end{array}\right), \label{LMatrix}$$ and ${\bf a}$ and ${\bf c}$ are $2\times2$ symmetric matrices and ${\bf b}$ is a $2\times2$ matrix. The elements of the symplectic group $Sp(4, \mathbb{R})$ close to the identity can be obtained via the exponential map [@hall2018theory] of the Lie algebra element ${\bf L}$ as $$\widetilde{\bf M} = \exp{ \left[ \left( \begin{array}{cc} {\bf J} & 0 \\ 0 & {\bf J} \end{array} \right) {\bf L} \right]} = \left( \begin{array}{cc} \widetilde{\bf A} & \widetilde{\bf B} \\ \widetilde{\bf C} & \widetilde{\bf D} \end{array}\right). \label{MandS}$$ The aim of this section is to obtain the relation between the block matrices $\widetilde{\bf A}$, $\widetilde{\bf B}$, $\widetilde{\bf C}$ and $\widetilde{\bf D}$ and the Lie algebra element ${\bf L}$. What we will obtain is a relation between the block matrices ${\bf a}$, ${\bf b}$ and ${\bf c}$ of the Lie algebra element ${\bf L}$ and the matrices $\widetilde{\bf A}$, $\widetilde{\bf B}$, $\widetilde{\bf C}$ and $\widetilde{\bf D}$ of the group element $\widetilde{\bf M}$. Recall that we are using a group action which is different, although equivalent, to the one used for the calculation of the covariance matrix and consequently, for the geometric phase, that is to say, in this section we use $\widetilde{\bf M}$ instead of ${\bf M}$. To proceed, let us expand the exponential in (\[MandS\]) and collect together the even and odd terms of the expansion as follows $$\begin{aligned} \widetilde{\bf M} &=& \left[ {\bf 1} + \frac{1}{2!} {\bf S} + \dots + \frac{1}{(2n)!} {\bf S}^n + \dots \right] + \sqrt{{\bf S}} \left[ {\bf 1} + \frac{1}{3!} {\bf S} + \dots + \frac{1}{(2n+1)!} {\bf S}^n + \dots \right], \label{Expansion} \end{aligned}$$ where the matrix ${\bf S}$ is defined as $${\bf S} = \left[ \left( \begin{array}{cc} {\bf J} & 0 \\ 0 & {\bf J} \end{array} \right) \left( \begin{array}{cc} {\bf a} & {\bf b} \\ {\bf b}^T & {\bf c} \end{array} \right) \right]^{2} = \left( \begin{array}{cc} - (\det {\bf a} + \det {\bf b}) {\bf 1}_{2\times2} & {\bf J} {\bf d} \\ - {\bf J} {\bf d}^T & - (\det {\bf b} + \det {\bf c}) {\bf 1}_{2\times2} \end{array}\right), \label{SDefinition}$$ and the matrix ${\bf d}$ is given by ${\bf d} = {\bf a} {\bf J} {\bf b} + {\bf b} {\bf J} {\bf c}$. The notation used in (\[Expansion\]) for $\sqrt{\bf S}$ refers to the matrix $$\sqrt{\bf S} := \left( \begin{array}{cc} {\bf J} & 0 \\ 0 & {\bf J} \end{array} \right) \left( \begin{array}{cc} {\bf a} & {\bf b} \\ {\bf b}^T & {\bf c} \end{array} \right).$$ As can be seen from the expansion (\[Expansion\]), in order to obtain the expression for $\widetilde{\bf M}$ we need first to determine ${\bf S}^n$ and then, we have to insert the expression for ${\bf S}^n$ in (\[Expansion\]) and calculate both sums therein. Let us proceed in the next subsection with the first step: the calculation of ${\bf S}^{n}$. Calculation of ${\bf S}^{n}$ ---------------------------- The matrix ${\bf S}$ is formed by four $2 \times 2$ block matrices where the upper left and the lower right are multiples of the identity matrix ${\bf 1}_{2\times2}$. The upper right block is the matrix ${\bf J} {\bf d}$ whereas the lower left is $- {\bf J} {\bf d}^T$. Notably, we found that this block structure is preserved after exponentiating the matrix ${\bf S}$ an integer number of times. That is to say, the n-power of matrix ${\bf S}$ yields a new matrix ${\bf S}^n$ given as $${\bf S}^n = \left( \begin{array}{cc} \alpha_n{\bf 1}_{2\times2} & \beta_n {\bf J} {\bf d} \\ - \beta_n {\bf J} {\bf d}^T & \gamma_n {\bf 1}_{2\times2} \end{array}\right). \label{MatrixSDef}$$ The coefficients $\alpha_n$, $\beta_n$ and $\gamma_n$, to be determined, depend on the values of the matrices ${\bf a}$, ${\bf b}$, ${\bf c}$ and ${\bf d}$. For $n=1$, these coefficients are given by the factors in the block matrices of ${\bf S} $ in (\[SDefinition\]) and can be directly defined as $$\alpha_1 := - (\det {\bf a} + \det {\bf b}), \; \beta_1 := +1, \; \gamma_1 := - (\det {\bf c} + \det {\bf b}). \label{initialvalues}$$ To calculate these coefficients $\alpha_n$, $\beta_n$ and $\gamma_n$ for arbitrary $n$, first note that they can be generated with a linear operator ${\bf T}$ as $$\begin{aligned} \left( \begin{array}{c} \alpha_n \\ \beta_n \\ \gamma_n \end{array}\right) = {\bf T}^{n-1} \left( \begin{array}{c} \alpha_1 \\ \beta_1 \\ \gamma_1 \end{array}\right), \label{EqForCoeff}\end{aligned}$$ where the matrix ${\bf T}$ is given by $${\bf T} = \left( \begin{array}{ccc} \alpha_1 & \beta_1 \det {\bf d} & 0 \\ \beta_1 & \gamma_1 & 0 \\ 0 & \beta_1 \det {\bf d} & \gamma_1 \end{array}\right).$$ The calculation shows that the $n-1$ power of ${\bf T}$ is a matrix of the form $${\bf T}^{n-1} = \left( \begin{array}{cc} {\bf U}^{n-1} & \vec{0}^T \\ \vec{u}^T \gamma_1^{n-2} \sum^{n-2}_{j=0} \gamma^{-j}_1 {\bf U}^j & \gamma^{n-1}_1 \end{array}\right),$$ where $\vec{0} = (0,0)$ and $\vec{u} = (0, \beta_1 \det d)$ and matrix ${\bf U}$ is given by $$\begin{aligned} {\bf U} = \left( \begin{array}{cc} \alpha_1 & \beta_1 \det {\bf d} \\ \beta_1 & \gamma_1\end{array}\right).\end{aligned}$$ Then, using (\[EqForCoeff\]) we have the following relation for the coefficients $$\begin{aligned} \left( \begin{array}{c} \alpha_n \\ \beta_n \end{array}\right) &=& {\bf U}^{n-1} \left( \begin{array}{c} \alpha_1 \\ \beta_1 \end{array}\right), \\ \gamma_n &=& \gamma^n_1 + \vec{u}^T \gamma^{n-2}_1 \sum^{n-2}_{j=0} \gamma^{-j}_1 {\bf U}^j \left( \begin{array}{c} \alpha_1 \\ \beta_1 \end{array}\right). \label{FExpressionCoeff}\end{aligned}$$ In order to calculate ${\bf U}^{n-1}$ we need to diagonalize matrix ${\bf U}$ hence, let ${\bf P}$ be the matrix diagonalizing ${\bf U}$, then $${\bf U} = {\bf P} \, {\bf U}_d \, {\bf P}^{-1}, \label{EqForU}$$ where ${\bf U}_d$ is the diagonal matrix and the matrix ${\bf P}$ is $${\bf P} = \left( \begin{array}{cc} \frac{(\lambda_+ - \gamma_1)}{\beta_1 } k_1 & \frac{(\lambda_{-} - \gamma_1)}{\beta_1} k_2 \\ k_1 & k_2 \end{array}\right).$$ The real arbitrary parameters $k_1$ and $k_2$ result from the diagonalization procedure. Its values will be automatically cancelled as part of the calculation of ${\bf U}^{n-1}$ further below. The eigenvalues of ${\bf U}$, denoted by $\lambda_{\pm}$, have the following expressions $$\begin{aligned} \lambda_{\pm} &=& - \frac{\det {\bf a} + \det {\bf c} + 2 \det {\bf b} }{2} \pm \frac{1}{2} \sqrt{ (\det {\bf a} - \det {\bf c})^2 + 4 \det {\bf d} }, \label{eigenvalues}\end{aligned}$$ and the diagonal matrix ${\bf U}_d$ is $${\bf U}_d = \left( \begin{array}{cc} \lambda_+ & 0 \\ 0 & \lambda_{-} \end{array}\right).$$ We now take the $n-1$ power of ${\bf U}$ given in (\[EqForU\]) to obtain the following result $$\begin{aligned} {\bf U}^{n-1} = \left( \begin{array}{cc} \frac{(\lambda_+ - \gamma_1)}{\beta_1 } k_1 & \frac{(\lambda_{-} - \gamma_1)}{\beta_1} k_2 \\ k_1 & k_2 \end{array}\right) \left( \begin{array}{cc} \lambda^{n-1}_+ & 0 \\ 0 & \lambda^{n-1}_{-} \end{array}\right) \left( \begin{array}{cc} \frac{(\lambda_+ - \gamma_1)}{\beta_1 } k_1 & \frac{(\lambda_{-} - \gamma_1)}{\beta_1} k_2 \\ k_1 & k_2 \end{array}\right)^{-1},\end{aligned}$$ which, when combined with the result in (\[FExpressionCoeff\]) together with the expression for $\vec{u}$, gives $$\begin{aligned} \alpha_n &=& \frac{\left[ ( \lambda_+ - \gamma_1 ) \lambda^n_+ - ( \lambda_- - \gamma_1 ) \lambda^n_- \right]}{\sqrt{ (\alpha_1 - \gamma_1)^2 + 4 \beta^2_1 \det {\bf d} } } ,\label{alpha} \\ \beta_n &=& \frac{\left[ \lambda^n_+ - \lambda^n_- \right]}{\sqrt{ (\alpha_1 - \gamma_1)^2 + 4 \beta^2_1 \det {\bf d} } } ,\label{beta} \\ \gamma_n &=& \frac{\left[ ( \lambda_+ - \gamma_1 ) \lambda^n_- - ( \lambda_- - \gamma_1 ) \lambda^n_+ \right]}{\sqrt{ (\alpha_1 - \gamma_1)^2 + 4 \beta^2_1 \det {\bf d} } }. \label{gamma}\end{aligned}$$ These are the final expressions for the coefficients in ${\bf S}^n$. We are now ready to move to the second step: the analysis of the infinite series in (\[Expansion\]). Series analysis --------------- Using the expression for the $n$ power of matrix ${\bf S}$, defined in (\[MatrixSDef\]), the expression (\[Expansion\]) can be written as $$\begin{aligned} \widetilde{\bf M} &=& {\bf 1} + \sum^{+\infty}_{n=1} \frac{1}{(2n)!} \left( \begin{array}{cc} \alpha_n & \beta_n {\bf J} {\bf d} \\ - \beta_n {\bf J} {\bf d}^T & \gamma_n \end{array}\right) + \sqrt{{\bf S}}\left[ {\bf 1} + \sum^{+\infty}_{n=1} \frac{1}{(2n+1)!} \left( \begin{array}{cc} \alpha_n & \beta_n {\bf J} {\bf d} \\ - \beta_n {\bf J} {\bf d}^T & \gamma_n \end{array}\right) \right]. \label{GroupMatrix}\end{aligned}$$ After collecting the components of each block matrix we obtain the following coefficients $$\begin{aligned} \alpha^{(e)} := 1 + \sum^{+\infty}_{n=1} \frac{1}{(2n)!} \alpha_n, \qquad \beta^{(e)} := \sum^{+\infty}_{n=1} \frac{1}{(2n)!} \beta_n, \qquad \gamma^{(e)} := 1 + \sum^{+\infty}_{n=1} \frac{1}{(2n)!} \gamma_n, \label{EvenCoeff}\\ \alpha^{(o)} := 1 + \sum^{+\infty}_{n=1} \frac{1}{(2n+1)!} \alpha_n, \qquad \beta^{(o)} := \sum^{+\infty}_{n=1} \frac{1}{(2n+1)!} \beta_n, \qquad \gamma^{(o)} := 1 + \sum^{+\infty}_{n=1} \frac{1}{(2n+1)!}\gamma_n. \label{OddCoeff}\end{aligned}$$ We now insert (\[alpha\]), (\[beta\]) and (\[gamma\]) in the relations (\[EvenCoeff\]) - (\[OddCoeff\]) to obtain the final form of the coefficients $$\begin{aligned} \alpha^{(e)} &=& \frac{ (\lambda_+ + \det{\bf b} + \det{\bf c}) \cosh\sqrt{\lambda_+} - (\lambda_- + \det{\bf b} + \det{\bf c}) \cosh\sqrt{\lambda_- } }{\sqrt{ (\det{\bf a} - \det{\bf c})^2 + 4 \det {\bf d} } } , \label{alphapar} \\ \alpha^{(o)} &=& \frac{ (\lambda_+ + \det{\bf b} + \det{\bf c}) \frac{ \sinh\sqrt{\lambda_+}}{\sqrt{\lambda_+}} - (\lambda_- + \det{\bf b} + \det{\bf c}) \frac{ \sinh\sqrt{\lambda_- }}{\sqrt{\lambda_-}} }{\sqrt{ (\det{\bf a} - \det{\bf c})^2 + 4 \det {\bf d} } } , \label{alphaimpar} \\ \beta^{(e)} &=& \frac{ \cosh\sqrt{\lambda_+} - \cosh\sqrt{\lambda_- }}{\sqrt{ (\det{\bf a} - \det{\bf c})^2 + 4 \det {\bf d} } } , \label{betapar} \\ \beta^{(o)} &=& \frac{\frac{ \sinh\sqrt{\lambda_+} }{\sqrt{\lambda_+}} - \frac{ \sinh\sqrt{\lambda_- }}{\sqrt{\lambda_-}} }{\sqrt{ (\det{\bf a} - \det{\bf c})^2 + 4 \det {\bf d} } } , \label{betaimpar} \\ \gamma^{(e)} &=& \frac{ (\lambda_+ + \det{\bf b} + \det{\bf c}) \cosh\sqrt{\lambda_-} - (\lambda_- + \det{\bf b} + \det{\bf c}) \cosh\sqrt{\lambda_+ } }{\sqrt{ (\det{\bf a} - \det{\bf c})^2 + 4 \det {\bf d} } } , \label{gammapar} \\ \gamma^{(o)} &=& \frac{ (\lambda_+ + \det{\bf b} + \det{\bf c}) \frac{ \sinh\sqrt{\lambda_-}}{\sqrt{\lambda_-}} - (\lambda_- + \det{\bf b} + \det{\bf c}) \frac{\sinh\sqrt{\lambda_+ }}{\sqrt{\lambda_+}} }{\sqrt{(\det{\bf a} - \det{\bf c})^2 + 4 \det {\bf d} } } , \label{gammaimpar}\end{aligned}$$ where we have to recall the expressions for the eigenvalues $\lambda_\pm$ in (\[eigenvalues\]). We now use these results to provide the final expression for the matrix components in (\[GroupMatrix\]) $$\begin{aligned} \widetilde{\bf A}({\bf a}, {\bf b}, {\bf c}) &=& \alpha^{(e)} + (\alpha^{(o)} - \beta^{(o)} \det{\bf b}) \, {\bf J} \, {\bf a} + \beta^{(o)} {\bf J} \,{\bf b} \, {\bf J} \, {\bf c}\, {\bf J} \, {\bf b}^T, \label{TerminoA}\\ \widetilde{\bf B}({\bf a}, {\bf b}, {\bf c}) &=& (\gamma^{(o)} - \beta^{(o)} \det{\bf a}) \, {\bf J} \, {\bf b} + \beta^{(e)} ( {\bf J} \,{\bf a} \, {\bf J} \, {\bf b} + {\bf J} \,{\bf b} \, {\bf J} \, {\bf c} )+ \beta^{(o)} {\bf J} \,{\bf a} \, {\bf J} \, {\bf b} \, {\bf J} \, {\bf c}, \label{TerminoB} \\ \widetilde{\bf C}({\bf a}, {\bf b}, {\bf c}) &=& (\alpha^{(o)} - \beta^{(o)} \det{\bf c}) \, {\bf J} \, {\bf b}^T + \beta^{(e)} ( {\bf J} \,{\bf b}^T \, {\bf J} \, {\bf a} + {\bf J} \,{\bf c} \, {\bf J} \, {\bf b}^T )+ \beta^{(o)} {\bf J} \,{\bf c} \, {\bf J} \, {\bf b}^T \, {\bf J} \, {\bf a}, \label{TerminoC} \\ \widetilde{\bf D}({\bf a}, {\bf b}, {\bf c}) &=& \gamma^{(e)} + (\gamma^{(o)} - \beta^{(o)} \det{\bf b}) \, {\bf J} \, {\bf c} + \beta^{(o)} {\bf J} \,{\bf b}^T \, {\bf J} \, {\bf a}\, {\bf J} \, {\bf b}. \label{TerminoD}\end{aligned}$$ These expressions provide the relation between the components ${\bf a}$, ${\bf b}$ and ${\bf c}$ of the Lie algebra element ${\bf m}$ in (\[LieAlegebraElem\]) with the corresponding symplectic matrix $\widetilde{\bf M}$. To the best of the author’s knowledge, this result is new and have not been reported elsewhere. A particular case is when ${\bf a} = {\bf c} = 0$ which is connected with the form of the squeeze operators in section \[BerryPhase\]. Inserting these values on the previous expressions yield $$\widetilde{\bf M}({\bf 0}, {\bf b}, {\bf 0}) = \left( \begin{array}{cc} \cosh(\sqrt{- \det{\bf b}}) & \frac{\sinh(\sqrt{- \det{\bf b}})}{\sqrt{- \det{\bf b}}} {\bf J} {\bf b} \\ \frac{\sinh( \sqrt{- \det{\bf b}})}{\sqrt{- \det{\bf b}}} {\bf J} {\bf b}^T & \cosh^2(\sqrt{- \det{\bf b}}) \end{array}\right) . \label{FormulaForSQ}$$ Appendix: geometric phase for ${\bf B} = {\bf 0}$. {#SingularCase} ================================================== In this appendix we will briefly consider the special case when the symplectic matrix ${\bf M}$ has the block matrix ${\bf B} = {\bf 0}$. In this case, the block form of ${\bf M}$ is given as $${\bf M}({\bf B} = {\bf 0}) = \left( \begin{array}{cc} {\bf A} & {\bf 0} \\ {\bf C} & {\bf A}^{- T} \end{array} \right), \label{ScaseB0}$$ the representation of the operator $\widehat{C}_{\bf M}$ is now given as $$\widehat{C}_{\bf M} \Psi(\vec{x}) = \frac{1}{\det {\bf A}} e^{\frac{i}{2 \hbar} \vec{x}^T {\bf C} {\bf A}^{-1} \vec{x} } \Psi({\bf A}^{-1} \vec{x}), \label{SingularB}$$ hence it only depends on the matrices ${\bf A}$ and ${\bf C}$. The geometric phase in this case can be written as $$\gamma_{\bf M} = i \int_{\cal C} \left[ \langle 0 | \widehat{C}^{\dagger}_{\bf M} \frac{\partial }{\partial A_{ab}} \widehat{C}_{\bf M} | 0 \rangle d A_{a b} + \langle 0 | \widehat{C}^{\dagger}_{\bf M} \frac{\partial }{\partial C_{ab}} \widehat{C}_{\bf M} | 0 \rangle d C_{a b} \right], \label{BerryB0}$$ where $A_{ab}$ and $C_{ab}$ are the components of the matrices ${\bf A}$ and ${\bf C}$ respectively. We now use the representation in (\[SingularB\]) and calculate the amplitudes in (\[BerryB0\]) $$\begin{aligned} \langle 0 | \widehat{C}^{\dagger}_{\bf M} \frac{\partial }{\partial A_{ab}} \widehat{C}_{\bf M} | 0 \rangle &=& - \frac{i}{4\hbar} l^2_b A_{j b} C_{j k} A^{-1}_{k a} , \\ \langle 0 | \widehat{C}^{\dagger}_{\bf M} \frac{\partial }{\partial C_{ab}} \widehat{C}_{\bf M} | 0 \rangle &=& \frac{i}{4\hbar} l^2_b A_{a b},\end{aligned}$$ which when inserted in (\[BerryB0\]) gives $$\gamma_{\bf M} = \frac{1}{4 \hbar} \int_{\cal C} \mbox{Tr}\left[ {\bf L}^2 {\bf A}^T {\bf C} {\bf A}^{-1} d{\bf A} - {\bf L}^2 {\bf A}^T d{\bf C} \right] = - \frac{1}{4 \hbar} \int_{\cal C} \mbox{Tr}\left[ {\bf L}^2 \, {\bf A}^T \, d{\bf C} - {\bf L}^2 \, {\bf C}^T \, d{\bf A} \right].$$ This result coincides with (\[BerryPhaseFormula\]) for a symplectic matrix of the form given by (\[ScaseB0\]). It can be notice that when the block matrix ${\bf C}$ is null, the phase in (\[SingularB\]) is zero and consequently, the geometric phase is also zero independently of the expression for ${\bf A}$. [99]{} Bennett, C. H. and DiVincenzo, D. P., “Quantum information and computation”, *nature*, [**404**]{}, 6775, 2000. DiVincenzo, D. P., “Quantum computation”, *Science*, [**270**]{}, 5234, 1995. Pan, J., Simon, C., Brukner, [Č]{}. and Zeilinger, A., “Entanglement purification for quantum communication”, *Nature*, [**410**]{}, 6832, 2001. Axline, Christopher and others., “On-demand quantum state transfer and entanglement between remote microwave cavity memories”, *Nature Physics*, [**14**]{}, 7, 2018. Bouwmeester, D., Pan, J., Mattle, K., Eibl, M., Weinfurter, H. and Zeilinger, A., “Experimental quantum teleportation”, *Nature*, [**390**]{}, 6660, 1997. Adesso, G. and Illuminati, F., “Entanglement in continuous-variable systems: recent advances and current perspectives”, *Journal of Physics A: Mathematical and Theoretical*, [**40**]{}, 28, 2007. Horodecki, R., Horodecki, P., Horodecki, M. and Horodecki, K, “Quantum entanglement”, *Reviews of modern physics*, [**81**]{}, 2, 2009. Richens, J. G., Selby, J. H. and Al-Safi, S. W., “Entanglement is necessary for emergent classicality in all physical theories”, *Physical review letters*, [**119**]{}, 8, 2017. Peres, A., “Separability criterion for density matrices”, *Physical Review Letters*, [**77**]{}, 8, 1996. Horodecki, P., *Separability criterion and inseparable mixed states with positive partial transposition*, Physics Letters A, [**232**]{}, 5, 1997. Simon, R., “Peres-Horodecki separability criterion for continuous variable systems”, *Physical Review Letters*, [**84**]{}, 12, 2000. Duan, L., Giedke, G., Cirac, J. I. and Zoller, P., *Inseparability criterion for continuous variable systems*, Physical Review Letters, [**84**]{}, 12, 2000. Werner, R. F. and Wolf, M. M., *Bound entangled Gaussian states*, Physical review letters, [**86**]{}, 16, 2001. Giedke, G., Kraus, B., Lewenstein, M. and Cirac, J. I., *Separability properties of three-mode Gaussian states*, Physical Review A, [**64**]{}, 5, 2001. Braunstein, S. L. and Van Loock, P., “Quantum information with continuous variables”, *Reviews of Modern Physics*, [**77**]{}, 2, 2005. Simon, R. Sudarshan, E. C. G. and Mukunda, N., “Gaussian pure states in quantum mechanics and the symplectic group”, *Physical Review A*, [**37**]{}, 8, 1988. Walls, D. F. and Milburn, G. J., “Quantum optics”, 2007, *Springer Science & Business Media*. Adesso, G., Ragy, S. and Lee, A. R., “Continuous variable quantum information: Gaussian states and beyond”, *Open Systems & Information Dynamics*, [**21**]{}, 2014, World Scientific. Ma, X. and Rhodes, W., “Multimode squeeze operators and squeezed states”, *Physical Review A*, [**41**]{}, 9, 1990. Schnabel, R., “Squeezed states of light and their applications in laser interferometers”, *Physics Reports*, [**684**]{}, 2017, Elsevier. Pirandola, S., Serafini, A. and Lloyd, S., “Correlation matrices of two-mode bosonic systems”, *Physical Review A*, [**79**]{}, 2009. Weedbrook, C. et. al., “Gaussian quantum information”, *Reviews of Modern Physics*, [**84**]{}, 2, 2012. Arvind, B. Dutta, N. Mukunda and R. Simon, “The Real symplectic groups in quantum mechanics and optics,” Pramana [**45**]{}, 471 (1995). Moshinsky, M. and Quesne, C., “Linear canonical transformations and their unitary representations”, *Journal of Mathematical Physics*, [**12**]{}, 8, 1971. Torre, A., “Linear ray and wave optics in phase space: bridging ray and wave optics via the Wigner phase-space picture”, 2005, Elsevier. Wolf, K., “Development of linear canonical transforms: a historical sketch”, in *Linear Canonical Transforms*, 3-28, 2016, Springer. De Gosson, M. A. *Symplectic methods in harmonic analysis and in mathematical physics*, [**7**]{}, 2011, [Springer Science & Business Media]{}. P. Hariharan, “The geometric phase”, *Progress in Optics*, [**48**]{}, 2005. D. Chruscinski and A. Jamiolkowski, *Geometric phases in classical and quantum mechanics*, [**36**]{}, 2012, Springer Science & Business Media. S. Sachdev, *[Quantum phase transitions]{}, *Handbook of Magnetism and Advanced Magnetic Materials*, 2007, Wiley Online Library.* Daniel C. Cabra, Andreas Honecker, and P. Pujol, *Modern theories of many-particle systems in condensed matter physics*, [**843**]{}, 2012, Springer Science & Business Media. Chaturvedi, S., Sriram, M. S. and Srinivasan, V., “Berry’s phase for coherent states”, *Journal of Physics A: Mathematical and General*, [**20**]{}, L1071, 1987. R. Y. Chiao and T. F. Jordan, “Lorentz-group Berry phases in squeezed light”, *Physics Letters A*, [**132**]{}, 77, 1988. R. Simon and N. Mukunda, “Bargmann invariant and the geometry of the Güoy effect”, *Physical review letters*, [**70**]{}, 7, 1993. P. G. Kwiat and R. Y. Chiao, “Observation of a nonclassical Berry?s phase for the photon”, *Physical review letters*, [**66**]{}, 5, 1991. Mukunda, N. and Simon, R., “Quantum kinematic approach to the geometric phase. I. General formalism”, *Annals of Physics*, [**228**]{}, 2, 1993. J. Brendel, W. Dultz and W. Martienssen, “Geometric phases in two-photon interference experiments”, *Physical Review A*, [**52**]{}, 4, 1995. D. V. Strekalov and Y. H. Shih, “Two-photon geometrical phase”, *Physical Review A*, [**56**]{}, 4, 1997. S. Seshadri, S. Lakshmibala and V. Balakrishnan, “Geometric phases for generalized squeezed coherent states”, *Physical Review A*, [**55**]{}, 2, 1997. I. Mendas, “Pancharatnam phase for ordinary and generalized squeezed states”, *Physical Review A*, [**55**]{}, 2, 1997. I. Fuentes-Guridi, S. Bose, and V. Vedral, “Proposal for measurement of harmonic oscillator Berry phase in ion traps”, *Physical review letters*, [**85**]{}, 24, 2000. A. S. M. de Castro and V. V. Dodonov “Squeezing exchange and entanglement between resonantly coupled modes”, *Journal of Russian Laser Research*,[**23**]{}, 2, 2002. Hall, M., “The theory of groups”, 2018, *Courier Dover Publications* N. Mukunda, S. Chaturvedi and R. Simon, “Classical light beams and geometric phases”, *JOSA A*, [**31**]{}, 6, 2014, *[Optical Society of America]{}.*
{ "pile_set_name": "ArXiv" }
--- abstract: 'We study static and transport properties of Skyrmions living within a finite spatial volume in a flat (3+1)-dimensional spacetime. In particular, we derive an explicit analytic expression for the compression modulus corresponding to these Skyrmions living within a finite box and we show that such expression can produce a reasonable value. The gauged version of these solitons can be also considered. It is possible to analyze the order of magnitude of the contributions to the electrons conductivity associated to the interactions with this Baryonic environment. The typical order of magnitude for these contributions to conductivity can be compared with the experimental values of the conductivity of layers of Baryons.' author: - | F. Canfora$^{1}$, N. Dimakis$^{2}$, A. Paliathanasis$^{3,4}$\ $^{1}$*Centro de Estudios Científicos (CECS), Casilla 1469, Valdivia, Chile.*\ $^{2}$*Center for Theoretical Physics, College of Physical Science and Technology,*\ *Sichuan University, Chengdu 610065, China.*\ $^{3}$*Instituto de Ciencias Físicas y Matemáticas, Universidad Austral de Chile, Valdivia, Chile.*\ $^{4}$*Institute of Systems Science, Durban University of Technology,*\ *PO Box 1334, Durban 4000, Republic of South Africa.*\ [[email protected], [email protected], [email protected]]{} title: 'Analytic Studies of Static and Transport Properties of (Gauged) Skyrmions' --- Introduction ============ The appearance of Skyrme theory [@skyrme] disclosed very neatly the fundamental role of topology in high energy physics (see for instance [nuc0,nuc1,nuc2,nuc3,nuc4,nuc5]{}). First of all, the low energy QCD is very well described by the Skyrme theory [@witten0]. Secondly, the solitons of this Bosonic theory (*Skyrmions*) describe Baryons. Thirdly, the Baryon charge is the winding number of the configuration (see [witten0,finkrub,manton,skyrev1,giulini,bala0,ANW,guada]{} and references therein). These arguments are more than enough to justify a profound analysis of the Skyrme model. Indeed, extensive studies of the latter can be found in literature (as the previous references clearly show). Not surprisingly[^1], the Skyrme field equations are a very hard nut to crack and, until very recently no analytic solution was available. Nevertheless, many numerical studies have shown that the Skyrme model provides results in good agreement with experiments. Despite the success of the model and the existence of several solutions among different contexts, the analysis of their phenomenological aspects seldom can be carried out in an analytic manner. For an analytic solution and a relevant study in compact manifolds see [@newref1]. The gauged Skyrme model (which describes the coupling of a $U(1)$ gauge field with the Skyrme theory) has also very important applications in the analysis of electromagnetic properties of Baryons, in the decay of nuclei in presence of defects (see [witten0,Witten,gipson,goldstone,dhoker,rubakov]{} and references therein). Obviously, from the point of view of constructing analytic solutions, the $U(1)$ gauged Skyrme model is even worse than the original Skyrme theory. Until very recently, no explicit topologically non-trivial solution was available. Thus, topological configurations of this theory have been deeply analyzed numerically (see [@gaugesky1; @gaugesky2] and references therein). Here we list three relevant problems in the applications of (gauged) Skyrme theory to high energy phenomenology which will be the focus of the present paper. **1)** *Finite density effects and the compression modulus*: Finite density effects (and, in general, the phase diagrams) in the Skyrme model have been historically a very difficult topic to analyze with analytic methods. The lack of explicit solutions with topological charge living within a finite flat box with the spherical Skyrme ansatz is the origin of the problem. Some numerical results with the use of the spherical Skyrme ansatz are presented in [klebanov,chemical1,chemical2,chemical3,chemical4]{} and references therein. Due to the fact that both finite volume effects and isospin chemical potential break spherical symmetry it is extremely difficult to improve the pioneering results in [@klebanov; @chemical1; @chemical2; @chemical3; @chemical4] without changing the original Skyrme ansatz. The main problem in this group is certainly the *compression modulus* [probcompression1,probcompression2,probcompression3]{} (to be defined precisely in the next section) which, roughly speaking, has to do with the derivative of the total energy of the Skyrmions with respect to the volume. The experimental value is different from the value derived using the original spherical hedgehog ansatz. The usual way to compute the compression modulus is to assume the Derrick rescaling for the reaction of nuclear matter to the action of external pressure (see the detailed discussion in [@Adam]). The resulting value is higher than the experimental value[^2]. A closely related technical difficulty is that, if one uses the original hedgehog ansatz for the Skyrmion, it is very unclear even *how to define* the compression modulus since the original Skyrme ansatz describes a spherical Skyrmion living within an infinite volume so that to compute the derivatives of the energy with respect to the volume becomes a subtle question. The best way out of this difficulty would be, of course, to have a consistent ansatz for a Skyrmion living within a finite volume. Relevant numerical results in the literature on that problem are presented in [@ref1; @ref2; @ref3; @ref4] where non-spherical ansätze have been considered. **2)** *Existence of Skyrmion-antiSkyrmion bound states/resonances*: multi-Skyrmionic bound states of Baryon charge higher than 1 are known to exist and they have been successfully constructed numerically (see, for instance, [@manton] and references therein). However, until very recently, the problem of the existence of Skyrmion-antiSkyrmion bound states and resonances did not possess the place it deserved in the literature on the Skyrme model and despite its importance. We can refer to an early work on the subject in [@newref2]. Here we shall study analytic results over the properties of such configurations. Experimentally,  Baryon-antiBaryon bound states and resonances do exist [@expbab1; @expbab2; @expbab3]: these should correspond to Skyrmion-antiSkyrmion bound states. Such bound states are very difficult to find since the corresponding classical solutions are not static. Indeed, at a semi-classical level, Skyrmion-antiSkyrmion bound states should look like time-periodic solutions in which a Skyrmion and an antiSkyrmion moves periodically around the center of mass of the system.  These kinds of time-dependent configurations are difficult to analyze even numerically. **3)** *Conductivities*: the analysis of electrons transport through gauged Skyrmions is a very interesting open issue. At semi-classical level, one should solve the Dirac equation for the electron in the background of the gauged Skyrmion and, from the solution of the Dirac equation, one could compute the conductivity. It would be especially interesting to be able to describe complex structures assembled from neutrons and protons interacting with electromagnetic fields (such as slabs of Baryons interacting with the corresponding Maxwell field). In nuclear physics and astrophysics these structures are called *nuclear pasta* and they are very relevant in a huge variety of phenomena (see, for instance, [@nuclearpasta1; @nuclearpasta2; @nuclearpasta3; @nuclearpasta4] and references therein). On the other hand, there are very few “first principles" computations of the transport properties of these complex structures (see [@conductivity] and references therein). At a first glance, one could think that this kind of complex structure is beyond the reach of the gauged Skyrme model. In order to achieve a deeper understanding of the above open issues, it is mandatory to be able to construct analytic examples of gauged multi-Skyrmionic configurations. In [canfora2,canfora3,canfora4,canfora4.5,yang1,canfora6,canfora6.5,cantalla4,cantalla5]{} a strategy has been developed to generalize the usual spherical hedgehog ansatz to situations without spherical symmetry both in Skyrme and Yang-Mills theories (see [@canYM1; @canYM2; @canYM3] and references therein). Such a framework also allows to analyze configurations living within a finite region of space. As far as the three open issues described above are concerned, this tool (which will be called here “generalized hedgehog ansatz") gave rise to the first derivation not only of the critical isospin chemical potential beyond which the Skyrmion living in the box ceases to exist, but also of the first explicit Skyrmion-antiSkyrmion bound states. Thus, this approach appears to be suitable to deal with the problems mentioned previously. Interestingly enough, the generalized hedgehog ansatz can be adapted to the $U(1)$ gauged Skyrme model [@Fab1; @gaugsk]: it allowed the construction of two types of gauged solitons. Firstly, gauged Skyrmions living within a finite volume. Secondly, smooth solutions of the $U(1)$ gauged Skyrme model whose periodic time-dependence is protected by a topological conservation law (as they cannot be deformed to static solutions). Here we demonstrate that by using this strategy it is possible to derive an explicit expression of the compression modulus. The transport properties of these gauged Skyrmions can also be analyzed. In this work we also present a simple estimate of the order of magnitude of the correction to the electron conductivities due to the interactions of the electrons with the baryonic environment. As far as transport properties are concerned, we will work at the level of approximation in which the electrons perceive the gauged Skyrmions as a classical background. Large **N** arguments strongly suggest that this is a very good approximation[^3] (see for a detailed review chapter 4 and, in particular, section 4.2 of the classic reference [@skyrev0]). This paper is organized as follows: in the second section the action for the gauged Skyrme model and our notations will be introduced. In the third section, the method to deal with Skyrmions at finite density will be described: as an application, a closed formula for the compression modulus of Skyrmions living within a cube will be derived. In the fourth section, the gauged Skyrmions at finite density will be considered. In the fifth section, the transport properties associated to electrons propagating in the Baryonic environment corresponding to the finite-density Skyrmions are analyzed. In section \[conclusions\], we draw some concluding ideas. The $U(1)$ Gauged Skyrme Model {#model} ============================== We consider the $U(1)$ gauged Skyrme model in four dimensions with global $SU(2)$ isospin internal symmetry and we will follow closely the conventions of [@Fab1; @gaugsk]. The action of the system is $$\begin{aligned} S& =\int d^{4}x\sqrt{-g}\left[ \frac{K}{2}\left( \frac{1}{2}\mathrm{Tr}\left( R^{\mu }R_{\mu }\right) +\frac{\lambda }{16}\mathrm{Tr}\left( G_{\mu \nu }G^{\mu \nu }\right) \right) -\frac{1}{4}F_{\mu \nu }F^{\mu \nu }\right] \ , \label{sky1} \\ R_{\mu }& =U^{-1}D_{\mu }U\ ,\ \ G_{\mu \nu }=\left[ R_{\mu },R_{\nu }\right] \ ,\ D_{\mu }=\nabla _{\mu }+\kappa A_{\mu }\left[ t_{3},\ .\ \right] \ , \label{sky2} \\ U& \in SU(2)\ ,\ \ R_{\mu }=R_{\mu }^{j}t_{j}\ ,\ \ t_{j}=\mathbbmtt{i}\sigma _{j}\ , \label{sky2.5}\end{aligned}$$where $\sqrt{-g}$ is the (square root of minus) the determinant of the metric, $F_{\mu \nu }=\partial _{\mu }A_{\nu }-\partial _{\nu }A_{\mu }$ is the electromagnetic field strength, $\nabla _{\mu }$ is the partial derivative, the positive parameters $K$ and $\lambda $ are fixed experimentally, $\kappa $ the coupling for the $U(1)$ field and $\sigma _{j}$ are the Pauli matrices. In our conventions $c=\hbar =\mu _{0}=1$, the space-time signature is $(-,+,+,+)$ and Greek indices run over space-time. The stress-energy tensor is $$T_{\mu \nu }=-\frac{K}{2}\mathrm{Tr}\left[ R_{\mu }R_{\nu }-\frac{1}{2}g_{\mu \nu }R^{\alpha }R_{\alpha }\right. \,+\left. \frac{\lambda }{4}\left( g^{\alpha \beta }G_{\mu \alpha }G_{\nu \beta }-\frac{g_{\mu \nu }}{4}G_{\sigma \rho }G^{\sigma \rho }\right) \right] +\bar{T}_{\mu \nu }, \notag \label{timunu1}$$with $$\bar{T}_{\mu \nu }=F_{\mu \alpha }F_{\nu }^{\;\alpha }-\frac{1}{4}F_{\alpha \beta }F^{\alpha \beta }g_{\mu \nu }.$$The field equations are $$D^{\mu }\left( R_{\mu }+\frac{\lambda }{4}\left[ R^{\nu },G_{\mu \nu }\right] \right) =0\ , \label{nonlinearsigma1}$$$$\nabla _{\mu }F^{\mu \nu }=J^{\nu }\ , \label{maxwellskyrme1}$$where $J^{\nu }$ is the variation of the Skyrme action (the first two terms in Eq. (\[sky1\])) with respect to $A_{\nu }$ $$J^{\mu }=\frac{\kappa K}{2}Tr\left[ \widehat{O}R^{\mu }+\frac{\lambda }{4}\widehat{O}\left[ R_{\nu },G^{\mu \nu }\right] \right] \ , \label{current}$$where$$\widehat{O}=U^{-1}t_{3}U-t_{3}\ .$$ In the following sections, *gauged Skyrmions* and *gauged time-crystals* will be terms describing to the two different kinds of gauged topological solitons appearing as solutions of the coupled system expressed by Eqs. (\[nonlinearsigma1\]) and (\[maxwellskyrme1\]). The aim of the present work is to show that the Skyrme model and its gauged version are able to give good predictions for important quantities such as the compression modulus and the conductivity. Topological charge ------------------ The proper way to define the topological charge in the presence of a minimal coupling with a $U(1)$ gauge potential has been constructed in [@Witten] (see also the pedagogical analysis in [@gaugesky1]):$$\begin{split} W=& \frac{1}{24\pi ^{2}}\int_{\Sigma }\epsilon ^{ijk}Tr\left\{ \left( U^{-1}\partial _{i}U\right) \left( U^{-1}\partial _{j}U\right) \left( U^{-1}\partial _{k}U\right) \right. - \\ & \left. \partial _{i}\left[ 3\kappa A_{j}t_{3}\left( U^{-1}\partial _{k}U+\partial _{k}UU^{-1}\right) \right] \right\} . \end{split} \label{new4.1}$$ In the literature one usually only considers situations where $\Sigma $ is a space-like three-dimensional hypersurface. In these situations $W$ is the Baryon charge. In fact it has been recently shown [@Fab1] [@gaugsk] that it is very interesting to also consider cases in which $\Sigma $ is time-like or light-like. Indeed, (whether $\Sigma $ is light-like, time-like or space-like) configurations with $W\neq 0$ cannot decay into the trivial vacuum $U=\mathbb{\mathbf{I}}$. Hence, if one is able to construct configurations such that $W\neq 0$ along a time-like $\Sigma $, then the corresponding gauged soliton possesses a topologically protected time-dependence as it cannot be continuously deformed into static solutions (since all the static solutions have $W=0$ along a time-like $\Sigma $). The natural name for these solitons is “(gauged) time-crystals" [Fab1,gaugsk]{}. We can adopt the standard parametrization of the $SU(2)$-valued scalar $U(x^{\mu }) $ $$U^{\pm 1}(x^{\mu })=Y^{0}(x^{\mu })\mathbb{\mathbf{I}}\pm Y^{i}(x^{\mu })t_{i}\ ,\ \ \left( Y^{0}\right) ^{2}+Y^{i}Y_{i}=1\,, \label{standnorm}$$where $\mathbb{\mathbf{I}}$ is the $2\times 2$ identity and $$\begin{aligned} Y^{0}& =\cos C\ ,\ Y^{i}=n^{i}\cdot \sin C\ , \label{pions1} \\ n^{1}& =\sin F\sin G\ ,\ \ n^{2}=\sin F\cos G\ ,\ \ n^{3}=\cos F\ . \label{pions2}\end{aligned}$$ with the help of which the standard baryon density (in the absence of a $U(1) $ field) reads $\rho _{B}=12\sin ^{2}C\sin F\ dC\wedge dF\wedge dG$. If we want a non-vanishing topological charge in this setting we have to demand $dC\wedge dF\wedge dG\neq 0$. Skyrmions at finite volume ========================== In the present section, the Skyrmions living within a finite flat box constructed in [@Fab1] will be slightly generalized. These explicit Skyrmionic configurations allow the explicit computations of the total energy of the system and, in particular, of its dependence on the Baryon charge and on the volume. Hence, among other things, one can arrive at a well-defined closed formula for the compression modulus. The following anstatz for the representation of the $SU(2)$ group is the starting point of the analysis $$G=\frac{q\phi - p\gamma}{2},\; \tan F=\frac{\tan H}{\sin A}, \; \tan C= \tan A \sqrt{1+\tan ^{2}F}\ , \label{pions2.25}$$where $$A=\frac{p\gamma +q\phi }{2\,}\ ,\ \ H=H\left( r,z\right) \ ,\ \ p,q\in \mathbb{N} \ . \label{pions2.26}$$ Moreover, it can be verified directly that, the topological density $\rho _{B}$ is non-vanishing. From the standard parametrization of $SU(2)$ [Shnir]{} it follows that $$0\leq \gamma \leq 4\pi ,\quad 0\leq \phi \leq 2\pi \ , \label{domain}$$while the boundary condition for $H$ will be discussed below; in any case, its range is in the segment $H \in [0,\frac{\pi}{2}]$, while for $r$ we assume $0\leq r\leq 2 \pi$. With the parametrization introduced by and the $SU(2)$ field assumes the form $$U = \pm \begin{pmatrix} \cos (H) e^{\frac{1}{2} i (p \gamma+q \phi )} & \sin (H) e^{\frac{1}{2} i (p \gamma -q \phi )} \\ -\sin (H) e^{-\frac{1}{2} i (p \gamma -q \phi )} & \cos (H) e^{-\frac{1}{2} i (p\gamma +q \phi )}\end{pmatrix}.$$ Hereafter, we just consider the plus expression for $U$ throughout all the range of the variables $\gamma$ and $\phi$, which makes it a continuous function of the latter. Skyrmions in a rectangular cuboid --------------------------------- We can extend the results presented in [@Fab1] by considering a cuboid with three different sizes along the three axis instead of a cube. Thus, we will use three - different in principle - fundamental lengths characterizing each direction, $l_{1}$, $l_{2}$ and $l_{3}$, inside the metric. The corresponding line element is $$ds^{2}=-dz^{2}+l_{1}^{2}dr^{2}+l_{2}^{2}d\gamma ^{2}+l_{3}^{2}d\phi ^{2}\ . \label{line3l}$$The profile function that we consider depends only on one variable[^4], $H=H(r)$. We note that in this section we do not take into account the effects of an electromagnetic field, hence we have $A_{\mu }=0$ in the relations of the previous sections. Under the aforementioned conditions the profile equation reduces to $$\label{profnoA} H^{\prime \prime }= \frac{\lambda l_1^2 p^2 q^2}{4 \left(l_2^2 \left(4 l_3^2+\lambda q^2\right)+\lambda l_3^2 p^2\right)} \sin (4 H).$$ It is impressive that such a system, in flat space, can lead to an integrable equation for the profile. This is owed to the existence of a first integral of that is given by $$\label{intofmol3} (H^{\prime})^2 \left(l_2^2 \left(4 l_3^2+\lambda q^2\right)+\lambda l_3^2 p^2\right)+\frac{\lambda l_1^2 p^2 q^2}{8} \cos (4 H) = I_0 .$$ The above relation can be written as $$\label{firstint2} (\tilde{H}^{\prime})^2 - k \sin (\tilde{H})^2 = \tilde{I}_0,$$ where $$\tilde{H}= 2 H, \quad k = \frac{\lambda l_1^2 p^2 q^2}{l_2^2 \left(4 l_3^2+\lambda q^2\right)+\lambda l_3^2 p^2}, \quad \tilde{I}_0= \frac{8 I_0-\lambda l_1^2 p^2 q^2}{8 l_2^2 l_3^2+2 \lambda l_2^2 q^2+2 \lambda l_3^2 p^2}.$$ Subsequently, we can bring into the form $$\label{firstint3} \frac{d \tilde{H}}{d r} = \pm \sqrt{\tilde{I}_0} \left(1 - \tilde{k}(\sin \tilde{H})^2 \right)^{\frac{1}{2}}$$ where we have set $\tilde{k} = - k/\tilde{I}_0$. The last expression leads to $$\label{finalsolH} \sqrt{\tilde{I}_0} \int_0^r d \bar{r} = \pm \int_0^{\tilde{H}} \left(1 - \tilde{k}(\sin \bar{H})^2 \right)^{-\frac{1}{2}} d \bar{H},$$ where we have introduced the bars in order to distinguish the variables that are integrated from the $r$ and $\tilde{H}(r)$ which are the boundaries of the two integrals. Of course we consider $\tilde{I}_0 > 0$. As a starting point for the integration we take $r=0$, $\tilde{H}(0)=0=H(0)$, although we could also set $r=0$, $\tilde{H}=\pi$ ($H(0)=\frac{\pi}{2}$). The difference between the two boundary choices is just in the sign of the topological charge. These boundary values, for $H$ and those that we have seen in for $\gamma$ and $\phi$ lead to a topological charge $W= p q$ in (for $A_\mu =0$). In the right hand side of we recognize the incomplete elliptic integral defined as $$F(\tilde{H}|\tilde{k}) = \int_0^{\tilde{H}} \left(1 - \tilde{k}(\sin \bar{H})^2 \right)^{-\frac{1}{2}} d \bar{H} .$$ The solution to the differential equation is just the inverse of this function, which is called the Jacobi amplitude $\mathrm{am}=F^{-1}(\tilde{H}|\tilde{k})$. So, in terms of our original equation the solution reads $$\label{soltodifeq} H(r) = \pm \frac{1}{2} \mathrm{am}(\tilde{I}_0^{1/2} r|\tilde{k}).$$ Finally, by considering the positive branch, the value of the constant of integration $\tilde{I}_0$ is governed by the boundary condition $H(2\pi)=\frac{\pi}{2}$. In the special case when $l_1=l_2=l_3=l$ we obtain the particular case which was studied in [@Fab1]. Here, we give emphasis to this general case and, especially, we want to study the most energetically convenient configurations and the way in which they are affected by the anisotropy in the three spatial directions. In Fig. \[Fig0\] we see a schematic representation of the finite box we are considering for this Skyrmionic configuration with a baryon number $B =p q$. The physical configuration that we try to reproduce with this model is the structure of matter in nuclear pasta. The latter is a dense form of matter that is encountered inside the crusts of neutron stars. Thus, we make this crude" (but analytic in its results) model trying to imitate with these $p$ and $q$ Skyrmionic layers a particular form of this matter that is encountered in nature. The dimensions of the configuration are governed by the three numbers $l_{1}$, $l_{2}$ and $l_{3}$. Of course we do not expect the binding energies of such a configuration to be at the same level with those produced by the usual spherically symmetric ansatz. This is something that we examine thoroughly in the next section. ![The finite box of the Skyrmionic system.[]{data-label="Fig0"}](box.eps){width=".40\textwidth"} ### The energy function We proceed to study the energy function for the solution that we previously introduced. The constant of motion $I_0$ in can be expressed in terms of the other constants of the model if we consider the boundary values $H(0)=0$ and $H(2\pi)=\pi/2$. By solving with respect to $H^{\prime }$ and integrating the resulting relation with respect to $r$ we obtain $$\label{intofint} 2\sqrt{2} \int_{a}^{b} \left(\frac{l_2^2 \left(4 l_3^2+q^2\right)+l_3^2 p^2}{8 I_0-l_1^2 p^2 q^2 \cos (4 H)}\right)^{1/2} dH = \int_{0}^{2\pi} dr$$ which leads to $$\label{l1tox} l_1 = \frac{x \mathrm{K}\left(-x^2\right) \sqrt{l_2^2 \left(4 l_3^2+q^2\right)+l_3^2 p^2}}{\pi p q},$$ where $\mathrm{K}$ is the complete elliptic integral of the first kind and $x $ is related to $I_0$ through $$\label{I0tox} I_0 = \frac{l_1^2 p^2 q^2 \left(x^2+2\right)}{8 x^2}.$$ The pure time component of the energy momentum tensor in our case is $$T_{00} = \frac{K}{8 V^2} \left[\left(l_2^2 \left(4 l_3^2+\lambda q^2\right)+\lambda l_3^2 p^2\right) H^{\prime 2 }+ \frac{\lambda l_1^2 p^2 q^2}{4}\sin ^2(2 H) + V^2 \left(\frac{p^2}{l_2^2}+\frac{q^2}{l_3^2}\right)\right].$$ As a result we can calculate the energy from the expression $$E= \int_{\Sigma} \sqrt{-{}^{(3)}g} T_{00}d^3 x = 8\pi^2 V \int_0^{\frac{\pi}{2}} \frac{T_{00}}{H^{\prime }} dH .$$ We can write the integrand as a pure function of $H$ with the help of and obtain - in principle - the energy as a function of the $l_i$’ s, $p$ and $q$. However, due to the fact that relation cannot be straightforwardly inverted so as to substitute $I_0$ as a function of $l_1$ (through and ) we choose to express the energy function in terms of $x$ instead of $l_1$. In what follows, we assume the values $K=2$ and $\lambda=1$ for the coupling constants [@skyrev1], so that lengths are measured in fm and the energy in MeV. In this manner we get $$\label{energyfull} E(x,l_2,l_3,p,q) = \frac{\pi ^2 p q \sqrt{l_2^2 \left(4 l_3^2+q^2\right)+l_3^2 p^2}}{l_2 l_3} \frac{K(-x^2) \left(\frac{4 l_2^2 x^2 K(-x^2)}{p^2}-\frac{\mathrm{K}(-x^2) \left(q^2-4 l_3^2 x^2\right)}{q^2}+2 \mathcal{E}(-x^2)\right)}{x |K(-x^2)|},$$ where $\mathcal{E}$ is the complete elliptic integral of the second kind. The $x$, as we discussed, is linked - with the help of the boundary conditions of the problem - through to $l_1$. If we fix all variables apart from $x$ and plot the energy as a function of the latter we get what we see in Fig. \[Fig1\]. In this graph, we observe that the minimum of the energy is “moving" to smaller values of $x$ as the box is being enlarged in the two directions of $l_2$ and $l_3$. However, we have to keep in mind that the other of the lengths, namely $l_1$, depends also on the values of $l_2$ and $l_3$ through . For the particular set of values used in the figure we can see that as $l_2$ and $l_3$ rise, $l_1$ is also relocated to larger values. In the next section we study more thoroughly the function $E(x,l_2,l_3,p,q)$ and its derivatives near the values that correspond to the most energetically convenient configurations. ![The plots of $E(x)$ (in MeV) for three sets of values: (a) $p=q=3$, $l_2=l_3=1$ fm (dashed line), (b) $p=q=3$, $l_2=l_3=2$ fm (dotted line) and (c) $p=q=3$, $l_2=l_3=3$ fm (continuous line). The minimum of the energy corresponds to $l_1=0.227$ fm, $l_1=0.323$ fm and $l_1=0.42$ fm respectively. []{data-label="Fig1"}](Energy_of_x.eps){width=".40\textwidth"} ### The energy as a function of the three $l_i$’s Let us see how the energy behaves in terms of the three fundamental lengths $l_1$, $l_2$ and $l_3$ under the condition that we fix $p$ and $q$ to specific values. In the table \[tab1\] we can observe the location of the minimum of the energy for specific values of $p$ and $q$. $E_{min}$ (MeV) $p$ $q$ $l_1$ (fm) $l_2$ (fm) $l_3$ (fm) ----------------- ----- ----- ------------ ------------ ------------ 167 1 1 0.251 0.413 0.413 334 1 2 0.251 0.413 0.826 669 2 2 0.251 0.826 0.826 835638 100 50 0.251 41.306 20.653 835638 50 100 0.251 20.653 41.306 : Minimum of the energy for values of $p$ and $q$.[]{data-label="tab1"} First, we have to note that the interchange of $p$ and $q$ makes no significant difference, so weather you take $p=100$ and $q=50$ or $p=50$ and $q=100$, the only thing that happens is that the values of the corresponding lengths $l_{2}$ and $l_{3}$ are also interchanged. However, the arithmetic value that the energy assumes remains the same. Another thing that we have to notice is that, if we calculate the percentage difference of the minimum of the energy from the topological bound $E_{0}=12\pi ^{2}|B|=12\pi ^{2}pq$; in all cases we get $\Delta (\%)=\frac{E-E_{0}}{E_{0}}(\%)=41.11\%$. Thus, we see that the minimum of the energy $E(l_{1},l_{2},l_{3})$ has a fixed deviation from the Bogomol’nyi bound irrespectively of the $p$, $q$ configuration. We also observe that this most energetically convenient situation arises when the box has convenient lengths. In particular we see that the relation $\frac{l_{2}}{l_{3}}=\frac{p}{q}$ is satisfied in all cases, while $l_{1}$ remains fixed in a single optimal" value. By comparing with the usual spherically symmetry Skyrmionic configuration in an infinite volume, this higher deviation from the Bogomol’nyi bound may be anticipated due to the “compression" of the system into a finite volume. It is also interesting to study the first derivatives of the energy with respect to the three lengths of the box. To this end, and since we have $E$ in terms of $x$ which also involves $l_{1}$, $l_{2}$ and $l_{3}$ we need to write $$\begin{split} dE(x,l_{2},l_{3})& =\frac{\partial E}{\partial x}dx+\frac{\partial E}{\partial l_{2}}dl_{2}+\frac{\partial E}{\partial l_{3}}dl_{3} \\ & =\frac{\partial E}{\partial x}\frac{\partial x}{\partial l_{1}}dl_{1}+\left( \frac{\partial E}{\partial x}\frac{\partial x}{\partial l_{2}}+\frac{\partial E}{\partial l_{2}}\right) dl_{2}+\left( \frac{\partial E}{\partial x}\frac{\partial x}{\partial l_{3}}+\frac{\partial E}{\partial l_{3}}\right) \\ & =d\tilde{E}(l_{1},l_{2},l_{3}). \end{split}$$In Fig. \[Fig2\] we can see the general behavior of three $\frac{\partial \tilde{E}}{\partial l_{i}}$ for fixed $l_{1}=0.251$ in terms of $l_{2}$ and $l_{3}$ near the values where the energy assumes its minimum. On the other hand, in Fig. \[Fig3\] we plot the derivatives of the energy with respect to $x$ after fixing $l_{2}$ and $l_{3}$ to their minimum value for various $p $, $q$ configurations. We can see that $\frac{\partial E}{\partial l_{2}}$ and $\frac{\partial E}{\partial l_{3}}$ are indistinguishable when $p=q$. On the other hand if $q>p$ the $\frac{\partial E}{\partial l_{3}}$ line runs closer to the vertical axis than $\frac{\partial E}{\partial l_{2}}$ and vice versa when $p>q$. Finally, before proceeding to study the energy as a function of $p$ and $q$, we give in Fig. \[Fig4\] its graph in terms of $l_{2}$ and $l_{3}$ when $l_{1}$ assumes the value that corresponds to the minimum of the energy. ![Plot of the energy $E$ in the $l_{2}-l_{3}$ plane when $l_{1}$ takes the value that corresponds to the minimum of $E$.[]{data-label="Fig4"}](El2l3.eps){width=".40\textwidth"} The energy of the symmetric configuration ----------------------------------------- Due to using in the previous section so as to write the energy as a function of $x$, $l_{2}$ and $l_{3}$, it is not straightforward from that expression to derive what happens in the case where one considers a symmetric box $l_{1}=l_{2}=l_{3}=l$. In this section we treat this situation from the very beginning by setting all fundamental lengths as equal in Eq. . We have to note that throughout this section we also make use of the system of units $K=2$, $\lambda =1$. The expression relative to , from the resulting integral of motion, leads to $$l=\frac{\sqrt{\pi ^{2}p^{2}q^{2}-x^{2}\mathrm{K}\left( -x^{2}\right) ^{2}\left( p^{2}+q^{2}\right) }}{2x\mathrm{K}\left( -x^{2}\right) }, \label{ltox}$$where $x$ is defined as in the previous section by relation , with $l_{1}=l$. By following the exact same steps as before we are led to the following expression for the energy $$E_{c}(x,p,q)=\frac{2\pi ^{3}\left( 2p^{2}q^{2}\mathrm{K}\left( -x^{2}\right) \mathcal{E}\left( -x^{2}\right) -\mathrm{K}\left( -x^{2}\right) ^{2}\left( p^{4}x^{2}+p^{2}q^{2}\left( 2x^{2}+1\right) +q^{4}x^{2}\right) +\pi ^{2}p^{2}q^{2}\left( p^{2}+q^{2}\right) \right) }{x^{2}\mathrm{K}\left( -x^{2}\right) \sqrt{\frac{\pi ^{2}p^{2}q^{2}}{x^{2}}-\mathrm{K}\left( -x^{2}\right) ^{2}\left( p^{2}+q^{2}\right) }}. \label{encube}$$ It is easy to note that the energy is symmetric under the mirror change $p\leftrightarrow q$. We verify that the for a bigger baryon number, the most optimal configuration corresponds also to a larger box. In Fig. \[Fig4new\] we can see the plot of the energy with respect to various configurations demonstrating the aforementioned fact. The second thing that we can note is that the deviation $\Delta= \frac{E-E_0}{E_0}$ from saturating the bound also increases for larger baryonic configurations. In table \[tab2\] we provide some basic examples. Surprisingly we can see that the configuration $p=q=2$ is slightly more convenient than the one corresponding to $p=2$, $q=1$. As long as we know, this is the only case where this is happening. In general it can be seen that the $p=q$ construction requires more energy than the $p$, $q-1$, with an exception in the $p=q=2$ case. $p$ $q$ $l$ (fm) $\Delta(\%)$ ----- ----- ---------- -------------- 1 1 0.322 53 2 1 0.369 105 3 1 0.385 177 2 2 0.463 104 3 2 0.505 138 3 3 0.571 148 : Deviation from the topological bound for several values of $p$ and $q$.[]{data-label="tab2"} ![Plot of the energy of the cubic configuration $E_c$ with respect to $x$. The dashed line corresponds to $p=q=1$, the dotted to $p=2$, $q=1$ and the continuous line to $p=3$, $q=1$. The minimum of the energy in terms of the size of the cube $l$ is: $l=0.322$, $l=0.369$ and $l=0.385$ respectively. []{data-label="Fig4new"}](Energy_cube.eps){width=".40\textwidth"} The compression modulus for the rectangular box ----------------------------------------------- From the technical point of view, it is worth to emphasize here that the very notion of compression modulus would require to put the Skyrmions within a finite flat box of volume $V$: then the compression modulus is related to the second derivative of the total energy of the system with respect to $V$. As it has been already mentioned, this requires to generalize the hedgehog ansatz to situations without spherical symmetry. On the other hand, if one insists in defining the compression modulus for the spherical hedgehog, it becomes a rather subtle issue (see the nice analysis in [@Adam]) how to define the derivative of the energy with respect to the volume. Here we are using the generalized hedgehog ansatz [@Fab1; @gaugsk] which is well suited to deal with situations without spherical symmetry. In this way we can analyze Skyrmions living within a region of flat space-time of finite spatial volume avoiding all the subtleties mentioned above. In particular, in the present case the “derivative with respect to the volume” means, literally, the derivative (of the total energy of the system) with respect to the spatial volume of the region in which the Skyrmions are living. As we obtained the general behavior of the three $\frac{\partial E}{\partial l_{i}}$ functions in the previous sub-sections, we are also able to derive an analytic expression of the compression modulus [@Brown; @Co] $$\mathcal{K}=\frac{9V}{B\beta }\approx 210\pm 30MeV$$where $\beta =-\frac{1}{V}\frac{\partial V}{\partial P}$ is the compressibility. By using $P=\frac{dE}{dV}$ we acquire $$\mathcal{K}=-\frac{9V^{2}}{B}\frac{d^{2}E}{dV^{2}}, \label{compmod}$$where $B$ is the baryon charge and $V$ the finite volume in which we confine the system; in our case this volume is $V=16\pi ^{3}l_{1}l_{2}l_{3}$. The difference in the sign of in comparison to other expressions in the literature [@Blaizot] is owed to the metric signature that we follow here and which affects the derivation of $E$ from $T_{00}$. In order to express the energy that we obtain from as a function of the volume, we introduce the following reparametrization of the $l_{i}$’s into three new variables $$l_{1}=c_{1}\left( \frac{V}{16\pi ^{3}}\right) ^{1/3},\quad l_{2}=c_{2}\left( \frac{V}{16\pi ^{3}}\right) ^{1/3},\quad \text{and}\quad l_{3}=\frac{1}{c_{1}c_{2}}\left( \frac{V}{16\pi ^{3}}\right) ^{1/3},$$so that $l_{1}l_{2}l_{3}=\frac{V}{16\pi ^{3}}$. We can substitute the above expressions into both and . By solving the first with respect to $V$ and substituting to the second we obtain the energy as a pure function of $x$ which is associated through with the volume $V$. We can thus calculate the first and second derivatives of the energy with respect to the volume by just taking $\frac{dE}{dV}=\left( \frac{dV}{dx}\right) ^{-1}\frac{dE}{dx}$ and $\frac{d^{2}E}{dV^{2}}=\left( \frac{dV}{dx}\right) ^{-1}\frac{d}{dx}\left[ \left( \frac{dV}{dx}\right) ^{-1}\frac{dE}{dx}\right] $. The first derivative of $E(V)$ with respect to the volume defines the pressure of the system, i.e. $P=\frac{dE}{dV}$. In Fig. \[Fig5\] we see the graphs of the pressure the compression modulus and the energy with respect to the volume for specific regions of the variable $V$. Due to the complicated nature of the relation between $x$ and $V$ it is not easy to put in this parametric plot the behavior of $P$ and $E$ near the region where $V\rightarrow 0$. However, one can calculate through the relations that as one shrinks the volume to zero, the pressure suddenly falls and changes sign becoming negative. The same happens to the compression modulus $\mathcal{K}$ as well, for even smaller values of $V$, while the energy remains positive for all $V$. Unfortunately the expressions are too cumbersome to present them analytically in this work, but the graphs in Fig. \[Fig5\] demonstrate the general behavior. In the case of a finite cube with $l_{1}=l_{2}=l_{3}$ the situation is a lot simpler as we can see in the following section. ### Compression modulus in the symmetric case The most natural case corresponds to choose $l_{1}=l_{2}=l_{3}=l$. In this way, we can derive a closed analytic formula for the compression modulus of the Skyrmions living within such a cuboid. To the best of our knowledge, this is the first case in which one can derive an analytic formula (Eqs. (\[Kbox\]) and (\[Vbox\]) below) for the compression modulus in a highly interacting theory such as the low energy limit of QCD. Indeed, by expressing the fundamental length as $l=\left( \frac{V}{16\pi ^{3}}\right) ^{1/3}$ we can easily use to relate the volume $V$ with the variable $x$ on which the energy depends . In this manner we can get an analytical expression for the compression modulus of the cube in terms of the variable $x$, which is $$\begin{split} \mathcal{K}(x)=& -\frac{36}{pq}\Bigg[\left( x^{2}+1\right) \mathrm{K}\left( -x^{2}\right) ^{3}\left( \pi ^{2}p^{2}q^{2}-x^{2}\mathrm{K}\left( -x^{2}\right) ^{2}\left( p^{2}+q^{2}\right) \right) ^{2} \\ & +x^{2}\mathrm{K}\left( -x^{2}\right) ^{3}\left( p^{2}+q^{2}\right) \mathcal{E}\left( -x^{2}\right) ^{2}\left( 5\pi ^{2}p^{2}q^{2}-x^{2}\mathrm{K}\left( -x^{2}\right) ^{2}\left( p^{2}+q^{2}\right) \right) \\ & +\pi ^{4}p^{2}q^{2}\mathcal{E}\left( -x^{2}\right) \left( \mathrm{K}\left( -x^{2}\right) ^{2}\left( x^{2}\left( p^{2}+q^{2}\right) ^{2}-2p^{2}q^{2}\right) -\pi ^{2}p^{2}q^{2}\left( p^{2}+q^{2}\right) \right) \Bigg]. \end{split} \label{Kbox}$$It can be shown that the parametric plots with respect to the volume which is $$V=2\pi ^{3}\frac{\left( \pi ^{2}p^{2}q^{2}-x^{2}\mathrm{K}\left( -x^{2}\right) ^{2}\left( p^{2}+q^{2}\right) \right) ^{3/2}}{x^{3}\mathrm{K}\left( -x^{2}\right) ^{3}} \label{Vbox}$$lead to the same behavior for the pressure, the energy and the compression modulus that has being derived in the previous section. For various values of $p$ and $q$ the behavior of the before mentioned quantities is described by the same graphs as given in Fig \[Fig5\]. A baryon density ($n=\frac{B}{V}$) of $0.04$ fm$^{-3}$ $\lesssim n\lesssim 0.07$ fm$^{-3}$ is assumed [@Caplan] to be appropriate for characterizing nuclear pasta and in particular lasagna. Within this range densities we can see that with expressions and we can achieve a compression modulus around $\mathcal{K}\sim 230$MeV (which is quite reasonable [@Adam; @Dutra]). For instance in table \[tab3\] one can observe various examples of configurations involving baryon densities $n$ and the corresponding baryon numbers $B$, whose compression modulus - as calculated with the help of - is $\mathcal{K}\sim 230$MeV. In all cases presented in the table we have considered $p=q$, thus $B=p^{2}$. $B$ 144 196 225 324 ----------------- ------- ------- ------- ------- $n$ (fm$^{-3}$) 0.044 0.048 0.051 0.057 : Examples of configurations corresponding to a compression modulus $\mathcal{K}\sim 230$MeV.[]{data-label="tab3"} Gauged solitons =============== Here we will shortly describe (a slight generalization of) the gauged solitons constructed in [@gaugsk]. Gauged Skyrmions ---------------- As in [@gaugsk], we introduce an electromagnetic potential of the form $$\label{empot} A_{\mu }=(b_{1}(r),0,b_{2}(r),b_{3}(r)),$$to be coupled to the multi-Skyrmionic system under consideration. The Maxwell equations reduce to $$\label{Maxbs} b_{i}^{\prime \prime }=\kappa^2 M_{ij}b_{j}+ \kappa N_{i}$$with the nonzero components of $M$ and $N$ being $$\begin{aligned} M_{11}& =-K\sin ^{2}(H)\left[ l_{1}^{2}\left( 4+\lambda \left( \frac{p^{2}}{l_{2}^{2}}+\frac{q^{2}}{l_{3}^{2}}\right) \cos ^{2}(H)\right) +4\lambda H^{\prime 2}\right] \\ M_{23}& =\frac{K\lambda l_{1}^{2}pq}{4l_{3}^{2}}\sin ^{2}(2H) \\ M_{32}& =\frac{l_{3}^{2}}{l_{2}^{2}}M_{23} \\ M_{22}& =M_{11}+\frac{p}{q}M_{32} \\ M_{22}& =M_{11}+\frac{q}{p}M_{23} \\ N_{2}& =\frac{p}{4}M_{11}+\frac{1}{4}\left( \frac{l_{3}^{2}p^{2}}{l_{2}^{2}q}-q\right) M_{23} \\ N_{3}& =-\frac{q}{4}M_{11}-\frac{1}{4}\left( \frac{l_{2}^{2}q^{2}}{l_{3}^{2}p}-p\right) M_{32}.\end{aligned}$$A direct computation shows that, using the line element in Eq. (\[line3l\]), the three coupled gauged Skyrme equations (namely, $\mathit{E}^{j}=0$, $j=1$, $2$, $3$) in Eq. (\[nonlinearsigma1\]) $$D^{\mu }\left( R_{\mu }+\frac{\lambda }{4}\left[ R^{\nu },G_{\mu \nu }\right] \right) =\mathit{E}^{j}t_{j}=0$$reduce to only one Skyrme field equation (since the third Skyrme equation is identically satisfied while the first and the second are proportional):$$\begin{aligned} \mathit{E}^{3} &=&0\ , \\ \mathit{E}^{1} &=&I_{1}P\left[ H\right] \ ,\ \mathit{E}^{2}=I_{2}P\left[ H\right] \ ,\ \ I_{1}\neq 0\ ,\ I_{2}\neq 0\ ,\end{aligned}$$where $I_{j}$ are real and non-vanishing. Thus, the Skyrme field equations reduce to $P\left[ H\right] =0$ namely $$\begin{split} & 4\left[ X\sin ^{2}(H)-\lambda \left( l_{2}^{2}q^{2}+l_{3}^{2}p^{2}\right) -4l_{2}^{2}l_{3}^{2}\right] H^{\prime \prime }+2X\sin (2H)H^{\prime 2}+4\sin ^{2}(H)X^{\prime }H^{\prime } \\ & +\Big[\lambda \kappa \left( l_{3}^{2}pb_{2}+l_{2}^{2}qb_{3}\right) \left( -\frac{4l_{1}^{2}p}{l_{2}^{2}}\kappa b_{2}-\frac{4l_{1}^{2}q}{l_{3}^{2}}\kappa b_{3}+2l_{1}^{2}\left( \frac{q^{2}}{l_{3}^{2}}-\frac{p^{2}}{l_{2}^{2}}\right) \right) -\frac{1}{4}l_{1}^{2}X\left( \frac{p^{2}}{l_{2}^{2}}+\frac{q^{2}}{l_{3}^{2}}\right) \\ & +\lambda l_{1}^{2}p^{2}q^{2}\Big]\sin (4H)-\frac{2l_{1}^{2}}{\lambda }X\sin (2H)=P\left[ H\right] =0\ , \end{split} \label{profl3}$$where $$X(r)=8\lambda \kappa \left( 2l_{2}^{2}l_{3}^{2}\kappa b_{1}^{2}-l_{3}^{2}b_{2}(2\kappa b_{2}+p)+l_{2}^{2}b_{3}(q-2\kappa b_{3})\right) .$$ Quite remarkably, if we demand that $$X(r)=\lambda \left( l_{2}^{2}q^{2}+l_{3}^{2}p^{2}\right) ,\quad b_{2}(r)=-\frac{l_{2}^{2}q}{l_{3}^{2}p}b_{3}+\frac{1}{\kappa }\left( \frac{l_{2}^{2}q^{2}}{4l_{3}^{2}p}-\frac{p}{4}\right) \ , \label{Xcon1}$$then the equation for the profile $H(r)$ can be solved explicitly. More importantly, the above algebraic conditions in Eq. (\[Xcon1\]) are consistent with the Maxwell equations written above. Indeed, if one plugs the two algebraic conditions in Eq. (\[Xcon1\]) into the three Maxwell equations one obtains a single Maxwell equation for $b_{3}(r)$: $$b_{3}^{\prime \prime }=\frac{\kappa K}{8l_{2}^{2}l_{3}^{2}}(q-4\kappa b_{3})\left[ 8\lambda l_{2}^{2}l_{3}^{2}H^{\prime 2}+l_{1}^{2}\left( \lambda \cos (2H)\left( l_{2}^{2}q^{2}+l_{3}^{2}p^{2}\right) +l_{2}^{2}\left( 8l_{3}^{2}+\lambda q^{2}\right) +\lambda l_{3}^{2}p^{2}\right) \right] \sin ^{2}(H)\ , \label{b3XYcon1}$$while for the profile $H(r)$ we have a decoupled (from $b_{3}$) equation that reads $$\left[ \lambda \cos (2H)\left( l_{2}^{2}q^{2}+l_{3}^{2}p^{2}\right) +l_{2}^{2}\left( 8l_{3}^{2}+\lambda q^{2}\right) +\lambda l_{3}^{2}p^{2}\right] H^{\prime \prime }+\left( l_{2}^{2}q^{2}+l_{3}^{2}p^{2}\right) \left( l_{1}^{2}-\lambda H^{\prime 2}\right) \sin (2H)=0. \label{profXYcon1}$$ Thus, the big technical achievement of the present approach is that the three coupled gauged Skyrme equations in Eq. (\[nonlinearsigma1\]) and the corresponding four Maxwell equations in Eq. (\[maxwellskyrme1\]) with exactly the Skyrme ansatz in Eqs. (\[pions2.25\]) and (\[pions2.26\]) and the gauge potential in Eq. (\[empot\]) reduce to Eqs. (\[b3XYcon1\]) and (\[profXYcon1\]) when the two algebraic conditions in Eq. (\[Xcon1\]) are satisfied. We want to stress that the aforementioned relations provide an exact solution and they are not a product of an approximation. As for the boundary conditions that are needed to be set, we have to keep in mind that the system is confined to a finite box. Thus, the easiest way to realize this is by imposing periodic boundary conditions in $\gamma$ and $\phi$ and Dirichlet in $r$ Interestingly enough, Eq. (\[profXYcon1\]) can be solved explicitly by observing that it has the following first integral$$Y\left( H\right) \frac{H^{\prime 2}}{2}+V(H)=E_{0}\ , \label{firstint}$$with $$\begin{aligned} Y\left( H\right) &=&2\lambda \left( l_{2}^{2}q^{2}+l_{3}^{2}p^{2}\right) \cos ^{2}(H)+8l_{2}^{2}l_{3}^{2}, \label{firstint2} \\ V\left( H\right) &=&-\frac{1}{2}l_{1}^{2}\left( l_{2}^{2}q^{2}+l_{3}^{2}p^{2}\right) \cos (2H) \label{firstint3}\end{aligned}$$and where $E_{0}$ is an integration constant to be determined by requiring that the boundary conditions to have non-vanishing topological charge are satisfied. Thus, Eq. (\[profXYcon1\]) can be reduced to a quadrature (which defines a generalized elliptic integral). Eq. (\[b3XYcon1\]) for $b_{3}$ is linear (since $H(r)$ can be found explicitly), however its integration is not a trivial task. In any case, integration of ([b3XYcon1]{}) that results in an expression for $b_{3}$ makes trivial the determination of the other two components of $A_{\mu }$ since both $b_{1}$ and $b_{2}$ are given algebraically in terms of $b_{3}$ through conditions . Nevertheless, even without the explicit expressions, it is still possible to analyze the generic features of the transport properties electrons passing through the above gauged Skyrmions. Gauged time-crystals -------------------- In order to have a time periodic solution with a non vanishing topological charge, that can be characterized as a time-crystal (for the introduction to the notion of time crystals see [@timec1; @timec2; @timec3; @timecr]) we start by considering the line element $$ds^2 = - d\gamma^2 +l_1 dr^2 +l_2 dz^2 +l_3 d\phi^2,$$ where $\gamma$ in the new ansatz $$G=\frac{q\phi - \omega \gamma}{2}, \; A=- \frac{q\phi + \omega \gamma}{2}$$ is the time variable, making the ensuing solution a time periodic configuration. The constant $\omega$ is the frequency of the time-crystal characterizing the periodicity of the system. Again we consider a finite box, where this time we take $$0 \leq r \leq 2\pi, \quad 0 \leq z \leq 4\pi, \quad 0 \leq \phi \leq 2\pi .$$ We adopt a similar form for the electromagnetic potential as the one given in . However, we have to note now that the index of the coordinates is changed into $x^\mu = (\gamma, r, z, \phi)$. Thus, the vector potential is $$A_\mu = (b_2(r),0,b_1(r),b_3(r)),$$ making $b_2(r)$ the electrostatic potential instead of $b_1(r)$ that we had in the Skyrmion case. The Maxwell equations retain same form as with $$\begin{aligned} M_{11} & = -\frac{K}{2 l_3^2} \sin ^2(H) \left[8 \lambda l_3^2 H^{\prime 2}+l_1^2 \left(2 \lambda \cos ^2(H) \left(q^2-l_3^2 \omega ^2\right)+8 l_3^2\right)\right] \\ M_{23} & = \frac{K \lambda l_1^2 q \omega}{4 l_3^2} \sin ^2(2 H) \\ M_{32} & = - l_3^2 M_{23} \\ M_{22} & = M_{11} + \frac{\omega}{q} M_{3,2} \\ M_{33} & = M_{11} + \frac{q}{\omega} M_{2,3} \\ N_2 & = \frac{\omega}{4} M_{11} - \frac{1}{4}\left(\frac{l_3^2 \omega ^2}{q} +q\right) \\ N_3 & = -\frac{q}{4}M_{11} + \frac{1}{4} \left(\frac{q^2}{l_3^2 \omega } + \omega \right),\end{aligned}$$ while the rest of the components of $M$ and $N$ are zero. As also happened in the Skyrme case, again here, the field equations reduce to a single ordinary differential equation for the profile function $H(r)$. In this case the relative equation reads $$\begin{split} & 4\left( X\sin ^{2}(H)+l_{2}^{2}\left( l_{3}^{2}\left( \lambda \omega ^{2}-4\right) -\lambda q^{2}\right) \right) H^{\prime \prime }+2X\sin (2H)(H^{\prime })^{2}+4\sin ^{2}(H)X^{\prime }H^{\prime } \\ & +\frac{l_{1}^{2}}{4l_{3}^{2}}\left[ 4\lambda l_{2}^{2}\left( 2\kappa qb_{3}-l_{3}^{2}\omega (2\kappa b_{2}+\omega )\right) \left( 2\kappa l_{3}^{2}\omega b_{2}+q(q-2\kappa b_{3})\right) -X\left( q^{2}-l_{3}^{2}\omega ^{2}\right) \right] \sin (4H) \\ & -\frac{\left( 2l_{1}^{2}\right) }{\lambda }X\sin (2H)=0, \end{split} \label{profeqTC}$$where $$X(r)=-8\kappa \lambda \left[ 2\kappa l_{3}^{2}b_{1}^{2}-l_{2}^{2}\left( l_{3}^{2}b_{2}(2\kappa b_{2}+\omega )+b_{3}(q-2\kappa b_{3})\right) \right] .$$Once more, profile equation can be reduced to an integrable one that is decoupled from the Maxwell field. Let us assume the following conditions for the components $b_{1}$ and $b_{3}$ of the electromagnetic potential $A_{\mu }$: $$X(r)=\lambda l_{2}^{2}\left( q^{2}-l_{3}^{2}\omega ^{2}\right) ,\quad b_{3}(r)=\frac{l_{3}^{2}\omega }{q}b_{2}(r)+\frac{l_{3}^{2}\omega ^{2}}{4\kappa q}+\frac{q}{4\kappa }.$$Then, the remaining Maxwell equation that needs to be satisfied for $b_{2}$ is $$b_{2}^{\prime \prime }=-\frac{\kappa K}{8l_{3}^{2}}(4\kappa b_{2}+\omega )\left[ 8l_{3}^{2}\left( \lambda (H^{\prime })^{2}+l_{1}^{2}\right) +2\lambda l_{1}^{2}\cos ^{2}(H)\left( q^{2}-l_{3}^{2}\omega ^{2}\right) \right] \sin ^{2}(H) \label{matc}$$and the profile equation is reduced to $$\left( 2\lambda \cos ^{2}(H)\left( q^{2}-l_{3}^{2}\omega ^{2}\right) +8l_{3}^{2}\right) H^{\prime \prime }+\sin (2H)\left( q^{2}-l_{3}^{2}\omega ^{2}\right) \left( l_{1}^{2}-\lambda H^{\prime 2}\right) =0.$$Obviously it exhibits a first integral of the form where now $$\begin{aligned} Y(H)& =2\lambda \cos ^{2}(H)\left( q^{2}-l_{3}^{2}\omega ^{2}\right) +8l_{3}^{2} \\ V(H)& =\frac{l_{1}^{2}}{2}\left( l_{3}^{2}\omega ^{2}-q^{2}\right) \cos (2H).\end{aligned}$$We can notice the similarities with the expressions derived for the Skyrmion in the previous case. In [@gaugsk] there has been presented an extensive discussion on the extended duality" that exists between two such systems. Topological Current for the gauged Skyrmion ------------------------------------------- The topological current [@Witten] of the gauged Skyrme model can be divided into two terms $$J_{\mu }^{B}=J_{\mu }^{Sk}+J_{\mu }^{B-em}$$with the first term $J_{\mu }^{Sk}$ being the usual Baryonic current, while second term is the correction to the latter, owed to the coupling with the electromagnetic field. For the first term we have $$J_{\mu }^{Sk}=\frac{1}{24\pi ^{2}}E_{\mu \alpha \beta \nu }Tr\left( R^{\alpha }R^{\beta }R^{\nu }\right) , \label{Dirac}$$which in our case has a single nonzero component $$J_{0}^{Sk}=-\frac{pq}{8\pi ^{2}l_{1}l_{2}l_{3}}H^{\prime }\sin (2H)=-2\pi \widehat{n}_{B}H^{\prime }\sin (2H)\ ,\ V=16\pi ^{3}l_{1}l_{2}l_{3}, \label{Bden1}$$where $V=16\pi ^{3}l_{1}l_{2}l_{3}$ is the volume of the box and $\widehat{n}_{B}$ is the Baryon density ($\widehat{n}_{B}=pq/V$) of the system. Note that in we make use of the Levi-Civita tensor $E_{\mu \alpha \beta \nu }=\sqrt{-g}\,\epsilon _{\mu \alpha \beta \nu }$ instead of the Levi-Civita symbol $\epsilon _{\mu \alpha \beta \nu }$ so that $J_{\mu }^{Sk} $ transforms covariantly and the topological charge results in a pure number. If for instance we apply the boundary conditions $H(0)=0$, $H(2\pi )=\frac{\pi }{2}$ we obtain $$B=\int_{\Sigma }\!\!\sqrt{-g}J_{Sk}^{0}drd\gamma d\phi =pq.$$The correction $J_{\mu }^{B-em}$ to the baryonic current, due to the electromagnetic field, is $$J_{\mu }^{em}=-\frac{\kappa }{8\pi ^{2}}E_{\mu \alpha \beta \nu }\nabla ^{\alpha }\left[ A^{\beta }Tr\left( t_{3}(U^{-1}\nabla ^{\nu }U-\nabla ^{\nu }UU^{-1})\right) \right] \label{curBem}$$and the total gauged Baryonic current reads $$\label{fullcurB} \begin{split} J_{\mu }^{B}=\Big\{-\frac{pq\pi }{V}\partial _{r}\left( \cos (2H)\right) +\frac{4\pi \kappa }{V}\partial _{r}\left( \cos ^{2}(H)(qb_{2}-pb_{3})\right) ,0,& \\ -\frac{4\pi q\kappa }{V}\partial _{r}\left( b_{1}\cos ^{2}(H)\right) ,\frac{4\pi p\kappa }{V}\partial _{r}\left( \cos ^{2}(H)\right) & \Big\}, \end{split}$$From what we see, the total baryon number when the Skyrmion is coupled to the electromagnetic field depends also on the boundary conditions that one may impose on the latter ($b_{2}$ and $b_{3}$ in particular). Baryonic current for the Time-Crystal ------------------------------------- The topological current of the time-crystal can be calculated with the use of the same relations and . Here we just give the result for the full current of the Gauged Time Crystal (GTC) which is $$\label{JGTC} \begin{split} J_\mu^{GTC} = \Big\{ &-\frac{4\pi q \kappa }{V} \partial_r \left(b_1 \cos^2(H)\right), 0, \\ & -\frac{l_2^2 q \pi \omega }{ V} \partial_r (\cos(2H)) + \frac{4 \pi \kappa }{V} \partial_r\left[\cos ^2(H) (q b_2-\omega b_3)\right], \frac{4\pi \kappa \omega }{V} \partial_r \left(b_1 \cos^2(H) \right) \Big\} . \end{split}$$ In the absence of the coupling with the electromagnetic field, $\kappa=0$, we can see that the expression for the non-zero topological current of the time-crystal is simplified to $$J_\mu^{TC} = \Big\{ 0,0, -\frac{\pi l_2^2 q \omega }{V} \partial_r (\cos(2H)),0\Big\} .$$ On the conductivity of gauged solitons ====================================== At semi-classical level, the transport properties of electrons travelling through the above gauged Skyrmions can be determined by analyzing the corresponding Dirac equation. Obviously, the electrons interact directly both with the gauge field and with the Baryons. The fermion couples to $A_{\mu }$, as QED dictates. However, there are further effects due to the coupling with the baryonic current. Here, we follow a very simple toy model interaction just to make a qualitative description of such effects. At this level of approximation in which the electrons perceive the gauged Skyrmions as a classical background, both interactions can be described as current-current" interactions in the Dirac Hamiltonian. The interaction of the electronic Dirac field $\Psi $ with the gauge potential $A_{\mu }$ corresponds to the following interaction Hamiltonian$$\begin{aligned} H_{int}^{U(1)} &=&\kappa J_{\mu }^{e}A^{\mu }\ , \label{u1corr} \\ J_{\mu }^{e} &=&\overline{\Psi }\gamma _{\mu }\Psi \ , \notag \\ \overline{\Psi } &=&\Psi ^{\dag }\gamma ^{0}, \notag\end{aligned}$$where $\kappa $ is the Maxwell coupling$$\kappa \approx \left( \frac{1}{137}\right) ^{\frac{1}{2}}, \label{u1coupling}$$$\gamma _{\mu }$ are the Dirac gamma-matrices (the conventions are collected in the appendix \[appA\]), $\Psi ^{\dag }$ is the conjugate transpose of $\Psi $ and $\overline{\Psi }$ the adjoint spinor. On the other hand, a simple way to describe the interactions of the electronic Dirac field with the baryonic current $J_{B}^{\mu }$ is with the following Hamiltonian$$H_{int}^{B}=g_{eff}J_{\mu }^{e}J_{B}^{\mu }\ , \label{barcorr}$$where $g_{eff}$ is the effective coupling constant of the electron-Baryon interaction. At the present level of approximation (in which the energy scale is not high enough to disclose the parton structure of the Baryon) a reasonable assumption is: $$g_{eff}\approx G_{F},$$where $G_{F}$ is the Fermi constant. In order to evaluate the relative strength of the two contributions to the conductivity (a brief analysis is given in Appendix \[appB\]), one arising from the term owed to the coupling with the $U(1)$ field (the $\kappa A_{\mu }$ in Eq. (\[Diraceq\]), see section \[appB1\] of Appendix \[appB\]) and the other arising from the term produced from the baryon current (the $G_F J_{\mu }^{B}$ in Eq. (\[Diraceq\])) one needs to evaluate the relative strength of the $U(1)$ coupling with respect to the interactions with the Skyrmionic current. There are two competing factors in the interactions with the Skyrmionic current. The first factor is the electro-weak coupling constant (which is obviously weaker than the $U(1)$ coupling). The second factor is related with the Skyrmions profile $H$ and can be evaluated explicitly thanks to the present analytic solutions. Assuming that both $\sin (2H)$ and $H^{\prime }$ are of order 1 (since both quantities are adimensional and the solitonic solutions we are considering are smooth and regular) one can see that the effective adimensional coupling $\widehat{g}$ measuring the strength of the contributions to the conductivity due to the interactions of the electrons with the Skyrmionic current is: $$\widehat{g}= l_1 G_{F}\widehat{n}_{B}\ . \label{efcoupl}$$Given that $G_F \sim 1.166$ GeV$^{-2}$ or $G_F \sim 4.564$ fm$^2$ in natural units we can see that the contribution of the interaction with $J_{\mu }^{B}$ remains small in comparison to the coupling with $A_{\mu }$ - at least for baryon densities $\widehat{n}_{B}$ and lengths $l_1$ of the box that can be characterized as natural. The “Baryonic" correction $\delta \Psi $ to the wave function in Eq. (\[perturbedpsi\]) depends on the effective coupling $\widehat{g}$ defined in Eq. (\[efcoupl\]) and on the Fourier transform of quantities related with the background Skyrmion. For completeness, in sections \[appB2\] and \[appB3\] of Appendix [appB]{} we have included the Dirac equations for the electrons propagating in the gauged solitons background described above. Although these Dirac equations cannot be solved analytically (due to the fact that Eqs. ([b3XYcon1]{}) and (\[matc\]) are not integrable in general), they can be useful starting points for numerical analysis of transport properties of the present gauged solitons. Conclusions and perspectives {#conclusions} ============================ In the present paper we have studied (gauged) Skyrmionic configurations in a finite box. We provided the reduced field equations under the adopted ansatz and distinguished the conditions over the potential functions $A_\mu$ for which the aforementioned equations can be characterized as integrable. Additionally, we have presented analytic expressions for the energy and studied its general behaviour in relation to the baryon number and the possible sizes of the box under consideration. We also managed to demonstrate and analyze the cases where the more energetically convenient configurations emerge in relations to these variables. What is more, we have derived an explicit analytic expression for the compression modulus corresponding to Skyrmions living within a finite volume in flat space-times. This is the first case in which one can derive an analytic formula (Eqs. (\[Kbox\]) and (\[Vbox\]) in the previous section) for such an important quantity in a highly interacting theory such as the low energy limit of QCD. This expression produces a reasonable value with a correct order of magnitude. The gauged version of these solitons living within a finite volume can be also considered. Using these gauged solitons, it is possible to analyze the contributions to the electrons conductivity associated to the interactions with this Baryonic environment (which represents a slab of baryons which can be very large in two of the three spatial directions). To the best of authors knowledge, the present is the first concrete setting in which it is possible to perform analytic computations of these relevant quantities in the original version of the Skyrme model (and its gauged version). Acknowledgements {#acknowledgements .unnumbered} ---------------- The authors would like to thank A. Zerwekh for useful discussions. This work has been funded by the Fondecyt grants 1160137, 1161150 and 3160121. The Centro de Estudios Científicos (CECs) is funded by the Chilean Government through the Centers of Excellence Base Financing Program of Conicyt. Conventions {#appA} =========== Throughout the paper we use the metric signature $(-,+,+,+)$. The ordering of the space-time coordinates is $x^{\mu }=(z,r,\gamma ,\phi )$ for the Skyrmion and $x^{\mu }=(\gamma ,r,z,\phi )$ for the time-crystal. The four Dirac matrices are $$\begin{aligned} \gamma ^{0}& =\left( \begin{array}{cccc} 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & -1 & 0 \\ 0 & 0 & 0 & -1\end{array}\right) ,\quad \gamma ^{1}=\left( \begin{array}{cccc} 0 & 0 & 0 & 1 \\ 0 & 0 & 1 & 0 \\ 0 & -1 & 0 & 0 \\ -1 & 0 & 0 & 0\end{array}\right) \\ \gamma ^{2}& =\left( \begin{array}{cccc} 0 & 0 & 0 & -\mathbbmtt{i} \\ 0 & 0 & \mathbbmtt{i} & 0 \\ 0 & \mathbbmtt{i} & 0 & 0 \\ -\mathbbmtt{i} & 0 & 0 & 0\end{array}\right) ,\quad \gamma ^{3}=\left( \begin{array}{cccc} 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & -1 \\ -1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0\end{array}\right) .\end{aligned}$$ Dirac equation {#appB} ============== Here we include, for completeness, the Dirac equation for an electron propagating in the two gauged solitons described in the main text. Although, in these cases, the Dirac equation cannot be solved analytically, it shows clearly that the present framework provides with a concrete setting to attack computations which, at a first glance, could appear very difficult (like the conductivities associated to gauged solitons at finite densities). Qualitative Analysis {#appB1} -------------------- The Dirac equation which describes the propagation of the electron through the above gauged Skyrmion is $$\left[ \gamma ^{\mu }\left( \mathbbmtt{i}\nabla _{\mu }-\kappa A_{\mu }-G_{F}J_{\mu }^{B}\right) +m\right] \Psi (z,r,\gamma ,\phi )=0\ , \label{Diraceq}$$where $m$ is the electron mass and $J_{\mu }^{B}$ is given by . It is convenient to write the above Dirac equation as follows:$$\begin{aligned} \left( H_{0}+H_{int}\right) \Psi &=&0\ , \label{int0} \\ H_{0} &=&\left[ \mathbbmtt{i}\gamma ^{\mu }\nabla _{\mu }+m\right] \ , \label{int1} \\ H_{int} &=&\left[ \gamma ^{\mu }\left( -\kappa A_{\mu }-G_{F}J_{\mu }^{B}\right) \right] \footnotemark . \label{int3}\end{aligned}$$We will work to first order in perturbation theory and we will consider $H_{int}$ as a small perturbation. The main goal of our analysis is to take the first order corrections to the conductivity and make a comparison between the part that is owed to the interactions with the solitons and the usual contributions arising from electromagnetic sources other than the soliton itself. The last *ingredient* we need is the Kubo formula for the conductivity associated to electrons moving in a medium (for a detailed review see chapter 4 of [@Dressel]). Following the usual steps one arrives at the following expression for the conductivity $\sigma _{\mu \nu }\left( \overrightarrow{q},\Omega \right) $ (where $\overrightarrow{q}$ and $\Omega$ the wave vector and frequency respectively of the incident electromagnetic wave): $$\sigma _{\mu \nu }\left( \overrightarrow{q},\Omega \right) =\sum_{s}\frac{1}{\hbar \Omega }\int dt\left\langle s\right\vert J_{0\mu }^{e}\left( \overrightarrow{q},0\right) J_{0\nu }^{\ast e}\left( \overrightarrow{q},\Omega \right) \left\vert s\right\rangle \exp \left[ -i\Omega t\right]$$where $\left\vert s\right\rangle $ and $J_{0\mu }^{e}$ are the eigenstate of the free Dirac Hamiltonian and the corresponding current in the box where the gauged solitons live. Due to the interaction Hamiltonian $H_{int}$ defined[^5] in Eqs. (\[int0\]), (\[int1\]) and (\[int3\]), the electron currents $J_{\mu }^{e}=\overline{\Psi }\gamma _{\mu }\Psi $ changes$$J_{0\mu }^{e}\rightarrow J_{0\mu }^{e}+\left( \delta \overline{\Psi }\right) \gamma _{\mu }\Psi +\overline{\Psi }\gamma _{\mu }\left( \delta \Psi \right) =J_{0\mu }^{e}+\delta J_{\mu }^{e}\ ,$$where $\delta \Psi $ can be computed using first order perturbation theory. In particular, if $\Psi _{0}$ is a solution of the un-perturbed equation$$H_{0}\Psi _{0}=E\Psi _{0}\ ,$$then the eigenstate $\Psi $ of the interacting case can be written as$$\Psi =\Psi _{0}-H_{0}^{-1}\left( H_{int}\Psi _{0}\right) ,$$where $H_{0}^{-1}$ is the inverse Dirac operator defined as the Green function $H_{0}^{-1}=G(x-x^{\prime })$ satisfying $$H_{0}G_{0}(x-x^{\prime })=\delta (x-x^{\prime }).$$We now from the free particle case that the Green function in space-time variables is expressed as $$H_{0}^{-1}=G_{0}(x-x^{\prime })=\int \!\!\frac{d^{4}k}{(2\pi )^{4}}e^{-\mathbbmtt{i}k_{\mu }(x^{\mu }-x^{\prime \mu })}\frac{m-\gamma ^{\mu }k_{\mu }}{k_{\mu }k^{\mu }+m^{2}}$$(of course in our case, for the finite box, the integral is to be substituted by series). Consequently, we have a perturbation of the form $$\delta \Psi =\int \!\!\frac{d^{4}k}{(2\pi )^{4}}\int d^{4}x^{\prime -\mathbbmtt{i}k_{\mu }(x^{\mu }-x^{\prime \mu })}\frac{m-\gamma ^{\mu }k_{\mu }}{k_{\mu }k^{\mu }+m^{2}}\left( \kappa A_{\mu }+G_{F}J_{\mu }^{B}\right) \Psi _{0}(x^{\prime }) \label{perturbedpsi}$$owed to two contributions; the Maxwell field $A_{\mu }$ and the baryon current $J_{\mu }^{B}$. As for the free particle solution $\Psi _{0}$, it is easy to see that $$\Psi _{0}(x)=\begin{pmatrix} \psi _{1} \\ \psi _{2} \\ \psi _{3} \\ \psi _{4}\end{pmatrix}e^{-\mathbbmtt{i}k_{\mu }x^{\mu }}$$with $$\psi _{1}=\frac{k_{3}\psi _{3}+(k_{1}-\mathbbmtt{i}k_{2})\psi _{4}}{k_{0}+m},\quad \psi _{2}=\frac{(k_{1}+\mathbbmtt{i}k_{2})\psi _{3}-k_{3}\psi _{4}}{k_{0}+m},\quad k_{0}^{2}=\vec{k}^{2}+m$$satisfies $H_{0}\Psi _{0}=0$. Consequently, $$\begin{aligned} \sigma _{\mu \nu } &\rightarrow &\sigma _{\mu \nu }+\delta \sigma _{\mu \nu }\ , \\ \delta \sigma _{\mu \nu } &=&\sum_{s}\frac{1}{\hbar \Omega }\int dt\left\langle s\right\vert \left[ \delta J_{0\mu }^{e}\left( \overrightarrow{q},0\right) J_{0\nu }^{\ast e}\left( \overrightarrow{q},\Omega \right) +J_{0\mu }^{e}\left( \overrightarrow{q},0\right) \delta J_{0\nu }^{\ast e}\left( \overrightarrow{q},\Omega \right) \right] \left\vert s\right\rangle \exp \left[ -i\Omega t\right] \ .\end{aligned}$$ Dirac equation for the gauged Skyrmion {#appB2} -------------------------------------- The symmetries of the problem allow to search for a separated solution of the form $$\label{Danz1} \Psi (z,r,\gamma ,\phi )=e^{-\mathbbmtt{i}(\omega _{1}z-k_{2}\gamma -k_{3}\phi )}\{\psi _{1}(r),\psi _{2}(r),\psi _{3}(r),\psi _{4}(r)\}.$$By introducing into the Dirac equation we obtain the following set of equations for the components of $\Psi$: $$\begin{aligned} \psi _{1}^{\prime }=& \left( k_{2}+\kappa b_{2}-\kappa g\frac{q}{4\pi ^{2}V}\phi _{1}^{\prime }\right) \psi _{1}+\mathbbmtt{i}\left( k_{3}+\kappa b_{3}+\kappa g\frac{p}{4\pi ^{2}V}\phi _{1}^{\prime }\right) \psi _{2} \notag \\ & +\mathbbmtt{i}\left( \omega _{1}-m-\kappa b_{1}-\frac{g}{16\pi ^{2}V}\left( 4\kappa \phi _{2}^{\prime }+\phi _{3}^{\prime }\right) \right) \psi _{4} \\ \psi _{2}^{\prime }=& -\mathbbmtt{i}\left( \kappa b_{3}+k_{3}+\kappa g\frac{p}{4\pi ^{2}V}\phi _{1}^{\prime }\right) \psi _{1}-\left( \kappa b_{2}+k_{2}-\kappa g\frac{q}{4\pi ^{2}V}\phi _{1}^{\prime }\right) \psi _{2} \notag \\ & +\mathbbmtt{i}\left( \omega _{1}-m-\kappa b_{1}-\frac{g}{16\pi ^{2}V}\left( 4\kappa \phi _{2}^{\prime }+\phi _{3}^{\prime }\right) \right) \psi _{3} \\ \psi _{3}^{\prime }=& \mathbbmtt{i}\left( \omega _{1}+m-\kappa b_{1}-\frac{g}{16\pi ^{2}V}\left( 4\kappa \phi _{2}^{\prime }+\phi _{3}^{\prime }\right) \right) \psi _{2} \notag \\ & +\left( k_{2}+\kappa b_{2}-\kappa g\frac{q}{4\pi ^{2}V}\phi _{1}^{\prime }\right) \psi _{3}+\mathbbmtt{i}\left( k_{3}+\sigma b_{3}+\kappa g\frac{p}{4\pi ^{2}V}\phi _{1}^{\prime }\right) \psi _{4} \\ \psi _{4}^{\prime }=& \mathbbmtt{i}\left( \omega _{1}+m-\kappa b_{1}-\frac{g}{16\pi ^{2}V}\left( 4\kappa \phi _{2}^{\prime }+\phi _{3}^{\prime }\right) \right) \psi _{1} \notag \\ & -\mathbbmtt{i}\left( \kappa b_{3}+k_{3}+\kappa g\frac{p}{4\pi ^{2}V}\phi _{1}^{\prime }\right) \psi _{3}-\left( \kappa b_{2}+k_{2}-\kappa g\frac{q}{4\pi ^{2}V}\phi _{1}^{\prime }\right) \psi _{4},\end{aligned}$$ where $$\begin{aligned} \phi _{1}(r)& =b_{1}(r)\cos ^{2}(H(r))\ , \\ \phi _{2}(r)& =\cos ^{2}(H(r))(qb_{2}(r)-pb_{3}(r))\ , \\ \phi _{3}(r)& =pq\cos (2H(r))\ .\end{aligned}$$ Dirac equation for the gauged time-crystal {#appB3} ------------------------------------------ By using the expression for $J_{\mu }^{GTC}$ as given by inside , instead of $J_{\mu }^{B}$ that we had for the Skyrmion, and by considering a separable solution of the form $$\Psi (z,r,\gamma ,\phi )=e^{-\mathbbmtt{i}(\omega _{1}\gamma -k_{2}z-k_{3}\phi )}\{\psi _{1}(r),\psi _{2}(r),\psi _{3}(r),\psi _{4}(r)\},$$we obtain a system of equations given by $$\begin{aligned} \psi _{1}^{\prime }=& \left( \kappa b_{2}-\omega _{1}-\frac{gq\kappa }{4\pi ^{2}V}\phi _{1}^{\prime }\right) \psi _{1}+\mathbbmtt{i}\left( \kappa b_{3}+k_{3}+\frac{g\kappa \omega }{4\pi ^{2}V}\phi _{1}^{\prime }\right) \psi _{2} \notag \\ & -\mathbbmtt{i}\left( \kappa b_{1}+k_{2}+m+\frac{g}{16\pi ^{2}V}\left( 4\kappa \phi _{2}^{\prime }-l_{2}^{2}\phi _{3}^{\prime }\right) \right) \psi _{4} \\ \psi _{2}^{\prime }=& -\mathbbmtt{i}\left( \kappa b_{2}+k_{3}+\frac{g\kappa \omega }{4\pi ^{2}V}\phi _{1}^{\prime }\right) \psi _{1}+\left( \omega _{1}-\kappa b_{2}+\frac{gq\kappa }{4\pi ^{2}V}\phi _{1}^{\prime }\right) \psi _{2} \notag \\ & -\mathbbmtt{i}\left( \kappa b_{1}+k_{2}+m+\frac{g}{16\pi ^{2}V}\left( 4\kappa \phi _{2}^{\prime }-l_{2}^{2}\phi _{3}^{\prime }\right) \right) \psi _{3} \\ \psi _{3}^{\prime }=& \mathbbmtt{i}\left( m-\kappa b_{1}-k_{2}-\frac{g}{16\pi ^{2}V}\left( 4\kappa \phi _{2}^{\prime }-l_{2}^{2}\psi _{3}^{\prime }\right) \right) \psi _{2} \notag \\ & \left( \kappa b_{2}-\frac{gq\kappa }{4\pi ^{2}V}\phi _{1}^{\prime }-\omega _{1}\right) \psi _{3}+\mathbbmtt{i}\left( \kappa b_{3}+k_{3}+\frac{g\kappa \omega }{4\pi ^{2}V}\phi _{1}^{\prime }\right) \psi _{4} \\ \psi _{4}^{\prime }=& \mathbbmtt{i}\left( m-\kappa b_{1}-k_{2}-\frac{g}{16\pi ^{2}V}\left( 4\kappa \phi _{2}^{\prime }-l_{2}^{2}\phi _{3}^{\prime }\right) \right) \psi _{1} \notag \\ & -\mathbbmtt{i}\left( \kappa b_{3}+k_{3}+\frac{g\kappa \omega }{4\pi ^{2}V}\phi _{1}^{\prime }\right) \psi _{3}+\left( \omega _{1}-\kappa b_{2}+\frac{gq\kappa }{4\pi ^{2}V}\phi _{1}^{\prime }\right) \psi _{4}.\end{aligned}$$ The functions $\phi _{1}(r)$, $\phi _{2}(r)$ and $\phi _{3}(r)$ are the same as before, only now we have $\omega $ appearing in them in place of $p$, i.e. $$\begin{aligned} \phi _{1}(r)& =b_{1}(r)\cos ^{2}(H(r)) \\ \phi _{2}(r)& =\cos ^{2}(H(r))(qb_{2}(r)-\omega b_{3}(r)) \\ \phi _{3}(r)& =q\omega \cos (2H(r)).\end{aligned}$$ [999]{} T. Skyrme, *Proc. R. Soc. London A* **260**, 127 (1961); *Proc. R. Soc. London A* **262**, 237 (1961); *Nucl. Phys.* **31**, 556 (1962). T.S. Walhout, *Nucl. Phys. A* **531**, 596 (1991) C. Adam, M. Haberichter and A. Wereszczynski, *Phys. Rev. C* **92**, 055807 (2015) J.-i Fukuda and S. Žumer, *Nature Comm.* **2**, 246 (2011) H. Stefan et al. *Nature Physics* **7**, 713 (2011) D. Fostar and S. Krusch, *Nuc. Phys. B* **897**, 697 (2015) M. Gillard,* Nucl. Phys. B* **895**, 272 (2015) E. Witten, *Nucl. Phys. B* **223** (1983), 422; *Nucl. Phys. B* **223**, (1983) 433. D. Finkelstein, J. Rubinstein, *J. Math. Phys*.**9**, 1762–1779 (1968). N. Manton and P. Sutcliffe, *Topological Solitons*, (Cambridge University Press, Cambridge, 2007). V. G. Makhanov, Y. P. Rybakov, V. I. Sanyuk, *The Skyrme model*, Springer-Verlag (1993). D. Giulini, *Mod. Phys. Lett. A***8**, 1917–1924 (1993). A.P. Balachandran, A. Barducci, F. Lizzi, V.G.J. Rodgers,A. Stern, *Phys. Rev. Lett.* **52** (1984), 887. G. S. Adkins, C. R. Nappi, E. Witten, *Nucl. Phys. B* **228** (1983), 552-566. E. Guadagnini, *Nucl. Phys. B* **236** (1984), 35-47. A. D .Jackson, N. S. Manton and A. Wirzba *Nucl. Phys. A* **495** (1989) 499. C. G. Callan Jr. and E. Witten, *Nucl. Phys. B* **239** (1984) 161-176. J.M. Gipson and H.Ch. Tze, *Nucl. Phys. B* **183** (1981) 524. J. Goldstone and F. Wilczek, *Phys. Rev. Lett.* **47** (1981) 986. E. D’Hoker and E. Farhi, *Nucl. Phys. B* **241** (1984) 109. V.A. Rubakov, *Nucl. Phys. B* **256** (1985) 509. B.M.A.G. Piette, D. H. Tchrakian, *Phys.Rev. D* **62** (2000) 025020. E. Radu, D. H. Tchrakian, *Phys. Lett. B* **632** (2006) 109-113. I. Klebanov, *Nucl. Phys. B* **262** (1985) 133. A. Actor, *Phys. Lett. B* **157** (1985) 53. H. A. Weldon, *Phys. Rev. D* **26** (1982) 1394. M. Loewe, S. Mendizabal, J.C. Rojas, *Phys. Lett. B* **632** (2006) 512. J. A. Ponciano, N. N. Scoccola, *Phys. Lett. B* **659** (2008) 551. M. Kugler and S. Shtrikman, Phys. Rev. D 40 3421 (1989) M. Kugler and S. Shtrikman, Phys. Lett. B 208 491 (1988) M. Harada, Y.L. Ma, H.K. Lee and M. Rho, Fractionized Skyrmions in Dense Compact-Star Matter, The Multifaceted Skyrmion (second edition), ed. M. Rho and I. Zahed, World Scientific, Singapore (2016) Y.-L. Ma and M. Rho, Effective field theories for Nuclei and Compact-Star matter, World Scientific, Singapore (2018) L. C. Biedenharn, Y. Dothan, and M. Tarlini, *Phys. Rev. D* **31**, 649 (1985). U.-G. Meissner and I. Zahed, *Adv. Nucl. Phys.* **17**, (1986) 143. Th. Meissner, F. Grummer, K. Goeke, and M. Harvey, *Phys. Rev.* **D 39**, (1989) 1903. C. Adam, C. Naya, J. Sanchez-Guillen, J. M. Speight and A. Wereszczynski, *Phys. Rev. D* **90** (2014) 045003. H. Walliser, A. Hayashi and G. Holzwarth, *Nucl. Phys. A* **456**, (1986) 717. BES Collaboration, J. Z. Bai et al., Phys. Rev. Lett. 91, 022001 (2003); BES Collaboration, M. Ablikim et al., hep-ex/0503030; BES Collaboration, M. Ablikim et al., Phys. Rev. Lett. 93, 112002 (2004). S. Olsen, hep-ex/0305048, invited talk at the International Symposium on Hadron Spectroscopy, Chiral Symmetry and Relativistic Description of Bound States, Tokyo (2003). Wang, Z.-G., *Eur. Phys. J. A* **47**, (2011) 71. D. G. Ravenhall, C. J. Pethick, J. R. Wilson, *Phys. Rev. Lett*. **50**, 2066 (1983). D. Page, S. Reddy, *Annu. Rev. Nucl. Part. Sci.* (2006) **56**: 327–74. W. G. Newton, *Nature Physics.* **9**: 396 (2013). J. A. Pons, D. Viganò, N. Rea, *Nature Physics.* **9**: 431 (2013). D. G. Yakovlev, *MNRAS* **453**, 581 (2015). F. Canfora, H. Maeda, *Phys. Rev. D* **87**, 084049 (2013). F. Canfora, *Phys. Rev. D* **88**, 065028 (2013). F. Canfora, F. Correa, J. Zanelli, *Phys. Rev. D* **90**, 085002 (2014). F. Canfora, A. Giacomini and S. Pavluchenko, *Phys. Rev. D* **90**, 043516 (2014). S. Chen, Y. Li, Y. Yang, *Phys. Rev. D* **89** (2014), 025007. E. Ayon-Beato, F. Canfora, J. Zanelli, *Phys. Lett. B* **752**, (2016) 201-205. F. Canfora, M. Di Mauro, M. A. Kurkov, A. Naddeo, *Eur. Phys. J.* **C** (2015) **75**: 443. F. Canfora, G. Tallarita, *Nucl. Phys. B* **921** (2017) 394. F. Canfora, G. Tallarita, *Phys. Rev. D* **94**, 025037 (2016). F. Canfora, G. Tallarita, *Phys. Rev. D* **91**, 085033 (2015). F. Canfora, G. Tallarita, *JHEP* **1409**, 136 (2014). F. Canfora, Seung Hun Oh, P. Salgado-Rebolledo, *Phys. Rev. D* **96** (2017), 084038. P. D. Alvarez, F. Canfora, N. Dimakis and A. Paliathanasis, *Phys. Lett B*. **773**, (2017) 401-407. L. Aviles, F. Canfora, N. Dimakis, D. Hidalgo, *Phys. Rev. D* **96**, (2017) 125005. H. Weigel, Chiral Soliton Models for Baryons, Lecture Notes in Physics (Springer, 2008). F. Wilczek, *Phys. Rev. Lett.* **109**, 160401 (2012). A. Shapere, F. Wilczek, *Phys. Rev. Lett.* **109**, 160402 (2012). F. Wilczek, *Phys. Rev. Lett*. **111**, 250402 (2013). K. Sacha, J. Zakrzewski, “*Time crystals: a review*” arXiv: 1704.03735. Y. M. Shnir, *Magnetic Monopoles* (Springer-Verlag, Berlin, Heidelberg, 2005) pp.500 G. E. Brown and E. Osnes *Phys. Lett. B* **159** (1985) 223-227 G. Co’ and J. Speth *Phys. Rev. Lett.* **57** (1986) 547-550 J. P. Blaizot, *Phys. Rep.* **64** (1980) 171-248 M. E. Caplan and C. J. Horowitz, *Rev. Mod. Phys.* **89** (2017) 041002. M. Dutra, O. Lourenço, J. S. Sà Martins, A. Delfino, J. R. Stone and P. D. Stevenson, *Phys. Rev. C* **85** (2012) 035201 M. Dressel and G. Grüner, *Electrodynamics of Solids* (Cambridge University Press, Cambridge, 2002). [^1]: At least taking into account that it is reasonable to expect that the theory describing the low energy limit of QCD should be a quite complicated one. [^2]: The following analysis suggests that this “uniform rescaling” assumption could be too strong. Indeed, the results at the end of section 3 shows that Skyrme theory, when analyzed at finite density, provides with values of the compression modulus which are close to the experimental one. [^3]: In the leading ’t Hooft approximation, in meson-Baryon scattering, the heavy Baryon (the Skyrmion in our case) is unaffected and, basically, only the meson can react. This is even more so in the electron-Baryon semiclassical interactions due to the huge mass difference between the Skyrmion and the electron. In this approximation, electrons perceive the Skyrmions as an effective medium. [^4]: On the other hand, when the coupling with Maxwell field is neglected, the profile can depend on time as well. In this case, one gets an effective sine-Gordon theory for the profile $H(t,r)$ [@Fab1]. [^5]: The gauge potential $A_{\mu }$ and the Baryon current $J_{\mu }^{B}$ in the interaction Hamiltonian are the ones corresponding to the gauged Skyrmion and to the gauged time-crystal defined in the previous section.
{ "pile_set_name": "ArXiv" }
--- address: | Department of Physics, University of Florida, Gainesville FL 32611, USA\ E-mail: [email protected] $\bf{\ \bullet\bullet}$ WWW: http://www.phys.ufl.edu/\~kennedy author: - 'Dallas C. Kennedy' title: Cosmic Ray Antiprotons --- General Properties of Cosmic Rays ================================= Cosmic rays are high-energy $p$, $\pbar$, nuclei, and $e^\pm$ in interplanetary and interstellar (IS) space. The dominant component consists of protons (hydrogen H) with a smaller admixture of heavier nuclei, especially He (Figure \[fig:pSpec\]). Antiprotons $(\pbar)$ occur at an abundance of $10^{-4,-5}$ times that of $p$. These energetic particles, with kinetic energy $K >$ 10 MeV, are Galactic in origin, not to be confused with the much denser solar wind plasma, with much lower $K$, streaming from the Sun.[^1] The relative element abundances in cosmic rays (CRs) indicate they originate in the IS medium, where they are ionized and accelerated, probably by supernova shocks. (Recent measurements all but rule out an origin in supernova ejecta proper.) Such accelerated, pre-existing nuclei are CR [*primaries*]{}. [@ACE] Once accelerated to high energies, the primaries induce the production of further CRs, the [*secondaries*]{}, in the IS medium and at local sites in the Galaxy. (The terms “primaries” and “secondaries” are also used in a completely different sense: CR primaries are the CRs that strike the top of the Earth’s atmosphere, the secondaries the induced CR shower propagating into the lower atmosphere.) Secondaries include $e^+$, $\pbar$, and certain nuclear isotopes. As some of these isotopes are unstable, their populations must be continually replenished to maintain their observed abundances. Cosmic Ray Antiprotons ---------------------- Standard secondary CR antiprotons are produced by the process $pA\rightarrow\pbar X$, with $p$ = high-energy CR, $A$ = IS medium nucleus of atomic weight $A$, and $X$ = anything consistent with charge and baryon number $(B)$ conservation. The threshold channel is $pA\rightarrow p\pbar pA$, with threshold $E_p$ = $(3+4/A)m_p$. (The nucleus $A$ can break up without significantly changing the dynamics.) The dominant case is H, $A$ = 1, with threshold $E_p$ = $7m_p$. The only other significant contribution comes from He target nuclei. By number, the IS medium is $\simeq$ 93% H and 7% He. [@GaisLevy; @Stephens; @WebbPot; @GaisSch; @Gaisser] The secondary $\pbar$’s subsequently propagate in the Galaxy and are subject to a variety of elastic (scattering, including energy-loss) and inelastic (annihilation and extra-Galactic leakage) processes. Leakage is the dominant loss; the Galactic storage time $\sim$ 13 Myr as inferred from the abundance of unstable CR isotopes. Energy loss shifts the $\pbar$ spectrum without changing their number. Some uncertainty is unavoidable in models of Galactic propagation, including H and He abundances, as well as the Galaxy’s highly tangled, stochastic magnetic field $B_{\rm Gal}\sim$ 0.3 nT and small wind $V_{\rm Gal}\lesssim$ 20 km sec$^{-1}$ (a superposition of many stellar and supernova winds) . The field and wind control the diffusion of CRs into intergalactic space and are fairly well constrained by measurements of unstable CR isotopes. But more complicated transport mechanisms are possible, including reacceleration shocks and variation of the Galactic geometry. [@Gaisser; @WebberLeeGupta; @Longair; @Chard] The CR fluxes measured at the top of the Earth’s atmosphere are modulated by their transport through the heliosphere, the Sun’s magnetic sphere of influence (Figure \[fig:helioA\]). The heliosphere consists of a solar wind $(V_W\simeq$ 400 km sec$^{-1}$ along the ecliptic plane, 700-800 km sec$^{-1}$ along the solar axes) of $e^-$ and nuclei (mainly $p)$ carrying the embedded magnetic field $B_\odot$ (Figure \[fig:helioA\]). Near the Sun, the field falls off with heliocentric distance $r$ as $r^{-2}$, arising from frozen field flux transported radially outwards. Since the Sun rotates, however, the field is twisted into an Archimedean or [*Parker spiral*]{}, and the field is predominantly azimuthal in the outer solar system, falling off more softly as $r^{-1}$. At one AU (AU = 149.5 Mkm, the Earth’s orbit), the heliomagnetic field strength $B_\odot\simeq$ 5 nT. [@Parker] The CRs gyrate around the local ${\bf B}$ field lines. The solar field is not fully deterministic, however: it is modified by episodic (practically random) shocks that cause the CRs to diffuse along and across field lines, especially at times of solar magnetic maximum (currently 2000-01 and periodically about every 11 years, when the heliomagnetic field changes sign). In addition, the wind both imposes a macroscopic convective drift and performs work on the CRs (adiabatic deceleration), lowering their energies as they fight “upstream” into the inner solar system. A realistic prediction of Earth-measured CR fluxes must include these mechanisms, which are particularly important at lower $K$ and affect oppositely-charged CRs differently. [@transportOld; @transportNew; @Longair] Heliospheric [*in situ*]{} measurements have been ongoing for four decades and have recently become much better with the [*IMP*]{} and [*Ulysses*]{} space probes. [@IMP; @Ulysses] Exotic Sources of Cosmic Ray Antiprotons ---------------------------------------- The density of IS matter $n_{\rm H}\sim$ 1 H atom cm$^{-3}$ and the known spectrum and abundance of CR $p$ primaries fix the predicted spectrum and abundance of $\pbar$ secondaries, if the $pA\rightarrow\pbar X$ cross section $\sigma (\pbar )$ is known. [@Stephens] Let $Q_{\pbar}(K)$ be the differential production rate (antiprotons cm$^{-3}$ MeV$^{-1}$ sec$^{-1}$); schematically, Q\_(K) =  dK\^ n\_p (K\^)v(p)n\_[H]{}d(pp; K,K\^)/dK\^. The $pp$ process has been measured in laboratory experiments, and the $p+$He case can be inferred from the $pp$ cross section (but see subsection \[sec:FutureTheo\]). The differential $\pbar$ abundance $n_{\pbar}(K)$ (antiprotons cm$^{-3}$ MeV$^{-1}$) is related to $Q_{\pbar}(K)$ by $n_{\pbar}(K)$ = $\tau_{\rm eff}(K)\cdot Q_{\pbar}(K)$, where the effective Galactic residence time = + +, summing over all loss mechanisms. The sum is dominated by the first term, the extra-Galactic diffusion rate. The measured CR $\pbar$ flux is then related to $n_{\pbar}(K)$ through the transformation by heliospheric transport. Variation of Galactic transport mechanisms modifies $\tau_{\rm eff}(K)$. This picture is the basis for the simple [*Leaky Box Model*]{}. A more complex picture, with explicit spatial dependence on Galactic geometry (inhomogeneous leaky disk model = ILDM), is possible and indeed necessary, because of measurements of Galactic plane cosmic ray synchrotron radiation mapping the IS CR distribution. [@CRgammas; @WebberLeeGupta; @Chard] The cross section $\sigma (\pbar )$ has a crucial property arising from its threshold at $E_p$ = 7$m_p$ (Figure \[fig:pbarProd\]). The spectrum of outgoing $\pbar$’s rises sharply from $K_{\pbar}$ = 0. Since $n_p(K)$ falls off rapidly (as $K^{-2.75}_p)$, $n_{\pbar}(K)$ falls off similarly at high $K_{\pbar}$, leaving a $\pbar$ secondary spectrum with a sharp rise to a peak at $K\sim$ 2 GeV and falling off above that. The lower threshold for He targets enhances the low-$K$ spectrum somewhat. Although the secondary $\pbar$ spectrum must be there, its presence does not rule out non-standard $\pbar$ sources, so-called “exotic primaries”. These would add to the predicted secondary flux in total number. More crucially, they can also change the [*shape*]{} of the $\pbar$ spectrum, particularly at low $K$, as well the fall-off for $K\gtrsim$ 3 GeV. Cosmologically significant amounts of antimatter are strongly disfavored. [@CosmicAntimatter] Instead, the most logical sources for trace amounts of exotic primary $\pbar$ would be annihilating or decaying dark matter remnants in the halo of our Galaxy. Popular models feature annihilating supersymmetric (SUSY) dark matter [@CDMgeneral; @LSP; @CDMneutralino; @CDMdetect] (WIMPs, assumed to be the LSP = lightest SUSY particle, usually neutralinos $\tilde{\chi}^0$) or decaying primordial black holes. [@Hawking; @Stephens] The predictions depend on model details, [@Chard; @Ullio; @PBHs] but both have roughly flat $\pbar$ production spectra as $K\rightarrow$ 0 and a non-standard fall-off with $K$ at high energies. Such signals can only be seen if the exotic primaries compete in number with standard secondaries. A general range of exotic SUSY $\pbar$ production (Figure \[fig:ullio\]) exhibits the dramatic modification of the low-$K$ $\pbar$ spectrum possible in SUSY CDM models of the Galactic halo for smaller neutralino mass. SUSY halo dark matter $\pbar$’s (from $\tilde{\chi}^0\tilde{\chi}^0\rightarrow q\overline{q}$, Figure \[fig:susyCDM\]a) requires sufficient abundance and a large enough annihilation cross section to be seen, in turn implying WIMP masses $\lesssim$ few 100 GeV and $\sigma ({\rm ann}) v({\rm WIMP})\sim$ 0.1 pb. (Production of heavier WIMPs is suppressed in the Big Bang with with increasing mass.) The hadronic shower evolves finally into $p$’s, $\pbar$’s, $e^\pm$, $\nu$’s, and $\gamma$’s. WIMP annihilation is natural in minimal SUSY models with conserved $R$-parity. An extension of minimal SUSY allows $R$-parity violation, in turn allowing the LSPs to decay to ordinary matter, violating lepton and/or baryon number (Figure \[fig:susyCDM\]b). This mechanism could open another source of CR $\pbar$’s. In a semi-realistic scenario, [@Dreiner] lepton number violation is dominant, leading in the end to excess (anti)neutrinos. An exciting possible signal of annihilating or decaying CDM in the Galactic halo is suggested by the Galactic gamma ray maps of the orbiting Compton Gamma Ray Observatory’s EGRET telescope (Figure \[fig:egret\]). Primordial black holes (PBHs) are postulated to have been produced very early in the hot Big Bang, in the quantum gravity era. [@Hawking] They evaporate in turn by the Hawking process, as their temperatures rise, and can produce significant $p$’s and $\pbar$’s at a late time when $T_{\rm BH}\gtrsim$ $\Lambda_{\rm QCD}$. The relic PBH density and $\pbar$ production rate have been estimated. [@PBHs] Intrinsic Properties of Antimatter: CPT Symmetry {#sec:cpt} ------------------------------------------------ Cosmic ray $\pbar$’s also give us a window on the intrinsic properties of antimatter. These properties should be the same or charge-conjugated from the corresponding matter by the CPT (charge-conjugation, parity- and time-reversal) symmetry of local relativistic quantum field theory (LRQFT). Some $\pbar$ properties have been checked in the laboratory directly. These include the mass, charge, magnetic moment, and the neutrality of hydrogen and antihydrogen. [@PDBppbar] More difficult to limit is the decay lifetime of $\pbar$’s. Not enough antimatter can be gathered into a detector for long enough to produce lifetime limits on antimatter competitive with the limits for matter. Astrophysical processes partially ameliorate this difficulty. The Galactic storage time for $\pbar$’s $\sim$ 10 Myr, and intrinsic $\pbar$ decay would modify the Galactic residence time $\tau_{\rm eff}$ . If the decay lifetime is short enough (taking Lorentz dilation into account), the $\pbar$ spectrum is significantly distorted. The shape and normalization of the $\pbar$ spectrum then place a lower limit on $\tau_{\pbar}$. [@GeerKenn] Laboratory limits have been obtained for the $\pbar$ lifetime. Earlier limits include the LEAR Collaboration at the CERN $\pbar$ storage ring $(\tau_{\pbar} >$ 0.08 yr) and the antihydrogen Penning trap of Gabrielse [*et al.*]{} $(\tau_{\pbar}>$ 0.28 yr). [@PDBppbar] The best current laboratory limit is that of the APEX Collaboration at the Fermilab $\pbar$ storage ring $(\tau_{\pbar}>$ 50 kyr for $\pbar\rightarrow\mu^- X$ and 300 kyr for $e^-\gamma$). [@APEX] A proposed APEX II experiment would be able to reach $\pbar$ lifetime limits of 1–10 Myr, comparable to the cosmic ray limit.[^2] Since the CPT symmetry holds in LRQFT under the assumptions of Poincar' e invariance, locality, microcausality, and vacuum uniqueness, modification of basic physics would be necessary to break it. [@CPTthm; @LorentzCPT] Within QFT, an extensive formalism and phenomenology of Lorentz and CPT violation has been developed by Kosteleck' y and collaborators. [@KostelCPT] String theory at first glance might seem to provide a natural way to violate locality, but perturbative string dynamics has been shown to preserve CPT in the field theory target space after compactification. [@KostelStrings] [*Non-perturbative*]{} string effects associated with compactification may evade this result. [@Polchin] Extended quantum mechanics, with non-unitary time evolution, violates CPT in general, by violating locality and/or Poincar' e symmetry. Controversial proposals of non-unitary evolution have been put forward as natural consequences of quantum gravity and information loss in the presence of spacetime horizons. [@EHNScpt] Non-unitary effects have been powerfully limited in the very well-measured $K^0$–$\overline{K}^0$ system [@CPLEAR] (to a few parts in $10^{16}$), but not well at all in other systems, particularly baryons. [@KennCPT] The most plausible source of CPT violation lies beyond the Planck scale, based on strings or some other quantum theory of gravity, because of the necessary generalization beyond global Poincar' e symmetry. Typically such effects are thought of as suppressed by the large Planck mass $M_{\rm Pl}\sim 10^{19}$ GeV. But if gravity is fundamentally associated with “large” extra dimensions acting at mass scales as low as 1 TeV, [@LargeExtraDim] the CPT-violating mechanisms may not be that suppressed at accessible energies. Measurements of Cosmic Ray Antiprotons ====================================== Detection of CR antiprotons has gone through three distinct phases, following the proposal of Gaisser and Levy to search for $\pbar$ secondaries. [@GaisLevy] All but recent space-based experiments have been mounted on high-altitude balloons. The measurements are conventionally quoted as the $\pbar /p$ ratio of fluxes, convenient because a number of theoretical and experimental uncertainties cancel: the overall IS primary $p$ flux normalization uncertainty, the overall detector flux normalization uncertainty, and (at $K\gtrsim$ 500 MeV) diffusive modulation of both fluxes (see below). The first two Western experiments (those of Golden [*et al.*]{} and Buffington [*et al.*]{}) detected $\pbar$ signals at a level higher than the standard secondary prediction. [@Golden; @Buffington] These early experiments detected $\pbar$’s by energy calorimetry (the deceleration and annihilation of the $\pbar$’s in the balloon), but lacked definite identification by a magnetic spectrometer. Of particular concern is the background of kaons in the detector, as $m_K\lesssim m_p$. Stimulated by the possibility of an excess of CR $\pbar$’s, a number of groups completed measurements in the 1970s and 1980s with better particle identification. The PBAR and LEAP groups established upper limits on the CR $\pbar$ flux contradicting the first-generation experiments. [@PBAR; @LEAP] Roughly contemporaneous, the Soviet group of Bogomolov [*et al.*]{} reported three flux measurements (from the periods 1972-77, 1984-85, and 1986-88) consistent with standard secondary predictions. [@Bogo1; @Bogo2; @Bogo3] Abundance & Spectrum of Antiprotons ----------------------------------- The third generation of experiments came in the 1990s and included markedly better particle detection by magnetic spectrometer, of quality comparable to accelerator experiments. From 1991 to 1997, the MASS (1991), IMAX (1992), CAPRICE (1994), and BESS (1993, 1995, 1997) collaborations have made clean measurements of the CR $\pbar$ flux with low backgrounds. [@MASS91a; @IMAX; @BESS93; @CAPRICE94; @BESS95; @BESS97] ----------------------------------- -- ---------------- ----------- ------------- -------- ----------- ---------------------------------------- --------------------- Experiment Field Flight KE Range Cand- Back- Observed Predict- Pol.$^{\rm a}$ Date (GeV) idates ground $\pbar /p$ Ratio tion$^{\rm b}$ Golden et al. 1979$^\dag$ $+$ June 1979 5.6 – 12.5 46 18.3 $(5.2\pm 1.5)\times 10^{-4}$ – Bogomolov et al. 1979$^\dag$ $+$ 1972-1977 2.0 – 5.0 2 – $(6\pm 4)\times 10^{-4}$ – Bogomolov et al. 1987$^\ddag$ $-$ 1984-1985 0.2 – 2.0 1 – $(6^{+14}_{-5})\times 10^{-5}$ – Bogomolov et al. 1990$^\ddag$ $-$ 1986-1988 2.0 – 5.0 3 – $(2.4^{+2.4}_{-1.3})\times 10^{-4}$ – MASS91 [@MASS91a] $+$ Sep. 1991 3.70–19.08 11 3.3 $(1.24^{+0.68}_{-0.51})\times 10^{-4}$ $1.3\times 10^{-4}$ IMAX$^{\ddag}$ [@IMAX] $+$ July 1992 0.25 – 1.0 3 0.3 $(3.14^{+3.4}_{-1.9})\times 10^{-5}$ $1.5\times 10^{-5}$ IMAX [@IMAX] $+$ July 1992 1.0 – 2.6 8 1.9 $(5.36^{+3.5}_{-2.4})\times 10^{-5}$ $6.5\times 10^{-5}$ IMAX [@IMAX] $+$ July 1992 2.6 – 3.2 5 1.2 $(1.94^{+1.8}_{-1.1})\times 10^{-4}$ $1.1\times 10^{-4}$ BESS93$^{\ddag}$ [@BESS93] $+$ July 1993 0.20 – 0.60 7 $ $(5.2^{+4.4}_{-2.8})\times 10^{-6}$ $8.9\times 10^{-6}$ \sim 1.4$ CAPRICE [@CAPRICE94] $+$ Aug. 1994 0.6 – 2.0 4 1.5 $(2.5^{+3.2}_{-1.9})\times 10^{-5}$ $3.5\times 10^{-5}$ CAPRICE [@CAPRICE94] $+$ Aug. 1994 2.0 – 3.2 5 1.3 $(1.9^{+1.6}_{-1.0})\times 10^{-4}$ $1.1\times 10^{-4}$ BESS95$^{\ddag}$$^\ast$ [@BESS95] $+$ July 1995 0.175 – 0.3 3 0.17 $(7.8^{+8.3}_{-4.8})\times 10^{-6}$ $-$ BESS95$^{\ddag}$$^\ast$ [@BESS95] $+$ July 1995 0.3 – 0.5 7 0.78 $(7.4^{+4.7}_{-3.3})\times 10^{-6}$ $1.1\times 10^{-5}$ BESS95$^{\ast}$ [@BESS95] $+$ July 1995 0.5 – 0.7 7 1.4 $(7.7^{+5.3}_{-3.7})\times 10^{-6}$ $5.5\times 10^{-6}$ BESS95$^{\ast}$ [@BESS95] $+$ July 1995 0.7 – 1.0 11 2.8 $(1.01^{+5.7}_{-4.3})\times 10^{-5}$ $1.3\times 10^{-5}$ BESS95$^{\ast}$ [@BESS95] $+$ July 1995 1.0 – 1.4 15 3.5 $(1.99^{+0.91}_{-0.73})\times 10^{-5}$ $3.1\times 10^{-5}$ ----------------------------------- -- ---------------- ----------- ------------- -------- ----------- ---------------------------------------- --------------------- : Summary of cosmic ray $\pbar$ measurements not contradicted by later experiments. Adapted from Geer and Kennedy. [@GeerKenn]\[tab:Summary\] \ \ \[balloon\_tab\] The analysis presented here in based on all refereed and published measurements not contradicted by later measurements with better detectors (Table \[tab:Summary\]). [@GeerKenn] Figure \[fig:unmodRatio\] shows the selected measurements compared with the ILDM prediction for IS fluxes. The disagreement evident in the figure is explicable by heliospheric modulation. Figure \[fig:modRatio\] compares the modulated ILDM predictions with the measured fluxes. This figure makes the comparison by renormalizing the measured fluxes to a single epoch (July 1995, chosen as roughly the most recent heliomagnetic minimum) and using the prediction for that epoch. Our analysis did not use measurements with $K <$ 500 MeV because of the large and difficult-to-calculate diffusion modulation in that energy range. Implications [@GeerKenn; @KennCPT] ---------------------------------- The most basic result implied by Table \[tab:Summary\] and Figure \[fig:modRatio\] is the standard $\pbar$ secondary flux alone, from a realistic ILDM, can account for the observed flux in the relevant energy range, within uncertainties. If variant Galactic transport mechanisms (such as reacceleration or shrouded sources [@Gaisser]) or exotic $\pbar$ sources are at work in this $K$ range, their effects are too small to see at this time. (A hint of reacceleration may be visible in the range $K\simeq$ 2–5 GeV by distortion of the spectrum evident in Figure \[fig:modRatio\], but the effect is not significant within uncertainties.) A second, less obvious, result is a limit on the intrinsic decay lifetime of the antiproton: $\tau_{\pbar} >$ 0.8 Myr, the best limit currently feasible. While the exclusion of the $K <$ 500 MeV spectrum does [*not*]{} significantly affect the $\tau_{\pbar}$ limit, it does limit conclusions about the absence of exotic $\pbar$ sources, as these would have their largest effect relative to the standard secondaries precisely at such low $K$. $n$ $M_X$ (GeV) $k$ $M_Y$ (GeV) ----- ------------------- ----- ------------------- 5 $2\times 10^{19}$ 1 $2\times 10^{63}$ 6 $4\times 10^9$ 2 $5\times 10^{31}$ 7 $3\times 10^6$ 3 $1\times 10^{21}$ 8 $6\times 10^4$ 4 $7\times 10^{15}$ 9 $7\times 10^3$ 5 $5\times 10^{12}$ 10 $2\times 10^3$ 6 $4\times 10^{10}$ : $CPT$– and $B$–violating scale limits associated with $p$ lifetime $\taup$ = $10^{32}$ yr and $\pbar$ lifetime $\tpbar$ = $10^7$ yr (see text). [@KennCPT] \[tab:CPTscales\] A short $\pbar$ lifetime $\tau_{\pbar}\lesssim$ 10 Myr (Galactic CR storage time) would of course indicate CPT violation. The two pictures of CPT violation introduced in subsection \[sec:cpt\] are: modification of LRQFT within ordinary quantum mechanics, and non-standard quantum mechanics (NSQM) with non-unitary time evolution. If only one new mass scale is relevant to the CPT violation, lower limits can be placed on such scales. In Table \[tab:CPTscales\], the limiting CPT-violating scales associated with modified QFT $(M_X)$ and NSQM $(M_Y)$ are shown, assuming $\tau_{\pbar}$ = 10 Myr. The $\pbar$ lifetime is assumed related to each scale by simple mass dimensions. For modified QFT, $\Gamma_{\pbar}$ = $m_p(m_p/M_X)^n$; while for NSQM, $\Gamma_{\pbar}$ = $(m_p/2)(m_p/M_Y)^k$. It is interesting to note that the [*largest*]{} $M_X$ lower bound is ${\cal O}(M_{\rm Pl})$, while the scales of order the “intermediate” scale $(10^8$–$10^{12}$ GeV) are possible, as well as scales $\sim$ TeV. The last scale may not be unreasonably low in the context of “large” extra dimensional gravity. [@LargeExtraDim] Future Developments and Prospects ================================= Uncertainties intrinsic to cosmic ray analysis will probably limit deduction of antimatter properties to about the level already achieved. But the search for exotic sources of primary $\pbar$’s is still open, especially at low energy. More and Better Measurements ---------------------------- Future measurement of the medium energy range $(K$ = 0.5–10 GeV) will define that part of the spectrum better, but it the spectral shape at the two extremes that is critical for exotic $\pbar$ searches. A number of experiments have already taken recent data not yet published. These include the CAPRICE (1998) and HEAT (1999) balloons, as well as the prototype AMS (1998) and PAMELA (1995 and 1997) systems tested on Space Shuttle STS-91 and the Mir space station, respectively. [@SoonPubExps; @UpcomingExps] These experiments can and have searched for positrons and $A >$ 1 antinuclei as well. The HEAT-$\pbar$99 data are especially of interest because of their large energy range ($K$ = 4–50 GeV). The PAMELA instrument, after being tested in prototype on the Mir space station, is scheduled to fly on an unmanned satellite (the Russian-Italian Resurs-Arktika 4) for three years, starting in 2002. [@UpcomingExps] It can detect $e^+$, $\pbar$, and $\overline{\rm He}$ at a relative sensitivity of better than one part in $10^7$ over a range $K$ = 0.1–150 GeV. The full AMS instrument is scheduled for the International Space Station Alpha starting in 2005, also for three years, with an antiparticle/antinucleus sensitivity of one part in $10^6$ for $E >$ 5 GeV. The MASS91 collaboration have also reanalyzed their data and released a new version divided into three energy bins, instead of one. [@MASS91b] These three experiments (MASS91, HEAT, and PAMELA) will decisively address the paucity of data at the highest energies and define the spectrum in that range. Production & Propagation: Importance of the Low-Energy Spectrum {#sec:FutureTheo} --------------------------------------------------------------- The low-energy range is already being mapped out by the BESS experiment, in particular in the 1995 and 1997 data sets. Repeated, reliable measurement of the low-energy spectrum is the most important task in the contemporary period of $\pbar$ measurements, followed closely by reliable measurement of the high-energy fall-off. The presence of exotic primary $\pbar$’s in this range should be detectable with the current or next generation of experiments. The main obstacles to conclusive limits on a non-standard $\pbar$ flux at low energy are now theoretical. There are two crucial effects needing clarification for such a signal to be found or ruled out. The first is the “subthreshold” $\pbar$ production on IS He-4 target nuclei. The status of previous estimates of this effect [@WebbPot; @GaisSch] has been changed in the last decade by laboratory measurements of the $\pbar$ production on heavy target nuclei, [@Sibir] providing evidence for a scaling relation between the $A >$ 1 and $A$ = 1 cases. Recent calculations [@Ullio] have begun to take account of these data, and further work is under way to develop a simple nuclear model. [@KennMillWP] The second is providing a complete heliospheric modulation calculation that includes diffusion, as well as the wind and magnetic drift. The present gap in the literature is defined on one side by thorough modulation calculations applied to low-energy CRs $(K <$ 100 MeV) [@transportNew] and on the other by accurate modulation done at higher energies $(K >$ 500 MeV) without diffusion. [@GeerKenn] Approximate calculations without magnetic drift are available, [@transportOld] but the charge-dependent magnetic drift is essential to predicting the $\pbar /p$ ratio correctly. A full calculation covering $K\lesssim$ 100 MeV to 500 MeV is essential to proper interpretation of the BESS data. [@GeerKennWP] Cosmic ray antimatter measurements are undergoing exciting developments that will define much of our future understanding of the composition of our Galaxy and of basic symmetries of Nature. Perhaps within 10 years, precise cosmic ray measurements will be a mature subject, along with the ripening of other types of particle astrophysics. Acknowledgments {#acknowledgments .unnumbered} =============== The author thanks the PASCOS 99 meeting organizers at UC Davis for the opportunity to present these results, based on work done in collaboration with Stephen Geer (Fermilab). This work was supported by the U.S. DOE under grants DE-FG02-97ER41029 (Univ. Florida) and DE-AC02-76CH03000 (Fermilab), NASA under grant NAG5-2788 (Fermilab), and the Institute for Fundamental Theory at the Univ. Florida, and was greatly enhanced by conversations with J. .R. Jokipii (Lunar & Planetary Laboratory, Univ. Arizona), E. J. Smith (Jet Propulsion Laboratory), V. A. Kosteleck' y (Indiana Univ.), and M. Kamionkowski (CalTech). The author is also grateful to the NASA/Fermilab Theoretical Astrophysics group and the Telluride Summer Research Center for their hospitality. [99]{} P. C. Frisch, ; ACE spacecraft: http://www.srl.caltech.edu/ACE/ . PAMELA experiment: http://msia02.msi.se/group\_docs/astro/research/ PAMELA.html . T. K. Gaisser and E. H. Levy, . S. A. Stephens, ; S. A. Stephens and R. L. Golden, . W. R. Webber and W. S. Potgieter, . T. K. Gaisser and R. K. Schaefer, . T. K. Gaisser, [*Cosmic Rays and Particle Physics*]{} (Cambridge University Press, Cambridge, 1990) chapter 10; J. F. Ormes [*et al*]{}, in [*Proc. 1994 Snowmass Summer Study: Particle and Nuclear Astrophysics and Cosmology in the Next Millenium*]{}, ed. E. W. Kolb and R. D. Peccei (World Scientific, Singapore, 1995) 312. W. R. Webber, M. A. Lee, and M. Gupta, . M. .S. Longair, [*High Energy Astrophysics*]{}, 2nd ed. (Cambridge University Press, Cambridge, 1992), 2 vols. P. Chardonnet [*et al*]{}, ; A. Bottino [*et al*]{}, . E. N. Parker, [*Interplanetary Dynamical Processes*]{} (Interscience Publishers, New York, 1963). L. A. Fisk and W. I. Axford, ; L. A. Fisk, . J. R. Jokipii and B. Thomas, ; J. K' ota and J. R. Jokipii, ; L. A. Fisk [*et al.*]{}, . IMP-8 spacecraft: http://nssdc.gsfc.nasa.gov/space/imp-8.html . Ulysses spacecraft: http://ulysses.jpl.nasa.gov/ . H. Bloemen, . G. Steigman, ; W. H. Kinney, E. W. Kolb, and M. S. Turner, . E. W. Kolb and M. S. Turner, [*The Early Universe*]{} (Addison-Wesley, Redwood City, CA, 1990) chapter 9; D. Spergel, in [*Unsolved Problems in Astrophysics*]{}, ed. J. N. Bahcall and J. P. Ostriker (Princeton University Press, Princeton, 1997) chapter 11. J. Ellis, J. S. Hagelin, D. V. Nanopoulos, K. Olive, and M. Srednicki, . J. Silk and M. Srednicki, . G. Jungman and M. Kamionkowski, ; G. Jungman, M. Kamionkowski, and K. Griest, . S. W. Hawking, . L. Bergstr" om, J. Edsj" o, and P. Ullio, LANL astro-ph/9902012; and in [*Proc. 26th Intl. Cosmic Ray Conf. (ICRC), Salt Lake City*]{}, ed. D. Kieda, M. Salamon, and B. Dingus (American Institute of Physics, College Park, MD, 2000), vol. 2, 285. P. Kiraly [*et al*]{}, . E. Perez, Y. Sirois, and H. Dreiner, in [*Proc. Workshop Future Physics at HERA*]{}, ed. G. Ingelman, A. De Roeck, and R. Klanner (DESY, Hamburg, 1997), vol. 1, 295, LANL hep-ph/9703444. EGRET telescope: http://cossc.gsfc.nasa.gov/code661/code661.html . C. Caso [*et al*]{} (Particle Data Group), and http://pdg.lbl.gov/ . S. H. Geer and D. C. Kennedy, LANL astro-ph/9809101; and . S. Geer [*et al*]{} (APEX Collaboration), ; M. Hu [*et al*]{} (APEX Collaboration), ; S. Geer [*et al*]{} (APEX Collaboration), . R. F. Streater and A. S. Wightman, [*PCT, Spin and Statistics, and All That*]{} (W. A. Benjamin, New York, 1964). , ed. V. A. Kosteleck' y (World Scientific, Singapore, 1999). D. Colladay and V. A. Kosteleck' y, and . V. A. Kosteleck' y and R. Potting, . J. Polchinski, [*String Theory*]{} (Cambridge University Press, Cambridge, 1998) chapter 18. S. W. Hawking, ; D. N. Page, ; T. Banks, L. Susskind, and M. E. Peskin, ; J. Ellis [*et al*]{}, ; P. Huet and M. E. Peskin, . R. Adler [*et al*]{} (CPLEAR Collaboration) and J. Ellis [*et al*]{}, . D. C. Kennedy, . J. D. Lykken, ; N. Arkani-Hamed, S. Dimopoulos, and G. Dvali, ; L. Randall and R. Sundrum, . R. L. Golden [*et al*]{}, . A. Buffington [*et al*]{}, . S. P. Ahlen [*et al*]{} (PBAR Collaboration), ; M. H. Salamon [*et al*]{} (PBAR Collaboration), . R. E. Streitmatter [*et al*]{} (LEAP Collaboration), in [*Proc. 21st Intl. Cosmic Ray Conf. (ICRC), Adelaide*]{} (Univ. Adelaide, Adelaide, 1990), vol. 3, 277. E. A. Bogomolov [*et al*]{}, in [*Proc. 16th Intl. Cosmic Ray Conf. (ICRC), Kyoto*]{} (Inst. Cosmic Ray Research, Tokyo, 1979), vol. 1, 330. E. A. Bogomolov [*et al*]{}, in [*Proc. 20th Intl. Cosmic Ray Conf. (ICRC), Moscow*]{} (Nauka, Moscow, 1987), vol. 2, 72. E. A. Bogomolov [*et al*]{}, in [*Proc. 21st ICRC*]{}, vol. 3, 288. M. Hof [*et al*]{} (MASS Collaboration), . J. W. Mitchell [*et al*]{} (IMAX Collaboration), . K. Yoshimura [*et al*]{} (BESS Collaboration), ; A. Moiseev [*et al*]{} (BESS Collaboration), . M. Boezio [*et al*]{} (CAPRICE Collaboration), . H. Matsunaga [*et al*]{} (BESS Collaboration), . S. Orito [*et al*]{} (BESS Collaboration), . Flown experiments soon to be published: C. R. Bower [*et al*]{} (HEAT Collaboration), M. L. Ambriola [*et al*]{} (CAPRICE Collaboration), and J. Alcaraz [*et al*]{} (AMS Collaboration), in [*Proc. 26th ICRC*]{}, vol. 5, resp. 13, 17, and 88. Experiments soon to fly: S. Ahlen [*et al*]{} (AMS Collaboration), ; O. Adriani [*et al*]{} (PAMELA Collaboration), in [*Proc. 26th ICRC*]{}, vol. 5, 96. G. Basini [*et al*]{} (MASS Collaboration), in [*Proc. 26th ICRC*]{}, vol. 3, 77. A. Sibirtsev [*et al*]{}, D. C. Kennedy and G. Miller, work in progress. S. H. Geer and D. C. Kennedy, work in progress. [^1]: Also ignored here are “pickup ions” or anomalous Galactic cosmic rays, neutral IS atoms which drift into the solar system and are then ionized by solar UV radiation. [^2]: All lifetime limits quoted here are at 90% C.L.
{ "pile_set_name": "ArXiv" }
--- abstract: 'A/B testing refers to the statistical procedure of conducting an experiment to compare two treatments, A and B, applied to different testing subjects. It is widely used by technology companies such as Facebook, LinkedIn, and Netflix, to compare different algorithms, web-designs, and other online products and services. The subjects participating these online A/B testing experiments are users who are connected in different scales of social networks. Two connected subjects are similar in terms of their social behaviors, education and financial background, and other demographic aspects. Hence, it is only natural to assume that their reactions to the online products and services are related to their network adjacency. In this paper, we propose to use the conditional auto-regressive model to present the network structure and include the network effects in the estimation and inference of the treatment effect. A D-optimal design criterion is developed based on the proposed model. Mixed integer programming formulations are developed to obtain the D-optimal designs. The effectiveness of the proposed method is shown through numerical results with synthetic networks and real social networks.' bibliography: - 'projectREF.bib' --- [**Keywords:**]{} A/B testing; Conditional auto-regressive model; D-optimal design; Mixed integer programming; Social network. Introduction ============ The theory of A/B testing dates back to Ronald Fisher’s experiments at the Rothamsted Agricultural Experimental Station in England in the 1920s [@yates1964sir]. A standard statistical testing framework is the Rubin causal model [@rubin1974estimating] usually used to conduct and analyze A/B testing experiments. A key assumption made in the Rubin causal model is the Stable Unit Treatment Value Assumption (SUTVA), which states that the behavior of each test subject in the experiment depends only on the individual treatment and not on the treatments of others, i.e., the test subjects are independent. Recently, A/B testing has been widely used online to test which alternative or treatments out of the two, A or B, leads to better outcomes. The treatments could be options of online commercial, web page designs, different recommendation algorithms, or any new online features that need to be evaluated so that the companies can make informed decisions. The response measures of the experiments can be numerical values of profits, sales, return on investment, click through rate, etc. Usually, the participants of the experiments are sampled from a much bigger population of users. Then the experimenter randomly assigns those subjects to either the treatment or control group. This procedure works well when the subjects can be considered independent of each other. However, in a social network environment, a user is more likely to adopt a new product or service if people around him/her adopt it too. An individual’s behavior can have a non-trivial effect on his/her social network. This effect is called network effect, also known as social interactions, peer influence, or social interference [@aronow2012estimating; @eckles2017design; @athreya2017statistical]. In an A/B testing experiment, this implies that if the treatment has a significant impact on a subject, the effect would reach his/her social circles, regardless whether his/her neighbors are in the treatment or control group. To account for network connections in causal analyses, researchers usually work with two specific settings, network interference, and network-correlated outcomes. When the network interference is present, the outcome of node $i$ (or user $i$) is a function of the treatment assigned to node $i$ and the treatment assigned to other nodes that are related to node $i$ through the network, and possibly the observed outcomes of these related nodes. For the network-correlated outcomes, the outcomes of the neighboring nodes are correlated because the features of the two connected nodes are more similar than those of the unconnected nodes [@basse2017limitations; @basse2018model]. This paper focuses on network-correlated outcomes. Under this setting, we assume that the A/B testing outcomes of adjacent nodes are positively correlated due to their similarities. An important question in A/B testing is how to allocate treatments to the subjects. Different from the SUTVA assumption, a random design which randomly assigns the treatment settings with equal probability to each user may not be efficient in estimating the treatment effect in the presence of the network-correlated outcomes. Cluster-based randomized treatment allocation has been used to block the effect of network correlation in A/B testing experiments. One such example can be found in [@xu2015infrastructure]. Also, [@saveski2017detecting] and [@pouget2017testing] used the cluster-based random design to determine the existence of the network effect. [@basse2018model] proposed the restricted randomization approach to minimize the mean squared error of the estimated treatment effect. Based on a normal-sum model, the analytical decomposition of the mean square error provided insights to develop the restricted randomization strategies in the absence of a detailed network structure. Although these cluster-based random designs are simple to use, they might not be able to achieve a perfect balance between the two treatment groups in terms of their network structures. If a reasonable model can be assumed for the effects of the treatment and network [@chen2018sequential], the classic model-based optimal design [@atkinson2007optimum; @wu2011experiments] can also be used for A/B testing experiments. Unfortunately, there has not been much development in this direction and we decide to fill the gap. In this paper, we focus on the construction of A/B testing experimental designs for network-correlated outcomes when users who are connected in a network share some common social and demographic backgrounds. We propose a spatial network model for A/B testing, called conditional auto-regressive model or CAR [@schmidt2014conditional] to incorporate the correlated network structure in the analysis. To accurately estimate the treatment effect, we use the D-optimal criterion [@sitter1995d], which seeks to maximize the determinant of the information matrix of the linear regression model of the response with respect to the treatment effects and other potential variables. Mixed integer programming formulations are developed to optimize the D-optimal criterion and construct the design. Finally, we conduct simulation studies on synthetic and real social networks to demonstrate the performances of the proposed method compared to the random designs, which do not consider the network structure. D-Optimal Design for CAR Model ============================== Network A/B Testing with CAR Model ---------------------------------- We consider an A/B testing experiment conducted on a social network with $n$ nodes. The social network is considered to be an undirected graph in the context of this paper. The edges of this network are recorded by an $n\times n$ adjacent matrix $W$ whose $(i,j)$-th entry is $w_{ij}$. The diagonal entries $w_{ii}$’s of this matrix is 0, whereas off-diagonal entries are $$w_{ij}=\begin{cases} 1, & \text{if node $i$ and node $j$ are adjacent}\\ 0, & \text{otherwise}. \end{cases}$$ Two adjacent nodes are the ones connected by an edge. The experimental design is the plan to allocate A or B treatment to each node. Let $x_i\in \{1,-1\}$ for $i=1, \ldots, n$ be the design of the $i$-th node and the two settings $\{1,-1\}$ represent A and B treatments. Denote the scalar response observation of the $i$-th node by $y_i$. In this paper, we focus on the case that the response is continuous. Assume a linear regression model for the response as follows. $$\label{eq:gcar} y_i=\beta_0+x_i\beta+\delta_i,$$ where $\beta_0$ is the intercept, $\beta$ represents the treatment effect, and $\delta_i$ is a zero mean random variable. Under the SUTVA assumption, $\delta_i$’s are assumed to be the random noise and independent with each other. But for the experiments on networks, two connected users share similarities in their social behaviors and other backgrounds, and thus their responses are often correlated. To incorporate this social correlation, we model $\delta_i$ in by the conditional auto-regressive (CAR) model [@besag1974spatial] $$\label{eq:delta} \delta_i|\boldsymbol{\delta}_{-i}\sim N\left(\rho\sum_{j\neq i}\frac{w_{ij}\delta_j}{m_i}, \frac{\sigma^2}{m_i}\right),$$ where $\boldsymbol{\delta}_{-i}=\{\delta_1, \ldots, \delta_{i-1}, \delta_{i+1}, \ldots, \delta_n\}$, $m_i=\sum^n_{j=1}w_{ij}$ is the number of nodes adjacent to the $i$-th node, $\sigma^2$ is the variance parameter that is assumed to be a constant in our scope, and $|\rho|<1$ is the correlation parameter of the CAR model. If $\rho=0$, $\delta_i$’s are independent with each other, which corresponds to the extreme case when the network only has $n$ nodes but without any edges. As noted in the Introduction, the connected users tend to have similar reactions to the same treatment. Hence, without loss of generality, we restrict that the correlation parameter is non-negative, i.e., $0\leq\rho<1$. By the Brook’s Lemma [@brook1964distinction], $\boldsymbol{\delta}=(\delta_1, \ldots, \delta_{n})^\top$ follows a multivariate normal distribution: $$\label{eq:mvn} \boldsymbol{\delta}\sim \mathcal{MVN}_n(0, \sigma^2(D-\rho W)^{-1})$$ where $D=\mathrm{diag}(m_1, \ldots, m_n)$. The detailed derivation to is deferred to Appendix \[sec:A1\]. The maximum likelihood method can be used to fit the model with $\delta_i$ from and estimate the model parameters. The goal of A/B testing is to accurately assess the treatment effect $\beta$. Next, we determine values of $x_i$’s using D-optimal design to improve the accuracy of the estimate $\hat\beta$. D-optimal Design for CAR Model ------------------------------- Given the correlation parameter $\rho$, the parameters $\beta_0$ and $\beta$ in can be estimated by $$(\hat\beta_0, \hat\beta)^\top=(X^\top V^{-1} X)^{-1} X V^{-1} {\mbox{\boldmath$ y $\unboldmath}},$$ where $X$ is a $n\times 2$ design matrix with $i$-th row $(1, x_i)$, ${\mbox{\boldmath$ y $\unboldmath}}=(y_1, \ldots, y_n)^\top$, and $V=(D-\rho W)^{-1}$, and the variance matrix of $(\hat\beta_0, \hat\beta)^\top$ is $$\label{eq:varm} \mathrm{Var}\left\{(\hat\beta_0, \hat\beta)^\top\right\}=\sigma^2(X^\top V^{-1}X)^{-1}=\sigma^2(X^\top (D-\rho W)X)^{-1}.$$ Under our model assumption in , the D-optimal design is determined by maximizing the determinant of the matrix $X^\top (D-\rho W)X$ in with respect to ${\mbox{\boldmath$ x $\unboldmath}}=(x_1, \ldots, x_n)^\top\in \{-1, 1\}^n$, i.e., $$\label{eq:dopt} \mathrm{argmax}_{{\mbox{\boldmath$ x $\unboldmath}}\in \{-1,1\}^n} \{D({\mbox{\boldmath$ x $\unboldmath}}):=|X^\top (D-\rho W)X|\}.$$ Notice that, $$\left\{X^\top (D-\rho W)X\right\}^{-1} = \begin{bmatrix} \sum\limits_i m_i-\rho \sum\limits_i \sum\limits_j w_{ij} & \sum\limits_i m_i x_i-\rho \sum\limits_i \sum\limits_{j>i} w_{ij} (x_i + x_j)\\ \sum\limits_i m_i x_i-\rho \sum\limits_i \sum\limits_{j>i} w_{ij} (x_i + x_j) & \sum\limits_i m_i x_i^2-\rho \sum\limits_i \sum\limits_j w_{ij} x_i x_j \end{bmatrix}^{-1}$$ $$= \frac{1}{D({\mbox{\boldmath$ x $\unboldmath}})} \begin{bmatrix} \sum\limits_i m_i x_i^2-\rho \sum\limits_i \sum\limits_j w_{ij} x_i x_j & -(1-\rho)\sum\limits_i m_i x_i\\ -(1-\rho)\sum\limits_i m_i x_i & (1-\rho)\sum\limits_i m_i \end{bmatrix}.$$ Then we obtain the following $$\label{eq:dx} D({\mbox{\boldmath$ x $\unboldmath}}) = (1-\rho)\sum\limits_i m_i (\sum\limits_i m_i -\rho \sum\limits_i \sum\limits_j w_{ij} x_i x_j) - (1-\rho)^2(\sum\limits_i m_i x_i)^2,$$ and $$\mathrm{Var}(\hat{\beta})=\frac{\sigma^2(1-\rho)\sum\limits_i m_i}{D({\mbox{\boldmath$ x $\unboldmath}})}.$$ Given a network, $\sigma^2(1-\rho)\sum\limits_i m_i$ is a constant. Therefore, the D-optimal design also minimizes the variance of the treatment effect. Based on , Proposition \[prop:dopt\] gives a simplified objective function to obtain the D-optimal design. \[prop:dopt\] Under model assumptions and , the D-optimal design ${\mbox{\boldmath$ x $\unboldmath}}$ is the solution of the following optimization problem. $$\label{eq:obj} \mathrm{argmax}_{{\mbox{\boldmath$ x $\unboldmath}}\in \{-1,1\}^n} D({\mbox{\boldmath$ x $\unboldmath}}) = \mathrm{argmin}_{{\mbox{\boldmath$ x $\unboldmath}}\in \{-1,1\}^n} \left\{a \sum\limits_i \sum\limits_j w_{ij} x_i x_j + \left(\sum\limits_i m_i x_i\right)^2\right\},$$ where $a=\frac{\rho}{1-\rho}\sum^n_{i=1} m_i$. The proof of this proposition is provided in the Appendix A2. Since $a$ is non-negative, a lower bound of the objective function in can be attained by minimizing $\sum\limits_i \sum\limits_j w_{ij} x_i x_j$ and $\left(\sum\limits_i m_i x_i\right)^2$, respectively. The lower bound of $\sum\limits_i \sum\limits_j w_{ij} x_i x_j$ is $-\sum\limits_i\sum\limits_j w_{ij}$ if $x_i\neq x_j$ for all $(i,j)$ with $w_{ij}=1$. The lower bound of $\left(\sum\limits_i m_i x_i\right)^2$ is zero if $\sum\limits_i m_i x_i=0$, which represents that the nodes allocated with -1 and the nodes allocated with 1 have equal number of first order neighborhoods. Similarly, we can obtain an upper bound for the D-optimality measure: $$D({\mbox{\boldmath$ x $\unboldmath}})\leq (1-\rho)(\sum^n_{i=1}m_i)^2+(1-\rho)\rho\sum^n_{i=1}m_i\sum^n_{i=1}\sum^n_{j=1}w_{ij}.$$ According to this upper bound, we are able to define a D-efficiency measure: $$\label{eq:d-eff} \frac{D({\mbox{\boldmath$ x $\unboldmath}})}{(1-\rho)(\sum^n_{i=1}m_i)^2+(1-\rho)\rho\sum^n_{i=1}m_i\sum^n_{i=1}\sum^n_{j=1}w_{ij}}.$$ This D-efficiency measure ranges from 0 to 1, which evaluates the quality of the design without concerning the scale of the $D({\mbox{\boldmath$ x $\unboldmath}})$. A larger value of D-efficiency corresponds to a better design. The definition of the D-efficiency is different from the conventional version in literature [@atkinson2007optimum], which would be the $1/p$th root of for $p$ experimental factors. But there is no need to calculate the $1/p$th root because only one experimental factor and no other covariates are involved in the CAR model. Mixed Integer Programming Formulations for D-optimal Design {#sec3} =========================================================== Since the decision space in is $\{-1, 1\}^n$, the optimization problem is an integer programming problem [@nemhauser1988integer]. To solve this problem, this section formulates the D-optimal design problem into mixed integer programming problems with the original D-optimal objective function in and a modified D-optimal objective function. A Mixed Integer Programming Formulation for D-optimal Design {#sec:mip} ------------------------------------------------------------ By observing that $(\sum\limits_i m_i x_i)^2 = \sum\limits_i m_i^2 x_i^2 + \sum\limits_i \sum\limits_{j \ne i} m_i m_j x_i x_j = \sum\limits_i m_i^2 + \sum\limits_i \sum\limits_{j \ne i} m_i m_j x_i x_j$, we express the objective function in by $$a \sum\limits_i \sum\limits_j w_{ij} x_i x_j + \sum\limits_i \sum\limits_{j \ne i} m_i m_j x_i x_j = \sum\limits_i \sum\limits_{j \ne i} b_{ij} x_i x_j,$$ where $b_{ij} = a w_{ij} + m_i m_j $. By introducing new variables $v_i=(x_i+1)/2$ and $u_{ij}=v_iv_j$, we can formulate the original optimization problem into the following mixed integer linear program (MIP) problem . Although $u_{ij}$ is not restricted to be binary, its definition makes $u_{ij}$ can only be 0 or 1. $$\begin{aligned} \label{eq:omip} \mathrm {min} & \left[\sum\limits_i \sum\limits_{j \neq i} b_{ij} u_{ij} - \sum\limits_i \sum\limits_{j \neq i} b_{ij} v_i \right] \\\nonumber \textrm{subject to }& u_{ij} \leq v_i, \quad u_{ij} \leq v_i\\\nonumber & u_{ij} \geq v_i+v_j-1, \\\nonumber & u_{ij} \in \mathbb{R}, \quad u_{ij} \geq 0,\\\nonumber & v_i \in \{0, 1\}, \textrm{ for }i=1,\ldots, n, j=1, \ldots, n. \end{aligned}$$ ![image](Network_Plot_1_v2.pdf) \[fig:plot\] Figure \[fig:plot\] depicts an example of solving this MIP problem to construct design to a network of size 20. Notice that $b_{ij}$ depends on the value of $a$, which is a function of the correlation parameter $\rho$ in . It is impractical to assume that $\rho$ can be accurately estimated before data collection. Using the Bayesian framework, we can derive the Bayesian D-optimal design criterion which is the expectation of the D-optimality $D({\mbox{\boldmath$ x $\unboldmath}})$ with respect to a user-specified prior distribution of $\rho$. From , we can see that the expected D-optimality only depends on the expected value of the ratio $\rho/(1-\rho)$, which essentially is a tuning parameter. Instead of the Bayesian approach, we decide to take an equivalent route by modifying the objective function and moving this tuning parameter into the constraint. A Mixed Integer Programming Formulation for a Modified D-optimality Criterion {#sec:modified-mip} ----------------------------------------------------------------------------- As we point out above, reducing values of $\sum\limits_i \sum\limits_j w_{ij} x_i x_j$ or $(\sum\limits_i m_i x_i)^2$ would improve D-efficiency defined in . To remove the parameter $a$ from the objective function in , an alternative solution is to use $\sum\limits_i \sum\limits_j w_{ij} x_i x_j$ as the objective function, and bound the value of $(\sum\limits_i m_i x_i)^2$ by a constraint. We modify the optimization problem in to be $$\label{prob:mod1} \text{min}_{{\mbox{\boldmath$ x $\unboldmath}}\in \{-1, 1\}^n} \sum\limits_i \sum\limits_j w_{ij} x_i x_j$$ $$\text{s.t. } -\delta \leq\sum\limits_i m_i x_i \leq \delta \text{, and } x_i\in \{-1,1\}~\mathrm{for}~i=1, \ldots, n,$$ where $\delta>0$ is a tuning parameter. For different networks, the ranges of $\sum\limits_i m_i x_i$ can be different. Now we discuss how to specify the value of $\delta$ by normalizing it to a unified range for different networks. Assume that $x_i$ is a random variable taking value from $\{-1, 1\}$ with equal weights. Hence, $$\mathrm{E}\left(\sum^n_{i=1} m_i x_i\right)=\sum^n_{i=1} m_i \mathrm{E} x_i=0 ~\mathrm{and}~ \mathrm{Var}\left(\sum^n_{i=1} m_i x_i\right)=\sum^n_{i=1} m^2_i \mathrm{Var}(x_i)=\sum^n_{i=1} m^2_i.$$ If $n^{-2}\sum^n_{i=1} m^2_i<\infty$ and $n^{-1}m_i\rightarrow 0$ for $i=1,\ldots, n$, we have that $\sum^n_{i=1} m_i x_i/\sqrt{\sum^n_{i=1} m^2_i}$ asymptotically follows a standard normal distribution by the Lindeberg’s Central Limit Theorem. Therefore, as $n\rightarrow\infty$, $$\mathrm{P}\left(-\Phi^{-1}(\alpha)\sqrt{\sum\limits_i m^2_i}<\sum m_i x_i<\Phi^{-1}(\alpha)\sqrt{\sum\limits_i m^2_i}\right)=2\alpha-1,$$ where $\Phi(\cdot)$ is the cumulative distribution function of the standard normal distribution, and $\alpha\in(0.5, 1)$. By $$\delta =\Phi^{-1}(\alpha)\sqrt{\sum\limits_i m^2_i},$$ $\delta$ is increasing with $\alpha$. Define $v_i$ and $u_i$ for $i=1,\ldots,n$ in the same way as in Section \[sec:mip\], the problem in becomes $$\begin{aligned} \label{prob:mod2} \text{min } & \left[ \sum\limits_i \sum\limits_{j \neq i} w_{ij} u_{ij} - \sum\limits_i \sum\limits_{j \neq i} w_{ij} v_i\right] \\\nonumber \textrm{subject to } & \sum\limits_i m_i v_i \leq \frac{1}{2}\left(\sum\limits_i m_i + \delta\right) \\\nonumber & \sum\limits_i m_i v_i \geq \frac{1}{2}\left(\sum\limits_i m_i - \delta\right) \\\nonumber & u_{ij} \leq v_i, \quad u_{ij} \leq v_i\\\nonumber & u_{ij} \geq v_i+v_j-1, \\\nonumber & u_{ij} \in \mathbb{R}, \quad u_{ij} \geq 0,\\\nonumber & v_i \in \{0, 1\}, \textrm{ for }i=1,\ldots, n, j=1, \ldots, n. \end{aligned}$$ Here are some remarks on the two reformulated optimization problems and . First, solving the MIP in Section \[sec:mip\] gives the exact D-optimal design, whereas solving the MIP for the modified problem in Section \[sec:modified-mip\] does not guarantee that the exact D-optimal design can be found. Second, the MIP requires that the correlation parameter $\rho$ to be known, whereas the modified MIP does not. Third, the original numbers of decision variables in both formation are $n(n-1)/2+n$. However, $u_{ij}$ can be removed if the corresponding coefficient ($b_{ij}$ in or $w_{ij}$ in ) is 0. Since $w_{ij}$ is more likely to be zero than $b_{ij}$, we expect that the number of decision variables of the formulation for the modified problems is often much smaller than that of the original one. Although both programming formulations are NP-hard, our observation based on the simulation study in Section \[sec:synthetic\] is that the reduction of the number of decision variables often lead to less computation in solving MIP. Numerical Study on Synthetic Networks {#sec:synthetic} ===================================== This section compares three methods: - Original-MIP: the mixed integer programming formulation for the original D-optimality objective function in Section \[sec:mip\]. - Modified-MIP: the mixed integer programming formulation for the modified D-optimality objective function in Section \[sec:modified-mip\]. - Random: randomly allocate -1 or 1 with equal weights to each node. We generate random networks to compare the three methods. For a network with $n$ nodes, the $n\times n$ adjacent matrix records the edges of this network. We randomly assign 0 and 1 to the upper or lower off-diagonal entries of this matrix. The proportion of ones is specified to be $p$. As defined in Section 2, the zero entry means that the two corresponding nodes are not adjacent, whereas one entry means the opposite. The proportion $p$ is referred to as the density of this network. Once we construct the designs using the three methods for a given network, we compare the designs on three aspects, computational efficiency, D-efficiency, and the empirical variance of the estimated $\beta$ in . In terms of computational efficiency, we compare the computational time of solving the objective functions of original-MIP and modified-MIP with GUROBI solver (<http://www.gurobi.com/>). We set the network correlation coefficient to be $\rho=0.2$ for the original-MIP and the tuning parameter to be $\alpha=0.6$ for the modified-MIP. For both methods, the longest allowable running time is 24 hours. Exceeding that limit, the solver is terminated whether it reaches the optimal solution or not. As pointed out earlier, the number of decision variables of the modified-MIP is likely much smaller than that of original-MIP, thus the modified-MIP would take less time to run than original-MIP. To confirm this, the running time (in seconds) of these two methods with networks of different sizes are given in Figure \[fig:density\]. It shows that the run time required to solve original-MIP increases dramatically as both network size and density increase. For the small networks in this simulation, the original-MIP is already so time-consuming thus it is not practical to be applied to the real-world social networks whose size is usually of thousands. Since the tuning parameter $\alpha$ determines the feasible region of the modified-MIP, we conduct an additional simulation whose results are shown in Table \[tab:alpha\] in Appendix A.3. It implies that the value of $\alpha$ does not affect the computational time of modified-MIP significantly. ![image](figure422.pdf){width="99.00000%"} \[fig:density\] Next, we compare the three design methods using D-efficiency, whose value in depends on the true correlation parameter $\rho$ of the CAR model. Thus, we pick four different values for $\rho$ to 0, 0.1, 0.2, and 0.3, respectively. We randomly generate a network of size $n=50$ and density $p=0.1$. For the modified-MIP, we vary the value of the tuning parameter $\alpha$ from 0.6, 0.7, to 0.8, and the running time of each case is less than one minute in GUROBI. For the random design, because $\mathrm{P}(x_i=1)=\mathrm{P}(x_i=-1)=1/2$, the expected D-efficiency is given by $$\frac{(1-\rho)(\sum^n_{i=1} m_i)^2-(1-\rho)^2\sum^n_{i=1} m^2_i}{(1-\rho)(\sum^n_{i=1}m_i)^2+(1-\rho)\rho\sum^n_{i=1}m_i\sum^n_{i=1}\sum^n_{j=1}w_{ij}}.,$$ which only depends on the network. For the parameter $\rho$ in the original-MIP, we use $\rho=0.2$ (not necessarily equal to the true value of $\rho$ used in computing the D-efficiency) to the generate design. Since the original-MIP can be extremely time-consuming to obtain, we report the solution after 24 hours with an optimality gap of 10.0863%. This optimality gap is defined as $(ub-lb)/ub$, where $ub$ and $lb$ are upper and lower bounds of the objective function. A smaller optimality gap indicates that the objective value corresponding to the current solution is closer to the true optimal objective value. Table \[tab:deff\] gives the D-efficiency values. Both MIP based methods are better than the random design, especially when the correlation parameter is not zero. Also, the D-efficiency of modified-MIP under different $\alpha$ values is comparable with the original-MIP method. Method $\rho = 0$ $\rho = 0.1$ $\rho = 0.2$ $\rho = 0.3$ ------------------------------- ------------ -------------- -------------- -------------- Random 0.98 0.90 0.82 0.76 Original-MIP 1.00 0.96 0.93 0.90 Modified-MIP ($\alpha = 0.6$) 1.00 0.96 0.93 0.90 Modified-MIP ($\alpha = 0.7$) 1.00 0.96 0.92 0.90 Modified-MIP ($\alpha = 0.8$) 1.00 0.96 0.92 0.90 : The D-Efficiency measures of different methods for a network of size $n=50$ with density $p=0.1$[]{data-label="tab:deff"} Since the aim of A/B testing is to accurately estimate the treatment effect $\beta$ in , we assess the empirical variance of $\hat\beta$ to compare the three designs. Using the random network we investigated in Table \[tab:deff\], and a pre-specified correlation coefficient $\rho$, we generate $\pmb\delta$ from the multivariate normal distribution as expressed in . Set the values of the parameters as $\beta_0=0$, $\beta=2$ and $\sigma^2=1$ in the model . Given the design of $x_i$, we generate the response $y_i$ for each node as in . An estimate of $\beta$ can be obtained by fitting a linear regression model (LM) or the CAR model. For the CAR model, the parameters $\beta_0$, $\beta$ and $\rho$ are estimated by the maximum likelihood method using `R` package `spdep` [@bivand2005spdep]. By repeating this procedure 500 times, we measure the accuracy of an estimate by calculating its sample variance : $$\hat{\mathrm V}(\hat\beta)=\frac{1}{499}\sum^{500}_{l=1}\left(\hat\beta_l-\frac{1}{500}\sum^{500}_{l=k}\hat\beta_k\right)^2,$$ where $\hat\beta_1, \ldots, \hat\beta_{500}$ are the 500 copies of the estimates for $\beta$. Under our response generation scheme, the bias of the estimate $\hat\beta$ is tiny for all the methods. Hence, the mean squared error of $\hat\beta$ is dominated by the variance of $\hat\beta$, and the bias is negligible. So we ignore the bias and report empirical variance only. Notice that the modified-MIP and the original-MIP generate deterministic designs for a given network, to the contrary of the random design. To make the three approaches comparable, 100 random designs are generated and we report the average value of $\hat{\mathrm V}(\hat\beta)$ over the 100 random designs. The empirical variance values of $\hat\beta$ for each method are given in Table \[tab:var50\]. As in Table \[tab:deff\], we also vary the value of the tuning parameter $\alpha$ from 0.6, 0.7, to 0.8 for the modified-MIP, and fix $\rho=0.2$ for the original-MIP to generate design. The value of $\rho$ used in design generation is not equal to its real value for generating responses when $\rho=0$, 0.1, or 0.3 in Table \[tab:var50\]. For the random design, we consider the $\beta$ estimated from both CAR and LM, although the data are generated from the CAR model. We summarize our main observations from Table \[tab:var50\] as follows: - For a non-zero correlation coefficient (i.e.,$\rho>0$), the variances of MIP based methods are smaller than the random design. - The variances resulted from modified-MIP are comparable with those from original-MIP. Method **$\rho = 0$** **$\rho = 0.1$** **$\rho = 0.2$** **$\rho = 0.3$** ------------------------------- ---------------- ------------------ ------------------ ------------------ Original-MIP 0.0046 0.0043 0.0041 0.0038 Modified-MIP ($\alpha = 0.6$) 0.0044 0.0043 0.0040 0.0037 Modified-MIP ($\alpha = 0.7$) 0.0049 0.0043 0.0043 0.0038 Modified-MIP ($\alpha = 0.8$) 0.0047 0.0042 0.0042 0.0038 Random (CAR) 0.0047 0.0047 0.0046 0.0046 Random (LM) 0.0045 0.0045 0.0044 0.0045 : Empirical variances of $\hat\beta$ based on designs generated from different methods for a network of size $n=50$ and density $p=0.1$[]{data-label="tab:var50"} In summary, the performance of modified-MIP is comparable with that of original-MIP for small networks. But it is much more practical because it takes less time to generate a design using GUROBI and does not require the correlation coefficient value $\rho$. Even though it does involve the tuning parameter $\alpha$, its the performance is robust to the value of $\alpha$. So we choose $\alpha=0.6$ in remaining numerical examples and only compare the random design and modified-MIP for the real-world networks in Section \[sec:real\]. Numerical Study on Real-world Networks {#sec:real} ====================================== This section studies two examples of real-world networks. Section \[sec:ego\] considers examples of ego networks from Facebook. Section \[sec:cluster\] considers examples of large networks containing multiple clusters. Ego Networks {#sec:ego} ------------ We consider five ego networks extracted from Facebook [@leskovec2012learning]. These networks are collected from survey participants who used the Facebook App. Each ego network consists of a few focal nodes (i.e., egos) and the nodes that are directly connected to the egos. The sizes of these five ego networks are $n=52$, 61, 168, 333 and 224, respectively. The densities of them range from 0.05 to 0.15. We use the modified-MIP with $\alpha=0.6$ to construct the designs. As in Section \[sec:synthetic\], the maximum running time is set to be 24 hours in GUROBI. The optimality gaps of these five networks are 0%, 0%, 10.77%, 36.87%, and 39.45%. Figure \[fig:egonet\] depicts the designs generated from the modified-MIP for the ego networks of size 52 and 61. ![image](Network_Plot_52_v2.pdf) ![image](Network_Plot_61_v2.pdf) \[fig:egonet\] After obtaining designs for each ego network, we evaluate the empirical variances of $\hat\beta$, which is calculated the same as in Section \[sec:synthetic\]. The true correlation parameter $\rho$ is specified to be 0.1, 0.2 or 0.3. Three methods are compared: 1) the modified-MIP with $\hat\beta$ estimated using the CAR model; 2) the random design with $\hat\beta$ estimated using the CAR model; 3) the random design with $\hat\beta$ estimated using a linear regression model (LM). The results are provided in Figure \[fig:var\_comp\]. We can see that in most cases the modified-MIP gives smaller variances than the random design when $\rho$ equals to 0.2 and 0.3. However, the advantage of modified-MIP is not significant when the correlation parameter $\rho$ is as small as $0.1$. ![image](figure5.pdf){width="99.00000%"} \[fig:var\_comp\] Large Networks with Disjoint Clusters {#sec:cluster} ------------------------------------- We consider some sub-networks of the Facebook network available at <https://snap.stanford.edu/data/>. To sample the sub-networks, we randomly selected $K$ clusters, each cluster with approximately $50$ nodes. The $K$ clusters form a network on which we apply our design. We pick $K=20$, $30$, and $40$ and the corresponding networks consist approximately 1000 to 2000 nodes. Then, we simulate the responses for each cluster of the large network. For each cluster, the correlation parameter $\rho$ is randomly generated from a uniform distribution $U(0.15, 0.25)$. Other parameters are specified the same as in Section \[sec:synthetic\]. The empirical variances are computed the same way as in Section \[sec:synthetic\]. The estimate of $\beta$ can be obtained by fitting the CAR model or a linear regression model. Repeating this procedure 100 times, we assess the accuracy of the estimates by their sample variances. For the random design, we obtain the results from 100 random designs and record the average of the sample variances over these 100 random designs. We now discuss how to generate the MIP based designs for such large networks with disjoint clusters. For a network of size over 500, it is nearly impossible to directly obtain the MIP based designs. Since the large network is constructed by disjoint clusters, we can alternatively generate the MIP based design for each smaller cluster, and then combine them to obtain the design for the entire large network. Let $W_1, \ldots, W_K$ be the adjacent matrices of $K$ disjoint clusters. For each of them, we generate the two-level design using the modified-MIP described in Section 3.2. We denote the resulted design as $x_{ik}$ for $k=1,\ldots, K$ and $i=1, \ldots, n_k$, where $x_{ik}\in \{-1, 1\}$ is the treatment assignment for the $i$-th node from the $k$-th network. Therefore, a combined design for the large network can be $\{c_k x_{ik}: i=1, \ldots, n_k, k=1, \ldots K\}$, where $c_k\in \{-1, 1\}$. By varying the choices of $c_k$ for $k=1, \ldots, K$, the total number of combined designs for the large network is $2^K$. We take an additional step to choose the optimal value of $c_k$s by minimizing $$\mathrm{argmin}_{{\mbox{\boldmath$ c $\unboldmath}}\in \{-1, 1\}^K }(\sum^K_{k=1}c_k\sum^{n_k}_{i=1}m_{ik}x_{ik})^2,$$ where ${\mbox{\boldmath$ c $\unboldmath}}=(c_1, \ldots, c_K)$, and $m_{ik}$ is the total number of neighbors for the $i$-th node from the $k$-th network, which leads that $\sum^{n_k}_{i=1}m_{ik}x_{ik}$ is a known constant. The optimal solution of ${\mbox{\boldmath$ c $\unboldmath}}$ gives the smallest value of $(\sum_i m_i x_i)^2$ in for the large network. This optimization problem is the same as original-MIP in Section 3.1. For $K=20$, 30 or 40, this problem can be solved by GUROBI efficiently. For $K=20$, 30 or 40, we generate 100 large networks with disjoint clusters as described above and compare the empirical variances of $\hat\beta$ of the three methods as in Section \[sec:ego\]. The empirical variances of the 100 large networks for each $K$ are depicted in Figure \[fig:box\]. We see that the median of empirical variances of modified-MIP is the smallest among the three methods for all three cases. According to the results from the Wilcox’s rank sum test, the median of modified-MIP is significantly smaller than that of the random design for each case. ![image](cluster.pdf){width="99.00000%"} \[fig:box\] Conclusion ========== This paper proposes using the D-optimal design for A/B testing conducted on a social network. We use the CAR model to characterize the dependence of the responses from the adjacent nodes in a network. Mixed integer programming formulations are proposed to solve the D-optimal objectives and construct the designs. Numerical examples are provided to show the effectiveness of the proposed method. We plan to address the topics including how to apply the D-optimal design on the network for categorical responses, how to include the covariates information in the design on the network and how to construct sequential design on the network in the future. Appendix ======== The derivation of the distribution $\pmb{\delta}$ {#sec:A1} -------------------------------------------------- According to Brook’s Lemma, for any ${\mbox{\boldmath$ Y $\unboldmath}} \in \mathbb{R}^n$, such that ${\mbox{\boldmath$ Y $\unboldmath}} = (Y_1, Y_2, ... , Y_n)$ and any ${\mbox{\boldmath$ Y $\unboldmath}}_0 \in \mathbb{R}^n$, such that ${\mbox{\boldmath$ Y_0 $\unboldmath}} = (Y_{01}, Y_{02}, ... , Y_{0n})$, where $Y_1, Y_2, ... , Y_n $ are random variables, and $Y_{01}, Y_{02}, ... , Y_{0n}$ is a copy of realizations of $Y_1, Y_2, ... , Y_n $, we have the following result: $$\frac{P({\mbox{\boldmath$ Y $\unboldmath}})}{P({\mbox{\boldmath$ Y $\unboldmath}}_0)} \propto \prod_{i=1}^{n} \frac{P(Y_i|Y_{01}, ..., Y_{0i-1}, Y_{i+1}, ..., Y_n)}{P(Y_{0i}|Y_{01}, ..., Y_{0i-1}, Y_{i+1}, ..., Y_n)}$$ Let $\tilde{w}_{ij}=w_{ij}/m_i$ and $\sigma^2_i=\sigma^2/m_i$. Under , we have that $$\begin{aligned} \frac{P({\mbox{\boldmath$ \delta $\unboldmath}})}{P({\mbox{\boldmath$ \delta $\unboldmath}}_0)} & \propto \prod_{i=1}^{n} \frac{P(\delta_i|\delta_{01}, ..., \delta_{0i-1}, \delta_{i+1}, ..., \delta_n)}{P(\delta_{0i}|\delta_{01}, ..., \delta_{0i-1}, \delta_{i+1}, ..., \delta_n)} \propto \prod_{i=1}^{n} \frac{\exp\{-\frac{1}{2\sigma_i^2}(\delta_i-\rho\sum_{j> i}\tilde{w}_{ij}\delta_j)^2\}}{\exp\{-\frac{1}{2\sigma_{0i}^2}(0-\rho\sum_{j> i}\tilde{w}_{ij}\delta_j)^2\}}\\ & \propto \exp\{ \sum_{i=1}^{n}[-\frac{1}{2\sigma_i^2}(\delta_i-\rho\sum_{j> i}\tilde{w}_{ij}\delta_j)^2 + \frac{1} {2\sigma_i^2} (\rho\sum_{j> i}\tilde{w}_{ij}\delta_j)^2] \} \\ & \propto \exp\{ \sum_{i=1}^{n}[-\frac{1}{2\sigma_i^2}(\delta_i^2-2\rho\delta_i\sum_{j> i}\tilde{w}_{ij}\delta_j + (\rho\sum_{j> i}\tilde{w}_{ij}\delta_j)^2 -(\rho\sum_{j> i}\tilde{w}_{ij}\delta_j)^2)] \} \\ & \propto \exp\{ \sum_{i=1}^{n}[-\frac{1}{2\sigma_i^2}(\delta_i^2-2\rho\delta_i\sum_{j> i}\tilde{w}_{ij}\delta_j)] \} \propto \exp\{ -\frac{1}{2\sigma^2} \sum_{i=1}^{n} m_i (\delta_i^2-2\rho\delta_i\sum_{j> i}\tilde{w}_{ij}\delta_j) \} \\ & \propto \exp\{ -\frac{1}{2\sigma^2} (\sum_{i=1}^{n} m_i\delta_i^2-2\rho\sum_{i=1}^{n}\sum_{j> i}\delta_im_i\tilde{w}_{ij}\delta_j)\} \\ & \propto \exp\{ -\frac{1}{2\sigma^2} (\sum_{i=1}^{n} m_i\delta_i^2-2\rho\sum_{i=1}^{n}\sum_{j> i}\delta_i\frac{m_i{w}_{ij}}{m_i}\delta_j) \} \\ & \propto \exp\{ -\frac{1}{2\sigma^2} (\sum_{i=1}^{n} m_i\delta_i^2-2\rho\sum_{i=1}^{n}\sum_{j> i}\delta_iw_{ij}\delta_j) \} \propto \exp\{ -\frac{1}{2\sigma^2} {\mbox{\boldmath$ \delta $\unboldmath}}^T (D-\rho W) {\mbox{\boldmath$ \delta $\unboldmath}} \} \end{aligned}$$ Therefore, $$\boldsymbol{\delta}=(\delta_1, \ldots, \delta_{n})^\top\sim \mathcal{MVN}_n(0, \sigma^2(D-\rho W)^{-1})$$ Proof of Proposition \[prop:dopt\] ---------------------------------- Following the expression in , we have that $$\begin{aligned} &\mathrm{max}_{{\mbox{\boldmath$ x $\unboldmath}} \in \{-1,1\}^n}[D({\mbox{\boldmath$ x $\unboldmath}})] \\ & = \mathrm{max}_{{\mbox{\boldmath$ x $\unboldmath}} \in \{-1,1\}^n}[(1-\rho)\sum\limits_i m_i (\sum\limits_i m_i x_i^2-\rho \sum\limits_i \sum\limits_j w_{ij} x_i x_j) - (1-\rho)^2(\sum\limits_i m_i x_i)^2]\\ & = \mathrm{max}_{{\mbox{\boldmath$ x $\unboldmath}} \in \{-1,1\}^n}[\sum\limits_i m_i \sum\limits_i m_i - \rho \sum\limits_i m_i \sum\limits_i \sum\limits_j w_{ij} x_i x_j - (1-\rho)(\sum\limits_i m_i x_i)^2], \end{aligned}$$ which is equivalent to $$\begin{aligned} & = \mathrm{min}_{{\mbox{\boldmath$ x $\unboldmath}} \in \{-1,1\}^n}\left[\rho \sum\limits_i m_i\sum\limits_i \sum\limits_j w_{ij} x_i x_j + (1-\rho)(\sum\limits_i m_i x_i)^2\right]\\ & = \mathrm{min}_{{\mbox{\boldmath$ x $\unboldmath}} \in \{-1,1\}^n}\left[\frac{\rho}{1-\rho} \sum\limits_i m_i\sum\limits_i \sum\limits_j w_{ij} x_i x_j + (\sum\limits_i m_i x_i)^2\right]\\ & = \mathrm{min}_{{\mbox{\boldmath$ x $\unboldmath}} \in \{-1,1\}^n}\left[a \sum\limits_i \sum\limits_j w_{ij} x_i x_j + (\sum\limits_i m_i x_i)^2\right], \end{aligned}$$ where ${\mbox{\boldmath$ x $\unboldmath}}=(x_1, \ldots, x_n)$, $x_i\in \{-1,1\}$, and $a = \frac{\rho}{1-\rho} \sum\limits_i m_i$. Additional numerical results {#sec:A2} ---------------------------- Table \[tab:alpha\] gives the time (in seconds) took for GUROBI to solve modified MIP for different network sizes and different choices of $\alpha$. Similar to the original-MIP, the running times of modified-MIP increase for larger networks. $\alpha$ $n=20$ $n=30$ $n=40$ $n=50$ ---------- -------- -------- -------- -------- 0.55 0.47 1.63 3.61 13.26 0.6 0.58 1.31 3.95 24.80 0.7 0.57 1.40 3.67 13.65 0.8 0.60 1.43 2.12 21.65 0.9 0.58 1.47 4.36 17.08 : Running times (in sec) of modified-MIP for different values of $\alpha$ averaged across 20 network sizes $n$ keeping network density $p = 0.1$.[]{data-label="tab:alpha"}
{ "pile_set_name": "ArXiv" }
--- abstract: 'We used the mark weighted correlation functions (MCFs), $W(s)$, to study the large scale structure of the Universe. We studied five types of MCFs with the weighting scheme $\rho^\alpha$, where $\rho$ is the local density, and $\alpha$ is taken as $-1,\ -0.5,\ 0,\ 0.5$, and 1. We found that different MCFs have very different amplitudes and scale-dependence. Some of the MCFs exhibit distinctive peaks and valleys that do not exist in the standard correlation functions. Their locations are robust against the redshifts and the background geometry, however it is unlikely that they can be used as “standard rulers” to probe the cosmic expansion history. Nonetheless we find that these features may be used to probe parameters related with the structure formation history, such as the values of $\sigma_8$ and the galaxy bias. Finally, after conducting a comprehensive analysis using the full shapes of the $W(s)$s and $W_{\Delta s}(\mu)$s, we found that, combining different types of MCFs can significantly improve the cosmological parameter constraints. Compared with using only the standard correlation function, the combinations of MCFs with $\alpha=0,\ 0.5,\ 1$ and $\alpha=0,\ -1,\ -0.5,\ 0.5,\ 1$ can improve the constraints on $\Omega_m$ and $w$ by $\approx30\%$ and 50%, respectively. We find highly significant evidence that MCFs can improve cosmological parameter constraints.' author: - Yizhao Yang - Haitao Miao - Qinglin Ma - Miaoxin Liu - 'Cristiano G. Sabiu' - 'Jaime Forero-Romero' - Yuanzhu Huang - Limin Lai - Qiyue Qian - Yi Zheng - 'Xiao-Dong Li' bibliography: - 'cites.bib' title: Using the Mark Weighted Correlation Functions to Improve the Constraints on Cosmological Parameters --- , Introduction {#intro} ============ The discovery of cosmic acceleration [@riess1998observational; @perlmutter1999measurements] implies either the existence of a “dark energy” component in our Universe or the breakdown of general relativity on cosmological scales. The theoretical explanation and observational probes of cosmic acceleration have attracted tremendous attention, and are still far from being well understood or accurately measured [@weinberg1989cosmological; @miao2011dark; @YOO_2012; @weinberg2013observational]. On scales of a few hundred Megaparsecs (Mpc) the spatial distribution of galaxies forms a distinct, very complicated filamentary motif known as the ‘cosmic web’ [@1986Bardeen; @1986deLapparent; @Huchra_2012; @Tegmark_2004; @Guzzo_2014]. The distribution and clustering properties of galaxies in the cosmic web encodes a huge amount of information on the expansion and structure growth history of the Universe. In the next decade, several large scale surveys, including DESI[^1], EUCLID[^2], LSST[^3], WFIRST[^4], and CSST [@Gong_2019], will begin operations and map out an unprecedented large volume of the Universe with extraordinary precision. It is of essential importance to develop powerful tools that can comprehensively and reliably infer the cosmological parameters from large-scale structure (LSS). The most widely-adopted LSS analysis methods is still the 2-point correlation function (2pCF) or power spectrum measurements, which are sensitive to the geometric and structure growth history of the Universe [@kaiser1987clustering; @ballinger1996measuring; @Eisenstein_1998; @Blake_2003; @Seo_2003]. These methods have achieved tremendous success when applied to a series of galaxy redshift surveys such as the 2-degree Field Galaxy Redshift Survey (2dFGRS; [@2df:Colless:2003wz]), the 6-degree Field Galaxy Survey (6dFGS; [@beutler20116df]), the WiggleZ survey [@blake2011wigglez; @blake2011wigglezb], and the Sloan Digital Sky Survey (SDSS; [@york2000sloan; @Eisenstein:2005su; @Percival:2007yw; @anderson2012clustering; @sanchez2012clustering; @sanchez2013clustering; @anderson2014clustering; @samushia2014clustering; @ross2015clustering; @beutler2016clustering; @sanchez2016clustering; @alam2017clustering; @chuang2017clustering]. The main limitation of this method is that they are only sensitive to the gaussian part of the density field, while both the structure formation process or some primordial conditions can introduce non-gaussian features in the LSS. ![PDFs (probability distribution functions) of the $\log$ density and the smoothing scale $h_W$, in case of using $n_{NB}$=30 and 300, measured from the BigMD $z=0.102$ halo sample. A masscut is applied to the sample to maintain a number density of $10^{-3}(h^{-1}{\rm Mpc})^{-3}$. Very roughly we find $h_{W}\varpropto n_{NB}^{0.4}$. []{data-label="fig_hists"}](hists.eps){width="8cm"} ![image](black.eps){width="18cm"} Ongoing research seeks to go beyond the 2-point statistics includes the methods such as 3-point statistics [@Sabiu_2016; @Slepian_2017], 4-point statistics [@Sabiu_2019], cosmic voids [@ryden1995measuring; @lavaux2012precision], deep learning [@Ravanbakhsh17; @Mathuriya18], and so on. While many of them have proved useful, here we investigate another statistical tool, namely the [*mark weighted correlation function*]{} [MCF; @Beisbart2000; @Beisbart2002; @Gottl2002; @Sheth:2004vb; @Sheth:2005aj; @Skibba2006; @White_2009; @White_2016; @Satpathy:2019nvo; @2020arXiv200111024M; @PMS2020] which is simpler and computationally easier compared than the statistics mentioned above. By weighting each galaxy using a “mark” that depends on its local density, the MCFs provide density-dependent clustering information from the sample which is useful for data mining. The weights can be set to be proportional to the positive or negative power of the density, to allow the statistics to place more emphasis on dense or undense regions, where the clustered structures and the redshift space distortions (RSDs) are physically very different. It is expectable that in this manner we can obtain more information from the data compared with using the traditional 2pcf, which equally treat all galaxy pairs regardless of the difference in their physical properties and environments. This paper is arranged as follows. Section \[data\] outlines the parameters of the datasets we use. Section \[method\] represents the methods used for the principle and operation of marked correlation functions. Section \[basic\] explains the clustering statistics of different number density. Section \[discussion\] discusses more details by various parameters to test whether the standard ruler persists. In section \[conclusion\] we present our general conclusions. Data ==== The analysis in this work relies on the large N-body simulation: BigMultiDark[^5] (BigMD), and also a series of fast simulations generated using COLA (COmoving Lagrangian Acceleration). The Multiverse simulations are a set of cosmological N-body simulations designed to study how variations in cosmological parameters affect the clustering and evolution of cosmic structures. Among them, the BigMD simulation is produced using $3\,840^3$ particles in a volume of $(2.5h^{-1}\rm Gpc)^3$, assuming a $\Lambda$CDM cosmology with $\Omega_m = 0.307115$, $\Omega_b = 0.048206$, $\sigma_8 = 0.8288$, $n_s = 0.9611$, and $H_0 = 67.77\ {\rm km}\ s^{-1} {\rm Mpc}^{-1}$ [@BD]. The initial conditions, based on primordial Gaussian fluctuations, are generated via the Zel’dovich approximation at $z_{\rm init} = 100$. Its large volume and huge number of particles make this an ideal simulation for the purposes of this work. For purpose of estimating measurement covariance matrices, we also use 150 simulations with $600^3$ particles and a boxsize $(512 h^{-1}\rm Mpc)^3$, in the BigMD cosmology, generated using the COLA [@Tassev13] algorithm. Second order Lagrangian perturbation Theory (2LPT) is a computationally efficient and accurate method for describing the gravitational dynamics on large scales. COLA combines 2LPT, for time integration for large scale dynamical evolution, with a full-blown N-body Particle-Mesh (PM) algorithm to calculate the small scale dynamics. Compared with the other fast simulation algorithms in the market, COLA performs better in simulating the structures on non-linear scales [@Chuang14]. Finally, to check the dependence on cosmologies, we run five sets of COLA simulations using $\Lambda$CDM models of $(\Omega_m, 10^9 A_s, \sigma_8)$ = (0.2, 2.1, 0.5557), (0.31, 2, 0.7965), (0.31, 2.1, 0.8161), (0.31, 2.29, 0.8523) and (0.46, 2.1, 1.0576) [^6], respectively. The other parameters are taken as $\Omega_b = 0.048206$, $n_s = 0.9611$ and $H_0 = 67.77\ {\rm km}\ s^{-1} {\rm Mpc}^{-1}$, the same to their used values in the BigMD simulation. Each simulation was run using $1024^3$ particles in a $(1024 h^{-1}\rm Mpc)^3$ box. We identify gravitationally bound structures in each of the BigMD and COLA DM simulations using the <span style="font-variant:small-caps;">ROCKSTAR</span> halo finder [@ROCKSTAR]. <span style="font-variant:small-caps;">ROCKSTAR</span> is a halo finder based on adaptive hierarchical refinement of friends-of-friends groups in six phase-space dimensions and one time dimension, allowing for robust tracking of substructure. Both halos and subhalos are included in the analysis. To ensure the comparability, we maintain a halo number density $\bar n=$0.001 $(h^{-1}\rm Mpc)^{-3}$ in all simulations. Methodology {#method} =========== The MCF is a simple extension of the standard configuration space 2pCF by assigning a mark to each object. Following [@White:2016yhs], we use the local density as the mark, and weight each halo by $${\rm weight}=\rho_{n_{\rm NB}}^\alpha,$$ which is a simpler expression than that proposed in [@White:2016yhs]. Here $\rho_{n_{\rm NB}}$ is the density estimated using its $n_{\rm NB}$ nearest neighbours, $$\rho_{n_{\rm NB}}({\bf r}) = \sum_{i=1}^{n_{\rm NB}} W_k({\bf r-r_i},h_W),$$ where $\rho_{n_{\rm NB}}(\bf r)$ is the number density at position $r$, and $W_k$ is the smoothing kernel, for which we choose the 3rd order B-spline functions having non-zero value within a sphere of radius $2h_W$ $h^{-1}$ Mpc [@Gingold1977; @Lucy1977]. We adopt an adjustable radius of the smoothing kernel to ensure that the kernel always includes $n_{\rm NB}$ nearest neighbour halos within $2h_W$. The value of $n_{\rm NB}$ determines the smoothing scale that we applied to the sample. Figure \[fig\_hists\] shows the PDF (probability distribution function) of the $\log$ density and $h_W$ in the constructed fields when using $n_{\rm NB}$=30 and 300, respectively. Since the values of $h_W$ depend on the local density, they are not a constant number under a given $n_{\rm NB}$. Here we find the central and 1$\sigma$ width of $h_W$ is $8.3\pm 3.7$, $20.1\pm 4.7$ $h^{-1}$Mpc if using $n_{\rm NB}$=30, 300, respectively. Very roughly, the central value scales as $\varpropto (n_{\rm NB})^{0.4}$. A larger $n_{\rm NB}$ decreases both the mean and the variance of $\log_{10}\rho$. In the MCF, the objects in the high and low dense regions are assigned different weights. Figure \[fig\_scatter\] shows the weights of some halos distributed in a $200\times200\times20 (h^{-1}\rm Mpc)^3$ slice, selected from the $z=0.102$ BigMD snapshot. While $\alpha=1$ assign significantly larger weights to the objects in dense environment, the $\alpha=-1$ strategy does the opposite. From the dense to un-dense regions, the clustering patterns and redshift space distortions vary dramatically, so we expect very different results for MCFs when using the two weighting strategies. Apart from the weight that is assigned to each halo, the computational procedure to measure the MCF is exactly the same as that to measurement the standard 2pCFs. We use the most commonly adopted Landy-Szalay estimator $$\label{eq:Wsmu} W(s,\mu) = \frac{WW-2WR+RR}{RR},$$ where $WW$ is the weighted number of galaxy-galaxy pairs, $WR$ denote the galaxy-random pairs, and $RR$ denote the number of random-random pairs. They are separated by a distance defined by $s\pm \Delta s$ and $\mu\pm \Delta \mu$, where $s$ is the distance between the pair and $\mu=\cos(\theta)$, with $\theta$ being the angle between the line joining the pair and the line of sight (LOS) direction [^7]. For the random samples, we always use 10x more particles than the data samples, and fix the weights of all particles to be 1. Compared to the tradition CF which is defined as $\xi(\bf r)=\left<\delta({\bf x}) \delta({\bf x+r})\right>$, the MCF takes the form of $$\begin{aligned} \label{eq:W_delta_rho} W({\bf r}) &=& \left<\delta({\bf x}) \rho_{n_{NB}}({\bf x})^{\alpha} \delta({\bf x+r}) \rho_{n_{NB}}({\bf x+r})^{\alpha}\right> . $$ Notice the difference between $\delta$ and $\rho_{n_{\rm NB}}$. The latter one is the smoothed density field, while the former is the contrast of the [*point-like*]{} density $\rho$. effectively, $\rho$ is the special case of $\rho_{n_{\rm NB}}$ with $n_{\rm NB}$=1. A glance at the weighted CF {#basic} =========================== In what follows we present the MCFs measured from the BigMD halos, distributed in the redshift range of $0<z<1.45$. To guarantee the comparability of objects at different redshifts, and also to maintain a uniform smoothing scale, from each sample we select a number of most massive halos to build up a subsample having a constant number density $\bar n=10^{-3} (h^{-1} \rm Mpc)^{-3}$ for all samples. $W(s)$ Measurements {#sec:ws} ------------------- Figure \[fig\_xi\_s\] shows the MCFs as functions of clustering scale, i.e. the monopole $W(s)$ [^8]. They are computed by ignoring the $\mu$-dependence in Equation \[eq:Wsmu\] when counting the weighted number of pairs. We shows the results using $\alpha=-1,\ -0.5,\ 0,\ 0.5$, and $1$, at the redshifts of 0, 0.51, 1.0, and 1.45, respectively. In all plots, we use $n_{NB}=30$. ![image](xis.eps){width="16cm"} ![image](bigmd_vs_cola.eps){width="16cm"} A significant dependence on the weighting scheme is detected when comparing the MCFs using different $\alpha$. A larger $\alpha$ assigns more weights to the dense, clustered region, thus results in stronger correlation (higher magnitude). The clustering patterns in dense and undense regions are different from each other, so the shape of MCFs is also sensitive to $\alpha$. As shown in the Figure, when using $\alpha=-1,\ -0.5,\ 0,\ 0.5$ and $1$, $s^2 W(z=0)$ peaks at $s\approx2,\ 4,\ 17,\ 14$ and $18$ $h^{-1}\ $Mpc, with amplitudes of 16, 23, 80, 180, 400 $(h^{-1}\rm Mpc)^2$, respectively. The $\alpha=1$ result has a peak magnitude 5 times stronger than the $\alpha=0$ case, while the latter is again 5 times stronger if compared with the $\alpha=-1$ case; if comparing the clustering amplitude on the BAO scale, then the $\alpha=1$ case is 2/4/15/100 times stronger than $\alpha=0.5/0/-0.5/-1$ cases, respectively. The statistical error also increases with the decreasing of $\alpha$. For the $\alpha=-1$, $z=0$ case, the BAO peak is not very detectable, possibly due to the large noise therein. By enforcing $\bar n=10^{-3}$ at all redshifts both the clustering amplitude and the shape remain similar at all redshifts. Compared with low redshift result we find the BAO peak at higher redshift is more prominent, because there the smearing effect from the peculiar velocity and the non-linear structure formation is less significant. The shape of the MCF is changing persistently when we tune the value of $\alpha$. Several distinctive features, including a sharp peak (around 5-10 $h^{-1}$ Mpc) in the $\alpha=1,\ -0.5,\ -1$ results, are a valley (around $15 h^{-1}$ Mpc) in the $\alpha=-0.5,\ -1$ results, are detected. We will discuss their origins, implications and usabilities in the latter sections. Finally, a quick check presented in Figure \[fig\_bigmd\_vs\_cola\] shows that for most cases COLA achieves $\lesssim10\%$ accuracy in predicting the MCFs within the clustering range considered here [@Ma_2020]. Relative large discrepancy is detected at the $s\lesssim20$ $h^{-1}$Mpc regime in the $\alpha=0.5$ and 1 cases. This consistency may be resolved by measures such as increasing the time steps or enhancing the resolution of the simulations, but we will not study it in details. ![image](ximu_6to40.eps){width="16cm"} ![image](ximu_2to10.eps){width="16cm"} ![image](ximu_10to30.eps){width="16cm"} $W_{\Delta s}(\mu)$ Measurements {#sec:wmu} -------------------------------- The RSDs in high and low density regions are quite different. So we expect different anisotropic clustering features in the different MCFs. In what follows, we study $\mu$-dependence of the MCFs. By integrating $W(s,\mu)$ along the $s$ direction, we define $$W_{\Delta s}(\mu)\equiv \int_{s_{\rm min}}^{s_{\rm max}} W(s,\mu) ds,$$ as well as its normalized version $$\hat W_{\Delta s}(\mu)\equiv \frac{W_{\Delta s}(\mu)}{\int_{0}^{\mu_{\rm max}}W_{\Delta s}(\mu)\ d\mu}.$$ These two quantities describe the difference in the clustering strength in different directions [*w.r.t*]{} the LOS. They have been used to quantify the RSDs and the AP distortions in the tomographic Alcock-Pacyznski (AP) method [@LI14; @LI15; @LI16; @LI18; @LI19; @Park:2019mvn; @Zhang2019]. Figure \[fig\_ximu\] shows the measured $W_{\Delta s}(\mu)$ at redshifts of $z=$ 0, 0.51, 1 1.45, using $\alpha$=-1, -0.5, 0, 0.5, 1 and an integral range $s\in$ (6,40), (2,10), (10,30) $h^{-1}$ Mpc, respectively. In all curves, we see a sharp peak near $1-\mu=0.1$, which is produced by the small-scale, non-linear finger-of-god (FOG) effect [@Jackson_1972]; Also, we see a slope in the range of $1-\mu\gtrsim0.1$, as a consequence of the Kaiser effect [@kaiser1987clustering]. The amplitude of $W_{\Delta s}(\mu)$ is enhanced if we tune down $s_{\rm min}$ and include more small-scale clustering into the integration. In doing this, we also enhance the leftmost peak since FOG is stronger on smaller clustering scales. Similar to what we found with $W(s)$, the $W_{\Delta s}(\mu)$ has a larger amplitude and smaller statistical noise when using a larger value of $\alpha$. On the other hand, we do not detect any “violent” changes in the shape of $W_{\Delta s}(\mu)$ when tuning the value of $\alpha$. However, this does not necessarily mean that the information encoded in these different $W_{\Delta s}(\mu)$ are all the same. We will revisit this issue in the later. Distinctive Features in $W(s)$ {#sec:distinctive_features} ------------------------------ In the $W(s)$ curves there are several distinctive peaks and valleys which do not exist in the standard 2pCFs. In what follows, we briefly discuss their possible origins. ![ In the $\alpha$=1 MCF, we find a sharp peak around $s=5\ h^{-1}$ Mpc. Comparison between the measurements on different directions implies that this peak has something to do with the FOG effect (left panel). Its location and position is insensitive to the redshift (right panel). []{data-label="fig_rho1"}](rho1_fog.eps){width="8cm"} ![ The $\alpha$=-1 MCF has an unusual “S”-like shape. It bears a valley-like feature on scales of $\sim$15 ${h^{-1}}$ Mpc, arising from the difference between the point-like field $\rho$ and the smoothed field $\rho_{30}$. Its shape and the strength of remains robust against redshift.[]{data-label="fig_rho-1"}](rho-1.eps){width="8cm"} ### Sharp peak {#section_sharp_peak} In many plots of $s^{2}W(s)$ there exist a sharp peak located around $5-10\ h^{-1}$ Mpc (see Figure \[fig\_xi\_s\]). This means that on that scale there exist a large number of clustering pairs. In all plots we use the weight $\rho_{30}$, whose smoothing scale is $\approx 8$ $h^{-1}$ Mpc. That smoothing produces a correlation on that scale, so it is not surprising to see a peak on the corresponding scale. However, comparing with a random sample smoothed in the same way shows that the the amplitude of the peak also heavily depends on the intrinsic clustering property of the sample. A comparison between the measurements on different directions implies that this sharp clustering peak has something to do with the FOG effect, which produces $\sim5-10\ h^{-1}$ Mpc “spikes” like structure along the LOS As shown in the left panel of Figure \[fig\_rho1\], the peak along the LOS direction is far more prominent than what found in the transverse direction. The other panel of the Figure shows that the heights and locations of the peak is rather insensitive to the redshift. ### Distinctive Valley Besides the peak we also detect a valley located at $s\approx15\ h^{-1}$ Mpc in the $\alpha=-1,\ -0.5$ cases. In particular, the $\alpha=-1$ case possesses both the peak and the valley, so has an unusual “S”-like shape. Figure \[fig\_rho-1\] shows the valleys in the $\alpha=-1$ MCFs. Equation \[eq:W\_delta\_rho\] means that the features in the MCF should be highly related with the difference between the local density $\rho$ and its smoothed counterpart $\rho_{30}$. In the redshift range of $\sim$0-1.5, the location and the strength of this valley-like feature remains rather robust. Contrary to the situation of the peak, we find that the valley looks rather similar in both the LOS and transverse directions, leading us to believe that it has little or nothing to do with RSD effects. Implications for Cosmological Analysis {#discussion} ====================================== In this section we discuss the implications of the MCFs to the cosmological analysis. In the first part, we report our attempt to utilise the peak and the valley as standard rulers. In the second part, we adopt a more comprehensive approach by using the full shape of the MCFs. Usability of the distinctive features as standard rulers {#subsec:features} --------------------------------------------------------- The distinctive peaks and valleys as discovered in the $\alpha \neq 0$ MCFs are not found in the standard 2pCF. A remarkable feature is that, the locations of these peaks and valleys are rather robust against the redshift. This inspires us to consider using them as “standard rulers” to probe the expansion history. In galaxy surveys, the angular positions and redshifts of each galaxy is converted to 3D positions using the redshift-distance relation $r(z)$ adopted in an assumed cosmology. So wrongly adopted cosmology parameters lead to the following distortions of length in the directions parallel and perpendicular to the LOS, $$\label{eq:alpha} \begin{aligned} &\alpha_{\parallel}(z) = \frac{H_{\rm true}(z)}{H_{\rm wrong}(z)}, \\ &\alpha_{\perp}(z) = \frac{D_{A,\rm wrong}(z)}{D_{A,\rm true}(z)}, \end{aligned}$$ where “true” and “wrong” denote the values of quantities in the true and incorrectly assumed cosmologies, respectively. This leads to two effects in the wrong cosmology, - The changes in the size of structures, known as the “volume effect”. This changes the BAO peak location, shifts the clustering patterns [@Li2017], and changes the sizes of structures in the density field [@Park2010]. - Changes in the shape of structures, known as the Alcock-Paczynsk (AP) distortion [@AP1979; @ballinger1996measuring]. For an incomplete list of the methods based on this effect and their applications to the data, see [@Ryden1995; @Matsubara1996; @Outram2004; @Marinoni2010; @Blake2011; @Lavaux2012; @Alam2017; @Qingqing2016; @LI14; @LI16; @Ramanah2019]. In what follows, we mainly test the feasibility of using the distinctive features to probe the “volume effect”. To mimick the effect, we take Equation \[eq:alpha\] to convert the sample into backgrounds of two wrong cosmologies, $$(\Omega_m,\ w)_{\rm wrong} = (0.1,\ -1), \ \ (0.3071,\ -1.5).$$ Notice that, in doing this we just “re-observe” the simulation using the $r(z)$s of the new cosmologies, without running new simulations. That is exactly what one is doing when conducting the BAO or AP analysis on the observational data. The comparability of samples requires them having the same smoothing scale. Thus, when using a wrong background, we change the lower halo mass cut to maintain a constant number density $\bar n=1\times10^{-3} (h^{-1}\rm Mpc)^{-3}$. In the following subsections, we test the feasibility of using the $\alpha=-1$ and $\alpha=1$ MCFs, respectively. ### The background ![Comparing the $\alpha=-1$ MCFs in backgrounds of the fiducial cosmology and an extremely wrong cosmology. Regardless of the dramatic difference in the background geometry, there is little difference in the locations of the peaks or the valleys. So it is unlikely to use these features to probe the geometry of the Universe. []{data-label="fig_peak_wrong_cosmos"}](peak_diffNB_xis.eps){width="9cm"} ![ Comparing the $\alpha=1$ MCFs in different backgrounds. The conclusion is similar to what we found in Figure \[fig\_peak\_wrong\_cosmos\]. []{data-label="fig_peak_wrong_cosmos_rho1"}](diffmu_alpha_1.eps){width="9cm"} When adopting an incorrect expansion history for the background, we expect a scale-shift in the shape of the CF. Because of the nonuniform scaling at different redshifts, we expect a redshift-evolution of the CFs, determined by $\left(\alpha_{\perp}(z)^2\alpha_{\parallel}(z)\right)^{1/3}$ (see [@Li2017]). However, regardless of the strong volume effect in the two extremely incorrect cosmologies considered here, we do not detect any significant change in the scale of the MCFs. Figure \[fig\_peak\_wrong\_cosmos\] shows that, in the wrong cosmological backgrounds, the locations of the peaks or valleys remain the same to their fiducial values. The conclusion is unchanged when we try using $n_{\rm NB}=10$, 30 and 300. Considering the valleys of the $n_{\rm NB}$=30 measurements as an example. While in the fiducial cosmology the valley locates at $s=15-16$ $h^{-1}\rm Mpc$, in the $\Omega_m=0.1$ wrong cosmology it still shows up at $s\approx16$ $h^{-1}\rm Mpc$. For comparison, in this cosmology the comoving length is artificially rescaled by a rate of $\gtrsim$20% at $z\gtrsim0.6$, so we expect the valley appears near 18-19 $h^{-1}$ Mpc. While being insensitive to the background, the location of the peaks or valleys are rather sensitive to the choice of the smoothing scale. When changing $n_{\rm NB}$ from 30 to 10/300, the location of the valley is shifted to 26/13 $h^{-1}\rm Mpc$, respectively. In the $\alpha=1$ MCFs, again, we find that the locations of the peaks are, rather insensitive to the background change (see Figure \[fig\_peak\_wrong\_cosmos\_rho1\]). Moreover, it appears robust against changes in redshift in the wrong cosmology we chose. This means that it is impossible to make use of their redshift evolution as a signal to identity the wrong cosmologies [^9]. Here we point out that, actually, this FOG related pattern has been detected in other statistics. [@Fang2019] reported a detection of a peak around $\sim3h^{-1}{\rm Mpc}$ in the $\beta$-skeleton statistics. In Appendix A, we report that the peak in that statistics can not be used to conduct cosmological analysis, either. ### Dependence on bias and $\sigma_8$ {#sec_five_cola_sim} While being rather insensitive to the background, these features do have some dependence on the bias and $\sigma_8$. Figure \[fig\_diffsim\_xis\] shows the MCFs measured in five sets of COLA simulations, with the $\Lambda$CDM parameters of $(\Omega_m, 10^9 A_s, \sigma_8)$ = (0.2, 2.1, 0.5557), (0.31, 2, 0.7965), (0.31, 2.1, 0.8161), (0.31, 2.29, 0.8523) and (0.46, 2.1, 1.0576), respectively. Clearly, when adopting a smaller $\sigma_8$, the locations of the peaks shift towards small scales. Basically, a smaller $\sigma_8$ leads to a smaller peak scale, except that the $\sigma_8=0.8/0.82/0.85$ curves in the $\alpha=1$, $n_{\rm NB}=300$ case do not precisely obey this order. Possibly, becase $n_{\rm NB}=300$ corresponds to a smoothing scale much larger than 8$h^{-1}$Mpc, here $\sigma_8$ can not precisely describe what is happening. Although the basic trend is still correct, some complexities arise if we carefully investigate the details. Meanwhile, we also find they have some dependence on the halo bias. Figure \[fig\_diffbias\] shows the MCFs of three subsamples of BigMD $z=0$ halos, distributed in different mass range (we keep $\bar n=10^{-3}\ (h^{-1}\rm Mpc)^{-3}$ in all subsamples). The valleys are more affected compared with the peaks. In summary, our analysis shows that these peaks and valleys can not be used as “standard rulers” to probe the geometry of the Universe. But when changing the parameters related with the structure formation, we do observe shifts in the peaks or valleys. So these features maybe useful for the probing of those parameters related with the structure formation, e.g. the values of $\sigma_8$ and the halo/galaxy bias. Apart from $\hat W(s)$s, the $\hat W_{\Delta s}(\mu)$s are also sensitive to $\sigma_8$. We investigate this sensitivity further in Appendix B. ![ MCFs of five COLA simulations, with different values of $\sigma_8$. The peak is shifted towards to larger scales when using a larger $\sigma_8$. []{data-label="fig_diffsim_xis"}](peak_diffsim_xis.eps){width="9cm"} ![ MCFs of samples with different halo bias. The valleys are more affected compared with the peaks. []{data-label="fig_diffbias"}](peak_diffbias.eps){width="9cm"} Using the full shape of the MCFs -------------------------------- In what follows, we take a more comprehensive approach and use the full shape of the marked CF to predict the cosmological parameters. Figure \[fig\_cov\] shows the [*correlation coefficients* ]{} of $\hat W(s)$ and $\hat W_{\Delta s}(\mu)$ [^10]. They are estimated using the 150 COLA simulations. Among all MCFs, the $\alpha=-1$ case has weakest correlation with the others. The negative correlations are the consequence of the normalization. We choose the $\Omega_m=0.3071,\ w=-1$ cosmology as the [*fiducial geometrical background*]{}, and define a statistical function to distinguish the other backgrounds from it, $$\chi^2 = ({\bf p_{\rm fiducial}-p_{\rm target}}) \cdot {\bf Cov}^{-1} \cdot ({\bf p_{\rm fiducial}-p_{\rm target}}),$$ where $\bf p$ denotes $\hat W(s)$ or $\hat W_{\Delta s}(\mu)$. Considering that the number of mocks is not too many compared with the binning number of the $W$s, we use the formula suggested by @Hartlap:2006kj to correct the bias in the estimated covariance matrix. The MCFs of the halo catalogues embedded in the backgrounds of the incorrect cosmologies are obtained using the following coordinate transforms (see [@LI18] for details), $$\label{eq_smu_trans1} s_{\rm target} = s_{\rm fiducial} \sqrt{\alpha_{\parallel}^2\mu_{\rm fiducial}^2 + \alpha_{\bot}^2(1-\mu_{\rm \bot}^2)},$$ $$\label{eq_smu_trans2} \mu_{\rm target} = \mu_{\rm fiducial} \frac{\alpha_\parallel} {\sqrt{\alpha_{\parallel}^2\mu_{\rm fiducial}^2 + \alpha_{\bot}^2(1-\mu_{\rm \bot}^2)}}.$$ This is much more efficient compared with converting the samples into the different backgrounds and re-measuring the MCFs. A caveat is that equations \[eq\_smu\_trans1\] and \[eq\_smu\_trans2\] do not capture [*the change in the values of the weights*]{}. That definitely happens, since the Alcock-Paczynski effect non-uniformly distorts the geometry, so the set of $n_{\rm NB}$ nearest neighbors can differ from one cosmology to next. In Appendix C we check this caveat and show that if we neglect this issue it introduces only minor effect, and thus equations \[eq\_smu\_trans1\] and \[eq\_smu\_trans2\] are deemed precise enough for this proof-of-concept study. ![image](cov_xis_normed.eps){width="8cm"} ![image](cov_ximu_normed.eps){width="8cm"} Equations \[eq\_smu\_trans1\] and \[eq\_smu\_trans2\] only consider the [*background*]{} information of the cosmologies. A more comprehensive analysis should involve the information of the structure growth, but that would require many more numerical simulations. Constraints from $\hat W(s)$ ---------------------------- The conditional constraints on $\Omega_m$ and $w$ (fixing one of them as the fiducial and constrain the other one) using the full shape of $\hat W_{\Delta s}(\mu)$ are presented in the lower panel of Figure \[fig\_chisq\_1d\]. In the plots, we use the clustering range of $s\in(5,50)\ h^{-1}\rm Mpc$, divided into 15 bins. Including the results on larger scales does not further enhance the power of constraints. We find the the $\alpha=0$ results leads to the tightest constraints among all cases considered. Also, combining different MCFs can improve the constraint. Taking the $w=-0.4$ cosmology as an example. Compared with the fiducial cosmology, it is disfavored by $\chi^2=7.3/5.8/5.5/1.4$ when using the $\alpha=0/0.5/-0.5/1$ MCF, so the $\alpha=0.5/-0.5/1$ result is 20%/24%/81% worse than the $\alpha=0$ result, respectively. Combining the $\alpha=0$ and $\alpha=1$ MCFs, we get a 17% improvement compared with only using the $\alpha=0$ MCF. If we combine the $\alpha=0/0.5/1$ MCFs together, the $\chi^2$ is then enlarged to 15, a $\approx$100% improvement compared with only using the $\alpha=0$ MCF. The $\chi^2$ of the $\alpha=0,\ 0.5,\ 1$ combination is very close to the summation of the $\chi^2$s using the three MCFs separately. This means that, the cosmological information carried by the three MCFs is not strongly overlapping from each other. This is essentially important for the MCF statistics, meaning that we can significantly improve the cosmological constraints by combining different MCFs. Constraints from $\hat W_{\Delta s}(\mu)$ ----------------------------------------- ![Conditional constraints on $\Omega_m$ and $w$, derived using $W_{\Delta s}(\mu)$ and $\hat W_{\Delta s}(\mu)$. Estimation is based on the $\bar n=10^{-3}\ (h^{-1}\rm Mpc)^3$ halo samples of the 150 COLA samples having a boxsize of $(512 h^{-1}\rm Mpc)^3$. Among all MCFs, the $\alpha=0$ MCF has the largest statistical power. We can largely improve the statistical power by combining different MCFs. []{data-label="fig_chisq_1d"}](chisq_ws_norm.eps "fig:"){width="8cm"} ![Conditional constraints on $\Omega_m$ and $w$, derived using $W_{\Delta s}(\mu)$ and $\hat W_{\Delta s}(\mu)$. Estimation is based on the $\bar n=10^{-3}\ (h^{-1}\rm Mpc)^3$ halo samples of the 150 COLA samples having a boxsize of $(512 h^{-1}\rm Mpc)^3$. Among all MCFs, the $\alpha=0$ MCF has the largest statistical power. We can largely improve the statistical power by combining different MCFs. []{data-label="fig_chisq_1d"}](chisq_wmu_norm.eps "fig:"){width="8cm"} The conditional constraints on $\Omega_m$ and $w$ using the full shape of $\hat W_{\Delta s}(\mu)$ are presented in the lower panel of Figure \[fig\_chisq\_1d\], where the integration range of $s$ is taken as $(6,40)\ h^{-1}\rm Mpc$, and we use the shape of $\hat W_{\Delta s}(\mu)$ in the range of $\mu\in(0,0.97)$, divided into 12 bins. Remarkably, the constraints derived using the $\hat W_{\Delta s}(\mu)$s are much more powerful than those derived using the $\hat W(s)$s. Similarly, we find $\alpha=0$ achieves the best performance, and the results can be improved by combining different MCFs. In Table \[tab1\], we list the $\chi^2$s of the $w=-0.4$ cosmology using different $\alpha$s or their combinations. Compared with using $\alpha=0$ MCF, using $\alpha=0,0.5,1$ and $\alpha=0,-1,-0.5,0.5,1$ can improve the $\chi^2$ by 116% and 285%, respectively. Figure \[fig\_chisq\_contour\] shows the constraints in the 2-d $\Omega_m$-$w$ parameter space. The directions of degeneracy using different $\alpha$s are identical to each other, so combining the different MCFs does not help in breaking the degeneracy. But by doing this we do manage to shrink the contour size. Very roughly, compared with the $\alpha=0$ MCF, the $\alpha=0,0.5,1$ and $\alpha=0,-1,-0.5,0.5,1$ combinations can improve the constraints on the parameters by $\approx30\%$ and 50%, respectively. $\alpha$ 0 -1 -0.5 0.5 1 0,-1 0,-0.5 0,0.5 0,1 0,0.5,1 0,-1,-0.5,0.5,1 -------------------------------------- ------ ------ ------ ------ ------ ------ -------- ------- ------ --------- ----------------- $\chi^2$ 21.4 8.1 19.8 14.9 6.9 29.2 45.0 40.0 30.8 46.2 82.3 $\frac{\chi^2}{\chi^2_{\alpha=0}}-1$ 0% -62% -7% -31% -68% 36% 111% 87% 44% 116% 285% ![\[fig\_chisq\_contour\] 68.3% CL constraints on $\Omega_m$ and $w$, derived using $\hat W_{\Delta s}(\mu)$. Different MCFs have the same direction of degeneracy. Compared with only using one kind of MCF, the two combinations can significantly reduce the constrained area. ](contour2.eps){width="8cm"} Conclusion ========== We performed a detailed analysis on the MCFs for which the objects are weighted by $\rho^\alpha$. In this analysis, we considered five different MCFs, i.e. $\alpha=-1,\ -0.5,\ 0,\ 0.5$ and 1, and characterize their scale and angular dependence by using $W(s)$ and $W_{\Delta s}(\mu)$. When studying the scale dependence of the MCFs, i.e. the $W(s)$s, we find the different MCFs have very different amplitudes and scale-dependence. Especially, we found distinctive peaks and valleys in some $\alpha\neq 0$ MCFs, on scales around $\approx5$ and $15$ $h^{-1}$ Mpc, depending on the smoothing scale that we adopted to estimate the density. Their origin and properties are studied in detail. One particular point of interest is that the locations of these features are rather invariant with redshift. In studying the possibilities of using the MCFs in cosmological analysis, we find the locations of the peaks or valleys are rather insensitive to the background geometry. Thus, it is unlikely that they can be utilized as “standard rulers” to probe the geometry. However, their locations are affected by the value of $\sigma_8$ and the galaxy bias, so they could be useful for the determination of these parameters. Finally, we studied the power of the different MCFs in distinguishing the different cosmologies, by using the full shape of the $\hat W(s)$s and the $\hat W_{\Delta s}(\mu)$s. We find they have similar direction of degeneracy in constrained $\Omega_m$ and $w$, while the $\alpha=0$ MCF, corresponding to the standard CF, has the strongest power in distinguishing the background of the different cosmologies. Also, the constraint can be further improved by combining the different MCFs together. In particular, compared with the $\alpha=0$ $W_{\Delta s}(\mu)$, the $\alpha=0,\ 0.5,\ 1$ and $\alpha=0,\ -1,\ -0.5,\ 0.5,\ 1$ combinations achieve $\approx30\%$ and 50% improvement in reducing the constrained area, respectively. The reason why MCF can improve the constrain is easy to understand. The dense and under-dense regions have very different clustering patterns and RSDs features. The many MCFs provide different weighting schemes of the clustering information according to their local density. By using them together, we can separate the regions with different patterns and extracting more clustering information. While previous works regarding the MCF mainly focus on modified gravity theories[@2018Aguayo; @2018Armijo], our work suggests that they could be useful for probing any parameter that is related with the expansion and structure growth history. By using the MCFs, we can enlarge the obtained information by 3-4 more times. MCF are also computationally efficient compared with the high order statistics, like the 3pCF. [@PMS2020] used perturbation theory to study the marked power spectrum using perturbation theory, and found that the mark introduces a significant coupling between small-scale non-Gaussianities and large scale clustering. This explains why using this statistics we can get additional information, and provides further support to the findings of this work. We find that the statistical quantity $\hat W_{\Delta s}(\mu)$ is more powerful than the $\hat W(s)$ in constraining the cosmological parameters. It may be possible to use $\hat W_{\Delta s}(\mu)$s instead of just $\hat \xi_{\Delta s}(\mu)$ in the tomographic AP method to improve the performance. However we leave this issue for future works. While in our analysis the $\alpha=0$ MCF has the strongest power in distinguishing the background of the different cosmologies, the authors of [@2020arXiv200111024M] found that, a marked power spectrum can better constrain cosmological parameters than the power spectrum itself. This difference may be due to two reasons. 1) By using tens of thousands of simulations, [@2020arXiv200111024M] built an emulator to capture both the expansion history and structure formation of the Universe. In contrast, we just “re-observe” one simulation using different backgrounds to study effect of the expansion history. Very possibly, the sensitivity of the MCFs to the structure formation is more important than its dependence to the expansion history, but we do not have it quantified in this simple treatment. We need to conduct a more comprehensive study in future analysis. 2) While [@2020arXiv200111024M] used the power spectrum as the statistical discriminator, we used $W_{\Delta s}(\mu)$, the dependence of clustering strength on the direction. The two statistical quantities are physically quite different, and it is reasonable that the results derived using them are also different. There are still many issues regarding MCF that are important but that we chose not to address in the present work. Although we have shown that MCFs encode a lot of information, we did not detail specific methods to extract them. In particular, we did not check whether the MCFs is useful for improve the measuring of the BAO peaks. In studying the different weighting schemes we only explore the restricted range of $-1\leq\alpha\leq1$. Finally, we only considered the halo number density as the weight, while there are possibilities to use features computed directly on the connectivity graph of the halo distribution. Those graph features are related to topological characteristics of the cosmic web [@Suarez-Perez:2020], features that are in turn naturally correlated to $\sigma_8$ and the halo-galaxy bias. Appendix A. Usability of Peaks in the $\beta$-skeleton statistics {#appa .unnumbered} ================================================================= The $\beta$-skeleton is a novel statistical tool proposed in [@Fang2019] to study the cosmic web. In this statistic, the “spikes” produced by the FOG leads to a peak in the histogram of connections near $2.5{h^{-1}\rm Mpc}$, which is rather robust to the redshift. The origin and the properties of this peak is very close to to the peak we found in the $\alpha=1,\ 0.5,\ -1$ $W(s)$s. ![ Histograms of lengths and directions of connections in the $\beta=3$ cosmic web. Similar to the $W(s)$, there are peaks produced by the FOG on the small scales, which are robust against the redshift, but unlikely to be used in cosmological studies. The full shape of the histograms show some cosmological dependence. []{data-label="fig_beta_hists"}](beta_hists.eps){width="9cm"} We find that, the peak in the $\beta$-skeleton statistics also cannot be used to probe the cosmic expansion history. Figure \[fig\_beta\_hists\] shows the distribution of the lengths and directions of the connections in the $\beta=3$ web, measured in three different backgrounds. It is clear that the locations of the peaks are rather insensitive to the background geometry. The full shape of the histograms of the lengths and directions show some cosmological dependence. We will not go deep and discuss its usability in details. In other work we have found that the entropy and complexity of the $\beta$-skeleton graph actually correlates with $\sigma_8$ [@Torres-Guarin:2020], suggesting that the graph is more sensitive to the global tracer topology than to the more geometrical influence of $\Omega_m$ and $w$. Appendix B. $W_{\Delta s}(\mu)$ in the five COLA simulations {#appb .unnumbered} ============================================================ In section \[sec\_five\_cola\_sim\] we only discussed the $W(s)$ measured in simulations with different $\sigma_8$. Here we present the results of $\hat W_{\Delta s}(\mu)$s. ![ The $W_{\Delta s}(\mu)$s measured in the five sets of COLA simulations with different values of $\sigma_8$. The leftmost part of the curve is dominated by the FOG effect. There the amplitude is significantly enhanced if using a large $\sigma_8$. []{data-label="fig_ximu_diffcosmo"}](ximu_diffcosmo.eps){width="8cm"} As shown in Figure \[fig\_ximu\], the $1-\mu\lesssim0.1$ part, where the FOG should dominate, has a strong dependence on the value of $\sigma_8$. A larger $\sigma_8$ results in a stronger FOG effect, and thus a sharper peak. The $1-\mu\gtrsim0.1$ part, dominated by the Kaiser effect, seems to have a similar shape with different $\sigma_8$. Appendix C. Accuracy of the Approximately Estimated MCFs {#appendix-c.-accuracy-of-the-approximately-estimated-mcfs .unnumbered} ======================================================== ![ $W(s)$ and $W_{\Delta s}(\mu)$ of the BigMD $z=0.5$ halos, measured in the backgrounds of the fiducial cosmology and a wrong one. Using equations \[eq\_smu\_trans1\] and \[eq\_smu\_trans2\], we can estimate the wrong cosmology results in a fast speed while still maintaining an enough accuracy. []{data-label="fig_wrong_cosmos"}](wrong_cosmos.eps){width="8cm"} Figure \[fig\_wrong\_cosmos\] shows $W(s)$, $W_{\Delta s}(\mu)$ in the fiducial cosmology $(\Omega_m,w)=(0.3071,-1)$, and in the background of a wrong cosmology $(\Omega_m,w)=(0.5,-1)$. The results in the wrong background are computed in two ways, the precise measurement obtained by constructing the sample in the wrong background and then re-measuring the MCFs, and also the approximate results inferred using equations \[eq\_smu\_trans1\] and \[eq\_smu\_trans2\], Inspected by eye, we find the approximate results are very close to their precise correspondance. Since equations \[eq\_smu\_trans1\] and \[eq\_smu\_trans2\] do not capture the change of the weights in the different backgrounds, it is important to check its influence. Here we showed that it is minor compared with the cosmological effect. acknowledgments {#acknowledgments .unnumbered} =============== We thank Kwan-Chuen Chan and Xin Wang for helpful discussions. XDL acknowledges the supported from NSFC grant (No. 11803094), the Science and Technology Program of Guangzhou, China (No. 202002030360). CGS acknowledges financial support from the National Research Foundation of Korea (NRF; \#2020R1I1A1A01073494). J.E. F-R acknowledges support from COLCIENCIAS Contract No. 287-2016, Project 1204-712-50459. We acknowledge the use of the [*Kunlun*]{} cluster, a supercomputer owned by the School of Physics and Astronomy, Sun Yat-Sen University. The CosmoSim database used in this paper is a service by the Leibniz-Institute for Astrophysics Potsdam (AIP). The MultiDark database was developed in cooperation with the Spanish MultiDark Consolider Project CSD2009-00064. [^1]: https://desi.lbl.gov/ [^2]: http://sci.esa.int/euclid/ [^3]: https://www.lsst.org/ [^4]: https://wfirst.gsfc.nasa.gov/ [^5]: webpage: https://www.cosmosim.org [^6]: Notice that these three parameters are actually not independent; $\sigma_8$ is usually considered as a derived parameter crucially dependent on $\Omega_m$ and $A_s$. [^7]: Here we use $s$ instead of $r$ because the statistics is usually performed using the redshift space positions, due to the RSDs they are related with each other via $s = r+v/(aH)$. [^8]: Here we do not study the higher order multipoles, since the $\mu$-dependence is studied in the next Section using another statistical quantity. [^9]: Not only the locations of the peaks/valleys are insensitive to the background change, we find their heights also being rather insensitive to the background. The reason is that, by maintaining a same number density in all backgrounds, we are selecting objects with different bias; the change in the bias counteracts the effect of the background alteration on the clustering strength. [^10]: The covariance matrix of different $\alpha$s have very different magnitude, so we plot the correlation coefficients.
{ "pile_set_name": "ArXiv" }
--- abstract: 'UV-induced photodesorption of ice is a non-thermal evaporation process that can explain the presence of cold molecular gas in a range of interstellar regions. Information on the average UV photodesorption yield of astrophysically important ices exists for broadband UV lamp experiments. UV fields around low-mass pre-main sequence stars, around shocks and in many other astrophysical environments are however often dominated by discrete atomic and molecular emission lines. It is therefore crucial to consider the wavelength dependence of photodesorption yields and mechanisms. In this work, for the first time, the wavelength-dependent photodesorption of pure CO ice is explored between 90 and 170 nm. The experiments are performed under ultra high vacuum conditions using tunable synchrotron radiation. Ice photodesorption is simultaneously probed by infrared absorption spectroscopy in reflection mode of the ice and by quadrupole mass spectrometry of the gas phase. The experimental results for CO reveal a strong wavelength dependence directly linked to the vibronic transition strengths of CO ice, implying that photodesorption is induced by electronic transition (DIET). The observed dependence on the ice absorption spectra implies relatively low photodesorption yields at 121.6 nm (Lyman $\alpha$), where CO barely absorbs, compared to the high yields found at wavelengths coinciding with transitions into the first electronic state of CO (A$^1\Pi$ at 150 nm); the CO photodesorption rates depend strongly on the UV profiles encountered in different star formation environments.' author: - 'Edith C. Fayolle, Mathieu Bertin, Claire Romanzin, Xavier Michaut, Karin I. Öberg, Harold Linnartz, and Jean-Hugues Fillion' title: 'CO ice photodesorption: a wavelength-dependent study' --- Introduction ============ Carbon monoxide, the second most abundant molecule observed in the gas phase of the interstellar medium (ISM) after H$_2$, is also one of the most commonly detected molecules in the solid phase, condensed on the surface of sub-micron sizes dust grains [e.g., @Pontoppidan_03]. Mechanisms triggering CO phase transitions under densities and temperatures encountered in star forming environments are crucial to interpret observations of CO lines. In cold parts of the ISM, most of the CO is depleted onto the grains but non-thermal desorption mechanisms, such as UV photodesorption or cosmic ray sputtering, can maintain a part of the CO budget in the gas phase, explaining the presence of CO below its thermal desorption temperature [e.g., @Willacy_00]. Desorption of CO ice induced by UV photons under astrophysically relevant conditions has recently been studied by [@Oberg_07; @Oberg_09], and [@MunozCaro_10] using an H$_2$ based broadband microwave discharge lamp as photon source and monitoring the ice loss through infrared spectroscopy during irradiation. The derived photodesorption rates are substantially higher than previously assumed [@Greenberg_73], but differ by up to a factor of 20 between the two groups. Absolute photodesorption yield values have substantial uncertainties resulting from (1) the different photon flux calibration – [@Oberg_07; @Oberg_09] used a NIST-calibrated photodiode whereas [@MunozCaro_10] employed a chemical actionometry method – and (2) the IR measurement of the ice sublimation, performed in transmission in the case of [@MunozCaro_10], and in reflection in the case of [@Oberg_07; @Oberg_09]. Nonetheless, the combined uncertainties are estimated to a factor of a few and cannot fully account for the order of magnitude difference between the two groups. The different UV irradiation profiles of the two discharge lamps may instead be at the origin of this discrepancy if the photodesorption efficiency is wavelength dependent. From these studies, excitation of CO in its first electronic state was proposed to induce desorption after energy transfer to neighboring CO molecules and rearrangement of the ice surface (M.C. van Hemert 2010, private communication). Wavelength-resolved studies are required to confirm this mechanism and to investigate whether the photodesorption mechanism of CO is wavelength dependent. Wavelength-resolved photodesorption yields are also important for astrochemical networks. Different FUV field profiles are encountered at different star formation stages and around young stellar objects of different spectral types. The FUV can be dominated by the interstellar radiation field (ISRF) at the edge of molecular clouds, cosmic-rays excited H$_2$ emission in starless clouds [@Gredel_87], black body emission from protostars, and emission lines due to material accretion onto the protostar or pre-main-sequence star [@Bergin_03; @vDishoeck_06]. Thus if photodesorption yields are wavelength-dependent, the photodesorption yield per incident UV photon may vary significantly in different star-forming environments. In order to fully characterize the CO photodesorption process, the present work investigates for the first time the wavelength dependence of CO photodesorption yields for astrophysically relevant ices and conditions. The experimental techniques for ice preparation and synchrotron-based irradiations are explained in Section \[Sec\_exp\]. The results are described in Section \[Sec\_res\] and the mechanism and astrophysical implications are discussed in Section \[Sec\_dis\]. Experimental {#Sec_exp} ============ ![Decrease of the CO vibration RAIRS signal during irradiation of a CO ice 10 ML thick at 18 K by 9.2 eV photons. Inset shows the integrated area of the corresponding CO RAIRS band vs. UV fluence. []{data-label="Fig_RAIRS"}](fig1){width="47.00000%"} CO ices are grown and irradiated in the SPICES (Surface Processes in ICES) apparatus, described in detail elsewhere [@Bertin_11]. CO ice films are prepared under ultra high vacuum (UHV) conditions (base pressure: $5\times10^{-10}$ Torr) on a polycrystalline gold substrate which is cooled down to $18~ \rm K$ by a closed cycle helium cryostat. The CO ices are grown via background deposition, i.e., by exposing the cold Au surface to a partial pressure of CO gas (Air Liquide, 99% purity) during a given time. The amount of CO deposited on the surface is monitored through Reflection Absorption InfraRed Spectroscopy (RAIRS). The thickness of the deposited ice in monolayers (ML) is obtained by performing an isothermal desorption experiment of the CO ice monitored by RAIRS, resulting in the absorbance of 1 monolayer of CO ice [@Oberg_09]. Complementary estimations of the CO ice thickness have also been performed using Temperature Programmed Desorption (TPD) experiments and from the exposure, assuming a sticking coefficient of unity. All techniques gave similar results, yielding ice thicknesses between 9 and 10 ML. The set-up is equipped with two complementary photodesorption detection techniques. A Quadrupole Mass Spectrometer (QMS) allows the detection of gas phase molecules within the UHV chamber, while condensed molecules are probed by RAIRS. Simultaneous probing of the remaining condensed molecules and of the gas-phase photodesorbed molecules is therefore possible during ice irradiation. QMS measurements are used to determine the relative photodesorption rates for all employed UV fluxes, while RAIRS provides absolute yields for UV fluxes higher than 10$^{13}$ photons s$^{-1}$. The UHV set-up is connected to the undulator-based FUV beamline DESIRS at the SOLEIL synchrotron facility in Saint-Aubin (France). In a first irradiation mode, the photon beam from the undulator stage is used directly. The high harmonics of the photon energy are filtered using a rare-gas filled chamber, set in the pass of the UV beam. Absolute incident photon fluxes are measured using calibrated photodiodes. This setting yields high photon flux (10$^{14}$ photons s$^{-1}$) at selected energies with a well defined photon energy distribution (Lorentzian with FWHM of 1 eV). Irradiations at fixed photon energies (8.2, 9.2, 10.2, and 11.2 eV) are performed to monitor the CO ice loss with photon fluence using RAIR spectroscopy (see Figure \[Fig\_RAIRS\]). These experiments are used to derive absolute photodesorption rates for different spectral windows where CO ice absorption is known to differ (see Section \[Sec:kinetics\]). In a second irradiation mode, the photons enter a 6.65 m normal incidence monochromator equipped with a 200 grooves mm$^{-1}$ grating. In this configuration, a photon flux up to 10$^{12}$ photons s$^{-1}$ with a resolution of 40 meV at 10 eV for a 700 $\mu$m exit slit is achieved. The CO ice film is then irradiated by photons whose energy is linearly increased throughout the experiment, from 7 to 13.6 eV (90 to 170 nm). A relative photodesorption rate spectrum is obtained by recording the QMS signal of desorbing CO while tuning the photon energy (see upper panel of Figure \[Fig\_spec\] and Section \[Sec:spec\]). Results {#Sec_res} ======= Absolute Photodesorption Yields for 1 eV Spectral Windows {#Sec:kinetics} --------------------------------------------------------- [l c c]{} \[Tab\_rates\] 8.2 & $2.19 \times 10^{-19}$ & $2.8 \pm 1.7 \times 10^{-2}$\ 9.2 & $1.07 \times 10^{-19}$& $1.3 \pm 0.91 \times 10^{-2}$\ 10.2 & $0.54 \times 10^{-19}$& $6.9 \pm 2.4 \times 10^{-3}$\ 11.2 & $0.74 \times 10^{-19}$& $9.3 \pm 3.4 \times 10^{-3}$ Figure \[Fig\_RAIRS\] shows the decrease of the CO stretching feature during irradiation of a 10 ML thick ice at 18 K by 9.2 eV photons at a flux of $1.2 \times 10^{14} \rm photon s^{-1}$. The double-peaked structure of the infrared feature does not change during ice growth and is most likely due to the roughness of the gold substrate; the peak structure is known to be sensitive to the arrangement of adsorbed molecules on the surface [@Palumbo_06]. The evolution of the CO absorption band integrated area as a function of photon fluence is shown in the inset. The integrated band area is directly proportional to the amount of CO ice (from RAIR spectra acquired during ice growth). The loss of integrated band area during irradiation can therefore be used to quantify CO desorption. The observed linear relationship between CO ice integrated area and photon fluence is consistent with a zeroth-order process, similar to what has been seen for broadband irradiation experiments [@Oberg_07; @MunozCaro_10]. This zeroth-order kinetics is observed for the four explored irradiation energies (8.2, 9.2, 10.2, and 11.2 eV) until the ice coverage gets below 0.03 cm$^{-1}$ integrated absorbance ($\sim$6 ML). The same effect has been observed by [@MunozCaro_10]. Fitting straight lines to the linear parts of the ice loss at the four investigated irradiation energies provides energy-dependent photodesorption yields. The results from the fits are summarized in Table \[Tab\_rates\]. The relative photodesorption yields (absorbance per incident photon flux) are converted into absolute photodesorption yields (CO molecules per incident photon) using the absorbance to ML conversion factor for the present RAIRS set-up (Section \[Sec\_exp\]) and assuming that 1 ML = 10$^{15}$ molecules cm$^{-2}$. The uncertainties in the relative yields are due to the intrinsic RAIRS measurement uncertainties. The larger uncertainties on the absolute yields are due to ice thickness calibration and to differences in the FUV irradiated area versus IR probed area. Indeed, the sample area probed by the IR beam and the UV-beam spot at the surface are $1 \pm 0.1\, \rm cm^2$ and $0.7\pm 0.1\, \rm cm^2$, respectively. Caution has been taken to ensure that the UV irradiated area is integrally probed by the IR beam. The photodesorption yields vary between $\rm 6.9 \, and \, 28 \times 10^{-3}$ molecules photon$^{-1}$ for the investigated energy range and are thus clearly wavelength-dependent. The lowest yield is obtained at 10.2 eV, which is important considering that previous photodesorption results were based on broadband UV discharge lamps, which are often dominated by photons at this excitation energy. In addition, the strong variation with photon energy motivates a more detailed investigation on the energy dependence of the photodesorption yield. CO Photodesorption Yield Spectrum {#Sec:spec} --------------------------------- ![(a) Photodesorption spectrum (after background subtraction) of a 10 ML CO ice at 18 K using a resolution of 40 meV (at 10 eV) and a scan step of 12 meV every 0.5 s. (b) Absorption spectrum of solid CO at 10 K [@Lu_05]. The dashed line indicates photodesorption and absorption values at Ly$\alpha$, 121.6 nm.[]{data-label="Fig_spec"}](fig2){width="47.00000%"} A detailed investigation of the wavelength dependence is achieved in the second kind of experiment, where a 10 ML thick CO ice sample is irradiated by a monochromatic photon beam scanning 7–13.6 eV at a constant rate, while probing photodesorption with the QMS. The QMS current resulting from the desorbing gas phase CO ($m/z = 28$) versus photon energy is presented in Figure \[Fig\_spec\] together with a UV absorption spectra of CO ice from [@Lu_05]. This photodesorption spectrum is converted into molecules photon$^{-1}$ using the absolute photodesorption yields, obtained from the first series of experiments (Section \[Sec:kinetics\], Table \[Tab\_rates\]). The scaling takes into account the exact spectral profiles in the 1 eV spectral-window experiments, which was monitored carefully for the 8.2 eV window. Figure \[Fig\_spec\](a) shows that when probed at a high spectral resolution, the photodesorption yield varies even more across the investigated energy range compared to the 1 eV spectral-window measurements. The lowest yield is achieved below 7.8 eV and at 10.2 eV ($<6\times10^{-3}$ molecules photon$^{-1}$) and it is an order of magnitude lower than the peak value of $5\times10^{-2}$ molecules photon$^{-1}$ at $\sim$ 8.2 eV. The average UV profile below 10 eV results in a total photodesorption yield of $1.8 \times 10^{-2} \rm molecules\, photon^{-1}$. This value is a factor of seven higher than the previously reported yield of $2.7 \times 10^{-3} \rm molecules\, photon^{-1}$ from [@Oberg_09], and comparable within a factor of two to the yield of $3.5 \times 10^{-2} \rm molecules\, photon^{-1}$ found by [@MunozCaro_10]. The relative contribution of Ly$\alpha$ photons (and their inefficiency at inducing photodesorption) to the total photon flux in each experiment is a probable origin for the different yields, further stressing the importance of wavelength resolved studies. Discussion {#Sec_dis} ========== Mechanism --------- The photodesorption spectrum (Figure \[Fig\_spec\](a)) has a band structure below 10 eV that is almost identical to the absorption spectrum of pure CO ice obtained by [@Lu_05]. The observed spectral progression is attributed to the transition from CO ground state to vibrational levels of the first allowed electronic state ($\rm A ^1\Pi$). The striking similarity between the two spectra provides the first experimental evidence that desorption is a DIET (Desorption Induced by Electronic Transition) process, as previously suggested by [@Oberg_07; @Oberg_09]. Desorption of condensed molecules is probably induced by subsequent relaxation of the excited molecules via energetic transfer from electronic to vibrational degrees of freedom [@Avouris_89]. This process clearly dominates the spectrum below 10 eV. Above 10 eV, the main photodesorption mechanism is expected to be more complex since the second absorption band of CO is dissociative. In this case, photodesorption may result from several relaxation pathways, including chemical recombination and desorption, as proposed in the case of water [@Westley_95; @Watanabe_00; @Arasa_10]. It should be noted that electrons, resulting from the interaction of incident UV photons with the metallic substrate, may also contribute to the observed CO ice desorption. Most of the UV photons can be absorbed by the metallic substrate since the ice coverage is an order of magnitude below the coverage limit for which most of the photons would be absorbed by the ice before reaching the substrate. Hot electrons resulting from electron-hole pair excitation of gold are expected to be produced for all investigated irradiation energies and may induce desorption (e.g., [@Bonn_99]). Free electrons photo-emitted from the gold substrate may also induce CO desorption by electronic excitation or resonant electron attachment [@Schultz_73; @Rakhovskaia_95; @Shi_98; @Mann_95]. Only photons whose energy is above 10 eV can produce secondary free electrons energetic enough to trigger the latter process, since the CO-covered gold surface work function is $\sim 4.4 \rm eV$ [@Gottfried_03]. The importance of these electron-induced processes can be estimated from our experiment, since the production of electrons (primary photoelectrons and ineslatically diffused secondary electrons) that have enough kinetic energy to trigger the CO desorption process within the ice, should increase with the photon energy. This effect is barely observed; electrons are most likely the cause of the low-level continuously increasing desorption background seen in Figure \[Fig\_spec\](a). In other words, DIET is the main pathway leading to the observed photodesorption below 10 eV, resulting in a strong energy dependence of the photodesorption yields. Astrophysical implications -------------------------- The determined strong wavelength dependence of the photodesorption yield of CO ice presented in Figure \[Fig\_spec\](a) demonstrates that the CO photodesorption yield is expected to vary significantly in different astrophysical environments. Incorporating this information into astrochemical networks is important to accurately predict the partitioning of CO between the gas and ice. Table \[Tab\_env\] exemplifies how this wavelength dependent data can be used to predict the photodesorption efficiency in sources with characteristic UV fields. The listed photodesorption yields have been calculated by convolving UV field profiles from the literature with the CO photodesorption spectrum presented in Figure \[Fig\_spec\](a) between 90 and 180 nm. The UV profiles are the ISRF from [@Mathis_83] appropriate to model the UV field at the edges of molecular clouds, an emission spectrum of cosmic-ray-excited H$_2$ [@Gredel_87] to mimic UV field encountered in cloud cores, a 10,000 K black body for the UV field around an Herbig Ae star [@vDishoeck_06] and the emission of TW Hydrae from [@Bergin_03] to illustrate photodesorption in protoplanetary disks. Pure Ly$\alpha$ is also added for comparison. The calculated characteristic photodesorption yields vary by a factor of four, with the lowest yields around T Tauri stars and other environments where the UV field is dominated by Ly$\alpha$ photons, and the highest yields per incident photon at cloud edges and next to a 10,000 K black body. The photodesorption yields used here are valid for pure CO ice thicker than $\sim 6\, \rm ML$, as stated in Section \[Sec:kinetics\]. How the CO photodesorption efficiency and mechanism change with ice thickness and/or composition should certainly be studied in order to extend the applicability of the present results to thinner pure CO ices or mixed CO containing astrophysical ices. It is important to note that the UV spectral profile deeper into the cloud and disk may be radically different compared to the UV field incident on the cloud and disk surface because of wavelength-dependent radiative transfer. Determining accurate interstellar relevant photodesorption yields requires a combination of photodesorption spectrum and detailed UV radiative transfer. In the meantime, the presented photodesorption spectrum is recommended to calculate photodesorption yields for specific environments (Table \[Tab\_env\]). [l c c]{} \[Tab\_env\] Edges of clouds & $1.2 \times 10^{-2}$ & [@Mathis_83]\ Pre-stellar cores & $9.4 \times 10^{-3}$& [@Gredel_87]\ Black body 10,000K & $ 1.6 \times 10^{-2}$ & [@vDishoeck_06]\ Tw Hydrae & $6.6 \times 10^{-3}$ & [@Bergin_03]\ Pure ly$\alpha$ & $4.1 \times 10^{-3}$& 121.6 nm Conclusions =========== This study provides the first photodesorption spectrum of pure CO ice, obtained by tunable synchrotron UV irradiation of CO ice under astrophysically relevant conditions. A quantitative determination of the photodesorption yields at 8.2–13.6 eV has been achieved by the simultaneous probe of the ice- and gas- phase CO concentrations coupled to either narrowband (1 eV) excitation mode at selected energies or to continuous energy scanning by monochromatic UV radiation. The resulting photodesorption yields vary by an order of magnitude over the investigated wavelength range. The CO photodesorption process is dominated by the direct electronic excitation of the condensed CO molecules (DIET) below 10 eV. Other desorption mechanisms involving secondary electrons due to UV interactions with the substrate are present, but their contributions to the total measured desorption yield are minor. Consistent with a DIET mechanism, the photodesorption yield of CO ice at Ly$\alpha$ is low. This may result in different chemical evolutions in regions where the UV field is dominated by line emission compared to regions where the UV emission is due to black body radiation. The calibrated photodesorption spectrum presented here should therefore be used to determine the characteristic CO ice photodesorption yields by convolving the CO photodesorption spectrum to the UV profile found in a specific astrophysical environment. In general, the link between wavelength-dependent UV absorption of the ice and the resulting photodesorption and photochemistry is key to implement astrochemical gas-grain models and should be experimentally investigated for relevant ice species. We are grateful to Ewine van Dishoeck and Marc van Hemert for stimulating discussions. We thank Hsiao-Chi Lu, Roland Gredel, and Edwin Bergin for providing useful data. We acknowledge SOLEIL for provision of synchrotron radiation facilities, as well as Laurent Nahon for technical help on the beamline DESIRS. Financial supports from the French national program PCMI (Physique chimie du milieu interstellaire), the Dutch program NOVA (Nederlandse Onderzoekschool Voor Astronomie), the COST action CM0805 “The Chemical Cosmos”, and the Hubert Curien Partenership “Van Gogh”, are gratefully acknowledged. Support for K.I.Ö is provided by NASA through Hubble Fellowship grant awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract NAS 5-26555. Arasa C., Andersson S., Cuppen H. M., van Dishoeck E. F., & Kroes G.-J., 2010, J. Chem. Phys., 132, 184510 Avouris P., & Walkup R. E., 1989, Annu. Rev. Phys. Chem., 40, 173–206, *and references therein* Bergin E., Calvet N., D’Alessio P. & Herczeg G. J., 2003, , 591, L159–L162 Bertin M., Romanzin C., Michaut X., Jeseck P., & Fillion J.-H., 2011, J. Phys. Chem. C, 115, 12920-–12928 Bonn M., Funk S., Hess Ch., Denzler D. N., Stampfl C., Scheffler M., Wolf M. & Ertl G., 1999, Science, 285, 1042–1045 Gottfried J. M., Schmidt K. J., Schroeder S. L. M., & Christmann K., 2003, Surf. Sci., 536, 206–224 Gredel R., Lepp S. & Dalgarno A., 1987, , 323, L137–L139 Greenberg L. T., 1973, Interstellar Dust and Related Topics, J. M. Greenberg & H. C. van de Hulst, IAU Symposium, 1973, 52 Lu H.-C., Chen H.-K., Cheng B.-M., Kuo Y.-P. & Ogilvie J.F., 2005, J. Phys. B: At. Mol. Opt. Phys., 38, 3693-–3704 Mann A., Cloutier P., Liu D., & Sanche L., 1995, Phys. Rev. B, 51, 7200–7206 Mathis J. S., Mezger P. G. & Panagia N., 1983, , 128, 212–229 Muñoz Caro G.M., Jiménez-Escobar A., Martín-Gago J. A., Rogero C., Atienze C., Puertas S., Sobrado J.M. & Torres-Redondo J., 2010, , 522, A108 Öberg K. I., Fuchs G. W., Awad Z., Fraser H. J., Schlemmer S., van Dishoeck E. F. & Linnartz H., 2007, , 662, L23–-L26 Öberg K. I., van Dishoeck E. F. & Linnartz H., 2009, , 496, 281–-293 Palumbo M. E., Baratta G. A., Collings M. P. & McCoustra M. R. S., 2006, Phys. Chem. Chem. Phys., 8, 279–284 , K. M., [Fraser]{}, H. J., [Dartois]{}, E., [et al.]{} 2003, , 408, 981 Rakhovskaia O., Wiethoff P., & Feulner P., 1995, NIM B, 101, 169 Schulz G. J., 1973, Rev. Mod. Phys., 45, 423–486 Shi H., Cloutier P., & Sanche L., 1998, Low Temp. Phys., 24, 742 van Dishoeck E. F., Bastiaan J. & van Hemert M. C., 2006, Faraday Discuss., 133, 231–243 Watanabe N., Horii T. & Kouchi A., 2000, , 541, 772–778 Westley M. S., Baragiola R. A., Johnson R. E. & Baratta G. A., 1995, Nature, 373, 405–407 Willacy K. & Langer W. D., 2000, , 544, 903–920
{ "pile_set_name": "ArXiv" }
--- date: 'December, 2012' title: | The quiver approach to the BPS spectrum\ of a $4d$ $\cn=2$ gauge theory --- Introduction {#intt} ============ In the last few years many new powerful methods were introduced to compute the exact BPS spectrum of a four–dimensional $\cn=2$ supersymmetric QFT. We may divide the methods in two broad classes: *i)* geometric methods [@GMN09; @GMN10; @GMN11; @GMN12; @GMN12b] and *ii)* algebraic methods [@CNV; @CV11; @arnold; @ACCERV1; @ACCERV2; @cattoy; @half; @nons; @nonsX]. The geometric methods give a deep understanding of the non–perturbative physics, while the algebraic ones are quite convenient for actual computations. In the algebraic approach the problem of computing the BPS spectrum is mapped to a *canonical* problem in the Representation Theory (RT) of (basic) associative algebras. A lot of classical results in RT have a direct physical interpretation and may be used to make the BPS spectral problem ‘easy’ for interesting classes of $\cn=2$ theories. Besides, by comparing RT and physics a lot of interesting structures emerge which shed light on both subjects. From $\cn=2$ QFT to quiver representations {#qquiq} ------------------------------------------ To fix the notation, we recall how the BPS states are related to quiver representations, referring to [@ACCERV2] for more details. The conserved charges of the theory (electric, magnetic, and flavor) are integrally quantized, and hence take value in a lattice $\Gamma=\oplus_v\, \Z e_v$. On $\Gamma$ we have a skew–symmetric integral pairing, $\langle\gamma, \gamma^\prime\rangle_\text{Dirac}\in \Z$, given by the Dirac electro–magnetic pairing; the flavor charges then correspond to the zero–eigenvectors of the matrix $B_{uv}\equiv \langle e_u, e_v\rangle_\text{Dirac}\in \Z$. Following [@CV11] we say that our $\cn=2$ model has the *quiver property* if we may find a set of generators $\{e_v\}$ of $\Gamma$ such that the charge vectors $\gamma\in\Gamma$ of all the BPS particles satisfy $$\gamma\in \Gamma_+\quad\text{or}\quad-\gamma\in \Gamma_+,$$ where $\Gamma_+\equiv \oplus_v \,\Z_+\,e_v$ is the *positive cone* in $\Gamma$. Given a $\cn=2$ theory with the quiver property, we associate a $2$–acyclic quiver $Q$ to the data $(\Gamma_+, \langle\cdot,\cdot\rangle_\text{Dirac})$: to each positive generator $e_v$ of $\Gamma_+$ we associate a node $v$ of $Q$ and we connect the nodes $u$ $v$ with $B_{uv}$ arrows $u\rightarrow v$ (a negative number meaning arrows in the opposite direction). The positive cone $\Gamma_+\subset\Gamma$ is then identified with the cone of dimension vectors of the representations $X$ of $Q$ trough $\mathbf{dim}\,X\equiv \sum_v \dim X_v\, e_v$. The emergence of the quiver $Q$ may be understood as follows. Fix a particle with charge $\gamma= \sum_v N_v\,e_v\in\Gamma_+$; on its word–line we have a one dimensional supersymmetric theory with 4 supercharges, and the BPS particles correspond to states which are <span style="font-variant:small-caps;">susy</span> vacua of this 1d theory. The 1d theory turns out to be a quiver theory in the sense that its Kähler target space is the representation space of $Q$ of dimension $\sum_v N_v \,e_v$ $$\label{ppqD}\prod_{\text{arrows}\atop u\rightarrow v} \C^{N_{u} N_{v}}\Big/\Big/ \prod_{\text{nodes}\atop v} U(N_v)\qquad \begin{footnotesize}\text{(symplectic quotient)}\end{footnotesize}.$$ To completely define the 1d theory we need to specify a $\prod_\text{nodes} U(N_v)$–invariant superpotential $\cw$ (and the FI terms implicit in ); gauge invariance requires $\cw$ to be a function of the traces of the products of the bi–fundamental Higgs fields along the closed oriented loops in $Q$. In facts it turns out that this function must be linear (a sum of single–trace operators) and thus canonically identified with a linear combination (with complex coefficients) of the oriented cycles in $Q$. Thus $\cw$ is a potential for the quiver $Q$ in the sense of DWZ [@derksen1]. One shows [@ACCERV2] that a 1d configuration is a classical <span style="font-variant:small-caps;">susy</span> vacuum if and only if the bi–fundamental Higgs fields associated to the arrows of $Q$ form a *stable* module $X$ of the Jacobian algebra[^1] $$\mathscr{J}\!(Q,\cw):=\C Q/(\partial\cw),$$ and two field configurations are physically equivalent iff the corresponding modules are isomorphic. *Stability* is defined in terms of the central charge $Z$ of the $\cn=2$ <span style="font-variant:small-caps;">susy</span> algebra. Being conserved, $Z$ is a linear combinations of the various charges; hence may be seen as a linear map $Z\colon \Gamma\rightarrow \C$. We assume $\mathrm{Im}\,Z(\Gamma_+) \geq0$, so that we have a well–defined function $\arg Z\colon \Gamma_+\rightarrow [0,\pi]$. Then $X\in\mathsf{mod} \mathscr{J}\!(Q,\cw)$ is *stable* (with respect to the given central charge $Z$) iff, for all proper non–zero submodules $Y$, $\arg Z(Y)<\arg Z(X)$. In particular, *$X$ is stable* $\Rightarrow$ *$X$ is a brick,* $\Leftrightarrow \mathrm{End}\,X=\C$. The isoclasses of stable modules of given dimension $\gamma$ typically form a family parameterized by a Kähler manifold $\cm_\gamma$; from the viewpoint of the 1d theory the space $\cm_\gamma$ corresponds to zero–modes which should be quantized producing $SU(2)_\text{spin}\times SU(2)_R$ quantum numbers. In particular, a $d$–dimensional family corresponds (at least) to a BPS supermultiplet with spin content $(0,\tfrac{1}{2})\otimes \frac{d}{2}$ (thus rigid modules corresponds to hypermultiplets, $\mathbb{P}^1$–families to vector supermultiplets, and so on). Notice that the full dependence of the BPS spectrum from the parameters of the theory is encoded in the central charge $Z$, which depends on these parameters as specified by the Seiberg–Witten geometry. For a given $\cn=2$ theory $(Q,\cw)$ is *not* unique; indeed there may be several sets of generators $\{e_v\}$ with the above properties. Two allowed $(Q,\cw)$ are related by a Seiberg duality, which precisely coincides with the mutations of a quiver with potential in the sense of cluster algebras [@derksen1] (this, in particular, requires $\cw$ to be non–degenerate in that sense). Therefore, to a QFT we associate a full *mutation class* of quivers. This mutation class is finite iff the theory is complete [@CV11] which, in particular, implies that no BPS state has spin larger than $1$. \[t2duality\] The Seiberg duality/DWZ mutation is not the only source of quiver non–uniqueness. The quiver mutations preserve both the number of nodes and $2$–acyclicity. There are more general dualties which do not share these properties. As an example consider the Gaiotto theory corresponding to the $A_1$ $(6,0)$ theory on a sphere with 3 regular punctures (the $T_2$ theory) [@Gaiotto]. $T_2$ consists of 4 free hypermultiplets, carrying 4 flavor charges, which corresponds to a disconnected quiver with 4 nodes and no arrows. On the other hand, we may associate to it a quiver with only *three* nodes, each pair of nodes being connected by a pair of opposite arrows $\rightleftarrows$ [@ACCERV2]. We refer to the equivalence of the two quivers as ‘$T_2$–duality’. The $(Q,\cw)$ class associated to a $\cn=2$ theory ================================================== The BPS states correspond to the stable bricks of the Jacobian algebra. This reduces our problem to a standard problem in Representation Theory *provided* we know which $(Q,\cw)$ mutation class is associated to our $\cn=2$ theory. Determing the mutation class for several interesting gauge theories is the main focus of the present note. For $\cn=2$ models having a corner in their parameter space with a weakly coupled Lagrangian description, we have a very physical criterion to check whether a candidate pair $(Q,\cw)$ is correct. Simply use the category $\mathsf{mod}\mathscr{J}\!(Q,\cw)$ to compute the would–be BPS spectrum in the limit of vanishing YM coupling $g_\mathrm{YM}\rightarrow 0$ and compare the result with the prediction of perturbation theory. The weakly coupled spectrum should consists of - finitely many mutually–local states with bounded masses as $g_\mathrm{YM}\rightarrow 0$: 1. vector multiplets making of the adjoint representation of the gauge group $G$ (photons and $W$–bosons); 2. hypermultiplets making definite (quaternionic) representations $R_a$ of $G$ (quarks); - particles non–local relatively to the $W$–bosons with masses $O(1/g^2_\mathrm{YM})$ (heavy dyons). We ask which pairs $(Q,\cw)$ have such a property (the Ringel property [@cattoy]). Magnetic charge & weak coupling regime {#Dirrrac} -------------------------------------- Consider a quiver $\cn=2$ gauge theory having a weak coupling description with gauge group $G$ (of rank $r$). We pick a particular pair $(Q,\cw)$ in the corresponding Seiberg mutation–class which is appropriate for the weak coupling regime (along the Coulomb branch). $\mathsf{mod}\mathscr{J}\!(Q,\cw)$ should contain, in particular, one–parameter families of representations corresponding to the massive $W$–boson vector–multiplets which are in one–to–one correspondence with the positive roots of $G$. We write $\delta_a$ ($a=1,2,\dots, r$) for the charge (*i.e.* dimension) vector of the $W$–boson associated to the *simple–root* $\alpha_a$ of $G$. At a generic point in the Coulomb branch we have an unbroken $U(1)^r$ symmetry. The $U(1)^r$ electric charges, properly normalized so that they are integral for all states, are given by the fundamental coroots[^2] $\alpha_a^\vee\in\mathfrak{h}$ ($a=1,2,\dots,r$). The $a$–th electric charge of the $W$–boson associated to $b$–th simple root $\alpha_b$ then is $$q_a=\alpha_a(\alpha^\vee_b)=C_{ab}, \qquad \text{(the Cartan matrix of }G).$$ Therefore the vector in $\Gamma\otimes \mathbb{Q}$ corresponding to the $a$–th unit electric charge is $$\mathfrak{q_a}=(C^{-1})_{ab}\,\delta_b.$$ Then the magnetic weights (charges) of a representation $X$ are given by $$\label{magneticxxx} m_a(X)\equiv \langle \dim X, \mathfrak{q}_a\rangle_\text{Dirac}=(C^{-1})_{ab}\,B_{ij}\,(\dim X)_i\,(\delta_b)_j.$$ Dirac quantization requires the $r$ linear forms $m_a(\cdot)$ to be *integral* [@cattoy]. This integrality condition is quite a strong constraint on the quiver $Q$, and is our main tool to determine it. At weak coupling, $g_\mathrm{YM}\rightarrow 0$, the central charge takes the classical form [@cattoy] $$\label{zweak} Z(X)= -\frac{1}{g^2_\mathrm{YM}}\,\sum_i C_a\,m_a(X)+O(1),$$ where $C_a=-i \langle \varphi_a\rangle >0$ in the region of the Coulomb branch covered by the quiver $Q$. It is convenient to define the light category, $\mathscr{L}(Q,\cw)$, as the subcategory of the modules $X\in\mathsf{mod}\mathscr{J}\!(Q,\cw)$ with $m_a(X)=0$ for all $a$ such that all their submodules have $m_a(Y)\leq 0$. Comparing with the definition of stability in §.\[qquiq\], we see that all BPS states with bounded mass in the limit $g_\mathrm{YM}\rightarrow 0$ correspond to modules in $\mathscr{L}(Q,\cw)$, and, in facts, for a $\cn=2$ theory which has a weakly coupled Lagrangian description the stable objects of $\mathscr{L}(Q,\cw)$ precisely match the perturbative states. They are just the gauge bosons, making one copy of the adjoint of $G$, together with finitely many hypermultiplets transforming in definite representations of $G$. The detailed structure of $\mathscr{L}(Q,\cw)$ is described in [@cattoy]. **Remarks & Properties** 1. $\mathsf{mod}\mathscr{J}\!(Q,\mathcal{W})$ contains *many* ligh subcategories, one for each weakly coupled corner. *E.g.* $SU(2)$ $N_f=4$ has a $SL(2,\mathbb{Z})$ orbit of such subcategories; 2. $m(\Gamma_+)\not\geq 0$ $\Rightarrow$ the light category is *not* the restriction to a subquiver, and its quiver is *not* necessarily $2$–acyclic (as in the $T_2$ case [@ACCERV2; @cattoy]); 3. the category $\mathscr{L}(Q,\mathcal{W})$ is *tame* (physically: no light BPS state of spin $>1$); 4. *universality of the SYM sector*: for given gauge group $G$ $$\mathscr{L}(Q_\mathrm{SYM},\mathcal{W}_\mathrm{SYM})\subset \mathscr{L}(Q,\mathcal{W})$$ where $(Q_\mathrm{SYM},\mathcal{W}_\mathrm{SYM})$ is the pair for pure $G$ SYM. Only finitely many bricks $X\in \mathscr{L}(Q,\mathcal{W})$ and $ X\not\in \mathscr{L}(Q_\mathrm{SYM},\mathcal{W}_\mathrm{SYM})$, they correspond to ‘quarks’. First examples ============== As a warm–up we consider four classes of (simple) examples. Example 1: $SU(2)$ SQCD with $N_f\leq 4$ ---------------------------------------- These examples are discussed in detail in [@CV11; @ACCERV2; @cattoy]; here we limit ourselves to a description of the resulting categories. One shows [@cattoy] that the category $\mathsf{mod}\mathscr{J}\!(Q,\cw)$ is Seiberg–duality equivalent to the Abelian category $\mathrm{Coh}(\mathbb{P}^1_{N_f})$ where $\mathbb{P}^1_{N_f}$ is $\mathbb{P}^1$ with $N_f$ ‘double points’, that is, the variety in the weighted projective space $W\mathbb{P}(2,2,\dots,2,1,1)$ of equations $$X_i^2-\lambda_i\,X_{N_f+1}-\mu_i\,X_{N_f+2}=0,\qquad i=1,2,\dots, N_f,\quad (\lambda_i:\mu_i)\in\mathbb{P}^1.$$ In $\mathrm{Coh}(\mathbb{P}^1_{N_f})$ we have two quantum numbers, *degree* and *rank* $$\text{rank = magnetic charge,}\qquad \text{degree = $2\times$ electric charge}.$$ The light subcategory $\mathrm{Coh}(\mathbb{P}^1_{N_f})\supset\mathscr{L}=\{\text{sheaves of finite length}\}$ *a.k.a.* ‘skyscrapers’, while the dyons correspond to line bundles of various degree. For $N_f=4$ the curve $\mathbb{P}^1_4$ is Calabi–Yau, hence an elliptic curve $E$. The moduli space of the degree 1 skyscrapers, which is the curve $E$ itself, is isomorphic to its Jacobian $J(E)$ which parameterizes the line bundles of fixed degree. Quantization of $J(E)$ then produces magnetic charged vector–multiplets. Of course, $E\sim J(E)$ reflects the $S$–duality of the theory. See [@cattoy] for more details. Example 2: SYM with a simply–laced gauge group $G$ {#99ert} -------------------------------------------------- The quiver exchange matrix $B$ is fixed by the Dirac charge quantization [@cattoy] (cfr. §.\[Dirrrac\]). The standard quiver (the square form) corresponds to $$B_\square= C\otimes S, \qquad \text{where }\ \begin{cases}C\ \text{is the Cartan matrix of }G,\\ S\ \text{is the modular $S$--matrix.}\end{cases}$$ $$\begin{gathered}\xymatrix{\alpha_1^{(1)} \ar@<0.5ex>[d]\ar@<-0.5ex>[d] & \alpha_2^{(2)}\ar[l]\ar[r] & \alpha_3^{(1)} \ar@<0.5ex>[d]\ar@<-0.5ex>[d] & \alpha_4^{(2)}\ar[l]\ar[r] & \alpha_5^{(1)} \ar@<0.5ex>[d]\ar@<-0.5ex>[d]\\ \alpha_1^{(2)}\ar[r] &\alpha_2^{(1)}\ar@<0.5ex>[u]\ar@<-0.5ex>[u] & \alpha_3^{(2)}\ar[l]\ar[r] & \alpha_4^{(1)}\ar@<0.5ex>[u]\ar@<-0.5ex>[u] & \alpha_5^{(2)}\ar[l]}\end{gathered}$$ The [square]{} quiver is represented (for $G=SU(6)$) in figure \[yzzrre\]; it is supplemented by a quartic superpotential $\cw$ [@ACCERV2; @cattoy]. The charge vector of the $a$–th simple root $W$–boson is equal to $\delta_a\equiv \alpha_a^{(1)}+\alpha_a^{(2)}$, *i.e.* the $a$–th simple–root $W$ bosons corresponds to the $\mathbb{P}^1$–family of bricks associated with the minimal imaginary root of the $a$–th $\widehat{A}(1,1)$ affine subquiver $\upuparrows_a$. The $a$–th magnetic charge (weight) is (cfr. eqn.) $$m_a(X)= \dim X_{\alpha_a^{(1)}}-\dim X_{\alpha_a^{(2)}}.$$ From the discussion around eqn., the light subcategory $\mathcal{L}^\mathrm{YM}(G)$ containing the perturbative BPS spectrum is then given by the modules $X\in\mathsf{mod}\mathscr{J}\!(Q,\cw)$ with $m_a(X)=0$ such that all their submodules $Y$ satisfy $m_a(Y)\leq 0$, $\forall a$. We may break $G\rightarrow SU(2)_a\times U(1)^{r-1}$ at weak coupling and describe the Higgs mechanism perturbatively; that is, the gauge breaking should respect the light *sub*category. Mathematically, this gives the following result at the level of Abelian categories of modules $$X\in \mathscr{L}^\mathrm{YM}(G) \ \Rightarrow \ X\big|_{\upuparrows_a}\in \mathscr{L}^\mathrm{YM}(SU(2))\quad\forall\,a,$$ which may be checked directly. Then, if $X$ is indecomposable, in each Kronecker subquiver $\upuparrows_a$ we may set one of the arrows to $1$ with the result that the category $\mathscr{L}^\mathrm{YM}(G)$ gets identified with the category of modules of a Jacobian algebra $$\mathscr{L}^\mathrm{YM}(G)=\mathsf{mod}\mathscr{J}\!(Q^\prime,\mathcal{W}^\prime)$$ where $Q^\prime$ is the *double*[^3] of the $G$ Dynkin graph with loops $A_v$ attached at the nodes (the ‘$\mathcal{N}=2$ quiver’ of $G$), see figure \[0023c\] for the $SU(6)$ example. The reduced quiver $Q^\prime$ is equipped with the superpotential $$\label{eerrq} \mathcal{W}^\prime=\sum_{a:\;\overrightarrow{\text{edges}}\in G} \mathrm{tr}\big(\widetilde{\psi}_aA_{t(a)}\psi_{a}-{\psi}_aA_{h(a)}\widetilde{\psi}_{a}\big).$$ $$\begin{aligned} & \begin{gathered} \xymatrix@R=2.0pc@C=3.0pc{ &\ar@(ul,dl)[]_{A_{1}} \alpha_1 \ar@/_0.7pc/[r]_{\widetilde{\psi}_{1}} &\ar@/_0.7pc/[l]_{\psi_{1}} \ar@(dr,dl)[]^{A_2} \alpha_2 \ar@/_0.7pc/[r]_{\widetilde{\psi}_2} &\ar@/_0.7pc/[l]_{\psi_{2}} \ar@(dr,dl)[]^{A_3} \alpha_3 \ar@/_0.7pc/[r]_{\widetilde{\psi}_3}&\ar@/_0.7pc/[l]_{\psi_{3}} \ar@(dr,dl)[]^{A_4} \alpha_4 \ar@/_0.7pc/[r]_{\widetilde{\psi}_4}&\ar@/_0.7pc/[l]_{\psi_{4}} \ar@(ur,dr)[]^{A_5} \alpha_5 } \end{gathered}\end{aligned}$$ Given a module $X\in\mathsf{mod}\mathscr{J}\!(Q^\prime,\mathcal{W}^\prime)$, consider the linear map $$\ell\colon(X_{\alpha_1},X_{\alpha_2},\cdots, X_{\alpha_r})\mapsto (A_1 X_{\alpha_1},A_2 X_{\alpha_2},\cdots, A_r X_{\alpha_r}).$$ It is easy to check that $\ell\in \mathrm{End}\, X$, hence $X$ a brick $\Rightarrow$ $A_i=\lambda\in \mathbb{C}$ for all $i$ (in facts, $\lambda\in \mathbb{P}^1$). Fixing $\lambda\in \mathbb{P}^1$, the brick $X$ is identified with a brick of the double $\overline{G}$ of the Dynkin graph[^4] $$\begin{aligned} \overline{A_5} \hskip -0.8cm\begin{gathered} \xymatrix@R=2.0pc@C=3.0pc{ & \alpha_1 \ar@/_0.7pc/[r]_{\widetilde{\psi}_{1}} &\ar@/_0.7pc/[l]_{\psi_{1}} \alpha_2 \ar@/_0.7pc/[r]_{\widetilde{\psi}_2} &\ar@/_0.7pc/[l]_{\psi_{2}} \alpha_3 \ar@/_0.7pc/[r]_{\widetilde{\psi}_3}&\ar@/_0.7pc/[l]_{\psi_{3}} \alpha_4 \ar@/_0.7pc/[r]_{\widetilde{\psi}_4}&\ar@/_0.7pc/[l]_{\psi_{4}} \alpha_5 } \end{gathered}\end{aligned}$$ subjected to relations $$\label{rerrt} \sum_{t(a)=v} \psi_a\widetilde{\psi}_a-\sum_{h(a)=v}\widetilde{\psi}_a\psi_a=0.$$ The algebra defined by the double quiver $\overline{G}$ with the relations is known as the Gelfand–Ponomarev *preprojective algebra* of the graph $G$, written $\mathcal{P}(G)$ [@pre1]. There are three basic results on the preprojective algebra of a graph $L$: - Gelfand and Ponomarev [@pre1]: $\dim \mathcal{P}(L)<\infty$ if and only if $L$ is an $ADE$ Dynkin graph; - Crawley–Boevey [@CBlemma]: Let $C_L=2-I_L$ be the Cartan matrix of the graph $L$. Then for all $X\in \mathsf{mod}\,\mathcal{P}(L)$ $$\label{xxx4561} 2\, \dim\mathrm{End}\, X= (\mathbf{dim}\, X)^t\, C_L(\mathbf{dim}\, X)+\dim \mathrm{Ext}^1(X,X)$$ - Lusztig [@LU]: Let $X$ be an indecomposable module of $\mathcal{P}(L)$ belonging to a family of non–isomorphic ones parameterized by the (Kähler) moduli space $\cm(X)$. Then $$\label{xxx4562}\dim \mathcal{M}(X)=\tfrac{1}{2}\dim \mathrm{Ext}^1(X,X).$$ If $L$ is an $ADE$ graph $G$, the integral quadratic form $v^t\, C_G\, v$ is positive–definite and even; then $X\neq 0$ implies $(\mathbf{dim}\, X)^t\, C_L(\mathbf{dim}\, X)\geq 2$ with equality if and only if $\mathbf{dim}\,X$ is a positive root of $G$. From eqns. it follows that if $X$ is a brick of $\mathcal{P}(G)$ it must be rigid with $\mathbf{dim}\,X$ a positive root of $G$. Going back to $\mathscr{L}^\mathrm{YM}(G)$, we see that a module in the light category is a brick iff $\mathbf{dim}\, X$ is a positive root of $G$ and $\mathcal{M}(X)=\mathbb{P}^1$. By the dictionary between physics and Representation Theory, this means that the BPS states which are stable and have bounded mass as $g_\mathrm{YM}\rightarrow 0$ are *vector–multiplets in the adjoint of* the gauge group $G$. In facts, a more detailed analysis shows [@cattoy] that there is precisely one copy of the adjoint in each weakly coupled BPS chamber. This is, clearly, the result expected for pure SYM at weak coupling; in particular, is shows that the identification [@CNV] of $(Q,\mathcal{W})$ is correct. Example 3: SQCD with $G$ simply–laced and\ $N_a$ quarks in the $a$–th fundamental representation ----------------------------------------------------- We consider $\cn=2$ SQCD with a simply–laced gauge group $G=ADE$ coupled to $N_a$ *full* hypermultipletss in the representation $F_a$ with Dynkin label $[0,\cdots, 0, 1,0,\cdots,0]$ ($1$ in the $a$–th position, $a=1,2,\dots, r$). The prescription for the quiver is simple [@ACCERV2]: one replaces the $a$–th Kronecker subquiver $\downdownarrows_a$ of the pure $G$ SYM quiver (cfr. §.\[99ert\]) as follows $$\label{p223r} \begin{gathered}\xymatrix{\ar@{..}[r] &\bullet\ar@<0.4ex>[dd] ^{B_a}\ar@<-0.4ex>[dd]_{A_a}\ar@{..}[r] &\\ \\ \ar@{..}[r] &\bullet\ar@{..}[r] & }\end{gathered}\quad \xrightarrow{\phantom{mmmmmmmm}}\quad \begin{gathered}\xymatrix{\ar@{..}[r] &\bullet\ar@<0.4ex>[dd]^{B_a}\ar@<-0.4ex>[dd]_{A_a}\ar@{..}[r] &\\ && &\bullet\ar[ull]^{\phi_1} &\cdots& \bullet\ar[ullll]_{\phi_{N_a}} \\ \ar@{..}[r] &\bullet\ar[rru]^{\widetilde{\phi}_1}\ar[rrrru]_{\widetilde{\phi}_ {N_a}}\ar@{..}[r] & }\end{gathered}$$ and replaces the pure SYM superpotential $\cw_\mathrm{SYM}$ with $$\begin{gathered} \label{p223rs} \mathcal{W}\longrightarrow \mathcal{W}_\mathrm{SYM}+\sum_{i=1}^{N_i}\mathrm{tr}\big[(\alpha_i\,A_a-\beta_i\,B_a)\phi_i\,\widetilde{\phi}_i\big],\\ (\alpha_i:\beta_i)\equiv \lambda_a\in\mathbb{P}^1\ \text{pairwise distinct}.\end{gathered}$$ The exchange matrix of the resulting quiver, $B$, has $N_i$ zero eigenvalues corresponding to the $N_a$ flavor charges carried by the quarks. Formally [@ACCERV2], we may extendend this construction to the case in which we have quarks in several distinct fundamental representations, just be applying the substitutions to all the corresponding Kronecker subquivers of the (square) pure SYM quiver. Going trough the same steps as in §.\[99ert\], one sees that the light category $\mathscr{L}=\mathsf{mod}\mathscr{J}\!(Q^\prime, \mathcal{W}^\prime)$ with $Q^\prime$ the double of the graph $G[a,N_a]$ obtained by adding $N_a$ extra nodes to the Dynkin graph $G$ connected with a single hedge to the $a$–th node of $G$ and having loops only at all ‘old’ nodes of $G$ [@cattoy] (see figure \[ll451\] for a typical example) and superpotential $$\mathcal{W}^\prime= \mathcal{W}^\prime_\mathrm{SYM}+\sum\nolimits_i \mathrm{tr}\big[(\alpha_i\,A_a-\beta_i)\phi_i\,\widetilde{\phi}_i\big].$$ $$\begin{aligned} & \begin{gathered} \xymatrix@R=2.0pc@C=3.0pc{& & & h \ar@/_0.7pc/[dd]_{{\phi}_{1}} \\ \\ &\ar@(ul,dl)[]_{A_{1}} \alpha_1 \ar@/_0.7pc/[r]_{\widetilde{\psi}_{1}} &\ar@/_0.7pc/[l]_{\psi_{1}} \ar@(dr,dl)[]^{A_2} \alpha_2 \ar@/_0.7pc/[r]_{\widetilde{\psi}_2} &\ar@/_0.7pc/[l]_{\psi_{2}} \ar@(dr,dl)[]^{A_3} \alpha_3\ar@/_0.7pc/[uu]_{\widetilde{\phi}_{1}} \ar@/_0.7pc/[r]_{\widetilde{\psi}_3}&\ar@/_0.7pc/[l]_{\psi_{3}} \ar@(dr,dl)[]^{A_4} \alpha_4 \ar@/_0.7pc/[r]_{\widetilde{\psi}_4}&\ar@/_0.7pc/[l]_{\psi_{4}} \ar@(ur,dr)[]^{A_5} \alpha_5 } \end{gathered}\end{aligned}$$ As in §.\[99ert\], $X$ is a brick $\boldsymbol{\Rightarrow}$ $A_i=\lambda\in\mathbb{P}^1$. Now we have two distinct cases: 1. $\lambda$ is (*i.e.* $\lambda\neq \lambda_i$, $i=1,2,\dots, N_a$): the Higgs fields $\phi_i,\widetilde{\phi}_i$ are *massive* and may be integrated out. Then $X$ is a brick of $\mathcal{P}(G)$ and its charge vector $\mathbf{dim}\,X$ is a positive root of $G$. These are the same representations as for the light category of pure SYM and they correspond to $W$–bosons in the adjoint of $G$; 2. $\lambda=\lambda_a$, then $X$ is a brick of the preprojective algebra $\mathcal{P}(G[i,1])$. Right properties (finitely many, rigid, in right reprs. of $G$) $G[i,1]$ is also a Dynkin graph. By comparison one gets the following [@cattoy]: **Theorem.** *(1) Consider $\mathcal{N}=2$ SYM with simple simply–laced gauge group $G$ coupled to a hyper in a representation of the form $F_a=[0,\cdots, 0, 1,0,\cdots,0]$. The resulting QFT is *Asymptotically Free* if and only if the augmented graph $G[a,1]$ obtained by adding to the Dynkin graph of $G$ an extra node connected by a single edge to the $a$–th node of $G$ is also an $ADE$ Dynkin graph. (2) The model has a Type IIB engineering iff, in addition, the extra node is an extension node in the extended (affine) augmented Dynkin graph $\widehat{G[a,1]}$.* $$\begin{aligned} &SU(N)\ \text{with }\mathbf{N} &&\begin{gathered}\xymatrix{\bullet \ar@{-}[r] & \bullet\ar@{-}[r] & \bullet \ar@{-}[r] &\cdots \ar@{-}[r] & \bullet \ar@{..}[r] & 0}\end{gathered}\\ \\ &SU(N)\ \text{with }\mathbf{N(N-1)/2} &&\begin{gathered} \xymatrix{\bullet \ar@{-}[r] & \bullet\ar@{-}[r] & \bullet \ar@{-}[r] &\cdots \ar@{-}[r] & \bullet \ar@{-}[r] & \bullet\\ & & & & 0\ar@{..}[u]}\end{gathered}\\ &SU(6)\ \text{with }\mathbf{20} &&\begin{gathered} \xymatrix{\bullet \ar@{-}[r] & \bullet\ar@{-}[r] & \bullet \ar@{-}[r] & \bullet \ar@{-}[r] & \bullet\\ & & 0\ar@{..}[u]}\end{gathered}\\ &SU(7)\ \text{with }\mathbf{35} &&\begin{gathered} \xymatrix{\bullet \ar@{-}[r] & \bullet\ar@{-}[r] & \bullet \ar@{-}[r] & \bullet \ar@{-}[r] & \bullet\ar@{-}[r]& \bullet\\ & & 0\ar@{..}[u]}\end{gathered}\\ &SU(8)\ \text{with }\mathbf{56} &&\begin{gathered} \xymatrix{\bullet \ar@{-}[r] & \bullet\ar@{-}[r] & \bullet \ar@{-}[r] & \bullet \ar@{-}[r] & \bullet\ar@{-}[r]& \bullet\ar@{-}[r] &\bullet\\ & & 0\ar@{..}[u]}\end{gathered}\\ &SO(2n)\ \text{with }\mathbf{2n} &&\begin{gathered}\xymatrix{0 \ar@{..}[r] & \bullet\ar@{-}[r] & \bullet \ar@{-}[r] &\cdots \ar@{-}[r] & \bullet \ar@{-}[r] & \bullet\\ & & & & \bullet\ar@{-}[u]}\end{gathered}\\ &SO(10)\ \text{with }\mathbf{16} &&\begin{gathered} \xymatrix{0\ar@{..}[r] & \bullet\ar@{-}[r] & \bullet \ar@{-}[r] & \bullet \ar@{-}[r] & \bullet\\ & & \bullet\ar@{-}[u]}\end{gathered}\\ &SO(12)\ \text{with }\mathbf{32} &&\begin{gathered} \xymatrix{0 \ar@{..}[r] & \bullet\ar@{-}[r] & \bullet \ar@{-}[r] & \bullet \ar@{-}[r] & \bullet\ar@{-}[r]& \bullet\\ & & \bullet\ar@{-}[u]}\end{gathered}\\ &SO(14)\ \text{with }\mathbf{64} &&\begin{gathered} \xymatrix{0 \ar@{..}[r] & \bullet\ar@{-}[r] & \bullet \ar@{-}[r] & \bullet \ar@{-}[r] & \bullet\ar@{-}[r]& \bullet\ar@{-}[r] &\bullet\\ & & \bullet\ar@{-}[u]}\end{gathered}\\ &E_6\ \text{with }\mathbf{27} &&\begin{gathered} \xymatrix{\bullet \ar@{-}[r] & \bullet\ar@{-}[r] & \bullet \ar@{-}[r] & \bullet \ar@{-}[r] & \bullet\ar@{..}[r]& 0\\ & & \bullet\ar@{-}[u]}\end{gathered}\\ &E_7\ \text{with }\mathbf{56} &&\begin{gathered} \xymatrix{\bullet \ar@{-}[r] & \bullet\ar@{-}[r] & \bullet \ar@{-}[r] & \bullet \ar@{-}[r] &\bullet \ar@{-}[r] &\bullet\ar@{..}[r]& 0\\ & & \bullet\ar@{-}[u]}\end{gathered}\\\end{aligned}$$ See figure \[kk34\] for the full list of asymptotically free theories of this class. Note that in case (2) the light category automatically contains hypermultiplets in the right representation of $G$ since, if $a$ is an extension node in $\widehat{G[a,1]}$ we have $$\mathrm{Ad}(G[a,1])=\mathrm{Ad}(G)\oplus [0,\cdots, 0,1,0,\cdots, 0]\oplus \overline{[0,\cdots, 0,1,0,\cdots, 0]}\oplus \text{singlets}.$$ We close this subsection we two remarks: - besides those in figure \[kk34\] there is another asymptotically free pair (group, representation), namely $SU(N)$ with the two–index *symmetric* representation (which is not *fundamental*) whose augmented graph is identified with the non–simply–laced Dynkin graph of type $B_N$ [@nons]; - the asymptotically free representations of the form $R_1\oplus R_2$ correspond to the twice–augmented graphs which are Dynkin. See the table gauge group matter repre. aug. graph gauge group matter repr. aug. graph ------------- ----------------------------------------------------- ------------ ------------- ------------------------------------------------------ ------------ $SU(n)$ $\boldsymbol{n}\oplus\boldsymbol{n}$ $A_{n+1}$ $SU(n)$ $\boldsymbol{\frac{n(n-1)}{2}}\oplus\boldsymbol{n}$ $D_{n+1}$ $SU(n)$ $\boldsymbol{\frac{n(n+1)}{2}}\oplus\boldsymbol{n}$ $B_{n+1}$ $SO(8)$ $\boldsymbol{8}_i\oplus \boldsymbol{8}_j$, $i\neq j$ $E_6$ $SU(6)$ $\boldsymbol{20}\oplus\boldsymbol{6}$ $E_{7}$ $SU(7)$ $\boldsymbol{35}\oplus \boldsymbol{7}$ $E_8$ Example 4: $G$ non–simply–laced ------------------------------- The Dynkin graph of a non–simply laced Lie group $G$ arises by folding a parent simply–laced Dynkin graph $G_\text{parent}$ along an automorphism group $U$. Specifically, the $G_\text{parent}\rightarrow G$ foldings are $$\begin{aligned} &D_{n+1}\longrightarrow B_n &&A_{2n-1}\longrightarrow C_n\\ &D_{4}\longrightarrow G_2 &&E_{6}\longrightarrow F_4.\end{aligned}$$ $U=\Z_2$ in all cases except for $D_4\rightarrow G_2$ where it is $\Z_3$. To each node of the folded Dynkin diagrams there is attached an integer $d_a$, namely the number of nodes of the parent graph which were folded into it. This number corresponds to one–half the length–square of the corresponding simple co–root $\alpha_a^\vee$ $$\label{defdi} d_a= \frac{1}{2}(\alpha_a^\vee,\alpha_a^\vee)\equiv \frac{2}{(\alpha_a,\alpha_a)}\qquad a=1,2,\dots, r.$$ In general, the light category of a (quiver) $\cn=2$ gauge theory with group $G$ has the structure $$\mathscr{L}=\bigvee_{\lambda\in\mathbb{P}^1/U}\mathscr{L}_\lambda$$ with $U$ acting on the category $\mathscr{L}_\lambda$ trough monodromy functors $\mathscr{M}_u$ [@nons] $$\label{yyyq} \mathscr{L}_{u\cdot \lambda}= \mathscr{M}_u(\mathscr{L}_\lambda)\qquad u\in U.$$ Since the cylinder $\C^*\subset \mathbb{P}^1$ is identified with the Gaiotto plumbing cylinder associated to the gauge group $G$, this monodromical construction is equivalent to the geometric realization of the non–simply–laced gauge groups in the Gaiotto framework [@tackns] or in F–theory [@BIKMSV]. In the simply–laced case the light category was described in terms of the preprojective algebra of $G$; likewise, to each gauge group $G=BCFG$ we may associate a generalized ‘preprojective’ algebra of the form $\mathscr{J}\!(Q^\prime,\cw^\prime)$. $Q^\prime$ is the same reduced quiver as in the $A_r$ case (see figure \[0023c\] for the $r=5$ example) while the reduced superpotential is $$\label{msuper} \mathscr{W}=\sum_{a\xrightarrow{\alpha}b} \Big( \alpha\, A_{s(\alpha)}^{n(\alpha)}\,\alpha^*-\alpha^*\, A_{t(\alpha)}^{m(\alpha)}\,\alpha\Big),$$ where the sum is over the edges $\xymatrix{a \ar@{-}[r]^\alpha&b}$ of $A_r$ and $$\big(n(\alpha),\: m(\alpha)\big)= \left(\frac{d_a}{(d_a,d_b)},\: \frac{d_b}{(d_a,d_b)}\right).$$ One checks [@nons] that $\mathsf{mod}\mathscr{J}\!(Q^\prime,\cw^\prime)$ has the monodromic property and the dimension vectors of its bricks are the positive roots of $G$, so that the light category corresponds to vector multiplets forming a single copy of the adjoint of $G$, as required for pure SYM. From the light subcategory $\mathsf{mod}\mathscr{J}\!(Q^\prime,\cw^\prime)$ one reconstructs the full *non–perturbative* Abelian category $\mathsf{mod}\mathscr{J}\!(Q,\cw)$, which describes the model in all physical regimes, by using the Dirac integrality conditions described in §.\[Dirrrac\]. See ref.[@nons] for details. Half–hypers =========== Coupling *full* hypermutliplets to SYM {#fullH} -------------------------------------- The construction of the pairs $(Q_{N_f},\cw_{N_f})$ for $G=ADE$ SQCD coupled to $N_f$ fundamental *full* hypermultiplets of refs.[@ACCERV2; @cattoy] was relatively easy: each hypermultiplet has a gauge invariant mass $m_i$, and taking the *decoupling limit* $m_i\rightarrow \infty$ we make $N_f\rightarrow N_f-1$. At the level of modules categories this decoupling processes insets $$\begin{gathered} \label{reddc} \mathsf{mod}\mathscr{J}(Q_{N_f-1},\mathcal{W}_{N_f-1}) \xrightarrow{\subset} \mathsf{mod}\mathscr{J}(Q_{N_f},\mathcal{W}_{N_f})\end{gathered}$$ as an extension–closed, exact, full, *controlled* Abelian subcategory [@cattoy]. In general, a *control function* is a linear map $\eta\colon \Gamma\rightarrow \Z$, and the controlled subcategory is the full subcategory over the objects $X$ such that $\eta(X)=0$ while for all their subobjects $\eta(Y)\leq 0$. The light subcategory is an example of controlled one with control function the magnetic charge. All decoupling limits of QFT correspond to controlled subcategories in the RT language. For the decoupling limit $m_i\rightarrow \infty$ the control function $f_i\colon \Gamma\rightarrow \mathbb{Z}$ corresponds to the flavor charge dual to $m_i$. Choosing $f_i$ so that $f_i(\Gamma_+)\geq 0$, we realize $Q_{N_f-1}$ as a full subquiver of $Q_{N_f}$ missing one node, the functor $\mathsf{mod}\mathscr{J}(Q_{N_f-1},\mathcal{W}_{N_f-1})\rightarrow \mathsf{mod}\mathscr{J}(Q_{N_f},\mathcal{W}_{N_f})$ being the restriction. This gives a recursion relation in $N_f$ of the form $$\begin{gathered} \begin{gathered}\includegraphics[width=0.27\textwidth, height=0.11\textwidth]{nfm2.jpg}\end{gathered}\end{gathered}$$ By repeated use of this relation, we eventually get to pure $G$ SYM whose quiver is known, see §.\[99ert\]. The decoupling process may be easily inverted to get a recursive map $Q_{N_f-1}\rightarrow Q_{N_f}$. Indeed, to define such a map we have only to determine the black arrows connecting $Q_{N_f-1}$ to the extra node in the <span style="font-variant:small-caps;">rhs</span> of which correspond to an extra massive quark. Given the electric weights (*i.e.* the $G$–representation) of the quark, the black arrows are uniquely determined by the Dirac pairing. This strategy does work for SYM coupled to *half*–hypermultiplets: they carry no flavor symmetry, have no mass parameter. They are tricky theories, always on the verge of inconsistency: most of them are indeed quantum inconsistent, but there are a few consistent models which owe their existence to peculiar *‘miracles’*. The typical example being $G=E_7$ SYM coupled to half a $\mathbf{56}$. Coupling *half* hypermutliplets ------------------------------- We use yet another decoupling limit: *extreme Higgs*. Given a $\mathcal{N}=2$ gauge theory with group $G_r$, of rank $r$, we take a v.e.v. of the adjoint field $\langle \Phi\rangle\in \mathfrak{h}(G)$ such that[^5] $$\label{pp33a} \alpha_b(\langle \Phi\rangle)=\begin{cases} t\,e^{i\phi},\ \ t\rightarrow+\infty, & b=a\\ O(1) & \text{otherwise} \end{cases}$$ States having electric weight $\rho$ such that $\rho(\langle \Phi\rangle)=O(t)$ decouple, and we remain with a gauge theory with a gauge group $G_{r-1}$ whose Dynkin diagram is obtained by deleting the $a$–th node from that of $G_r$ (coupled to specific matter). *E.g.* starting from $G_7=E_7$ coupled to $\tfrac{1}{2}\,\mathbf{56}$ and choosing $a=1$ we get $G_6=\mathrm{Spin}(12)$ coupled to $\tfrac{1}{2}\,\mathbf{32}$ corresponding to deteling the black node in the Dynkin graph $$\begin{gathered} \xymatrix{\bullet \ar@{-}[r] & \circ\ar@{-}[r] & \circ \ar@{-}[r] & \circ \ar@{-}[r] &\circ \ar@{-}[r] &\circ\\ & & \circ\ar@{-}[u]}\end{gathered}$$ Again, the decoupling limit should correspond to a controlled Abelian *sub*category of the representations of $(Q_{G_r},\cw_{G_r})$. One can choose $(Q_{G_r},\mathcal{W}_{G_r})$ in its mutation–class and the phase $\phi$ in so that the control function $\lambda(\cdot)$ is *non–negative* on the positive–cone $\Gamma_+$. Then $Q_{G_{r-1}}$ is a full subquiver of $Q_{G_r}$ and $\mathcal{W}_{G_{r-1}}$ is just the restriction of $\mathcal{W}_{G_r}$. It is easy to see that the complementary subquiver is a Kronecker one $\upuparrows$ [@half]. Putting everything together, we get a recursion of the quiver with respect to the rank $r$ of $G_r$ of the form -8pt ![image](rm2.jpg){width="36.00000%" height="18.00000%"} -16pt If we know the simpler quiver $Q_{G_{r-1}}$, to get $Q_{G_r}$ we need just the fix the red arrows connecting the Kronecker to $Q_{G_{r-1}}$ in the above figure. Just as in §.\[fullH\], the [red]{} arrows are uniquely fixed by Dirac charge quantization. Indeed, by the recursion assumption, we know the representations $X_{\alpha_a}$ associated to all simple–root $W$–bosons of $G_r$; under the maximal torus $U(1)^r\subset G$ the simple–root $W$–bosons have charges $q_a(X_{\alpha_b})=C_{ab}$ (Cartan matrix), while the dual magnetic charges are given by eqn. which explicitly depends on the red arrows. It turns out [@half] that $m_a(X)\in {}^L\Gamma_\mathrm{root}$ for all $X$ for a *unique* choice of the arrows which are then fixed. Then $Q_{G_r}$ is uniquely determined if we know $Q_{G_{r-1}}$. $\mathcal{W}_{G_r}$ is also essentially determined, up to some higher–order ambiguity [@half]. Taking a suitable chain of such Higgs decouplings/symmetry breakings $$\label{44rr}G_r\rightarrow G_{r-1}\rightarrow G_{r-2}\rightarrow \cdots\cdots \rightarrow G_k,$$ we eventually end up with a *complete* $\mathcal{N}=2$ with gauge group $G_k=SU(2)^k$. The complete $\mathcal{N}=2$ quivers are known by classification [@CV11]. Inverting the Higgs procedure, we may construct the pair $(Q_{G_r},\mathcal{W}_{G_r})$ for the theory of interest by [‘pulling back’]{} trough the chain the pair $(Q_\text{max comp},\mathcal{W}_\text{max comp})$ of their [maximal complete]{} subsector. For the models of interest the *‘pull back’* chain is presented in figure \[ppq\]. The bottom model $SU(2)^3$ with $\tfrac{1}{2}(\mathbf{2},\mathbf{2},\mathbf{2})$ is complete [@CV11; @cattoy]. $$\xymatrix{ \text{\fbox{\;$E_7\phantom{\Big|}$ $\tfrac{1}{2}\,\mathbf{56}$\;}} & \text{\begin{small}$SU(2)\times SO(12)$\: $\tfrac{1}{2}(\mathbf{2},\mathbf{12})$\end{small}}\\ \text{\fbox{\;$SO(12)\phantom{\Big|}$ $\tfrac{1}{2}\,\mathbf{32}$\;}}\ar@{=>}[u] & \text{\begin{small}$SU(2)\times SO(10)$\: $\tfrac{1}{2}(\mathbf{2},\mathbf{10})$\end{small}}\ar@{..>}[ul]\ar@{..>}[u]\\ \text{\fbox{\;$SU(6)\phantom{\Big|}$ $\tfrac{1}{2}\,\mathbf{20}$\;}}\ar@{=>}[u] & \text{\fbox{\;$SU(2)\times SO(8)\phantom{\Big|}$ $\tfrac{1}{2}\,(\mathbf{2},\mathbf{8})$\;}}\ar@{=>}[ul]\ar@{..>}[u]\\ \text{\fbox{\;$SU(2)\times SU(4)\phantom{\Big|}$ $\tfrac{1}{2}\,(\mathbf{2},\mathbf{6})$\;}}\ar@{=>}[u]\ar@{=>}[ur]\\ \text{\fbox{\;$SU(2)^3\phantom{\Big|}$ $\tfrac{1}{2}\,(\mathbf{2},\mathbf{2},\mathbf{2})$\;}}\ar@{=>}[u] }$$ The pair $(Q_{E_7},\cw_{E_7})$ for the model $G=E_7$ coupled to $\tfrac{1}{2}\,\textbf{56}$ is given in figure \[rrrr4\]; the other models in figure \[ppq\] correspond to the restriction to suitable subquivers of $(Q_{E_7},\cw_{E_7})$ [@half]. The light category deduced from these pairs contains light vectors forming one copy of the adjoint of $G$ plus light hypermultiplets in the $G$–representation $\tfrac{1}{2}\,\mathbf{R}$, with $\mathbf{R}$ irreducible *quaternionic* [@half]. Indeed, the light category has again the form $\mathsf{mod}\mathscr{J}(Q^\prime,\cw^\prime)$ for a reduced pair $(Q^\prime,\cw^\prime)$. See figure \[kkaq\] for the the reduced pair for $G=E_7$ coupled to $\tfrac{1}{2}\mathbf{56}$; the other models are obtained by restriction of this one. Note that $Q^\prime_{E_7}$ (and hence all reduced quivers $Q^\prime_{G_r}$ in the Higgs chain) contains as a full subquiver the quiver of the Gaiotto $A_1$ theory on $S^2$ with 3 punctures (the $T_2$ theory) described in [@ACCERV2]. Hence for all these models the ‘$T_2$–duality’ of §.\[qquiq\] is operative; this duality is crucial — together with special properties of the relevant Dynkin graphs — to check the above claims on the BPS spectrum at weak coupling. Details may be found in [@half]. -24pt $$\begin{gathered} \xymatrix{ &&&*++[o][F-]{1} \ar[ld]_{\phi} \ar[rr]^{H_1} && *++[o][F-]{2}\ar[dl]_{H_2}\ar[dr]^{\widetilde{\psi}}&\\ *++[o][F-]{\tau_{-2}}\ar@<0.5ex>@{->}[ddd]^{A_{-2}}\ar@<-0.5ex>@{->}[ddd]_{B_{-2}}& *++[o][F-]{\omega_{-1}}\ar[l]_{\psi_{-2}} \ar[r]^{\psi_{-1}^{\prime}}&*++[o][F-]{\tau_0} \ar@<0.5ex>@{->}[ddd]^{A_0}\ar@<-0.5ex>@{->}[ddd]_{B_0} \ar[ddr]^{\phi^{\prime}}& & *++[o][F-]{3}\ar[lu]_{H_3} \ar[ll]_{\psi_0 \qquad} \ar[rr]^{\quad\widetilde{\phi}}&&*++[o][F-]{\widetilde{\tau}}\ar@<0.5ex>@{->}[ddd]^{\widetilde{A}}\ar@<-0.5ex>@{->}[ddd]_{\widetilde{B}}\ar[dddll]_{\widetilde{\phi}^{\prime}}\\ &&&&& & &&\\ &&&*++[o][F-]{5}\ar[uuu]_{V_1}\ar[dr]^{h_3} & & *++[o][F-]{6}\ar[ll]_{h_1\qquad}\ar[uuu]_{V_2}&\\ *++[o][F-]{\omega_{-2}}\ar[r]^{\psi_{-2}^{\prime}}&*++[o][F-]{\tau_{-1}} \ar@<0.5ex>@{->}[uuu]^{A_{-1}}\ar@<-0.5ex>@{->}[uuu]_{B_{-1}}&*++[o][F-]{\omega_0}\ar[l]_{\psi_{-1}} \ar[rr]^{\psi_0^{\prime}}& & *++[o][F-]{4}\ar[ur]^{h_2}\ar[uuu]_{V_3} &&*++[o][F-]{\widetilde{\omega}}\ar[ul]_{\widetilde{\psi}^{\prime}}\\ } \end{gathered}$$ -8pt $$\begin{gathered} \mathcal{W}_{E_7}= H_1 H_3 H_2 + h_3 h_1 h_2 + A \psi V_3 \psi^{\prime} + B \psi H_2 V_2 h_2 \psi^{\prime} + \phi V_1 \phi^{\prime} + \psi V_3 h_3 \phi^{\prime} +\\ \phi H_3 V_3 \psi^{\prime} B+ \widetilde{A} \widetilde{\psi} V_2 \widetilde{\psi}^{\prime} + \widetilde{B} \widetilde{\psi} H_1 V_1 h_1 \widetilde{\psi}^{\prime} + \widetilde{\phi} V_3 \widetilde{\phi}^{\prime} + \widetilde{\psi} V_2 h_2 \widetilde{\phi}^{\prime} + \widetilde{\phi} H_2 V_2 \widetilde{\psi}^{\prime} \widetilde{B}+\\ +A_0\psi^\prime_{-1}B_{-1}\psi_{-1} -B_0\psi^\prime_{-1}A_{-1}\psi_{-1}+A_{-1}\psi^\prime_{-2}B_{-2}\psi_{-2}-B_{-1}\psi^\prime_{-2}A_{-2}\psi_{-2}\end{gathered}$$ $$\begin{gathered} \xymatrix@R=2.0pc@C=3.0pc{ &&&&&&1 \ar@/_0.5pc/[dll]_{h_3} \ar@/_0.5pc/[dd]_{H_1}\\ &\ar@(ul,dl)[]_{A_{-2}} -2 \ar@/_0.7pc/[r]_{\psi_{-2}^{\prime}} &\ar@/_0.7pc/[l]_{\psi_{-2}} \ar@(dr,dl)[]^{A_{-1}} -1\ar@/_0.7pc/[r]_{\psi_{-1}^{\prime}}&\ar@/_0.7pc/[l]_{\psi_{-1}} \ar@(dr,dl)[]^{A_0} 0 \ar@/_0.7pc/[r]_{\psi_0^{\prime}} & 3 \ar@/_0.7pc/[l]_{\psi_0} \ar@/_0.5pc/[urr]_{H_3} \ar@/_0.5pc/[drr]_{h_2}&&\\ &&&&&&2 \ar@/_0.5pc/[ull]_{H_2} \ar@/_0.5pc/[uu]_{h_1}\ar@/_0.7pc/_{\widetilde{\psi}}[d]\\ &&&&&&\ar@(ld,rd)[]_{\widetilde{A}}\widetilde 0 \ar@/_0.7pc/_{\widetilde{\psi}^\prime}[u] } \end{gathered}$$ -24pt $$\begin{gathered} \mathcal{W}^\prime_{E_7}= H_1H_3H_2+h_3h_1h_2+\psi_0(H_2h_2+h_3H_3)\psi_0^\prime+ \widetilde{\psi}(H_1h_1+h_2H_2)\widetilde{\psi}^\prime+\\ +A_0\psi_0\psi_0^\prime+A_0\psi^\prime_{-1}\psi_{-1}-A_{-1}\psi_{-1}\psi^\prime_{-1}+A_{-1}\psi^\prime_{-2}\psi_{-2}-A_{-2}\psi_{-2}\psi_{-2}^\prime +\widetilde{A}\widetilde{\psi}\widetilde{\psi}^\prime.\end{gathered}$$ [50]{} D. Gaiotto, G. W. Moore, and A. Neitzke, “[Wall-crossing, Hitchin Systems, and the WKB Approximation]{},” [[arXiv:0907.3987 \[hep-th\]]{}](http://arxiv.org/abs/0907.3987). D. Gaiotto, G. W. Moore, and A. Neitzke, “[Framed BPS States]{},” [[arXiv:1006.0146 \[hep-th\]]{}](http://arxiv.org/abs/1006.0146). D. Gaiotto, G. W. Moore, and A. Neitzke, “[Wall–Crossing in Coupled 2d-4d Systems]{},” [[arXiv:1103.2598 \[hep-th\]]{}](http://arxiv.org/abs/1103.2598). D. Gaiotto, G. W. Moore, and A. Neitzke, “[Spectral networks]{},” [[arXiv:1204.4824 \[hep-th\]]{}](http://arxiv.org/abs/1204.4824). D. Gaiotto, G. W. Moore, and A. Neitzke, “[Spectral networks and Snakes]{},” [[arXiv:1209.0866 \[hep-th\]]{}](http://arxiv.org/abs/1209.0866). S. Cecotti, A. Neitzke, and C. Vafa, “[R-Twisting and 4d/2d Correspondences]{},” [[arXiv:1006.3435 \[hep-th\]]{}](http://arxiv.org/abs/1006.3435). S. Cecotti and C. Vafa, “[Classification of complete N=2 supersymmetric theories in 4 dimensions]{},” [[arXiv:1103.5832 \[hep-th\]]{}](http://arxiv.org/abs/1103.5832). S. Cecotti and M. Del Zotto, *On Arnold’s 14 ‘exceptional’ $\cn=2$ superconformal gauge theories,* JHEP **1110** (2011) 099, [[arXiv:1107.5747 \[hep-th\]]{}](http://arxiv.org/abs/1107.5747). M. [Alim]{}, S. [Cecotti]{}, C. [Cordova]{}, S. [Espahbodi]{}, A. [Rastogi]{}, and C. [Vafa]{}, “[BPS Quivers and Spectra of Complete N=2 Quantum Field Theories]{},” [[arXiv:1109.4941 \[hep-th\]]{}](http://arxiv.org/abs/1109.4941). M. [Alim]{}, S. [Cecotti]{}, C. [Cordova]{}, S. [Espahbodi]{}, A. [Rastogi]{}, and C. [Vafa]{}, “[$N=2$ Quantum Field Theories and their BPS QUivers]{},” [[arXiv:1112.3984 \[hep-th\]]{}](http://arxiv.org/abs/1112.3984). S. Cecotti, “Categorical tinkertoys for $N=2$ gauge theories”, [[arXiv:1203.6743 \[hep-th\]]{}](http://arxiv.org/abs/1203.6743). S. Cecotti and M. Del Zotto, “Half–Hypers and Quivers,” *JHEP* **09** (2012) 135, [[arXiv:1207.2275 \[hep-th\]]{}](http://arxiv.org/abs//1207.2275). S. Cecotti and M. Del Zotto, “4d N=2 Gauge Theories and Quivers: the Non-Simply Laced Case,” *JHEP* **10** (2012) 190, [[arXiv:1207.7205 \[hep-th\]]{}](http://arxiv.org/abs//1207.7205). S. Cecotti and M. Del Zotto, “Infinitely many N=2 SCFT with ADE flavor symmetry,” [[arXiv:1210.2886 \[hep-th\]]{}](http://arxiv.org/abs//1210.2886). H. Derksen, J. Wyman, and A. Zelevinsky, “Quivers with potentials and their representations I: Mutations,” Selecta Mathematica **14** (2008) 59–119. D. Gaiotto, “[N=2 dualities]{},” [[arXiv:0904.2715 \[hep-th\]]{}](http://arxiv.org/abs/0904.2715). I.M. Gelfand and V.A. Ponomarev, “Model algebras and representations of graphs,” Funktsional. Anal. i Prilozhen. **13** (1979) 1–12. W. Crawley–Boevey, “On the exceptional fibres of Kleinian singularities,” Amer. J. Math. **122** (2000) 1027–1037. G. Lusztig, “Quivers, perverse sheaves, and quantized enveloping algebras,” J. Amer. Soc. **4** (1991) 313–324; Y. Tachikawa, “N=2 $S$–duality via outer–automorphism twists”, [[arXiv:1009.0339 \[hep-th\]]{}](http://arxiv.org/abs/1009.0339). M. Bershadsky, K.A. Intriligator, S. Kachru, D.R. Morrison, V. Sadov, and C. Vafa, “Geometric singularities and enhanced gauge symmetries,” Nucl. Phys. **B481** (1996) 215 [[arXiv:hep-th/9605200 \[hep-th\]]{}](http://arxiv.org/abs/hep-th/9605200). [^1]:  A module $X\in\mathsf{mod}\mathscr{J}\!(Q,\cw)$ of dimension $\sum_v N_v e_v$ is specified by giving, for each arrow $u \xrightarrow{\ \alpha\ } v$, an $N_v\times N_u$ matrix $X_\alpha$ such that the matrices $\{X_\alpha\}$ satisfy the relations $\partial_{X_{\beta}}\cw(X_\alpha)=0$ for all arrows $\beta$ in $Q$. Two such representations are *isomorphic* if they are related by a $\prod_v GL(N_v,\C)$ transformation. [^2]:  $\mathfrak{h}$ stands for the Cartan subalgebra of the complexified Lie algebra of the gauge group $G$. [^3]:  Given an unoriented graph $L$, its *double quiver* $\overline{L}$ is obtained by replacing each edge $a$ of $L$ by a pair of opposite arrows $\xymatrix{\bullet \ar@<0.4ex>[rr]^{\psi_a}&&\bullet \ar@<0.4ex>[ll]^{\widetilde{\psi}_a}}$. To write eqn. we have picked an arbitrary orientation of $G$, the algebra $\mathscr{J}(Q^\prime,\cw^\prime)$ being independent of choices, up to isomorphism. [^4]:  The reduced quiver is $2$–acyclic: this is related to the fact that it describes a subset of states which are *all* mutually local, hence have trivial Dirac pairing. At the level of the quiver this means that the *net* number of arrows from node $i$ to node $j$ must vanish (while we need to have arrows since the perturbative sector is not a free theory). [^5]:  $\mathfrak{h}(G)$ stands for the Cartan subalgebra of $G$; $\alpha_a\in \mathfrak{h}(G)^\vee$ is the $a$–th simple root.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Let $r \ge 4$ be an integer and consider the following game on the complete graph $K_n$ for $n \in r \mathbb{Z}$: Two players, Maker and Breaker, alternately claim previously unclaimed edges of $K_n$ such that in each turn Maker claims one and Breaker claims $b \in \mathbb{N}$ edges. Maker wins if her graph contains a $K_r$-factor, that is a collection of $n/r$ vertex-disjoint copies of $K_r$, and Breaker wins otherwise. In other words, we consider a $b$-biased $K_r$-factor Maker–Breaker game. We show that the threshold bias for this game is of order $n^{2/(r+2)}$. This makes a step towards determining the threshold bias for making bounded-degree spanning graphs and extends a result of Allen et al. who resolved the case $r \in \{3,4\}$ up to a logarithmic factor.' author: - 'Anita Liebenau [^1]' - 'Rajko Nenadov [^2]' bibliography: - 'references.bib' title: 'The threshold bias of the clique-factor game' --- Introduction ============ In this paper we consider biased positional games played on the edge set of the complete graph $K_n$ in which the winning sets are spanning subgraphs. Biased positional games were introduced by Chvátal and Erdős [@ce1978] in 1978 and form a central part of positional games, see, for example, the monograph by Beck [@BeckBook], or [@hkss2014; @k2014] for a more recent treatment. Let $X$ be a finite set and let ${{{\mathcal F}}}\subseteq 2^X$ be a family of subsets. The set $X$ is called the [*board*]{} and ${{{\mathcal F}}}$ is referred to as the family of [*winning sets*]{}. In the $b$-biased *Maker–Breaker* game $(X,{{{\mathcal F}}})$, two players called Maker and Breaker play in rounds. In every round Maker claims one previously unclaimed element of $X$ and Breaker responds by claiming $b$ previously unclaimed elements of $X$. Maker wins if she claims all elements of some $F\in {{{\mathcal F}}}$, otherwise Breaker wins the game. By definition a draw is impossible and thus exactly one player has a winning strategy since Maker–Breaker games are perfect information games. A certain class of games that received particular attention are Maker–Breaker games played on the edge set of the complete graph on $n$ vertices, denoted by $K_n$, in which case $X$ is the set of all unordered 2-element subsets of $K_n$, denoted by $\binom{[n]}{2}$. In the [*connectivity game*]{}, the [*perfect matching game*]{}, the [*Hamiltonicity game*]{}, and the [*triangle game*]{}, for example, the winning sets are the edge sets of all spanning trees, all perfect matchings, all Hamilton cycles, and all copies of $K_3$, respectively. When $n$ is large enough these games are heavily in favour of Maker in the [*unbiased*]{} version when $b=1$. Chvátal and Erdős [@ce1978] therefore examined the biased variant for these games. Define the threshold bias $b^*$ of a game $(X,{{\mathcal F}})$ to be the largest integer $b$ such that Maker wins the $b$-biased Maker–Breaker game $(X,{{\mathcal F}})$. Note that Maker–Breaker games are bias-monotone, that is Maker wins for every $b\le b^*$ and Breaker wins for every $b>b^*$. Chvátal and Erdős found that the threshold bias $b^*$ is of the order $n\ln n$ for the connectivity, the perfect matching, and the Hamiltonicity game; and of order $\sqrt{n}$ for the triangle game. The order of the threshold bias for an $H$-game, the game in which winning sets correspond to copies of $H$ in $K_n$, was later determined by Bednarska and [Ł]{}uczak [@bl2000] for any fixed graph $H$. Except for the connectivity game, all the aforementioned games can be cast in the following common form. Given a graph $H=H_n$, what is the threshold bias $b^*$ of the Maker–Breaker game played on $K_n$ in which all winning sets are copies of $H_n$? In the case of Hamiltonicity we simply have $H_n = C_n$, a cycle of length $n$, and in the perfect matching game $H_n$ is a collection of $n/2$ vertex disjoint edges. There are choices of $H_n$ for which Maker cannot win even if $b = 1$. A trivial such example is $H_n = K_n$, however even for $H_n$ being a complete graph with only $2 \log n$ vertices Maker cannot win the $H_n$-game [@BeckBook Theorem 6.4]. It turns out that this can be avoided if we restrict our attention to graphs with maximum degree some constant $\Delta$, and let $n$ be sufficiently large. Furthermore, rather than asking for the threshold bias for a specific $H_n$, we seek a *universal* upper bound: given $\Delta$ and $n$, what is the largest $b_\Delta = b_\Delta(n)$ such that, on the one hand, for every graph $H_n$ with at most $n$ vertices and maximum degree at most $\Delta$ Maker can win a $b$-biased $H_n$-game with $b \le b_\Delta$, and on the other hand there exists at least one such $H_n$ for which Breaker can win with bias $b = b_\Delta$ + 1? Recently, Allen, Böttcher, Kohayakawa, Naves, and Person [@abknp2017] showed that $b_\Delta(n)$ is of order at least $\Omega((n/\log n)^{1/\Delta})$[^3]. The triangle-preventing strategy for Breaker due to Chvátal and Erdős [@ce1978] shows that this is tight up to a factor of $\sqrt{\log n}$ when $\Delta = 2$. Furthermore, when $\Delta = 3$ the authors of [@abknp2017] show that Breaker can win a *$K_4$-factor* game for some $b = \Omega(n^{1/3})$, which shows (almost) optimality in this case as well. Here the $K_4$-factor and, in general, a $K_r$-factor, corresponds to a graph $H_n$ which consists of $\lfloor n/r \rfloor$ vertex-disjoint copies of $K_r$. However, the authors of [@abknp2017] have expressed a belief that their lower bound of $\Omega((n/\log n)^{1/\Delta})$, in general, is not optimal. We provide evidence for this feeling by determining the order of the threshold bias for the $K_{\Delta + 1}$-factor game for all $\Delta \ge 3$. For $\Delta = 3$, the threshold bias matches the upper bound in [@abknp2017], while for $\Delta \ge 4$ the exponent of $n$ of the threshold bias is strictly larger than $1/\Delta$. \[thm:main\] For any integer $r \ge 4$ there exist $c, C > 0$ such that the following holds for every $n \in r \mathbb{Z}$. 1. \[main:M\] If $b < c n^{2/(r+2)}$ then Maker has a winning strategy in the $b$-biased $K_r$-factor game played on the edge set of $K_n$. 2. \[main:B\] If $b > C n^{2/(r+2)}$ then Breaker has a winning strategy in the $b$-biased $K_r$-factor game played on the edge set of $K_n$. By taking $c$ and $C$ to be sufficiently small and large, respectively, we have that the theorem vacuously holds for all $n < n_0$ for any chosen $n_0$. Therefore, we assume throughout the paper that $n$ is as large as needed for the calculations to be correct. For $b \ge C n^{2/(r+2)}$ we show that Breaker has a strategy to ‘isolate’ one particular vertex from being in a copy of $K_r$, which clearly prevents Maker’s graphs from containing a $K_r$-factor. Somewhat surprisingly, though not uncommon in extremal and probabilistic combinatorics, this turns out to be Breaker’s best strategy: as soon as he cannot achieve this Maker is able to build a $K_r$-factor. Theorem \[thm:main\] suggests the following. For all $\Delta \ge 3$, $b_\Delta = \Theta(n^{2/(\Delta + 3)}).$ In other words, we believe that it is not significantly harder for Maker to build any other graph of maximum degree $\Delta$ than a $K_{\Delta+1}$-factor. We take justification for this assumption from two similar settings in extremal graph theory and in random graph theory. The celebrated theorem of Hajnal and Szemerédi [@hs1970] states that every graph $G$ of minimum degree at least $(1-1/(\Delta+1))n$ contains a $K_{\Delta + 1}$-factor, and that condition is tight. Bollobás and Eldridge [@be1978], and independently Catlin [@c1976], conjectured that the condition $\delta(G)\ge (1-1/(\Delta + 1))n$ is in fact sufficient to contain every graph $H$ with $n$ vertices and maximum degree $\Delta$. A similar assumption is made on the threshold bias $p^*$ for the random graph $G(n,p)$ to contain a certain graph $H_n$. Johansson, Kahn and Vu [@jkv2008] showed that $p^*(n)=(n^{-1}\log^{1/\Delta}n)^{2/(\Delta+1)}$ is a threshold function for $G(n,p)$ to contain a $K_{\Delta + 1}$-factor. It is folklore belief that, for every graph $H_n$ on at most $n$ vertices and of maximum degree $\Delta$, the function $p^*(n)$ is in fact an upper bound on the threshold functions for $G(n,p)$ to contain $H_n$, see for example Conjecture 1.3 in [@fln2017]. Supporting evidence towards this conjecture is given by Ferber, Luh and Nguyen [@fln2017] who prove it when $H_n$ is almost-spanning, that is when $H_n$ occupies at most $(1-{\ensuremath{\varepsilon}})n$ vertices. [**Structure of the paper.**]{}\ In Section 2 we take a little detour and discuss the *probabilistic intuition*, also called the [*Erdős paradigm*]{}. While this paradigm in its basic form does not apply to the problem we consider here, a variation of it due to Allen et al. [@abknp2017] (Theorem \[thm:maker\_rg\]) turns out to give the correct answer. This result will also serve us to provide further intuition why the threshold bias in Theorem \[thm:main\] is of the order $n^{2/(r+2)}$, or, more precisely, why Breaker is not able to isolate a single vertex from being in a copy of $K_r$ for $b < cn^{2/(r+2)}$. In Section 3 we fix notation and state preliminary results. In Section 4, we provide Maker’s strategy and prove Theorem \[thm:main\] \[main:M\]. Section 5 is devoted to Breaker’s strategy, i.e. Theorem \[thm:main\] \[main:B\]. Probabilistic Intuition Revised =============================== Chvátal and Erdős [@ce1978] found a surprising connection between biased positional games and random graphs. Replace Maker and Breaker by RandomMaker and RandomBreaker, respectively, who choose their edges uniformly at random from all unclaimed edges. At the end of the game, the graph of RandomMaker has the same distribution as $G(n,m)$, a graph with $m$ edges chosen uniformly at random from all $\binom{n}{2}$ possible edges, where $m$ is roughly $\binom{n}{2}/(b+1)$ (we omit floor and ceiling signs unless crucial). It is well known [@JLRbook] that $G(n,m)$ is (a) connected, (b) has a perfect matching, or (c) has a Hamilton cycle with probability tending to 0 if $m\ll n \ln n$, and with probability tending to 1 if $m\gg n \ln n$. That is, the threshold biases of the random version of the connectivity, the perfect matching, and the Hamiltonicity game are of the order $n\ln n$. The results in [@ce1978] imply that the threshold bias $b^*$ in the game with clever players is of the same order of magnitude for the connectivity, the perfect matching, and the Hamiltonicity game. This phenomenon is often called the [*random graph intuition*]{}, or the [*Erdős paradigm*]{}. In fact, it turns out that the threshold biases for the random and the clever game are asymptotically equal in the connectivity game [@gs2009] and in the Hamiltonicity game [@k2011]. It is one of the central questions in positional games to classify games for which the random graph intuition applies. A game which does very much not obey the random graph intuition is the above-mentioned triangle game or, more generally, an $H$-game for a fixed graph $H$ which contains a cycle. It is well-known that the threshold for the appearance of a triangle in $G(n,m)$ is of the order $\Theta(n)$ (see, e.g., [@JLRbook]). Chvátal and Erdős [@ce1978], however, showed that Breaker can prevent a triangle in Maker’s graph when playing with a bias $b =\Theta(\sqrt{n})$. It follows from Beck’s winning criterion for Breaker [@b1982], a generalisation of the classical Erdős-Selfridge criterion to biased games, that Breaker can [*always*]{} play at least as good as RandomBreaker against RandomMaker. A result by Chvátal and Erdős [@ce1978] shows that in some cases Breaker can play in a smarter way than just claiming edges at random. Bednarska and Łuczak [@bl2000] verified that this is also the case for any $H$-game. However, the main message of their paper is not that the probabilistic intuition completely fails in these games, but rather that it has to be slightly adjusted. As mentioned before, if both players play at random then Maker’s graph is distributed as a random graph $G(n, m)$ for $m = \binom{n}{2} / (b + 1)$. If Breaker does not play at random then by Maker still playing uniformly at random from the set of all *available* elements we lose control over the distribution of its graphs. To circumvent this, Bednarska and Łuczak [@bl2000] suggested the following strategy for Maker: choose a next element uniformly at random from the set of *all* elements (even those that have been previously claimed) and take it only if it forms a valid move, i.e. if it has not been previously claimed. Observe that Maker’s graph obtained following this strategy is not a random graph but rather a subgraph obtained from a random graph after deleting a few edges. Thus, even though we might not have a fine control over the actual Maker’s graph, knowing that it is a subgraph of a random graph turns out to give sufficient information to win an $H$-game. In particular, they show that when $b$ is not too large, the random graph $G(n,\binom{n}{2}/(b+1))$ is [*globally robust*]{} with respect to containing a copy of $H$, which in turn implies that Maker has a winning strategy. That is, even after removing any small proportion of the edges the plucked random graph still contains a copy of $H$. For a precise definition of robustness we refer the reader to [@sv2008] where a systematic study of this concept was initiated. The next step in explaining a connection between Maker–Breaker games and random graphs was done by Ferber, Krivelevich and Naves [@fkn2015]. While the strategy of playing purely at random works well in the case of $H$-games for graphs $H$ of fixed size, it fails when the winning sets are spanning subgraphs of $K_n$ as Breaker can isolate a vertex before Maker is likely to claim an edge incident to that vertex. To manifest the connection between Maker–Breaker games and random graphs for these spanning-graph games, Ferber, Krivelevich and Naves [@fkn2015] provided a [*local-resilience*]{} analogue to the theorem in [@bl2000] and showed that in a $b$-biased game played on $K_n$, Maker can claim a subgraph of $G(n,m)$ for $m=\Theta(n^2/b)$ such that each vertex is incident to $\Omega(n/b)$ Maker’s edges. Thus, if Maker tries to achieve a graph property ${{\mathcal P}}$ that cannot be destroyed by deleting a fixed proportion of edges at each vertex then the strategy in [@fkn2015] yields a winning strategy for Maker. In particular, lower bounds on the threshold bias for several games like the perfect matching game, the connectivity and the Hamiltonicity game could be re-established this way, though with a sub-optimal constant factor. However, as the reader could guess, the approach via local resilience does not work for all spanning-structure Maker–Breaker games on $K_n$. The property of containing a $K_3$-factor, for example, is not locally resilient as all triangles in $G(n,m)$ containing a fixed vertex $v$ can be destroyed by removing a vanishing proportion of edges incident to every vertex, see e.g. [@huang2012bandwidth]. For the same reason, the property of containing a $K_r$-factor, $r\ge 4$, is not locally resilient and the approach in [@fkn2015] is not applicable. Circumventing the short-coming of the resilience-type approaches, Allen, Böttcher, Kohayakawa, Naves, and Person [@abknp2017] finally show that Maker can also assume not only that its graph is a subgraph of a random graph with minimum degree of order $n/b$, but also that the neighbourhood of each vertex has sufficiently many edges. The following theorem makes this precise. For a real $p \in [0, 1]$ and an integer $n$, we write $\Gamma \sim G(n,p)$ if $\Gamma$ is formed by starting with an empty graph on $n$ vertices and adding each possible edge with probability $p$, independently of all other edges. Furthermore, $\Gamma\sim G(n,p)$ satisfies a certain property ${{\mathcal P}}$ [*asymptotically almost surely (a.a.s)*]{} if the probability that $\Gamma$ satisfies ${{\mathcal P}}$ tends to 1 as $n{\rightarrow}\infty$. \[thm:maker\_rg\] For every $n$, $\gamma = \gamma(n) \in (0, 1)$, $p \ge 10^8 \gamma^{-2} n^{-1/2}$, and $b \le 10^{-24} \gamma^6 p^{-1}$ the following holds. In the $b$-biased Maker-Breaker game played on $K_n$, for any fixed strategy of Breaker, if Maker draws a random graph $\Gamma \sim G(n,p)$ then a.a.s. $\Gamma$ is such that Maker can claim a spanning subgraph $G$ of $\Gamma$ with $\delta(G) \ge (1 - \gamma)np$ and $e_G(N_\Gamma(v)) \ge (1 - \gamma)p^3 n^2 / 2$ for every $v \in V(\Gamma)$. Using Theorem \[thm:maker\_rg\] in combination with a sparse blow-up lemma from [@abhkp2016], Allen et al. [@abknp2017] show that, for some $b =\Omega ((n/\log n)^{1/\Delta})$, these [*neighbourhood properties*]{} are enough for $G$ to contain all graphs of maximum degree $\Delta$ on at most $n$ vertices. For which $p$ can we guarantee that the neighbourhood properties given by Theorem \[thm:maker\_rg\] guarantee that every vertex of $G$ is contained in a copy of $K_r$? The neighbourhood $N_G(v)$ of a vertex $v$ in $G$ has size roughly $pn$, and the graph induced on $N_G(v)$ is a subgraph of $G(n,p)$ that still contains about $(1-{\ensuremath{\varepsilon}})n^2p^3/2$ edges, i.e. all but a small proportion of edges of $G(n,p)$ in $N_G(v)$ are also edges of $G$. That is, the subgraph of $G$ induced by $N_G(v)$ has roughly the same distribution as the random graph $G(np,p)$, and for the latter to robustly contain a copy of $K_{r-1}$ it is enough to have $p > C(np)^{-2/r}$ for some constant $C$, which translates to $p>Cn^{-2/(r+2)}$. It turns out that this is the main obstacle for Maker to create a $K_r$-factor. Preliminaries ============= We use standard graph-theoretic notation. All considered graphs are finite and simple. Given a graph $G$, we let $e(G)$ and $v(G)$ denote its number of edges and vertices, respectively. Given a set $X \subseteq V(G)$, let $e_G(X)$ denote the number of edges of $G$ with both endpoints in $X$. Similarly, for disjoint subsets $X, Y \subseteq V(G)$ we let $e_G(X, Y)$ denote the number of edges of $G$ with one endpoint in $X$ and the other in $Y$. Given a vertex $v \in V(G)$, we let $N_G(v)$ denote its neighbourhood, and for a set $X$ let $N_G(X) = \bigcup_{v \in X} N_G(v)$. When $G$ is clear from the context, we omit the subscript. For brevity we also omit floors and ceilings, keeping in mind that all the calculations leave enough margin to accumulate all the rounding errors. All asymptotic statements refer to $n$, the number of vertices, tending to $\infty$. Following standard asymptotic notation we write in particular, $f\ll g$ when $f/g {\rightarrow}0$ as $n{\rightarrow}\infty$, and $f\gg g$ if $g\ll f$. Properties of random graphs --------------------------- The following well-known estimates on the likely discrepancy of edges and the concentration of degrees in random graphs follow immediately from Chernoff’s inequality and the union bound. \[lemma:disc\] Let $p = p(n)$ be such that $n^{-1} \le p \le 0.99$. Then a.a.s. $\Gamma \sim{\ensuremath{G(n,p)}}$ satisfies the following properties: - For all disjoint subsets $X, Y \subseteq V(\Gamma)$ such that $|X| \le |Y|$ we have $$e(X, Y) = |X||Y|p \pm O\left( |Y| \sqrt{|X| p \log (n/|Y|)} \right);$$ - For every subset $X \subseteq V(\Gamma)$ we have $$e(X) = |X|^2p/2 \pm O\left( |X|\sqrt{|X|p \log (n/|X|)} \right);$$ - For every vertex $v \in V(\Gamma)$ we have $$|N(v)| = np \pm O(\sqrt{np \log n})$$ In order to state the second result we first need some preparation. Given a graph $G$ and ${\ensuremath{\varepsilon}}\in [0,1]$, we say that a pair of disjoint subsets $V_1, V_2 \subseteq V(G)$ forms an *$({\ensuremath{\varepsilon}})$-regular pair* if for $i=1,2$ and for every $V_i' \subseteq V_i$ of size $|V_i'| \ge {\ensuremath{\varepsilon}}|V_i|$ we have $$\left| e(V_1', V_2') - |V_1'||V_2'|p \right| \le {\ensuremath{\varepsilon}}|V_1'||V_2'| p,$$ where $p = e(V_1, V_2)/|V_1||V_2|$. Note that Lemma \[lemma:disc\] implies that a.a.s. every pair of subsets of ${\ensuremath{G(n,p)}}$ of size, say, at least $\log n / p$, forms an $({\ensuremath{\varepsilon}})$-regular pair for every fixed ${\ensuremath{\varepsilon}}>0$. Let $H$ be a graph with vertex set $\{1, \ldots, k\}$. We denote by ${{\mathcal G}}(H, n, m, {\ensuremath{\varepsilon}})$ the collection of all graphs $G$ obtained in the following way: (i) The vertex set of $G$ is a disjoint union $V_1 \cup \ldots \cup V_k$ of sets of size $n$; (ii) For each edge $ij \in E(H)$, we add to $G$ an $({\ensuremath{\varepsilon}})$-regular bipartite graph with $m$ edges between the pair $(V_i, V_j)$. Let ${{\mathcal G}}^*(H, n, m, {\ensuremath{\varepsilon}})$ denote the family of all graphs $G \in {{\mathcal G}}(H, n, m, {\ensuremath{\varepsilon}})$ which do not contain a copy of $H$. The following result, originally conjectured by Kohayakawa, Łuczak, and Rödl [@kohayakawa1997onk], was proven by Balogh, Morris, and Samotij [@balogh2015independent] and, independently, Saxton and Thomason [@saxton2015hypergraph]. \[thm:KLR\] Let $H$ be a fixed graph and $\beta > 0$. Then there exist $C, {\ensuremath{\varepsilon}}> 0$ and a positive integer $n_0$ such that $$\left| {{\mathcal G}}^*(H, n, m, {\ensuremath{\varepsilon}}) \right| \le \beta^m \binom{n^2}{m}^{e(H)}$$ for every $n \ge n_0$ and every $m \ge Cn^{2 - 1/m_2(H)}$, where $$m_2(H) = \max\left\{ \frac{e(H') - 1}{v(H') - 2} \colon H' \subset H, \; v(H) \ge 3 \right\}.$$ Theorem \[thm:KLR\] states that a random element from ${{\mathcal G}}(H, n, m, {\ensuremath{\varepsilon}})$ is highly unlikely to be $H$-free. Even more, it implies that the random graph $G(n,p)$ is unlikely to contain any graph from ${{\mathcal G}}^*(H, {\widetilde{n}}, m, {\ensuremath{\varepsilon}})$ for appropriate ${\widetilde{n}}$ and $p$. This is made precise in the following lemma. \[lemma:KLR\_rg\] Let $H$ be a graph such that $m_2(H) \ge 2$. Then there exist ${\ensuremath{\varepsilon}}, B > 0$ such that for $n^{-1/m_2(H)} \le p = p(n) \le \ln^{-2} n$, the graph $\Gamma \sim {\ensuremath{G(n,p)}}$ a.a.s. has the property that, for every ${\widetilde{n}} \ge Bp^{-m_2(H)}$, every $m \ge {\widetilde{n}}^2 p / 2$ and every graph $G'\in{{\mathcal G}}(H, {\widetilde{n}}, m, {\ensuremath{\varepsilon}})$, if $G' \subseteq \Gamma$ then $G'$ contains a copy of $H$. Let ${\ensuremath{\varepsilon}}$, $C > 0$ be as given by Theorem \[thm:KLR\] for $H$ and $\beta = (1/(2e^2))^{e(H)}$, and set $B = (2C)^{m_2(H)}$. Let $\Gamma\sim{\ensuremath{G(n,p)}}$. In order to prove the lemma it suffices to show that $\mu$ vanishes, where $\mu$ is the expected number of subgraphs of $\Gamma$ that are isomorphic to an element in ${{\mathcal G}}^*(H, {\widetilde{n}}, m, {\ensuremath{\varepsilon}})$. Note that $$\begin{aligned} \mu &= \sum_{{\widetilde{n}} \ge B p^{-m_2(H)}} \sum_{m \ge {\widetilde{n}}^2 p/2} \sum_{G\in {{\mathcal G}}^*} \Pr(G{\subseteq}\Gamma), $$ where ${{\mathcal G}}^* ={{\mathcal G}}^*(H, {\widetilde{n}}, m, {\ensuremath{\varepsilon}})$. Now $$\Pr(G{\subseteq}\Gamma)\le \binom{n}{{\widetilde{n}} k} ({\widetilde{n}} k)! p^{e(H) m}$$ where $k=v(H)$ for brevity. Furthermore, $m \ge C {\widetilde{n}}^{2 - 1/m_2(H)}$ follows from $m \ge {\widetilde{n}}^2 p/2$ and ${\widetilde{n}} \ge B p^{-m_2(H)}$. Thus we can apply the bound on $|{{\mathcal G}}^*(H, {\widetilde{n}}, m, {\ensuremath{\varepsilon}})|$ given by Theorem \[thm:KLR\]. We therefore have that $$\begin{aligned} \label{aux332} \mu &\le \sum_{{\widetilde{n}} \ge B p^{-m_2(H)}} \sum_{m \ge {\widetilde{n}}^2 p/2} \beta^m \binom{{\widetilde{n}}^2}{m}^{e(H)} \; \binom{n}{{\widetilde{n}} k } ({\widetilde{n}} k)! p^{e(H) m} \nonumber\\ &\le \sum_{{\widetilde{n}} \ge B p^{-m_2(H)}} \sum_{m \ge {\widetilde{n}}^2 p/2} \binom{n}{{\widetilde{n}}k} ({\widetilde{n}} k)! \beta^m \left( \frac{{{\widetilde{n}}}^2 e}{m} \right)^{e(H) m} p^{e(H) m} \nonumber\\ &\le \sum_{{\widetilde{n}} \ge B p^{-m_2(H)}} \sum_{m \ge {\widetilde{n}}^2 p/2} n^{2 {\widetilde{n}}k} \left( \beta^{1/e(H)} 2 e \right)^{e(H) m} \nonumber\\ &\le \sum_{{\widetilde{n}} \ge Bp^{-m_2(H)}} \sum_{m \ge {\widetilde{n}}^2 p / 2} \exp\left(2 k{\widetilde{n}} \ln n - m \right), \end{aligned}$$ where the third inequality follows from $m\ge {\widetilde{n}}^2p/2$, and the last inequality follows from our choice of $\beta$ and $e(H)\ge 1$ since $m_2(H)\ge 2$. Now, for sufficiently large $n$, $$2 k {\widetilde{n}} \ln n - m \le {\widetilde{n}} (2k \ln n - {\widetilde{n}} p /2) \le {\widetilde{n}} (2k \ln n - Bp^{-1}/2) < - \ln^2 n,$$ from the lower bound on $m$, the lower bound on ${\widetilde{n}}$ and $m_2(H)\ge 2$, and from the upper bound on $p$, respectively. Thus the final expression in  tends to 0 as $n{\rightarrow}\infty$. The assertion of the lemma follows from Markov’s Inequality. We remark that the condition $m_2(H) \ge 2$ is purely for convenience, and in fact $m_2(H) > 1$ would work as well (having an impact only on the upper bound on $p$). It should be noted that the previous lemma could also be derived from a result of Conlon, Gowers, Samotij, and Schacht [@conlon2014klr]. Finally, to apply the previous result in our proof we make use of the following lemma (see, e.g., [@gerke_steger_2005 Lemma 4.3]). \[lemma:exact\_m\_edges\] Given a positive ${\ensuremath{\varepsilon}}< 1/6$, there exists a constant $C$ such that any $({\ensuremath{\varepsilon}})$-regular graph $B = (V_1 \cup V_2, E)$ contains a $(2{\ensuremath{\varepsilon}})$-regular subgraph $B = (V_1 \cup V_2,E')$ with $|E'| = m$ edges for all $m$ satisfying $C |V(B)| \le m \le |E(B)|$. System of disjoint hyperedges ----------------------------- Given a hypergraph $H$, we denote by $\tau(H)$ the size of a smallest *vertex cover* of $H$, that is the size of a smallest subset $X \subseteq V(H)$ that intersects all the edges of $H$. Note that if $H$ and $H'$ are hypergraphs on the same vertex set then $\tau(H \cup H') \le \tau(H) + \tau (H')$. We make use of the following generalisation of Hall’s theorem due to Haxell [@haxell1995condition]. \[thm:haxell\] Let $H_1, \ldots, H_t$ be a family of $r$-uniform hypergraphs on the same vertex set. If for every $I \subseteq [t]$ we have $\tau(\bigcup_{i \in I} H_i) \ge 2 r |I|$ then one can choose a hyperedge $h_i \in E(H_i)$ for each $i \in [t]$ such that $h_i \cap h_j = \emptyset$ for distinct $i, j \in [t]$. The theorem from [@haxell1995condition] gives a slightly better bound than $2 r |I|$, however for our purposes this is sufficient. Maker’s strategy {#sec:Maker} ================ Our proof strategy is to show that Maker can build a graph which has certain properties and then show that these properties imply the existence of a $K_r$-factor. The properties we need are summarised in the following definition. Given $\alpha, \beta, p \in [0,1]$ and $r \in \mathbb{N}$, we say that a graph $G$ with $n$ vertices is *$(\alpha, \beta, p, r)$-neat* if it has the following properties: 1. \[prop:expand\] For every $v \in V(G)$ we have $|N_G(v)| \ge np/2$ and for all disjoint $X, Y \subseteq V(G)$ of size $|X| \ge \log n/p$ and $|Y| \ge \alpha n$ there exists a vertex $v \in X$ with at least $|Y|p/2$ neighbours in $Y$; 2. \[prop:in\_nbr\] For every $v\in V(G)$ and every subset $X \subseteq N_G(v)$ of size $|X| \ge \alpha np$ the induced subgraph $G[X]$ contains a copy of $K_{r-1}$; 3. \[prop:chain\] For all disjoint subsets $V_1, \ldots, V_{r+1} \subseteq V(G)$ of size $|V_i| \ge n^{1 - \beta}$ each, there exists a copy of $K_{r+1}^{-}$ with one vertex in each $V_i$, where $K_{r+1}^{-}$ is a graph obtained by removing an edge from a complete graph with $r+1$ vertices. The next lemma is the heart of the proof of Theorem \[thm:main\] \[main:M\]. It shows that neat graphs contain $K_r$-factors under certain mild conditions on the parameters. \[lemma:neat\_factor\] For any integer $r \ge 4$ and a positive $\beta$ there exists positive $\alpha$ and $n_0 \in \mathbb{N}$ such that any $(\alpha, \beta, p, r)$-neat graph with $n_0 < n \in r \mathbb{Z}$ vertices and $p \ge n^{-1/3}$ contains a $K_r$-factor. The proof of Lemma \[lemma:neat\_factor\] follows an approach from [@nenadov18triangle] and we postpone it for the next subsection. We now show how Lemma \[lemma:neat\_factor\] implies the first part of Theorem \[thm:main\]. Let $\alpha > 0$ be as given by Lemma \[lemma:neat\_factor\] for $\beta = 1/(5 (r+2)^2(r-1))$, let $K=K(\alpha)$ be a sufficiently large integer, and suppose $n \in r \mathbb{Z}$ is sufficiently large. Let $p= K n^{-2/(r+2)}$. We show that Maker can build an $(\alpha, \beta, p, r)$-neat graph in the $b$-biased Maker–Breaker game where $b= cp^{-1}$ for some small constant $c$. Such a graph contains a $K_r$-factor by Lemma \[lemma:neat\_factor\]. Maker plays two games in parallel: she plays Game 1 in every odd round and Game 2 in every even round, where Game 1 and Game 2 are defined below. Thus both games can be viewed as $(2b)$-biased Maker–Breaker games and can be played completely independently. For the rest of the argument we assume that Breaker has some fixed strategy, that is, for every disjoint pair $(E_M, E_B)$ of subsets of $E(K_n)$, which represents the current set of Maker’s and Breaker’s edges, he has some fixed rule what to claim next. If we can show that Maker has a winning strategy against an arbitrary such *rulebook*, then she can win regardless of what Breaker plays. In Game 1, Maker’s goal is to build a graph $G_1$ that satisfies \[prop:expand\] and \[prop:in\_nbr\], and in Game 2 she builds a graph $G_2$ that satisfies \[prop:chain\]. Overall, this implies that $G_1 \cup G_2$ is an $(\alpha, \beta, p, r)$-neat graph. #### Game 1. Let $\gamma > 0$ be a constant that we specify later, and let $\Gamma$ be a graph on $n$ vertices that has the following properties. 1. \[game1:maker\] In the $(2b)$-biased Maker–Breaker game on $K_n$, Maker has a strategy to claim a spanning subgraph $G \subseteq \Gamma$ with $\delta(G) \ge (1 - \gamma)np$ and $e_G(N_\Gamma(v)) \ge (1 - \gamma)p^3n^2/2$ for every $v \in V(\Gamma)$. 2. \[game1:disc\] $\Gamma$ satisfies the assertion of Lemma \[lemma:disc\]. 3. \[game1:r\_1\] For every ${\widetilde{n}}\ge Bp^{-r/2}$, every $m\ge {\widetilde{n}}^2 p/2$ and every graph $G'\in {{\mathcal G}}(K_{r-1},{\widetilde{n}}, m,{\ensuremath{\varepsilon}})$, if $G'{\subseteq}\Gamma$ then $G'$ contains $K_{r-1}$ as a subgraph, where $B=B(K_{r-1})$ and ${\ensuremath{\varepsilon}}={\ensuremath{\varepsilon}}(K_{r-1})$ are the constants from Lemma \[lemma:KLR\_rg\] applied to $H=K_{r-1}$. We argue briefly that such a graph $\Gamma$ exists. Let $\Gamma \sim {\ensuremath{G(n,p)}}$. Then $\Gamma$ satisfies the assertion of \[game1:maker\] a.a.s. by Theorem \[thm:maker\_rg\] if we choose $K=K(\gamma)$ large enough and $c=c(\gamma)$ small enough. Furthermore, $\Gamma$ satisfies \[game1:disc\] a.a.s. by Lemma \[lemma:disc\], and it satisfies \[game1:r\_1\] by Lemma \[lemma:KLR\_rg\] applied to $H=K_{r-1}$ where we note that $p= K n^{-2/(r+2)} > n^{-2/r} = n^{-1/m_2(K_{r-1})}$ and $m_2(K_{r-1}) \ge 2$. Therefore, we can choose one particular graph $\Gamma$ which has these properties. Let $G \subseteq \Gamma$ be a spanning subgraph guaranteed by \[game1:maker\]. We show that $G$ satisfies \[prop:expand\] and \[prop:in\_nbr\]. For \[prop:expand\] note that $G$ can be obtained from $\Gamma$ by removing at most $2 \gamma np$ edges touching each vertex since the maximum degree of $\Gamma$ is at most $(1 + \gamma)np$, by \[game1:disc\], and since $\delta(G) \ge (1 - \gamma)np$, by \[game1:maker\]. Furthermore, let $X, Y{\subseteq}V(G)$ be disjoint subsets of size $|X|\ge \log n/p$ and $|Y|\ge \alpha n$, respectively. Then $e_\Gamma(X, Y) \ge (1 - \gamma)|X||Y|p$ by \[game1:disc\]. But then at most $|X| \cdot 2\gamma np$ of those edges are not present in $G$ by the preceding observation. By choosing $\gamma < \alpha / 8$, we have $$e_G(X, Y) \ge (1 - \gamma)|X||Y|p - |X| \cdot 2 \gamma np > |X||Y|p/2.$$ Therefore there exists a vertex $v \in X$ with at least $|Y|p/2$ neighbours in $Y$. For \[prop:in\_nbr\] we show that for every $v\in V(G)$, every subset $X{\subseteq}N_G(v)$ of size $|X|\ge \alpha np$ hosts a copy of some $G'\in {{\mathcal G}}(K_{r-1},{\widetilde{n}}, m,{\ensuremath{\varepsilon}})$, for suitable ${\widetilde{n}}$, $m$ and ${\ensuremath{\varepsilon}}$, which contains a copy of $K_{r-1}$ by \[game1:r\_1\]. Fix $v\in V(G)$ and note that we have $|N_{\Gamma}(v)| = (1\pm \gamma) np\gg \log n /p$ by \[game1:disc\] and assumption on $p$. Thus, again by \[game1:disc\], $$e_\Gamma(N_\Gamma(v)) \le (1 + \gamma) |N_\Gamma(v)|^2p/2 \le (1 + \gamma)^3 n^2p^3 / 2 < (1 + 4 \gamma) n^2 p^3 / 2,$$ where in the last inequality we assumed that $\gamma$ is sufficiently small. From \[game1:maker\] we conclude that $G[N_\Gamma(v)]$ is ‘missing’ at most $5 \gamma n^2 p^3 / 2$ edges. For brevity, let us upper bound this by $3 \gamma n^2 p^3$. More precisely, there exists a graph $R_v$ on the vertex set $N_\Gamma(v)$ such that $e(R_v) \le 3 \gamma n^2p^3$ and $G[N_\Gamma(v)] = \Gamma[N_\Gamma(v)] \setminus R_v$. Therefore, for all disjoint $X, X' \subseteq N_\Gamma(v)$ we have $$e_\Gamma(X, X') - 3 \gamma n^2 p^3 \le e_G(X, X') \le e_\Gamma(X, X').$$ Additionally, if $|X|,|X'|\gg \log n /p$ then $e_\Gamma(X, X') = (1\pm \gamma)|X||X'|p$ by \[game1:disc\]. Let ${\ensuremath{\varepsilon}}' = {\ensuremath{\varepsilon}}(K_{r-1})/4$, where ${\ensuremath{\varepsilon}}(K_{r-1})$ is given in \[game1:r\_1\]. It follows that for any two disjoint subsets $X,X'{\subseteq}N_{\Gamma}(v)$ of size at least ${\ensuremath{\varepsilon}}'\cdot (\alpha n p/r)$ we have $$\begin{aligned} e_G(X, X') &\ge (1 - \gamma)|X||X'| p - 3 \gamma n^2 p^3 \ge \left(1- {\ensuremath{\varepsilon}}'\right) |X||X'| p \end{aligned}$$ and $$e_G(X, X') \le (1+\gamma) |X||X'| p \le \left(1+{\ensuremath{\varepsilon}}'\right) |X||X'| p,$$ if we choose $\gamma$ small enough in terms of ${\ensuremath{\varepsilon}}'$, $\alpha$ and $r$. This implies that any two disjoint subsets $V_1, V_2 \subseteq N_\Gamma(v)$ of size $\alpha np/r$ form a $(2 {\ensuremath{\varepsilon}}')$-regular pair. Let now $X \subseteq N_G(v) \subseteq N_\Gamma(v)$ be of size $\alpha np$. Arbitrarily choose $r-1$ disjoint subsets $V_1, \ldots, V_{r-1} \subseteq X$ of size $ {\widetilde{n}} = \alpha np / r$. Note that $ {\widetilde{n}} \ge B p^{-r/2} = B p^{-m_2(K_{r-1})}$, where $B=B(K_{r-1})$ is the constant given by \[game1:r\_1\], since $p \ge K n^{-2/(r+2)}$ and $K$ is a sufficiently large constant. As previously observed, every $(V_i, V_j)$ forms a $(2{\ensuremath{\varepsilon}}')$-regular pair with $m_{i,j} = (1 \pm {\ensuremath{\varepsilon}}') {\widetilde{n}}^2 p$ edges, thus we can apply Lemma \[lemma:exact\_m\_edges\] to each pair $(V_i, V_j)$ in order to obtain a subset $E_{ij} \subseteq E_G(V_i, V_j)$ of size exactly $$m = {\widetilde{n}}^2 p /2$$ such that $(V_i, V_j)$ is $(4{\ensuremath{\varepsilon}}')$-regular, i.e. $({\ensuremath{\varepsilon}})$-regular, with respect to $E_{ij}$. This gives us a graph $G' \in {\ensuremath{\mathcal G}}(K_{r-1}, {\widetilde{n}}, m, {\ensuremath{\varepsilon}})$. As $G'$ is a subgraph of $\Gamma$, from \[game1:r\_1\] we conclude that it contains a copy of $K_{r-1}$. Thus, $G$ satisfies \[prop:in\_nbr\]. #### Game 2. The properties \[prop:in\_nbr\] and \[prop:chain\] are achieved in a fairly similar way. Thus, the analysis of Game 2 follows along the lines of the second part of Game 1. There are some crucial differences in the choice of parameters though. Recall that $\beta = 1/(5(r+2)^2(r-1))$ and that for property \[prop:chain\] we want to find a copy of $K_{r+1}^-$ in certain sets of size $n^{1-\beta}$. Let $H=K_{r+1}^-$, let now $\gamma = n^{-3\beta}$ and let $q \in (0,1)$ satisfy $$\label{qBounds} n^{-\frac{1-\beta}{m_2(H)}}\ll q \ll n^{-\frac{2}{r+2}} \gamma^6.$$ Note that this is possible since $m_2(K_{r+1}^-)= \frac{r+2}{2}-\frac{1}{r-1}$ and by choice of $\beta$. Delicate choices for parameters $\gamma$ and $q$ will become apparent soon. We claim that a random graph $\Gamma \sim {\ensuremath{G(n,q)}}$ has the following properties with high probability. 1. \[game2:maker\] In the $(2b)$-biased Maker–Breaker game on $K_n$, Maker has a strategy to claim a spanning subgraph $G \subseteq \Gamma$ with $\delta(G) \ge (1 - \gamma)nq$ and $e_G(N_\Gamma(v)) \ge (1 - \gamma)q^3n^2/2$ for every $v \in V(\Gamma)$. 2. \[game2:disc\] $\Gamma$ satisfies the assertion of Lemma \[lemma:disc\]; 3. \[game2:r\_1\] For every ${\widetilde{n}}\ge Bq^{-m_2(H)}$, every $m\ge {\widetilde{n}}^2 q/2$ and every graph $G'\in {{\mathcal G}}(H,{\widetilde{n}}, m,{\ensuremath{\varepsilon}})$, if $G'{\subseteq}\Gamma$ then $G'$ contains $H$ as a subgraph, where now $B=B(H)$ and ${\ensuremath{\varepsilon}}={\ensuremath{\varepsilon}}(H)$ are the constants from Lemma \[lemma:KLR\_rg\] applied to $H=K_{r+1}^-$. Let us verify that $\Gamma$ indeed has these properties a.a.s. For \[game2:maker\] let us verify that the conditions of Theorem \[thm:maker\_rg\] hold. Firstly, $q\ge 10^{8}\gamma^{-2}n^{-1/2}$ is implied by the lower bound in  since $r\ge 4$ and since $\beta< 1/50$, say. Secondly, recall that $b= cp^{-1}=O(n^{2/(r+2)})$. This together with the upper bound in  implies that $\gamma^{6} q^{-1} \gg b$, so that \[game2:maker\] holds a.a.s. by Theorem \[thm:maker\_rg\]. Just as in Game 1, $\Gamma$ satisfies \[game2:disc\] a.a.s. by Lemma \[lemma:disc\]. Finally, note that the lower bound in  implies in particular that $q\ge n^{-1/m_2(H)}$. Thus \[game2:r\_1\] holds a.a.s. by Lemma \[lemma:KLR\_rg\] applied to $H=K_{r+1}^-$. Fix $\Gamma$ with these three properties and let $G \subseteq \Gamma$ be a spanning subgraph satisfying \[game2:maker\]. Crucially, we have sacrificed the value of $q$, which is now significantly smaller than $n^{-2/(r+2)}$, in order to get a smaller error term $\gamma$. Note that in order to guarantee that $G$ satisfies property \[prop:in\_nbr\] in Game 1 we needed $p = \Omega(n^{-2/(r+2)})$. Here it will turn out that a smaller $p$ (which we denote by $q$) suffices provided $\gamma$ is sufficiently small. We now make this precise. For ${\ensuremath{\varepsilon}}' = {\ensuremath{\varepsilon}}(H) /4 >0$ consider disjoint subsets $X, X' \subseteq V(G)$ of size at least ${\ensuremath{\varepsilon}}' n^{1 - \beta}$. First note that $$\begin{aligned} e_{\Gamma} (X,X') = (1\pm \gamma)|X||X'|q, \end{aligned}$$ by \[game2:disc\] since $\gamma^2n^{1-\beta}q\gg \log n$ by choice of $\beta$ being small enough and . As before, we have that $G$ is obtained from $\Gamma$ by removing at most $2 \gamma nq$ edges touching each vertex, which sums to at most $\gamma n^2 q$ removed edges in total. This, together with the above estimate on $ e_{\Gamma} (X,X')$, implies that $$\begin{aligned} (1 + \gamma)|X||X'| q \ge e_G(X, X') &\ge (1 - \gamma)|X||X'|q - \gamma n^2 q \\ &\ge (1 - {\ensuremath{\varepsilon}}')|X||X'|q, \end{aligned}$$ since $\gamma < {\ensuremath{\varepsilon}}'/2$, say, and $\gamma n^2 = n^{2-3\beta}\ll {\ensuremath{\varepsilon}}' |X||X'|/2$. Note that it was crucial here that $\gamma \ll n^{- 2\beta}$, that is, $\gamma$ polynomially depends on $n$. Therefore, every pair of disjoint subsets $X, Y \subseteq V(G)$ of size $n^{1 - \beta}$ forms a $(2{\ensuremath{\varepsilon}}')$-regular pair. The rest of the argument is the same as in the previous case. Consider some disjoint $V_1, \ldots, V_{r+1} \subseteq V(G)$, each of size ${\widetilde{n}} = n^{1 - \beta}$. As observed, each pair $(V_i, V_j)$ forms a $(2{\ensuremath{\varepsilon}}')$-regular pair with $m_{ij} \ge (1 - {\ensuremath{\varepsilon}}'){\widetilde{n}}^2 q$ edges. By Lemma \[lemma:exact\_m\_edges\], there exists a subset $E_{ij} \subseteq E_G(V_i, V_j)$ of size exactly $$m = {\widetilde{n}}^2 q / 2$$ such that $(V_i \cup V_j, E_{ij})$ is $(4{\ensuremath{\varepsilon}}')$-regular, i.e. $({\ensuremath{\varepsilon}})$-regular. This gives us a subgraph $G' \subseteq G$ which belongs to ${\ensuremath{\mathcal G}}(K_{r+1}^-, {\widetilde{n}}, m, {\ensuremath{\varepsilon}})$. From the lower bound in  we infer that ${\widetilde{n}} \ge B(H) q^{-m_2(H)}.$ Thus \[game2:r\_1\] implies that $G'$ contains a copy of $K_{r+1}^-$ with one vertex in each $V_i$. Neat graphs contain $K_r$-factors (Lemma \[lemma:neat\_factor\]) ---------------------------------------------------------------- The proof of Lemma \[lemma:neat\_factor\] closely follows ideas from [@nenadov18triangle] which are, in turn, based on ideas of Krivelevich [@krivelevich1997triangle]. Recall that $K_{r+1}^-$ denotes the graph obtained from $K_{r+1}$ by removing an edge. The main building block in the proof is an *$(r, \ell)$-chain*, the graph obtained by sequentially ‘gluing’ $\ell \ge 0$ copies of $K_{r+1}^-$ on a vertex of degree $r-1$ (see Figure \[fig:chain\]). We define the $(r, 0)$-chain to be a single vertex. A graph is an *$r$-chain* if it is isomorphic to an $(r, \ell)$-chain, for some integer $\ell \ge 0$. = \[fill=black,circle,inner sep=1.7pt,minimum size=0.5pt\] = \[fill=black,rectangle,inner sep=2.5pt,minimum size=2.5pt\] in [1,...,4]{}[ (f) at (0 + 5 \* , 0) ; //in [1/2.5/1, 2/2.5/-1, 3/2/0.3, 4/3/-0.3]{} (v) at (0 + 5 \* + , ) ; ]{} (f5) at (25, 0) ; /in [1/1,2/2,3/3,4/4]{} in [1,...,4]{} (f) – (v); (v1) – (v2) – (v3) – (v4) – (v1); (v1) – (v3); (v2) – (v4); /in [2/1,3/2,4/3,5/4]{}[ in [1,...,4]{} (f) – (v); ]{} ; in [1,...,3]{} (-0.5 + 5 \* , 1.5) rectangle (0.5 + 5 \* + 3, -1.5); (-0.5 + 5 \* 4 + 2, 1.5) rectangle (0.5 + 5 \* 5, -1.5); An $(r, \ell)$-chain contains $\ell + 1$ vertices such that removing either of them (but exactly one!) results in a graph which contains a $K_r$-factor. We call such vertices *removable*. If a graph $H$ is an $r$-chain then we use $R(H)$ to denote the set of its removable vertices. We repeatedly use the following observation. \[obs:canonical\] Let $G$ be a graph and $C_1, \ldots, C_r \subseteq G$ be vertex disjoint $r$-chains. If there exists a copy of $K_r$ in $G$ which intersects each $R(C_i)$ then the subgraph of $G$ induced by $\bigcup_{i \in [r]} V(C_i)$ contains a $K_r$-factor. The following lemma together with property \[prop:chain\] ensures the existence of large $(r, \ell)$-chains. \[lemma:long\_chain\] Let $G$ be a graph with $n$ vertices such that for every disjoint $X, Y \subseteq V(G)$, each of size at least $\alpha n$, there exists a copy of $K_{r+1}^-$ in $G$ with one vertex of degree $r-1$ in $X$ and all other vertices in $Y$. Then $G$ contains an $(r, \ell$)-chain for every $\ell < (1 - (r+2)\alpha) n/r$. In the case when $r=3$ this is Lemma 3.1 in [@nenadov18triangle]. Trivial adjustments to that proof give Lemma \[lemma:long\_chain\]. We omit the proof. The following, somewhat technical looking lemma provides a crucial *absorbing* property of a collection of $r$-chains that we exploit in the proof of Lemma \[lemma:neat\_factor\]. \[lemma:absorbing\] Let $G$ be a graph with $n$ vertices which satisfies \[prop:chain\] for some $\beta > 0$ and $r$, where $n \ge n_0(\beta, r)$ is sufficiently large. Let $W \subseteq V(G)$ be a subset of size $|W| \ge n/8$, and let $\ell, t \in \mathbb{N}_0$ and $\ell' \ge \ell, t' \in \mathbb{N}$ be such that: - $(\ell + 1) t' > n^{1 - \beta/2}$, and - $(t + t')(r \ell + 1) < |W|/2$. Suppose we are given disjoint $(r, \ell')$-chains $C_1', \ldots, C'_{t'} \subset V(G) \setminus W$. Then there exist disjoint $(r, \ell)$-chains $C_1, \ldots, C_t \subset G[W]$ with the following property: for every $L \subseteq [t]$ there exists $L' \subseteq [t']$ such that the subgraph of $G$ induced by $$\left( \bigcup_{i \in L} V(C_i) \right) \cup \left( \bigcup_{i \in L'} V(C'_i) \right)$$ contains a $K_r$-factor. Before we prove Lemma \[lemma:absorbing\] it is instructive to first see how it is used to derive Lemma \[lemma:neat\_factor\]. Consider an equipartition $V(G) = V_1 \cup V_2$ chosen uniformly at random. As each vertex has $np/2 \gg \log n$ neighbours (follows from \[prop:expand\] and the bound on $p$), by Chernoff’s inequality and union-bound we have with high probability that every vertex has at least $np/8$ neighbours in $V_1$. Therefore there exists a partition for which this holds. Without loss of generality we may assume that $\beta = 1/k$ for some integer $k \ge 2$. For each $i \in \{1, \ldots, 4k-1\}$ set $\ell_i = n^{1 - (4k-i)/4k}$ and $t_i = n / (32 k (r \ell_i + 1))$. Note that $$\label{eq:ell_i_t_i} (\ell_i + 1) t_{i+1} = \Theta(n^{1 - \beta/4}).$$ By repeated application of Lemma \[lemma:long\_chain\] we can find a collection $C_1^{4k-1}, \ldots, C_{t_{4k-1}}^{4k-1} \subseteq G[V_2]$ of pairwise vertex-disjoint $(r, \ell_{4k-1})$-chains. Let us elaborate briefly why this is indeed possible. Such chains occupy $t_{4k - 1} \cdot (r \ell_{4k - 1} + 1) < n / 32$ vertices. Thus if we greedily choose them one by one the set $W \subseteq V_2$ of unoccupied vertices in $V_2$ after every step is of size at least, say, $|W| \ge n/4$. Therefore, by \[prop:chain\] we have that $G[W]$ satisfies the assumption of Lemma \[lemma:long\_chain\] for any constant $\alpha > 0$, and consequently it contains an $(r, \ell)$-chain for $\ell < (1 - (r+2)\alpha) |W| / r$. As $\ell_{4k-1} = o(n)$, this proves our claim. Let $U_{4k - 1} = \bigcup_{i \in [t_{4k - 1}]} V(C_i^{4k - 1})$. For each $i = 4k - 2, \ldots, 1$, iteratively, let $C_1^i, \ldots, C_{t_i}^i \subset G[V_2] \setminus U_{i+1}$ be disjoint $(r, \ell_i)$-chains given by Lemma \[lemma:absorbing\] for $C_1^{i+1}, \ldots, C_{t_{i+1}}^{i+1}$ (as $C_1', \ldots, C'_{t'}$) and $W_i = V_2 \setminus U_{i+1}$, and set $U_i = U_{i+1} \cup \bigcup_{j \in [t_i]} V(C_j^i)$. Let us verify that the conditions of Lemma \[lemma:absorbing\] are met. First, $|W_i| = n/2 - |U_{i+1}|$ and $$|U_{i+1}| = \sum_{j = i+1}^{4k-1} t_j \cdot (r \ell_j + 1) < n/4.$$ From we have $(\ell_i + 1) t_{i+1} > n^{1 - \beta/2}$, and as $$(t_i + t_{i+1})(r \ell_i + 1) < 2 t_i (r \ell_i + 1) < n / 16 < |W|/2$$ we can indeed apply Lemma \[lemma:absorbing\] in each iteration. Finally, let $W_0 = V_2 \setminus U_1$. Apply Lemma \[lemma:absorbing\] one last time with $\ell_0 = 0$, $t_0 = |W_0|/4$ and $C_1^1, \ldots, C_{t_1}^1$ (as $C_1', \ldots, C_{t'}')$. This is justified as $t_1=\Theta(n^{1-1/4k})$ and $t_0+t_1=o(n)$. The obtained $0$-chains are then just a set of vertices $C_0 \subseteq W_0$ with the property that for every $L_0 \subseteq C_0$ there exists a subset $L_1' \subseteq [t_1]$ such that the subgraph of $G$ induced by $$L_0 \cup \left( \bigcup_{j \in L_1'} V(C_j^1) \right)$$ contains a $K_r$-factor. Next, we show that the set $C_0 \cup U_1$ has a strong absorbing property. For any subset $L_0 \subseteq C_0$ such that $|L_0| + |U_1| \in r \mathbb{Z}$, the induced subgraph $G[L_0 \cup U_1]$ contains a $K_r$-factor. Consider one such $L_0$ and let $L_1' \subseteq [t_1]$ be a subset such that $$L_0 \cup \left(\bigcup_{j \in L_1'} V(C_j^1) \right)$$ contains a $K_r$-factor. We further take $L_1 = [t_1] \setminus L_1'$ and use the property guaranteed by Lemma \[lemma:absorbing\] to obtain a subset $L_2' \subseteq [t_2]$ such that the subgraph of $G$ induced by $$\left( \bigcup_{j \in L_1} V(C_j^1) \right) \cup \left( \bigcup_{j \in L_2'} V(C_j^2) \right)$$ contains a $K_r$-factor. Continuing this way, we obtain a subset $L_{4k-1}' \subseteq [t_{4k - 1}]$ such that the subgraph of $G$ induced by $$L_0 \cup \bigcup_{i = 1}^{4k - 2} \left( \bigcup_{j \in [t_i]} V(C_j^i) \right) \cup \left( \bigcup_{j \in L_{4k-1}'} V(C_j^{4k - 1}) \right) = (L_0 \cup U_1) \setminus \bigcup_{j \in L_{4k - 1}} V(C_j^{4k - 1})$$ contains a $K_r$-factor, where $L_{4k - 1} = [t_{4k - 1}] \setminus L_{4k - 1}'$. As $|V(C_j^{4k - 1})| \equiv 1 (\textrm{mod } r)$ and $|L_0| + |U_1| \in r \mathbb{Z}$ we necessarily have $|L_{4k - 1}| \in r \mathbb{Z}$. Therefore, to complete a $K_r$-factor in $G[L_0 \cup U_1]$ it suffices to partition $L_{4k - 1}$ into groups of size $r$ and for each such group $\{i_1, \ldots, i_r\}$ find a copy of $K_r$ with one vertex in each $R(C_{i_1}^{4k - 1}), \ldots, R(C_{i_r}^{4k - 1})$ (see Observation \[obs:canonical\]). The existence of such $K_r$ follows from \[prop:chain\] and $|R_j^{4k - 1}| = \ell_{4k - 1} + 1 > n^{1 - 1/4k} > n^{1 - \beta}$. We now use this absorbing property to find a $K_r$-factor in $G$. First, let $B \subseteq V_1 \cup (W_0 \setminus C_0)$ be the set of all vertices which are not part of chains and such that they have less than $|C_0|p/2$ neighbours in $C_0$. As $|C_0| \ge \alpha n$, we have $|B| < \log n / p \ll np$, by \[prop:expand\] and the lower bound on $p.$ By \[prop:in\_nbr\] and the assumption that every vertex has at least $np/8$ neighbours in $V_1$, we can iteratively take one vertex $v \in B$ at a time and find a copy of $K_r$ which contains $v$ and has all other vertices in $V_1 \setminus B$. This takes care of $B$. Furthermore, we can continue covering the remaining vertices in $V_1 \cup (W_0 \setminus C_0)$ (i.e. those which are not part of previously chosen $K_r$’s) with disjoint copies of $K_r$ as long as there are still at least $r n^{1 - \beta}$ vertices, by \[prop:chain\]. Let us denote the set of remaining vertices by $L$. With the absorbing property of $C_0 \cup U_1$ in mind, to find a $K_r$-factor of $G$ it now suffices to find vertex-disjoint copies of $K_r$, each of which contains one vertex from $L$ and the others from $C_0$. Whatever we are left with in $C_0$ is guaranteed to form a $K_r$-factor with $U_1$, thus we are done. Note that this is very similar with how we took care of $B$, however the main difference is that $L$ is significantly larger than $B$ and a simple greedy strategy might not work. Instead, we find the desired copies of $K_r$ using Haxell’s matching theorem (Theorem \[thm:haxell\]). For each $v \in L$ create an $(r-1)$-uniform hypergraph $H_v$ on the vertex set $C_0$ such that $\{v_1, \ldots, v_{r-1}\}$ forms a hyperedge if and only if $\{v, v_1, \ldots, v_{r-1}\}$ form $K_r$ in $G$. If we can find for each $v \in L$ a hyperedge $h_v \in E(H_v)$ such that all these hyperedges are pairwise vertex-disjoint, then we are done. To show that such edges exist it suffices to verify Haxell’s criterium: $$\label{eq:verify_haxell} \tau(\bigcup_{v \in I} H_v) \ge 2(r-1)|I|$$ for every $I \subseteq L$. Equivalently, for all subsets $I{\subseteq}L$ and all $Z \subseteq C_0$ of size $|Z| \le 2(r-1)|I|$ there exists a copy of $K_r$ with one vertex in $I$ and all other vertices in $C_0 \setminus Z$. We consider two cases. Consider first the case when $|I| \le \log n /p$ and let $Z$ be some subset of $C_0$ of size at least $2r \log n / p$. As $L \cap B = \emptyset$, every vertex $v \in I$ has at least $|C_0|p/2 > np / 32$ neighbours in $V_1$, thus the subset $X = (N_G(v) \cap C_0) \setminus Z$ is of size at least $np/16$ (we used $np \gg \log n /p$ which follows from the lower bound on $p$). By \[prop:in\_nbr\] there exists a copy of $K_{r-1}$ in $X$. Suppose now that $|L| \ge |I| > \log n /p$ and consider a subset $Z \subseteq C_0$ of size $2r |L| < n/32$. The set $Y = C_0 \setminus Z$ is then of size at least $n/16$. Thus, there exists a vertex $v \in I$ with at least $|Y|p/2 \ge np/32$ neighbours in $Y$, by \[prop:expand\]. By \[prop:in\_nbr\] such a neighbourhood contains a copy of $K_{r-1}$, which gives us a desired copy of $K_r$. This finishes the proof. It remains to prove Lemma \[lemma:absorbing\]. By repeated application of Lemma \[lemma:long\_chain\] we can find a collection of $t + t'$ disjoint $(r, \ell)$-chains $C_1, \ldots, C_{t + t'} \in G[W]$. Clearly, for this we could have allowed $W$ to be much smaller than $n/8$, thus this constraint is only for convenience. For each $i \in [t + t']$ we create an auxiliary $(r-1)$-uniform hypergraph $H_i$ on the vertex set $V' = [t']$ by adding a hyperedge $\{j_1, \ldots, j_{r-1}\}$ if and only if there exists a copy of $K_r$ in $G$ with one vertex in each $R(C_i), R(C_{j_1}'), \ldots, R(C_{j_{r-1}}')$. Note that for every such hyperedge the subgraph of $G$ induced by $$V(C_i) \cup V(C_{j_1}') \cup \ldots V(C_{j_{r-1}}')$$ contains a $K_r$-factor (see Observation \[obs:canonical\]). We first show that there exists a subset $B \subseteq [t + t']$ of size at most $|B| \le t'$ such that for every subset $J \subseteq [t + t'] \setminus B$ of size $|J| \le t'/8r$ we have $$\label{eq:2rJ} \tau(\bigcup_{i \in J} H_i) \ge 2r|J|.$$ Initially, set $q = 0$ and $B = \emptyset$. As long as $|B| < t'/8r$ and there exists a subset $J \subseteq [t + t'] \setminus B$ of size $|J| \le t'/8r$ that violates set $B = B \cup J$, $J_{q+1} = J$ and increase $q$ by 1. Suppose towards a contradiction that for some $q$, $|B| \ge t'/8r$, and let $q$ be the smallest such index. Then $|B| \le t'/4r$ as $|J_q| \le t'/8r$. Moreover, we have $$\tau(\bigcup_{i \in B} H_i) \le \sum_{j = 1}^q \tau(\bigcup_{i \in J_j} H_i) < \sum_{j = 1}^q 2 r |J_j| = 2r|B| \le t'/2.$$ This implies that there exists a set ${\widetilde{B}} {\subseteq}V'$ of size at most $t'/2$ such that every hyperedge $h\in\bigcup_{i\in B}H_i$ intersects ${\widetilde{B}}$. In other words, there exists $B' \subseteq V'$ of size $|B'| \ge t'/2$ such that there is no copy of $K_r$ in $G$ with one vertex in $\bigcup_{i \in B} R(C_i)$ and the others in each $R(C_{j_2}'), \ldots R(C_{j_{r}}')$ for some distinct $j_2, \ldots, j_{r} \in B'$. Split $B'$ arbitrarily into $r-1$ sets of nearly equal size, denoted by $B_2', \ldots, B_{r}'$, each of size at least $t'/2r$, and set $X_j = \bigcup_{i \in B'_j} R(C_i')$ for $j=2,\ldots,r$. Then each such $X_j$ is of size at least $$(\ell' + 1) \frac{t'}{2r} > (\ell + 1) \frac{t'}{2r}.$$ On the other hand, $X_1$ defined as $\bigcup_{i \in B} R(C_i)$ is of size at least $$|X_0| \ge (\ell + 1) |B| \ge (\ell + 1) \frac{t'}{8r}.$$ Thus we have $|X_i| \ge n^{1 - \beta}$ by the assumption of the lemma, with room to spare. By \[prop:chain\] there exists a copy of $K_r$ intersecting each $X_i$, which is a contradiction. Therefore we have that there exists a set $|B|$ of size less than $t'/8r$ and every subset $J \subseteq [t + t'] \setminus B$ of size $|J| \le t' / 8r$ satisfies . Take an arbitrary $t$-subset $I \subseteq [t + t'] \setminus B$ and relabel $\{C_i\}_{i \in I}$ as $\{C_i\}_{i \in [t]}$. We show that such $(r, \ell)$-chains have the desired property. Consider some $L \subseteq [t]$. First, let $S \subseteq L$ be a smallest subset such that the subgraph of $G$ induced by $$\bigcup_{i \in L \setminus S} V(C_i)$$ contains a $K_r$-factor. We claim that $|S| < t'/8r$. Suppose towards a contradiction that $|S| \ge t' / 8r$. Consider an equipartition $S = S_1 \cup \ldots \cup S_r$. Then each set $X_i = \bigcup_{j \in S_i} R(C_j)$ is of size $$|X_i| \ge (\ell + 1) \frac{t'}{8r^2} > n^{1 - \beta}$$ thus by \[prop:chain\] there exist a copy of $K_r$ intersecting each $X_i$. Therefore there exists distinct $i_1, \ldots, i_r \in S$ and a copy of $K_r$ intersecting each $R(C_{i_j})$. By Observation \[obs:canonical\] this is a contradiction with the minimality of $S$. Finally, as $|S| \le t'/8r$ we have that every subset $J \subseteq S$ satisfies thus we can choose $h_i \in H_i$ for each $i \in S$ such that these edges are pairwise vertex disjoint, by Theorem \[thm:haxell\]. Let $L' = \bigcup_{i \in S} h_i$. The construction of such hyperedges implies that the subgraph of $G$ induced $$\bigcup_{i \in L} V(C_i) \cup \bigcup_{i \in L'} V(C_i')$$ contains a $K_r$-factor, as desired. Breaker’s strategy ================== The idea behind the proof is that Breaker prevents a fixed vertex $v$ from being in a copy of $K_r$ in Maker’s graph. To illustrate why this could be possible, fix a vertex $v\in [n]$ and assume for now that Maker at first only claims edges incident to $v$, as long as there is at least one such unclaimed edge. Breaker responds by claiming $b$ edges incident to $v$ in every round as well, so that at the end of this first stage of the game the set of neighbours of $v$ in Makers graph, denoted by $N_M(v)$, has size roughly $n/b$. For the rest of the game, Breaker only needs to prevent Maker from claiming a copy of $K_{r-1}$ in $N_M(v)$, which is possible if $b\ge C (n/b)^{2/r}$ for some constant $C$ which is independent of $n$, by the result of Bednarska and Łuczak [@bl2000]; or equivalently if $b\ge C n^{2/(r+2)}$ (with a different constant $C$). If Maker indeed first claims as many edges incident to $v$ as possible, this would be the end of the proof. Of course, we cannot rely on this assumption. The way to counterfeit it is to divide the attention of Breaker into two: the first $b/2$ claimed edges are incident to $v$, thus preventing its neighbourhood in the Maker’s graph from becoming larger than $2n/b$; the second $b/2$ claimed edges lie inside its current neighbourhood and prevent a copy of $K_{r-1}$. Crucially, the board of the game where we want to use the strategy ${{\mathcal S}}$ from [@bl2000] will be revealed over time only (as the neighbourhood of $v$ in Maker’s graph increases). It turns out that the proof of a static version of the game (where the whole board is ‘visible’) can be turned into a proof of a suitable dynamic version (where the board is revealed over time). Unfortunately, none of the ingredients of the proof is black-boxable so we need to dig into each part. Let us introduce necessary notation in a bit more generality than needed for our application. Let ${{\mathcal H}}$ be a given hypergraph, say on vertex set $V({{\mathcal H}})$ and edge set $E({{\mathcal H}})$, and let $m$ and $b$ be integers. We define the [*dynamic-board $({{\mathcal H}},m,b)$-game*]{} as follows. Let $V_0 = \emptyset$. The two players Maker and Breaker play in rounds, with Maker going first. For $i\ge 0$, suppose that $i$ rounds have been played and that $V_i{\subseteq}V({{\mathcal H}})$ is defined. In round $i+1$, Maker may play either according to [*Option (a)*]{} in which she claims up to $m$ elements of $V_i$ and sets $V_{i+1}=V_i$ (in case there are less than $m$ elements she claims all of them), or according to [*Option (b)*]{} in which she chooses elements $v_1,\ldots,v_{\ell}$ (for some $\ell\ge 1$) from $V({{\mathcal H}}) {\setminus}V_i$ and sets $V_{i+1} = V_i\cup\{v_1,\ldots,v_{\ell}\}$ (but Maker does not claim edges in a round when she enlarges the board). In case Option (a) is not possible, Maker is forced to play Option (b), unless it is the end of the game. Afterwards, Breaker claims (up to) $b$ elements in $V_{i+1}$. In case there are less than $b$ unclaimed elements in $V_{i+1}$, Breaker claims all of them. Maker wins if at the end of the game she has claimed all elements of some hyperedge $H\in E({{\mathcal H}})$. Otherwise, Breaker wins. Given a (fixed) graph $H$ and a complete graph $K_n$, we define a [*dynamic $b$-biased $H$-game*]{} as the $({{\mathcal H}}, 1, b)$-game where the vertex set of ${{\mathcal H}}$ are the edges of $K_n$, and the hyperedges of ${{\mathcal H}}$ correspond to edge sets of $K_n$ which form a copy of $H$. The following theorem is a generalisation of the mentioned result by Bednarska and Łuczak [@bl2000] to the dynamic setting. For the definition of $m_2(H)$, see Theorem \[thm:KLR\]. For every graph $H$ which contains at least three non-isolated vertices there exists a constant $C>0$ such that Breaker has a winning strategy in the dynamic $b$-biased $H$-game played on $K_n$ if $b\ge Cn^{1/m_2(H)}$. We may take $C$ sufficiently large such that the theorem statement is true for small $n$, so that in the proof we can safely assume that $n$ is as large as needed. The proof of Theorem \[ourDynamicBreaker\] proceeds along the lines of [@bl2000]. We sketch the argument in the next section, leaving out calculations that are identical to those in [@bl2000]. Theorem \[ourDynamicBreaker\] is all we need to describe Breaker’s strategy for isolating a vertex $v$ from being in a copy of $K_r$. Let $r\ge 4$, let $C$ be a large enough constant, let $n$ be an integer and let $b\ge Cn^{2/(r+2)}$. Let $v$ be a fixed vertex of $K_n$. We show that Breaker has a strategy in the $b$-biased Maker–Breaker game played on the edge set of $K_n$ to prevent Maker from claiming a copy of $K_r$ that contains the vertex $v$. Consequently, Maker’s graph does not contain a $K_r$-factor. Before we present the strategy of Breaker we describe an auxiliary game that Breaker simulates in parallel. Let $T$ be a set of size $2n/b$, disjoint from $V(K_n)$. By , if $b/2\ge C'(2n/b)^{1/m(K_{r-1})}$ then Breaker has a winning strategy ${{\mathcal S}}$ in the dynamic $(b/2)$-biased $K_{r-1}$-game played on $K_T$, the complete graph on the vertex set $T$. Equivalently, $b\ge Cn^{2/(r+2)}$ for suitable $C$. We now describe the strategy of Breaker in the $b$-biased Maker–Breaker game played on the edge set of $K_n$. Suppose that $i \ge 0$ rounds have been played already. Let $M$ and $B$ denote the graphs formed by Maker’s edges and by Breaker’s edges, respectively (we suppress dependence on $i$ for clarity of presentation). Breaker maintains the property that every vertex $w\in N_M(v)$ has a (unique) corresponding vertex $t_w\in T$ such that an edge $uw$ in $N_M(v)$ belongs to Maker’s (Breaker’s) graph if and only if $t_ut_w$ belongs to Maker’s (Breaker’s) graph in the auxiliary dynamic $b$-biased $K_{r-1}$-game played on $K_T$. Clearly, this is the case before the first round of the game, and we show that Breaker can maintain such a correspondence throughout the game. Set $T_{i} = \{t_w:w\in N_M(v)\}$. Let $xy$ denote the edge that Maker claims in round $i+1$. Then Breaker claims up to $b/2$ edges incident to $v$ including $xv$ or $yv$ if those edges are not claimed yet by either of the players. If Breaker has claimed $b' < b/2$ edges and there are no more unclaimed edges incident to $v$, then he claims $b/2 - b'$ arbitrary edges (note that additional edges do not hurt Breaker). For the remaining $b/2$ edges in round $i+1$ we distinguish between three cases (where the latter two are similar). In Case 1, assume that $x\not\in N_M(v) \cup \{v\}$ or $y\not\in N_M(v) \cup \{v\}$. Then Breaker claims $b/2$ arbitrary edges. In Case 2.1, assume that $x=v$ (the case $y=v$ is analogous). Let $t\in T{\setminus}T_{i}$ and set $t_y= t$. In the auxiliary dynamic $K_{r-1}$-game, Breaker pretends that (the auxiliary) Maker plays according to Option (b) and adds the elements $\{t_y t_w: w\in N_M(v)\}$ to the board (recall that the vertices in the hypergraph corresponding to that game are the edges of $K_T$). In Case 2.2, assume that $x,y\in N_M(v)$. Then Breaker pretends that in the auxiliary dynamic $K_{r-1}$-game Maker plays according to Option (a) and claims the edge $t_xt_y$. In either of Case 2.1 or 2.2, the strategy ${{\mathcal S}}$ in the auxiliary game gives $b/2$ edges $e_1,\ldots, e_{b/2} \in E(K_T)$ for Breaker to claim in the auxiliary board. Let $f_1,\ldots,f_{b/2}$ be the corresponding edges in $N_M(v)$, that is $f_i$ is the edge with endpoints $w_i$ and $u_i$ such that $e_i$ has endpoints $t_{w_i}$ and $t_{u_i}$. Breaker then claims $e_1,\ldots, e_{b/2}$ in the auxiliary game and $f_1,\ldots,f_{b/2}$ in the real game. We claim that this is indeed a winning strategy. First note that $|N_M(v)|\le 2n/b$ since Breaker claims $b/2$ of the $n-1$ total edges incident to $v$ in every round. Thus, the set $T$ is large enough so that Breaker can indeed maintain an injective map $w\mapsto t_w$ for $w\in N_M(v)$. Furthermore, it is clear from the strategy description that a Maker/Breaker edge in $N_M(v)$ corresponds to a Maker/Breaker edge in the auxiliary game in $T$. Finally, since ${{\mathcal S}}$ is a strategy for Breaker to prevent Maker in the auxiliary $b/2$-biased game to claim a copy of $K_{r-1}$ this implies that Breaker can indeed prevent Maker from claiming a copy of $K_{r-1}$ in $N_M(v)$, i.e. the vertex $v$ is not in a copy of $K_r$ in Maker’s graph. Proof of Theorem \[ourDynamicBreaker\] (sketch) {#proof:ourLemma5} ----------------------------------------------- The following is a dynamic-board variant of [@bl2000 Lemma 5]. We switch notation from $m$ to $p$ and from $b$ to $q$ for the bias of Maker and Breaker, respectively, to be consistent with the literature. In every dynamic-board $({{\mathcal H}},p,q)$-game Breaker has a strategy such that at the end of the game at most $(1+q) f({{\mathcal H}},p,q)$ edges of the hypergraph ${{\mathcal H}}$ have all their vertices claimed by Maker, where $f({{\mathcal H}},p,q)=\sum_{H\in E({{\mathcal H}})}(1+q)^{-|H|/p}.$ The proof is a simple adaptation of the potential function technique as introduced by Erdős and Selfridge [@es1973] that was generalised by Beck [@b1982] to biased Maker–Breaker games. We are unaware of such a dynamical-board variant thus the full proof follows. We follow notation and strategy of the proof of [@BeckBook Theorem 20.1]. Let ${{\mathcal H}}$, $p$, $q$ be as in the lemma and let $\mu$ be defined by $1+\mu = (1+q)^{1/p}$. Given two disjoint subsets $M$ and $B$ of the board $V=V({{\mathcal H}})$ and an element $z\in V$ set $$\begin{aligned} \Phi(M,B) &= \sum_{H\in {{\mathcal H}}: H\cap B =\emptyset} (1+\mu)^{-|H{\setminus}M|},\text{ and}\\ \Phi(M,B,z) &= \sum_{z\in H\in {{\mathcal H}}: H\cap B =\emptyset} (1+\mu)^{-|H{\setminus}M|}\end{aligned}$$ and note straight away the following inequalities: $$\begin{aligned} \Phi(M\cup\{e\},B,z)&\le (1+\mu)\Phi(M,B,z),{\label{T1}}\\ \Phi(M,B\cup\{e\},z)&\le \Phi(M,B,z).{\label{T2}}.\end{aligned}$$ For integers $r$ and $j$, let $b_r^{(j)}$ be the $j^{\mathrm{th}}$ element that Breaker picks in round $r$, and let $m_r^{(j)}$ be the $j^{\mathrm{th}}$ element that Maker picks in round $r$ if she decides to play according to Option (a) and pick elements in $V_{r-1}$ rather than enlarging $V_r$ (Option (b)). Furthermore, let $M_r$ and $B_r$ be the set of all elements of Maker and of Breaker, respectively, [*after*]{} round $r$, and let $M_{r,j}=M_r\cup \{m_{r+1}^{(1)},\ldots,m_{r+1}^{(j)}\}$ and $B_{r,j}=B_r\cup \{b_{r+1}^{(1)},\ldots,b_{r+1}^{(j)}\}$ (assuming that Maker/Breaker has claimed at least $j$ elements in round $r+1$). We now describe Breaker’s strategy in round $r$. If there are less than $q$ unclaimed elements in $V_{r+1}$, then Breaker claims all of them. Otherwise, for every $1\leq j\le q$, sequentially, Breaker calculates $\Phi(M_r,B_{r-1,j-1},z)$ for every unclaimed element $z\in V_r{\setminus}(M_r\cup B_{r-1,j-1})$ and claims the element $b_r^{(j)}$ which maximises this expression. Note that here we chose the element $b_r^{(j)}$ in $V_r$, and not in the whole board $V$. If, for some $j$, there are no unclaimed elements, then it is the end of Breaker’s turn. The crucial part of the potential function technique in positional games is to show that the [*potential*]{} $\Phi(M_{r+1},B_r)$ is decreasing (if evaluated after Makers move). But this is now straight-forward along the lines of the proof in [@BeckBook]. The only thing we have to notice is that in round $r+1$, if Maker chooses to claim elements in $V_r$, then their choices are on the same sub-board where Breaker claimed their elements in round $r$. For all $r\ge 1$, $\Phi(M_{r+1},B_r)\le \Phi(M_r,B_{r-1})$. If Maker has played according to Option (b) in round $r+1$, then $\Phi(M_{r+1}, B_r) = \Phi(M_r, B_r) \le \Phi(M_r, B_{r-1})$, where the inequality follows from . Therefore, if Breaker was not able to claim $q$ elements in round $r$, then in round $r+1$ Maker is forced to play Option (b) and the claim follows. For the rest of the proof we can assume that Breaker is able to claim $q$ elements in round $r$. Without loss of generality, we can also assume that Maker claims $p$ elements in round $r+1$ (claiming fewer than $p$ elements only makes it easier for the desired inequality to hold). Let us denote these elements by $b_r^{(1)}, \ldots, b_r^{(q)}$ and $m_{r+1}^{(1)}, \ldots, m_{r+1}^{(p)}$, respectively. We first note that $\Phi(M_r,B_{r-1,j+1})=\Phi(M_r,B_{r-1,j})-\Phi(M_r,B_{r-1,j},b_r^{(j+1)})$ for all $0\le j<q$, and $\Phi(M_{r,j+1},B_{r})=\Phi(M_{r,j},B_{r})+\mu\Phi(M_{r,j},B_{r},m_{r+1}^{(j+1)})$ for all $0\le j<p$. Hence, $$\begin{aligned} \Phi(M_{r+1},B_r)&= \Phi(M_r,B_{r-1}) - \sum_{j=1}^q \Phi(M_r,B_{r-1,j-1},b_r^{(j)}) +\mu \sum_{j=1}^p \Phi(M_{r,j-1},B_{r},m_{r+1}^{(j)}).\end{aligned}$$ It suffices to show $$\begin{aligned} {\label{aux343}} &\mu \sum_{j=1}^p \Phi(M_{r,j-1},B_{r},m_{r+1}^{(j)}) \le \sum_{j=1}^q \Phi(M_r,B_{r-1,j-1},b_r^{(j)}).\end{aligned}$$ First note that for all $1\le j\le p$, $$\Phi(M_{r,j-1},B_{r},m_{r+1}^{(j)})\le (1+\mu)^{j-1}\Phi(M_{r},B_{r},m_{r+1}^{(j)})\le (1+\mu)^{j-1}\Phi(M_{r},B_{r-1,q-1},m_{r+1}^{(j)}),$$ where the first inequality follows from  and the second from . Furthermore, the element $b_r^{(q)}$ is chosen by Breaker [*before*]{} Maker claims any of $m_{r+1}^{(j)}$ and it is chosen to maximise the expression $\Phi(M_{r},B_{r-1,q-1},z)$ [*over all unclaimed $z\in V_{r+1}$*]{} (note that Maker sets $V_{r+1}$ after her move in round $r$). Hence, we deduce $$\begin{aligned} \Phi(M_{r,j-1},B_{r},m_{r+1}^{(j)})\le (1+\mu)^{j-1}\Phi(M_{r},B_{r-1,q-1},b_r^{(q)}), \end{aligned}$$ which readily implies that $$\begin{aligned} {\label{aux346}} \mu \sum_{j=1}^p \Phi(M_{r,j-1},B_{r},m_{r+1}^{(j)}) \le \mu \Phi(M_{r},B_{r-1,q-1},b_r^{(q)}) \sum_{j=1}^p (1+\mu)^{j-1}.\end{aligned}$$ To bound the right hand side of  we note that for all $1\le j < q$, $$\begin{aligned} &\Phi(M_{r},B_{r-1,j-1},b_r^{(j)})\ge \Phi(M_{r},B_{r-1,j-1},b_r^{(q)})\ge \Phi(M_{r},B_{r-1,q-1},b_r^{(q)}),\end{aligned}$$ where the first inequality follows again by $b_r^{(j)}$ maximising $\Phi(M_{r},B_{r-1,j-1},z)$ among all unclaimed elements $z$ in $V_{r+1}$ (and $b_r^{(j)}$ is chosen before $b_r^{(q)}$), and the second inequality follows from . This implies that $$\sum_{j=1}^q \Phi(M_{r},B_{r-1,j-1},b_r^{(j)})\ge q \Phi(M_r,B_{r-1,q-1},b_r^{(q)}).$$ This and  imply that the inequality in  follows from $$\mu \Phi(M_{r},B_{r-1,q-1},b_r^{(q)}) \sum_{j=1}^p (1+\mu)^{j-1} \le q \Phi(M_r,B_{r-1,q-1},b_r^{(q)}),$$ which is true since $\Phi(M_{r},B_{r-1,q-1},b_r^{(q)})\ge 0$ and $(1 + \mu)^p - 1 = q$. To finish the proof of note that $\Phi (\emptyset,\emptyset) = f({{\mathcal H}},p,q)$ and that after Maker’s move in the first round we have $\Phi(M_1,B_0) = \Phi(M_1,\emptyset) \le (1+\mu)^p f({{\mathcal H}},p,q) = (1 + q) f({{\mathcal H}}, p ,q)$. By the above claim, the function $f(M,B)$ is decreasing if evaluated after Maker’s move. Hence, at the end of the game, we have $\Phi(M,B)\le \Phi(M_1,B_0)\le (1+q) f({{\mathcal H}},p,q)$, where now $M$ and $B$ denote the final set of elements that Maker and Breaker claimed, respectively. On the other hand, if Maker occupies all elements of an edge $H\in E({{\mathcal H}})$ (at any point of the game), then the additive contribution to $\Phi(M,B)$ is 1 for each such set. Thus, if $(1+q)f({{\mathcal H}},p,q)<k$ then Breaker has a strategy in the dynamic Maker–Breaker game so that at the end of the game Maker fully occupies fewer than $k$ sets $H\in E({{\mathcal H}})$. A complete proof of including all details would involve repeating the proof of [@bl2000 Theorem 1 (ii)] and replacing [@bl2000 Lemma 5] by wherever it occurs. Instead, we give an overview of the argument in [@bl2000] and describe where the changes are needed. We leave out calculations that are identical. Without loss of generality we may assume that $H$ is such that $m_2(H)$ achieves maximum for $H' = H$, i.e. $m_2(H) = (e(H) - 1)/(v(H) - 2)$. If this is not the case, then Breaker can choose $H' \subseteq H$ which determines $m_2(H)$ (which might not be unique), and prevent Maker from creating $H'$. Call such an $H$ [*$m_2$-maximal.*]{} Bednarska and Łuczak first identify the dangerous structures that Maker can create during a $b$-biased $H$-game on $K_n$. We say that a graph $F$ is an $\bar H$-graph if it contains two vertices $v$, $w$ such that $F+vw$ is isomorphic to $H$. We shall write $F^{vw}$ to specify such vertices in $F$. Let ${{\mathcal F}}=\{F_1^{v_1w_1},\ldots,F_t^{v_tw_t}\}$ be a family of different $\bar H$-graphs, whose vertex sets may intersect. Then ${{\mathcal F}}$ is called a [*$t$-fan*]{} if $|\bigcap_{i=1}^t V(F_i^{v_iw_i})|\ge 2$, and it is called a [*$t$-flower*]{} if $|\bigcap_{i=1}^t V(F_i^{v_iw_i})|\ge 3$. Furthermore, a $t$-fan ${{\mathcal F}}$ is called [*simple*]{} if $|\bigcap_{i=1}^t V(F_i^{v_iw_i})|=2$. If at some point during the game Maker’s graph contains some $\bar H$-graph $F^{vw}$ such that the pair $\{v,w\}$ has not been claimed by Breaker, then $F^{vw}$ is [*dangerous*]{}. Similarly, a $t$-fan ${{\mathcal F}}$ is called [*dangerous*]{} if its elements are dangerous subgraphs of Maker’s graph. The following is the dynamic-board variant of [@bl2000 Lemma 9]. For every $m_2$-maximal graph $H$ that contains a cycle there exist positive constants $C$ and $n_0$ such that for every $n\ge n_0$ and $q\ge Cn^{1/m_2(H)}$ Breaker has a strategy such that at no stage of the dynamic $q$-biased $H$-game on $K_n$ Maker’s graph contains a dangerous $q$-fan. Before sketching the proof of this lemma let us briefly explain how it implies . When $H$ is a forest the argument is identical to the one given in [@bl2000]: The requirement $b\ge 2n^{1/m_2(H)}$ implies that Breaker has a simple strategy in either case of $H$ containing a path on three vertices or consisting of disjoint edges. When $H$ contains a cycle let $b\ge 2Cn^{1/m_2(H)}$ where $C$ is the constant given by . Breaker’s strategy in the dynamic $b$-based $H$-game played on $K_n$ is as follows. In every round, Breaker claims $b/2$ edges to follow the strategy of and thus prevents Maker from building a dangerous $(b/2)$-fan. The remaining $b/2$ choices are used to block the pairs $v_1w_1,\ldots,v_tw_t$ of a dangerous $t$-fan $\{F_1^{v_1w_1},\ldots,F_t^{v_tw_t}\}$, if there is one for some $t\le b/2$. Note that this is indeed a winning strategy. Suppose that Maker can claim a copy of $H$, say in round $r$ of the game. Then Maker has created a dangerous $t$-fan for some $t\ge 1$ in round $r-1$ (or earlier). Since Maker plays with bias 1, she cannot create more than one dangerous $t$-fan per round. By , $t\le b/2$, thus Breaker would block such a dangerous $t$-fan in round $r-1$. We now sketch the proof of . The strategy for Breaker is based, yet again, on two strategies that are followed each with bias $q/2$. The following two lemmata encapsulate these strategies, they are the dynamic-board variants of Lemma 6 and Lemma 7 of [@bl2000], respectively. For every $m_2$-maximal graph $H$ that contains a cycle there exist positive constants $C_1$, $n_1$ and $\delta <1$ such that for every $n\ge n_1$ and $q\ge C_1n^{1/m_2(H)}$ Breaker has a strategy such that at each moment of the dynamic $q$-biased $H$-game on $K_n$ there are no dangerous $s$-flowers in Maker’s graph, where $s=q^{1-\delta}$. For every $m_2$-maximal graph $H$ that contains a cycle and every positive $\delta <1$, there exist constants $C_2$ and $n_2$ such that for every $n\ge n_2$ and $q\ge C_2n^{1/m_2(H)}$ Breaker has a strategy in the dynamic $q$-biased $H$-game on $K_n$ which does not allow Maker to build $\frac12 \binom{q}{t}$ simple $t$-fans, where $t=q^{\delta/3}.$ To prove , let $C_1$, $n_1$, $\delta$ be given as in , let $C_2$ and $n_2$ be as in , and set $C=2\max\{C_1,C_2\}$ and let $n$ be large enough. Let ${{\mathcal S}}_1$ be the strategy given by for the dynamic $(q/2)$-biased $H$-game on $K_n$, and let ${{\mathcal S}}_2$ be the strategy given by for the dynamic $(q/2)$-biased $H$-game on $K_n$. Then, using $q/2$ edges to follow strategy ${{\mathcal S}}_1$ and $q/2$ edges to follow strategy ${{\mathcal S}}_2$, Breaker can ensure that at no point during the dynamic $q$-biased $H$-game on $K_n$ Maker’s graph contains a dangerous $s$-flower with $s= (q/2)^{1-\delta}$, nor $\frac12 \binom{q/2}{t}$ simple $t$-fans with $t=(q/2)^{\delta/3}.$ The reasoning why this implies that Maker’s graph never contains a dangerous $q$-fan is exactly the same as in [@bl2000], see the last paragraph of the proof of Lemma 9 therein. Finally, we explain the proofs of . For , define a [*$t$-cluster*]{} to be a graph consisting of $t$ copies of $H$ that intersect in at least 3 vertices (the difference to a $t$-flower is that here the building blocks are copies of $H$ rather than $\bar H$-graphs). Let ${{\mathcal H}}_1$ be the collection of all edge sets on $K_n$ that form a copy of a [*$t$-cluster*]{}, where $t=t(H)$ is a large enough constant. Further, let $q_1 = q^{1-\delta_1/2}$ where $\delta_1=\delta_1(H)$ is a suitable constant as defined at the beginning of the proof of [@bl2000 Lemma 6]. The choice of parameters is identical to those in [@bl2000 Lemma 6]. Hence, calculations identical to those in [@bl2000 Lemma 6] show that $(q+1)f({{\mathcal H}}_1,1,q_1)<1$. It follows from that Breaker has a strategy in the dynamic-board $({{\mathcal H}}_1,1,q)$-game to prevent Maker from claiming all edges of a $t$-cluster. We pause for a psychological note: This strategy does not forbid the creation of one copy of $H$ per se, which may seem contradictory to the final goal. For the strategy of this is irrelevant though and shall not concern us. It then follows along the lines of the last two paragraphs of [@bl2000 Lemma 6] that by preventing such a $t$-cluster for sufficiently large $t$ (depending only on $H$), Breaker can also prevent that, in the dynamic $q$-biased $H$-game on $K_n$, Maker claims the edge set of a dangerous $s$-flower, for $s=q^{1-\delta}$, $\delta = \delta_1/4$, and $q\ge n^{1/m_2(H)}.$ For , let ${{\mathcal H}}_2$ consist of the edge sets of simple $t$-fans in $K_n$, for $t$ as in the lemma statement. Following the calculations in the proof of [@bl2000 Lemma 7] we find that $(q+1)f({{\mathcal H}}_2,1,q)<\frac12\binom{q}{t}$ if $q\ge C_2n^{1/m_2(H)}$ where $C_2$ and $n$ are large enough. It follows that Breaker has a strategy in the dynamic-board $({{\mathcal H}}_2,1,q)$-game such that at the end of the game at most $\frac12\binom{q}{t}$ of all $t$-fans are fully occupied by Maker. [^1]: School of Mathematics and Statistics, UNSW Sydney, NSW 2052, Australia. Email: [[email protected].]{} Supported by the Australian research council (DE170100789 and DP180103684). [^2]: Department of Mathematics, ETH Zurich, Switzerland. Email: [[email protected].]{} Supported in part by SNSF grant 200021-175573. [^3]: All asymptotic statements refer to $n$, the number of vertices, tending to $\infty$.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We present new  photometry of 22 stars that host transiting planets, 19 of which were discovered by the WASP survey. We use these data together with 2MASS JHK$_{\rm S}$ photometry to estimate the effective temperature of these stars using the infrared flux method. We find that the effective temperature estimates for stars discovered by the WASP survey based on the analysis of spectra are reliable to better than their quoted uncertainties.' author: - | P.F.L. Maxted$^1$, Koen, C.$^2$, Smalley, B.$^1$\ $^1$Astrophysics Group, Keele University, Keele, Staffordshire ST5 5BG, United Kingdom\ $^2$Department of Statistics, University of the Western Cape, Private Bag X17, Bellville 7535, South Africa bibliography: - 'wasp.bib' date: Submitted 2011 title: 'UBV(RI)$_{\rm C}$ photometry of transiting planet host stars' --- \[firstpage\] planetary systems – techniques: spectroscopic, photometric Introduction ============ The transit of a planet across the face of its host star provides us with the opportunity to measure the properties of the planet in great detail. Essential to exploiting this opportunity is a good understanding of the host star itself. For example, a combined analysis of the transit lightcurve together with the spectroscopic orbit of the host star leads directly to a measurement of the host star density, $\rho_{\star}$, and the surface gravity of the planet [@2007MNRAS.379L..11S; @2003ApJ...585.1038S]. To estimate the mass and radius of the planet, an additional constraint is needed. The details of how the additional constraint is applied varies between different groups, but they generally share the common feature that an analysis of the spectrum is used to estimate the stellar effective temperature, , stellar surface gravity, , and metallicity, \[Fe/H\][^1], and these are combined with the estimate of $\rho_{\star}$ to estimate the mass and radius of the star using either an empirical calibration or stellar models [@2010MNRAS.408.1689S; @2010A+A...516A..33E; @2010ApJ...710.1724B]. The radius and mass of the planet then follow directly from the observed depth of the transit and Kepler’s Law, respectively. Irrespective of the details, it is clear that the accurate characterisation of the host star is essential for an accurate understanding of the planets that orbit it. Most transiting planets discovered to-date have been found using wide-angle, ground-based photometric surveys such as WASP [@2006PASP..118.1407P] and HATNet [@2004PASP..116..266B]. These surveys target stars with visual magnitudes in the approximate range 8.5–13. One obstacle to the accurate characterisation of these stars is the poor quality of the optical photometry that is generally available for stars of this brightness. Accurate photometry for bright stars is available at infrared wavelengths (JHK$_{\rm S}$) across the entire sky from the 2MASS survey [@2006AJ....131.1163S] and in the southern hemisphere from the DENIS survey [@2005yCat.2263....0T]. The DENIS survey extends the available photometry to the I-band. In the northern hemisphere the CMC14 catalogue provides r’-band photometry . In the optical regime, BV photometry is available for stars brighter than $V\approx12$ from the Tycho-2 catalogue , although this catalogue is only complete to $V\approx11$ and the photometric precision deteriorates rapidly for $V\ga9.5$. Accurate optical photometry provides flux measurements around the peak of the spectral energy distribution (SED) for solar-type stars, unlike infrared photometry that samples the Rayleigh-Jeans tail of the SED. Optical photometry is also sensitive to reddening and metalicity, both of which affect the blue end of the SED much more than the red end. Both these effects need to be accurately accounted for if the distance to the star is to be estimated from the photometric properties of the star, e.g., for kinematical studies. The combination of accurate optical and infrared photometry also makes it possible to make a robust and accurate estimate of the star’s effective temperature using the infrared flux method. (@1979MNRAS.188..847B). Accurate optical photometry is also useful for planning follow-up observations, e.g., for estimating optimum exposure times. In this paper we present new, high quality photoelectric optical photometry for 22 planet-host stars, mostly WASP discoveries in the southern hemisphere. We use this photometry to make an independent check on the accuracy of the effective temperature estimates published for these stars based on the analysis of their spectra. Observations ============ Observations were obtained with the SAAO 0.5-m telescope and modular photometer [@1988MNSSA..47...69K]. This is a very stable and well understood instrumental setup for obtaining standardised  photometry [@2005ARA+A..43..293B]. Observations were obtained in dark sky conditions over the course of two observing runs, 2010 September 9-12 and 2011 March 5-12. Reduction of the instrumental magnitudes to the standard photometric system defined by the E-region standards of [@1989SAAOC..13....1M] followed the methods described in the appendix of [@1988MNSSA..47...69K]. The observations are presented in Table \[phottable\]. Star $N$ ------------------- ------------ ------- ----------- ------- ------- ----- WASP-2 11.788 0.897 0.604 0.488 0.932 2 WASP-4 12.463 0.783 0.362 0.405 0.782 2 WASP-5 12.136 0.705 0.281 0.380 0.727 2 WASP-7 9.483 0.460 0.015 0.267 0.528 2 WASP-8 9.773 0.747 0.369 0.404 0.788 2 WASP-15 10.918 0.495 $ 0.008 $ 0.295 0.579 3 $\pm$0.003 0.001 0.008 0.013 0.012 WASP-16 11.309 0.741 0.283 0.380 0.763 2 WASP-17 11.500 0.496 0.040 0.268 0.585 2 WASP-18 9.273 0.484 0.013 0.278 0.548 2 WASP-19 12.312 0.785 $ 0.398$ 0.425 0.812 4 $\pm$0.017 0.004 0.052 0.007 0.009 WASP-22 11.708 0.603 0.139 0.323 0.638 2 WASP-25 11.848 0.727 0.253 0.380 0.764 2 WASP-26 11.099 0.621 0.145 0.344 0.690 2 WASP-28 12.148 0.596 0.033 0.329 0.690 2 WASP-29 11.207 1.087 1.061 0.620 1.119 2 WASP-31 11.937 0.513 $-0.009$ 0.297 0.593 3 $\pm$0.006 0.017 0.004 0.016 0.004 WASP-34 10.366 0.684 0.224 0.364 0.716 3 $\pm$0.012 0.003 0.011 0.013 0.015 WASP-37$^{\rm a}$ 12.717 0.628 0.022 0.337 0.699 2 WASP-39 12.100 0.803 0.370 0.426 0.852 3 $\pm$0.012 0.021 0.040 0.006 0.021 HAT-P-24 11.754 0.462 $-0.017$ 0.260 0.518 2 HAT-P-27/ 12.163 0.909 0.645 0.467 0.892 3 WASP-40 $\pm$0.011 0.010 0.009 0.009 0.016 CoRoT-7$^{\rm b}$ 11.718 0.849 $ 0.915$ 0.437 0.827 3 $\pm$0.003 0.026 0.012 0.027 : Photometry of 22 transiting planet host stars. $N$ is the number of observations obtained. Standard errors are given for stars with 3 or more observations.[]{data-label="phottable"} $^{\rm a}$One discrepant measurement ignored $^{\rm b}$Only one reliable U-band measurement was obtained. Analysis ======== We estimate the effective temperature of these stars using a simplified version of the infrared flux method (IRFM; @1979MNRAS.188..847B). The essence of this method is to find values of the effective temperature, , and angular diameter, $\theta$, for which stellar atmosphere models simultaneously satisfy the observed value of the bolometric flux at the Earth, , and the observed flux at infrared wavelengths. Photometry covering the peak of the SED is important for an accurate estimate of in solar-type stars. Some type of interpolation scheme is required to enable the integration of an SED that is only sparsely sampled by broadband photometry. In our method we use numerical integration of the best-fitting model SED from a grid of stellar atmosphere models. The infrared flux of a solar-type star predicted by stellar models is insensitive to parameters such as surface gravity, metalicity and reddening or to the details of the model, so the value of  is almost model independent. In principle, the value of  derived by our method could be improved by recalculating or interpolating the model SED used in the integration of  and iterating this process. In practice we find that further refinement of our estimate is not required. An estimate of the interstellar reddening is required for an accurate comparison of the model SED to the observed fluxes. In practice, we find that the effect of interstellar reddening is negligible for the majority of the stars we have studied. We use the data provided by the 2MASS survey [@2006AJ....131.1163S] to obtain the JHK$_{\rm S}$ magnitudes for our targets. For stars where we have access to the spectra, we have measured the equivalent width of the interstellar absorption features due to sodium near 589nm and used the calibration of [@1997A+A...318..269M] to convert this to an estimate of the interstellar reddening, E(B$-$V). We estimate the bolometric flux from the star by integrating the best-fitting model spectral energy distribution (SED) from grid of models at 250K intervals in  from [@1993KurCD..13.....K]. We use least-squares fitting of the model fluxes integrated over the appropriate bandpasses for all available optical and infrared magnitudes to determine the best-fitting model SED. For stars with 3 or more observations we use the standard errors on the means to determine the weights of the  data in the fit. We use linear regression on the apparent magnitudes and standard errors in each bandpass for these stars to determine a relationship between apparent magnitude and standard error for each bandpass, and then use this to assign a weight to the data for stars with fewer than 3 observations. For the conversion of magnitudes to fluxes we use the zero-point values from @1979PASP...91..589B and @2003AJ....126.1090C. We include the effect of interstellar reddening in the least-squares fit of the grid of model SEDs for stars where an estimate of E(B$-$V) from the spectrum is available. The spectral energy distributions were de-reddened using the analytical extinction expressions from @1983MNRAS.203..301H. For the two stars for which we do not have access to the spectra we assume E(B$-$V)$=0$. For a typical star with ${\mbox{${\rm T}_{\rm eff}$}}\approx6000$K and E(B$-$V)$\approx$0.01, neglecting the reddening results in an over-estimate of  by about 30K. The effective temperature of the star is then estimated using each of the available J, H and K$_{\rm S}$ magnitudes independently and the weighted average of these results is taken as the final value. We refer to this estimate of the effective temperature as . The values of E(B$-$V) and  are given in Table \[tefftable\] together with published estimates of the effective temperature for each star based on the analysis of the spectrum, . Star E(B$-$V) Spectrograph Reference ---------- ---------- --------------- ---------------- -------------------- ---------------- ------------------------ WASP-2 0.02 5110 $\pm$ 60 5200 $\pm$ 200 $\sim 0.0$ SOPHIE [@2007MNRAS.375..951C] 5150 $\pm$  80 $-0.08$ $\pm$ 0.08 HARPS [@2010A+A...524A..25T] WASP-4 0.00 5540 $\pm$ 55 5500 $\pm$ 150 $ 0.0 $ $\pm$ 0.2 CORALIE [@2008ApJ...675L.113W] 5500 $\pm$ 100 $-0.03$ $\pm$ 0.09 UVES [@2009A+A...496..259G] WASP-5 0.00 5770 $\pm$ 65 5700 $\pm$ 100 $ 0.0 $ $\pm$ 0.2 CORALIE [@2008MNRAS.387L...4A] 5700 $\pm$ 100 $ 0.09$ $\pm$ 0.09 UVES [@2009A+A...496..259G] WASP-7 0.00 6520 $\pm$ 70 6400 $\pm$ 100 $ 0.0 $ $\pm$ 0.1 CORALIE [@2009ApJ...690L..89H] WASP-8 0.00 5570 $\pm$ 85 5600 $\pm$  80 $0.17 $ $\pm$ 0.17 HARPS [@2010A+A...517L...1Q] WASP-15 0.00 6210 $\pm$ 60 6300 $\pm$ 100 $1.4 $ $\pm$ 0.1 CORALIE [@2009AJ....137.4834W] WASP-16 0.01 5550 $\pm$ 60 5700 $\pm$ 150 $0.01 $ $\pm$ 0.10 CORALIE [@2009ApJ...703..752L] WASP-17 0.05 6500 $\pm$ 75 6550 $\pm$ 100 $-0.25$ $\pm$ 0.09 CORALIE [@2010ApJ...709..159A] 6650 $\pm$  80 $-0.19$ $\pm$ 0.09 HARPS [@2010A+A...524A..25T] WASP-18 0.00 6455 $\pm$ 70 6400 $\pm$ 100 $0.00 $ $\pm$ 0.09 HARPS [@2009Natur.460.1098H] WASP-19 0.00 5440 $\pm$ 60 5500 $\pm$ 100 $0.02 $ $\pm$ 0.09 CORALIE [@2010ApJ...708..224H] WASP-22 0.01 6020 $\pm$ 50 6000 $\pm$ 100 $0.05 $ $\pm$ 0.08 HARPS [@2010AJ....140.2007M] WASP-25 0.00 5615 $\pm$ 55 5750 $\pm$ 100 $-0.05$ $\pm$ 0.10 CORALIE [@2011MNRAS.410.1631E] WASP-26 0.01 6015 $\pm$ 55 5950 $\pm$ 100 $-0.02$ $\pm$ 0.09 CORALIE [@2010A+A...520A..56S] WASP-28 0.04 6190 $\pm$ 60 6100 $\pm$ 150 $-0.29$ $\pm$ 0.10 CORALIE [@West2010] WASP-29 0.00 4875 $\pm$ 65 4800 $\pm$ 150 $ 0.11$ $\pm$ 0.14 CORALIE [@2010ApJ...723L..60H] WASP-31 0.00 6175 $\pm$ 70 6250 $\pm$ 150 $-0.29$ $\pm$ 0.11 CORALIE [@2010arXiv1011.5882A] 6300 $\pm$ 100 $-0.20$ $\pm$ 0.09 HARPS [@2010arXiv1011.5882A] WASP-34 0.00 5695 $\pm$ 65 5700 $\pm$ 100 $-0.02$ $\pm$ 0.10 CORALIE [@2011A+A...526A.130S] WASP-37 0.05 5940 $\pm$ 55 5800 $\pm$ 150 $-0.40$ $\pm$ 0.12 CORALIE+SOPHIE [@2011AJ....141....8S] WASP-39 0.04 5460 $\pm$ 55 5400 $\pm$ 150 $-0.12$ $\pm$ 0.10 CORALIE [@2011arXiv1102.1375F] HAT-P-24 6330 $\pm$ 65 6373 $\pm$  80 $-0.16$ $\pm$ 0.08 HIRES [@2010ApJ...725.2017K] HAT-P-27 5175 $\pm$ 70 5300 $\pm$  90 $0.29 $ $\pm$ 0.10 HIRES [@2011arXiv1101.3511B] =WASP-40 0.01 5200 $\pm$ 150 $0.14 $ $\pm$ 0.11 CORALIE+SOPHIE [@2011arXiv1101.4643A] CoRoT-7 5240 $\pm$ 55 5275 $\pm$  75 $0.03 $ $\pm$ 0.06 UVES [@2009A+A...506..287L] 5250 $\pm$  60 $0.12 $ $\pm$ 0.06 HARPS+UVES [@2010A+A...519A..51B] Discussion ========== For all 25 stars from the WASP survey a method based on an analysis of the  line has been used to measure  [@2011A+A...526A.130S]. There may be systematic errors in these  estimates due to instrumental effects such as scattered light, the normalization of the spectra, etc., as well as systematic errors in the model stellar atmospheres used to analyse the  line. Similar issues will affect the  estimates for stars from other sources. @2010MNRAS.405.1907B have compared  estimates for 23 nearby solar-type stars to  determined directly from interferometric angular diameters. They find that their  estimates are too hot by $40\pm20$K. report  estimates for GK-type stars using the IRFM based on BV(RI)$_{\rm C}$JHK$_{\rm S}$ photometry similar to ours. They argue that the main source of systematic error in the IRFM method is the conversion of magnitudes to fluxes, i.e. the zero-point of the optical and infrared magnitude scales. They estimate that a 2per cent error in the zero-point results in an error of approximately 40K in the . Although our implementation of the IRFM is not the same in detail as that of this estimate of the systematic error inherent in the method applies equally to our method. A comparison of the effective temperature estimates  and  is shown in Fig. \[teffplot\]. The mean value of ${\mbox{${\rm T}_{\rm spec}$}}-{\mbox{${\rm T}_{\rm IRFM}$}}$ is $(-13\pm17)$K. It can be seen that the agreement between the two temperature scales is very good. The $\chi^2$ value for the 1:1 relation shown in Fig. \[teffplot\] is 12.2 for 29 degrees of freedom. This level of agreement is much better than would be expected given the standard errors quoted for  and . However, the uncertainties quoted for  for all the WASP stars (25 of the 29  values) include some estimate of the systematic error in the estimate and these uncertainties are generally quoted to the nearest 50K, e.g., $\pm 100$K. If we assume that any systematic error in  is about 40K, this suggests that the systematic error in for WASP stars is likely to be $\la 50$K, i.e., similar to the level of systematic error found by @2010MNRAS.405.1907B for their estimates. Another quantity sometimes estimated from  is (B$-$V)$_0$, the intrinsic B$-$V colour. This is used to calculate the chromospheric activity index $\log R'_{\rm HK}$ [@1984ApJ...279..763N]. For WASP stars the calibration of [@2008oasp.book.....G] is used to estimate B$-$V from  (e.g., @2011PASP..123..547M). For the 19 WASP stars here, we find that this estimate is accurate to better than 0.03 magnitudes. This corresponds to an additional uncertainty of about 0.05 in the value of $\log R'_{\rm HK}$, which is small compared to intrinsic decadal variability in this quantity for these types of stars. ![Comparision of effective temperature estimates using the IRFM and from spectroscopy. \[teffplot\] ](teff.eps){width="46.00000%"} Conclusions =========== We have used  photometry combined with published infrared photometry to show that the effective temperature estimates for planet host stars discovered by the WASP survey based an the analysis of the spectrum are consistent with the infrared flux method effective temperature scale to better than the quoted standard errors, typically $\pm 100$K. Acknowledgments {#acknowledgments .unnumbered} =============== We thank Francois van Wyk for obtaining a few of the measurements in Table 1. \[lastpage\] [^1]: \[Fe/H\] is the iron abundance relative to the Sun, metallicity is normally estimated by assuming that the abundances of other elements scales with the iron abundance
{ "pile_set_name": "ArXiv" }
--- abstract: 'We give a one-dimensional quantum cellular automaton (QCA) capable of simulating all others. By this we mean that the initial configuration and the local transition rule of any one-dimensional QCA can be encoded within the initial configuration of the universal QCA. Several steps of the universal QCA will then correspond to one step of the simulated QCA. The simulation preserves the topology in the sense that each cell of the simulated QCA is encoded as a group of adjacent cells in the universal QCA. The encoding is linear and hence does not carry any of the cost of the computation. We do this in two flavours: a weak one which requires an infinite but periodic initial configuration and a strong one which needs only a finite initial configuration.' author: - Pablo Arrighi - Renan Fargetton - Zizhu Wang title: ' Intrinsically universal one-dimensional quantum cellular automata in two flavours ' --- Introduction ============ In this section we give a quick overview of QCA and explain why they matter. We then move on to explain why the notion of intrinsic simulation is so important in the CA community, we then list some related construction and provide an outline for the paper. Quantum cellular automata ------------------------- One-dimensional cellular automata (CA) consists of a line of cells, each of which may take one in a finite number of possible states. These evolve in discrete time steps according to a local rule, applied synchronously and homogeneously across space. Because they are a physics-like model of computation it seems very natural to study their quantum extensions. The flourishing research in quantum information and quantum computer science provides us with appropriate context for doing so, both in terms of the theoretical framework and potential applications. Indeed this field has already brought to light a number of theoretical results about one-dimensional quantum cellular automata (QCA). For instance Schumacher and Werner [@Schumacher] have proved that any local, translation-invariant unitary evolution over a line of finite-dimensional quantum systems takes the form of a pattern of more elementary unitary evolutions – repeated across time and space. And right from the very birth of the field with Feynman’s 1986 paper, it was hoped that QCA may prove an important path to realistic implementations of quantum computers [@Feynman] – mainly because they eliminate the need for an external, classical control and hence the principal source of decoherence. Other possible aims include providing models of distributed quantum computation, providing bridges between computer science notions and modern theoretical physics, or understanding the dynamics of some quantum physical system in discrete spacetime, i.e. from an idealized viewpoint. As we shall see in section \[definitions\] of the paper, any QCA can be put into the form given by Figure \[structure\]. For our purpose, we will take this particular space and time tiling of a scattering unitary $U$ as the actual definition of QCA. ![\[structure\] Partitioned one-dimensional QCA with scattering unitary $U$. Each line represents a quantum system, in this case a whole cell. Each square represents a scattering unitary $U$ which gets applied upon two cells. Time flows upwards.](structure.pdf) Intrinsic universality ---------------------- *The most popular cellular automaton is Conway’s ‘Game of Life’, a two-dimensional CA which has been proven to be universal for computation – in the sense that any algorithm can be encoded within its initial state and then be run by the cellular automaton’s evolution.* This was accomplished by simulating any Turing Machine (TM) within the automaton, and since Turing Machines have long been regarded as pretty much the best definition of ‘what an algorithm is’ in classical computer science, this could have meant the end of the story to many people. Yet researchers in CA have always been looking for more than just running any algorithm, seeking to run distributed algorithms in a distributed manner, model some other phenomena together with their spatial structure, or make use of the spatial parallelism which is inherent to the model – as these are the features which are modelled by CA and not by TM. And hence they have had to come back [@Albert] to the original meaning of the word ‘universal’, namely the ability for one instance of a computational model to be able to simulate all other instances of the very same computational model. Nowadays there is an impressive number results about intrinsically universal CA as reviewed for instance in [@Delorme; @Ollinger1] – i.e. results on cellular automata capable of simulating all others efficiently and directly. (Incidentally of course they also simulate those CA which are capable of simulating the TM.) Or to put things differently, most of the CA community now seems to consider that a good notion of simulation is one which preserves the topology and the parallelism of the simulated CA, in some simple and explicit fashion. In the same manner, studying QCA rather than QTM for instance means we bother about the spatial structure of things, whether for the purpose of describing a quantum protocol, modelling a quantum physical phenomena, or again taking into account the spatial parallelism inherent to the model. Hence we argue that the kind of universality we are looking for is in fact stronger than the ability to simulate any quantum Turing Machine. We seek an intrinsically universal QCA, i.e. a QCA which can simulate all others efficiently and directly. A good intuition of the notion of intrinsic simulation is given by the diagram in Figure \[UsimV\], but formal definitions follow in section \[definitions\]. In particular, we will distinguish two notions of intrinsic universality depending upon whether we allow preparation of an infinite periodic initial configuration or insist to have only finite configurations. ![Intrinsic simulation of a QCA by another.\[UsimV\] (The QCA defined by $U$ simulates the QCA defined by $V$. In this case we need two cells of the $U$-QCA in order to encode one cell of the $V$-QCA, and we need to run the $U$-QCA for four time steps in order to simulate one time step of the $V$-QCA. More generally the challenge is to come up with an initial configuration of the $U$-QCA so that it behaves just as the $V$-QCA with respect to the encoded initial configuration, after some fixed number of time steps. Clearly such an encoding will have to hold the configuration of the $V$-QCA as well as some way of describing the scattering unitary $V$.)](UsimV.pdf) Related results --------------- In the realm of classical computing Durand-Lose [@Durand], has described an intrinsically universal one-dimensional reversible cellular automaton. Our construction will turn out to be a little simpler and cannot be substituted for this previous one, because reversible circuit universality requires at least one $3$-bit gate if done without the help of quantum mechanics. In the realm of quantum computing Shepherd, Franz and Werner [@Shepherd] have defined a class of QCA where the scattering unitary $V_i$ changes at each step $i$ (CCQCA). Via this construct they have built a QCA of cell-dimension $12$ which is universal in the circuit-sense. Universality in the circuit-sense had already been achieved by Van Dam [@VanDam] and Raussendorf [@Raussendorf] – the latter uses a two-dimensional QCA but has this inspiring idea of programs crossing the data, with computation occurring in the interaction. Watrous [@Watrous] has proved that QCA are universal in the sense of Quantum Turing Machines. To our knowledge there is no previous work on intrinsically universal quantum cellular automata. Plan ---- In section \[definitions\] the reader is provided with the necessary theoretical background on QCA – and the notion of intrinsic simulation is transposed to this theory. As in the classical case intrinsic simulation comes in two flavours, with one stronger than the other. In section \[weak\] we construct a particular QCA, which we show is intrinsically universal, in the first sense. In section \[strong\] we augment this QCA, and show that this results in a particular QCA which is intrinsically universal, in the second stronger sense. We conclude in section \[conclusion\]. Theoretical background {#definitions} ====================== One-dimensional QCA {#subsecdef} ------------------- We now recall the fundamental definitions and properties of one-dimensional QCA. In what follows $\Sigma$ is a fixed finite set of symbols (i.e. ‘the alphabet’, describing the possible basic states each cell may take) and $q$ is a symbol such that $q\notin\Sigma$, which will be known as ‘the quiescent symbol’, which represents an empty cells. We write $q+\Sigma=\{q\}\cup\Sigma$ for short. **(Finite configurations)** A *(finite) configuration* $c$ over $q+\Sigma$ is a function $c: {\mathbb{Z}}\longrightarrow q+\Sigma$, with $i\longmapsto c(i)=c_i$, such that there exists a (possibly empty) finite interval $I$ verifying $i\in I\Rightarrow c_i\in q+\Sigma$ and $i\notin I\Rightarrow c_i=q$. The smallest such interval $I$ is called interval domain of $c$, and is denoted $\textrm{idom}(c)$. The set of all finite configurations over $q+\Sigma$ will be denoted $\mathcal{C}_{q+\Sigma}$, whilst the set of all finite configurations having interval domain included in a finite interval $J$ will be denoted $\mathcal{C}_{q+\Sigma}^J$. Whilst configurations hold the basic states of an entire line of cells, and hence denote the possible basic states of the entire QCA, the global state of a QCA may well turn out to be a superposition of these. The following definition works because $\mathcal{C}_{q+\Sigma}$ is a countably infinite set.\ **(Superpositions of configurations)**\[superp\] Let $\mathcal{H}_{\mathcal{C}_{q+\Sigma}}$ be the Hilbert space of configurations, defined as follows. To each finite configuration $c$ is associated a unit vector ${| c \rangle}$, such that the family ${\left({| c \rangle}\right)}_{c\in\mathcal{C}_{q+\Sigma}}$ is an orthonormal basis of $\mathcal{H}_{\mathcal{C}_{q+\Sigma}}$. A *superposition of configurations* is then a unit vector in $\mathcal{H}_{\mathcal{C}_{q+\Sigma}}$. We also denote by $\mathcal{H}_{\mathcal{C}_{q+\Sigma}^J}$ the subspace of $\mathcal{H}_{\mathcal{C}_{q+\Sigma}}$ spanned by the configurations in $\mathcal{C}_{q+\Sigma}^J$. Note that this space of QCA configurations is the same one as in [@Arrighi1; @Arrighi2; @Durr1; @Durr2; @Watrous]. It is isomorphic to the cyclic one considered in [@Meyer1], but fundamentally different from the finite, bounded periodic space of [@VanDam] and the infinite setting of [@Schumacher]. The infinite setting of [@Schumacher] is slightly more permissive, but would force us to abandon the traditional setting of Hilbert spaces and move towards $C^*-$algebras. We choose not to do so out of simplicity, but our results apply to that context also since QCA have the same form in both contexts, as explained in [@Arrighi2]. **(Unitarity)**\[unitarity\] A linear operator $G:\mathcal{H}_{\mathcal{C}_{q+\Sigma}}\longrightarrow\mathcal{H}_{\mathcal{C}_{q+\Sigma}}$ is *unitary* if and only if $\{G{| c \rangle}\,|\,c\in\mathcal{C}_{q+\Sigma}\}$ is an orthonormal basis of $\mathcal{H}_{\mathcal{C}_{q+\Sigma}}.$ **(Shift-invariance)**\[shift-invariance\] Consider the shift operation which takes configuration\ $c=\ldots c_{i-1}c_ic_{i+1}\ldots$ to $c'=\ldots c'_{i-1}c'_ic'_{i+1}\ldots$ where for all $i$ $c'_i=c_{i+1}$. Let $\sigma:\mathcal{H}_{\mathcal{C}_{q+\Sigma}}\longrightarrow\mathcal{H}_{\mathcal{C}_{q+\Sigma}}$ be its linear extension to superpositions of configurations. A linear operator $G:\mathcal{H}_{\mathcal{C}_{q+\Sigma}}\longrightarrow\mathcal{H}_{\mathcal{C}_{q+\Sigma}}$ is said to be *shift invariant* if and only if $G\sigma=\sigma G$. **(Causality)**\[locality\] A linear operator $G:\mathcal{H}_{\mathcal{C}_{q+\Sigma}}\longrightarrow\mathcal{H}_{\mathcal{C}_{q+\Sigma}}$ is said to be *causal* with with radius $\frac{1}{2}$ if and only if for any $\rho,\rho'$ two states over $\mathcal{H}_{\mathcal{C}_f}$, and for any $i\in{\mathbb{Z}}$, we have $$\rho|_{i,i+1}=\rho'|_{i,i+1}\quad \Rightarrow G\rho G^{\dagger}|_i=G\rho'G^{\dagger}|_i. \label{loceq}$$ Here we used the notation $\rho|_J$ to mean restriction of $\rho$ to the region $J$ in the sense of the partial trace, and $G^{\dagger}$ is the Hermitian adjoint of $G$. In the classical case, the definition would be that the letter to be read in some given cell $i$ at time $t+1$ depends only on the state of the cells $i$ and $i+1$ at time $t$. Transposed to a quantum setting, we get the above definition: to know the state of cell number $i$, we only need to know the states of cells $i$ and $i+1$ before the evolution.\ We are now ready to give the formal definition of one-dimensional quantum cellular automata. **(QCA)**\[lca\] A one-dimensional quantum cellular automaton (QCA) is an operator\ $G:\mathcal{H}_{\mathcal{C}_{q+\Sigma}}\longrightarrow\mathcal{H}_{\mathcal{C}_{q+\Sigma}}$ which is unitary, shift-invariant and causal. This is clearly the natural axiomatic quantization of the notion of cellular automata. It was first given in [@Arrighi2] but stems from equivalent definition in the literature, phrased in terms of homomorphism of a $C^*$-algebra [@Schumacher]. There are other definitions in [@Arrighi1; @Delgado; @Durr1; @Durr2; @Meyer1; @VanDam; @Watrous] which are not axiomatic, in the sense that they all make particular assumptions about the form of the local action of $G$, and $G$ is then defined as a composition of these actions. That of [@Delgado] turns out to be almost equivalent to ours [@Arrighi2; @Schumacher] in the end, as explained in [@Arrighi3]. Intrinsic simulation of one-dimensional QCA {#subsecsim} ------------------------------------------- The notions of intrinsic simulation of one CA by another arises with Banks in [@Banks], but is not formalized until Albert and Culik [@Albert]. Even then, this apparently simple concept gives rise to two different competing definitions, which we may call intrinsic simulations and strong intrinsic simulations respectively. Both are nicely explained in [@Delorme], but subsequent works tend to focus on weak intrinsic simulations, as is nicely reviewed in [@Ollinger1; @Ollinger2]. Hence let us start with an explanation of the more modern notion of intrinsic simulation.\ The basic intuition in order to say that $G'$ simulates $G$ is that we can translate the content of each cell of $G$ into cells of $G'$, run $G'$, and then reverse the translation – and that this three steps process will be equivalent to just running $G$. But first we must make it clear what we mean by ‘translate’. This translation should be simple (the idea is that the real cost of the computation is carried over by $G'$), it should preserve the topology (the idea is that each cell of $G$ is encoded into cells of $G'$ in a way which preserves whom neighbours whom), and it should be faithful (the idea is that no information should be lost in translation). This latter requirement translates into a precise notion in quantum theory, which is that of being an isometry, i.e. an inner product preserving evolution with $Enc^\dagger Enc=\mathbb{I}$. This same requirement also coincides with the fact that we would like this translation to be a physical process, i.e. that an actual translating machine can actually be built in theory. This is because quantum mechanics limits all physical evolutions to be isometries. With these observations in mind we reach the following definitions. **(Isometric coding)**\[isomcode\] Consider $q+\Sigma$ and $q''+\Sigma''$ two alphabets with distinguished quiescent states $q$ and $q''$, and such that $|q+\Sigma|\leq|q''+\Sigma''|$. Consider $\mathcal{H}_{q+\Sigma}$ and $\mathcal{H}_{q''+\Sigma''}$ the Hilbert spaces having these alphabets as their basis, and $\mathcal{H}_{\mathcal{C}_{q+\Sigma}}$, $\mathcal{H}_{\mathcal{C}_{q''+\Sigma''}}$ the Hilbert spaces of finite configurations over these alphabets.\ Let $E$ be isometric linear map from $\mathcal{H}_{q+\Sigma}$ to $\mathcal{H}_{q''+\Sigma''}$ which preserves quiescence, i.e. such that $E{| q \rangle}={| q'' \rangle}$. It trivially extends into an isometric linear map $Enc=(\bigotimes_\mathbb{Z} E)$ from $\mathcal{H}_{\mathcal{C}_{q+\Sigma}}$ into $\mathcal{H}_{\mathcal{C}_{q''+\Sigma''}}$, which we refer to as an isometric encoding.\ Let $D$ be isometric linear map from $\mathcal{H}_{q''+\Sigma''}$ to $\mathcal{H}_{q+\Sigma}\otimes\mathcal{H}_{q''+\Sigma''}$ which also preserves quiescence, in the sense that $D{| q'' \rangle}={| q \rangle}\otimes{| q'' \rangle}$. It trivially extends into an isometric linear map $Dec=(\bigotimes_\mathbb{Z} D)$ from $\mathcal{H}_{\mathcal{C}_{q''+\Sigma''}}$ into $\mathcal{H}_{\mathcal{C}_{q+\Sigma}}\otimes\mathcal{H}_{\mathcal{C}_{q''+\Sigma''}}$, which we refer to as an isometric decoding.\ The isometries $E$ and $D$ define an isometric coding if the following condition is satisfied:\ $$\forall {| \psi \rangle}\in \mathcal{H}_{\mathcal{C}_{q+\Sigma}},\,\exists {| \phi \rangle}\in \mathcal{H}_{\mathcal{C}_{q''+\Sigma''}}\quad/\quad{| \psi \rangle}\otimes{| \phi \rangle}=Dec{\left(Enc {| \psi \rangle}\right)}.$$ (The understanding here is that $Dec$ is morally inverse function of $Enc$, but we may leave out some garbage ${| \phi \rangle}$ in the way.) **(Direct simulation)**\[directsim\] Consider $q+\Sigma$ and $q''+\Sigma''$ two alphabets with distinguished quiescent states $q$ and $q''$, and two QCA $G$ and $G''$ over these alphabets. We say that $G''$ directly simulates $G$, if and only if there exists an isometric coding such that $$\forall i\in\mathbb{N},\,\forall {| \psi \rangle}\in \mathcal{H}_{\mathcal{C}_{q+\Sigma}},\,\exists {| \phi \rangle}\in \mathcal{H}_{\mathcal{C}_{q''+\Sigma''}}\quad/\quad (G^i{| \psi \rangle})\otimes{| \phi \rangle}=Dec {\left({G''}^i{\left(Enc {| \psi \rangle}\right)}\right)}.$$ Unfortunately this is not quite enough for intrinsic simulation. Often we want to say that $G'$ simulates $G$ even though the translation:\ - takes one cell of $G$ into several, not just one cell of $G'$;\ - hence the quiescent symbol $q$ becomes a quiescent word $q'$;\ - $G'$ needs be run $t$ times instead of just once.\ All of these changes are made formal via the notion of grouping, as in the following definitions. **(Grouping)**\[packmap\] Let $G'$ be a QCA over alphabet $q'+\Sigma'$. Let $s$ and $t$ be two integers, $q''$ a word in $(q'+\Sigma')^s$, and $\Sigma''=\Sigma'^s \backslash \{q'\}$. Consider the iterate global evolution $G'^t$ up to a grouping of each $s$ adjacent cells into one supercell. If this operator can be considered to be a QCA $G''$ over $q''+\Sigma''$, then we say that $G''$ is an $(s,t,q')$-grouping of $G'$. **(Intrinsic simulation)**\[intsim\] Consider $q+\Sigma$ and $q'+\Sigma'$ two alphabets with distinguished quiescent states $q$ and $q'$, and two QCA $G$ and $G'$ over these alphabets. We say that $G'$ intrinsically simulates $G$ if and only if there exists $G''$ some grouping of $G'$ such that $G''$ directly simulates $G$. In simple words $G'$ intrinsically simulates $G$ if and only if there exists some isometry $E$ which translates each cell of $G$ into $s$ cells of $G'$, such that if we then run $G'^t$ and translate back, the whole process is just equivalent to just a run of $G$. This understanding is captured by Fig. \[IntrinsicSim\]. ![The notion of intrinsic simulation made formal.\[IntrinsicSim\]](IntrinsicSim.pdf) Notice, however, that $q$ the quiescent symbol of $G$ gets encoded into words ${| q' \rangle}=E{| q \rangle}$, which in general may not be the same as ${| q'^s \rangle}$, i.e. $s$ quiescent cells of $G'$. Hence with this notion of weak intrinsic simulation, we are indirectly assuming that the initial state of simulating QCA $G'$ could be prepared in a non-finite configuration, i.e. one which does not end and begin with only $q'$ symbols, but repeated $q'$ words instead. Formally this is not a problem, since $G'$ the $(s,t,q')$-grouping of $G'$ remains a valid QCA with quiescent symbol $q'$. Yet, depending on the application, one may wonder whether this notion of intrinsic simulation is the appropriate notion. For instance, if the implementation of $G'$ cannot be fed with $q'$ words left and right as the computation unravels, and if we do not know when the computation is supposed to stop, then this notion of intrinsic simulation may fail. These considerations will become clearer in the beginning of Section \[strong\], once illustrated with an example. Nevertheless they are exactly the ones which motivated the notion of strong intrinsic simulation.\ **(Strong isometric coding)**\[strongisomcode\] Consider $q+\Sigma$ and $q''+\Sigma''$ two alphabets with distinguished quiescent states $q$ and $q''$, and such that $|q+\Sigma|\leq|q''+\Sigma''|$. Consider $\mathcal{H}_{q+\Sigma}$ and $\mathcal{H}_{q''+\Sigma''}$ the Hilbert spaces having these alphabets as their basis, and $\mathcal{H}_{\mathcal{C}_{q+\Sigma}}$, $\mathcal{H}_{\mathcal{C}_{q''+\Sigma''}}$ the Hilbert spaces of finite configurations over these alphabets.\ Let $E, I$ be isometric linear maps from $\mathcal{H}_{q+\Sigma}$ to $\mathcal{H}_{q''+\Sigma''}$, where $I$ preserves quiescence, i.e. is such that $I{| q \rangle}={| q'' \rangle}$. Let $L, R$ be isometric linear maps from $\mathcal{H}^{\otimes r}_{q+\Sigma}$ to $\mathcal{H}^{\otimes r}_{q''+\Sigma''}$. Given any finite interval $J$, these can be combined into an isometric linear map $Enc^J=(I\bigotimes L\otimes \bigotimes_J E\otimes R \bigotimes I)$ from $\mathcal{H}_{\mathcal{C}_{q+\Sigma}}$ into $\mathcal{H}_{\mathcal{C}_{q''+\Sigma''}}$, which we refer to as the isometric encoding for interval $J$.\ Let $D$ be isometric linear map from $\mathcal{H}_{q''+\Sigma''}$ to $\mathcal{H}_{q+\Sigma}\otimes\mathcal{H}_{q''+\Sigma''}$, in the sense that $D{| q'' \rangle}={| q \rangle}\otimes{| q'' \rangle}$. It trivially extends into an isometric linear map $Dec=(\bigotimes_\mathbb{Z} D)$ from $\mathcal{H}_{\mathcal{C}_{q''+\Sigma''}}$ into $\mathcal{H}_{\mathcal{C}_{q+\Sigma}}\otimes\mathcal{H}_{\mathcal{C}_{q''+\Sigma''}}$, which we refer to as a strong isometric decoding.\ The isometries $L,E,R$ and $D$ define a strong isometric coding if the following condition is satisfied:\ $$\forall {| \psi \rangle}\in \mathcal{H}_{\mathcal{C}^J_{q+\Sigma}},\,\exists {| \phi \rangle}\in \mathcal{H}_{\mathcal{C}_{q''+\Sigma''}}\quad/\quad{| \psi \rangle}\otimes{| \phi \rangle}=Dec{\left(Enc^J {| \psi \rangle}\right)}.$$ In this scheme we now allow ourselves to know in advance whether the non-quiescent cells are all within the region $J$. Over the $J$ region we apply our encoding $\bigotimes E$ as usual. Over left and right quiescent tails we do not do much, since $\bigotimes I$ sends ${| \ldots qq\ldots \rangle}$ to ${| \ldots q''q''\ldots \rangle}$ anyway. In the direct surroundings of the $J$ region, we do something special, namely we apply $L$ on the left and $R$ on the right. Intuitively $L/R$ will be used to work at ‘encoding more left/right cells’ as the configuration grows through the computation, leading to the following expected notion of strong direct simulation. **(Strong direct simulation)**\[strgdirectsim\] Consider $q+\Sigma$ and $q''+\Sigma''$ two alphabets with distinguished quiescent states $q$ and $q''$, and two QCA $G$ and $G''$ over these alphabets. We say that $G''$ strongly directly simulates $G$, if and only if there exists a strong isometric coding such that for all interval $J$ we have: $$\forall i\in\mathbb{N},\,\forall {| \psi \rangle}\in \mathcal{H}_{\mathcal{C}^J_{q+\Sigma}},\,\exists {| \phi \rangle}\in \mathcal{H}_{\mathcal{C}_{q''+\Sigma''}}\quad/\quad (G^i{| \psi \rangle})\otimes{| \phi \rangle}=Dec {\left({G''}^i{\left(Enc^J {| \psi \rangle}\right)}\right)}.$$ Again this is not quite enough for intrinsic simulation, since we want to say that $G'$ simulates $G$ even though the translation takes one cell of $G$ into $s$ cells of $G'$, and one step of $G$ into $t$ steps of $G'$. But this time we want to make sure that the quiescent tails ${| \ldots qq \ldots \rangle}$ are taken into quiescent tails ${| \ldots q'q' \ldots \rangle}$. So we restrict the notion of grouping. **(Strong grouping)**\[strongpackmap\] Let $G'$ be a QCA over alphabet $q'+\Sigma'$. Let $s$ and $t$ be two integers, and $\Sigma''=\Sigma'^s \backslash \{q'^s\}$. Consider the iterate global evolution $G'^t$ up to a grouping of each $s$ adjacent cells into one supercell. If this operator can be considered to be a QCA $G''$ over $q'^s+\Sigma''$, then we say that $G''$ is an $(s,t)$-strong-grouping of $G'$. **(Strong simulation)**\[strongdirectsim\] Consider $q+\Sigma$ and $q'+\Sigma'$ two alphabets with distinguished quiescent states $q$ and $q'$, and two QCA $G$ and $G'$ over these alphabets. We say that $G'$ strongly intrinsically simulates $G$ if and only if there exists $G''$ some strong grouping of $G'$ such that $G''$ strongly directly simulates $G$. Roughly we can say $G'$ strongly intrinsically simulates $G$ if and only if there exists some topology-preserving isometry $Enc$ which translates each finite configuration of $G$ into a finite configuration $G'$, such that if we then run $G'^t$and translate back, the whole process is just equivalent to just a run of $G$. The structure of one-dimensional QCA {#subsecstruc} ------------------------------------ As we have just seen one-dimensional QCA can be defined in quite general terms as a local, translation-invariant, global unitary evolutions $G$ acting over a line of finite dimensional quantum system. But contrary to its classical counterpart the definition does not yield an immediate way of constructing / enumerating all of the instances of this model. So if we just stick to that one definition QCA remain excessively abstract, hard-to-grasp mathematical object.\ Fortunately Schumacher and Werner [@Schumacher], and later Arrighi et al. [@Arrighi2] in the context of finite unbounded configurations, have proved that any such mathematical object is structured according to Figure \[werner\], i.e. that one-dimensional QCA admit a $2-$layered unitary block representation. These elementary unitary evolutions $U_0$ and $U_1$ are well-understood objects, i.e. they can be just any finite unitaries, as implemented for instance by some circuit of universal quantum gates for instance. Cheung and Perez-Delgado have proposed a definition of QCA directly in terms of local unitary evolutions, but thanks to this theorem the two are equivalent in one-dimension (the situation is similar in $n$-dimensions, see [@Arrighi3]). ![Werner-style Margolus neighbourhood QCA [@Schumacher].\[werner\] The elementary unitary evolutions $U_0$ and $U_1$ are alternated repeatedly as shown. The letters $p$, $q$, $r$, $s$ with $pq=rs$ denote the dimensions of the quantum systems.](werner.pdf) This seems to be as good as it gets if we are not willing to modify the space upon which $G$ is acting. However the notion we are interested in is intrinsic universality, and in this context we are quite willing to interleave some extra cells for the purpose of simulating one QCA by another. And so it is not difficult to see that any such $U_0 U_1$-structured QCA can be straightforwardly made into a $U U$-structured QCA, as made explicit by Figure \[Vsimwerner\]. ![Partitioned QCA simulating a Werner-style Margolus neighbourhood QCA.\[Vsimwerner\] The scattering unitary $U$ is a $(pq+1)^2\times(pq+1)^2$ matrix, essentially acting like either $U_0$ or $U_1$ on parts of the input according to the control qubit. The black box could be just any unitary, the lines upon which it acts are added just so that $U$ takes subsystems of an equal number of dimensions left and right.](Vsimwerner.pdf) Hence for the purpose of this paper, can really just restrict our attention to $U U$-structured QCA as in Figure \[structure\]. These are also referred to as Partitioned QCA for instance in [@VanDam], by analogy with classical Partitioned CA. **(Partitioned QCA)**\[pqca\] A partitioned one-dimensional quantum cellular automaton (PQCA) over $\mathcal{H}_{\mathcal{C}_{q+\Sigma}}$ is defined by a unitary $U$ an operator $U:\mathcal{H}_{q+\Sigma}\otimes\mathcal{H}_{q+\Sigma}\longrightarrow\mathcal{H}_{q+\Sigma}\otimes\mathcal{H}_{q+\Sigma}$, such that $U{| qq \rangle}={| qq \rangle}$, i.e. one which we take two cells into two cells and preserves quiescence. Let $G=(\bigotimes_{\mathbb{Z}} U)$ the operator over $\mathcal{H}_{\mathcal{C}_{q+\Sigma}}$. The induced global evolution is $G$ at odd time steps, and $\sigma G$ at even time steps, as in Fig. \[structure\]. The defining elementary unitary evolution $U$ will be referred to as the scattering unitary, by analogy with quantum field theory. Hence we will talk about ‘$U$-defined QCA’ in order to designate the PQCA having scattering unitary $U$. In this paper we demonstrate how a particular $U$-defined QCA can simulate any $V$-defined QCA, that is for any $V$. But again let us insist that this restriction to PQCA is without loss of generality, since we have just explained that any QCA can be put into the form of a PQCA. Notice that in Fig. \[werner\], two vertical lines were denoting two subcells making up one cell of the QCA. But in the definition of PQCAs, we have referred to each of these subcells as just one cell. This was simply out of convenience, but one should keep in mind that one cell of a general QCA really corresponds to two cells of a PQCA. Intrinsic universality {#weak} ====================== What do we need for universality? {#conventions} --------------------------------- In subsection \[subsecdef\] we have recalled the formal definition of One-dimensional QCA, and in subsection \[subsecstruc\] we have recalled they have a simple circuit-like structure. Ultimately the picture one needs to have in mind in order to follow this section is that a QCA is just that of Fig. \[structure\].\ In subsection \[subsecsim\] we have provided a formal definition for the notion of intrinsic simulation. But again the picture one needs to have in mind in order to follow this section just that of Fig. \[UsimV\].\ The purpose of this section is to find a particular $U$-defined QCA, and which is capable of intrinsically simulating any $V$-defined QCA, whatever the $V$. In order to describe that $U$-defined QCA we need to describe two things:\ - How its cells are like (i.e. what are the vertical lines of Fig. \[structure\] made of, what is the dimensionality of $U$?). By definition of QCA they are finite dimensional quantum systems of some fixed dimension $d$ of course, but for clarity we will decompose these into subsystems of dimension $d_i$, and we will give names to these subsystems according to their purpose (i.e. the vertical lines of Fig. \[structure\] are ‘buses’).\ - How $U$ acts upon a pair of these cells, and more precisely upon the subsystems making up the pair of cells. We will say this informally, but we will also provide a formal circuit descriptions of $U$, and check that such a $U$ is indeed a unitary.\ Before continuing with our detailed discussion of the QCA, it is useful for the sake of clarity to make precise some of the vocabulary and conventions that will be encountered in the rest of the paper. By a *subsystem* we mean a constituent of a cell that serves a specific function in our QCA. A subsystem may take many different *values*. Subsystems have names written in bold. The term *signal* is used to refer to the value of a certain subsystem when it consistently travels between a cell and its left/right neighbour. In the space-time diagram of the QCA, a signal looks like a ‘line’ propagating through cells. Usually we take the name of a signal the name of the subsystem that takes those values.\ This section will be organized as follows: first we give an intuitive idea about the mechanism we used to solve this problem, then we will explain different components of the QCA in detail, and finally we will see how the whole system fits together. Intuition --------- The cells of our universal, $U$-defined, QCA will have a subsystem called $\textbf{data}$ used for encoding one qubit of information about the state of a cell of the simulated, $V$-defined QCA. Hence one simulated cell (dotted oval in Fig. \[IntuitionWeak\]) will in general be encoded as several adjacent simulating cells (small grey ovals in Fig. \[IntuitionWeak\]). We also need to simulate the action of some arbitrary unitary $V$, and so the cells of our universal QCA will have a subsystem called $\textbf{program}$, holding some description of one of the elementary universal quantum gates that make up $V$. Hence $V$ (dotted box in Fig. \[IntuitionWeak\]) will in general be encoded as several adjacent simulating cells (small black ovals in Fig. \[IntuitionWeak\]). Because $V$ originally acts upon two cells, it is encoded in the surroundings of the two encoded cells.\ During a first phase, as time unravels, the information held in the $\textbf{data}$ subsystem of the encoded cells remain stationary, and so we can think of them as some stationary data signals. Meanwhile the information held in the $\textbf{program}$ subsystems of the surroundings travels at lightspeed, and we can think of them as moving program signals. The program signals cross the data signals, leaving them unchanged, until they collide between one another. When that happens an elementary universal quantum gate is applied on the $\textbf{data}$ subsystems (grey box in Fig. \[IntuitionWeak\]), thereby implementing $V$. Which elementary universal quantum gate gets applied depends on the value of the colliding program signals. Where the elementary universal quantum gate gets applied depends on where the collision takes place.\ During a second phase, we need to ‘reload’ this situation, with the added difficulty that $V$ gets applied in a shifted manner. Hence we need to arrange so that the left/right encoded cell travels left/right in order to meet with their right/left counterpart on the next site (travelling small grey ovals in Fig. \[IntuitionWeak\]).\ ![Outline of the simulation scheme.\[IntuitionWeak\]](IntuitionWeak.pdf) We will explain how to do this in three steps. First we draw a background pattern which allow us to synchronize this whole process. Second we let some data signals flow upon this background. Third we let the program signals move upon the background, crossing and acting upon the data signals. Note that whenever we write ${| dgp \rangle}$ for the state of a cell, we mean that the subsystem $\textbf{data}$ is in state ${| d \rangle}$, subsystem $\textbf{program}$ is in state ${| g \rangle}$, subsystem $\textbf{mode}$ is in state ${| p \rangle}$. Ternary background pattern {#tern} -------------------------- First we seek to draw the *ternary background pattern* of Figure \[ternary\], where the large squares cycle through the three colours Light grey, Middle grey or Dark grey. The reason why this is useful will become clear in the next paragraph, for now it suffices to know that this ternary background pattern will help us synchronize the flow of the data signals as in Figure \[IntuitionWeak\] and hence organize the computation as we add more interesting things to the initial configuration, i.e. the bottom of the diagram.\ In order to achieve this ternary background pattern each cell must contain a $3$-dimensional system to code for those three different colours. This is really the purpose of subsystem **mode**: when the **mode** equals $0$, $1$ or $2$ the background colour is Light grey, Middle grey or Dark grey respectively. We must then place some signals at regular intervals, travelling at lightspeed and telling mode signals that they must change colour, and this is really role played by state ${| 1 \rangle}$ of subsystem **program**.\ Let us show that the scattering unitary $U$ which is given in Figure \[Uascircuit\] does the job of generating Figure \[ternary\]. Observe Figure \[Uascircuit\] and notice that the content of the **mode** and **program** subsystems is always propagated unchanged by the scattering unitary $U$, to the right/left if it comes from the left/right. Moreover observe Figure \[structure\] and notice that at the next layer the content of **mode** and **program** will again come up from the left/right and hence be propagated again to the right/left. Hence whatever value is in the **mode** or in the **program** subsystem it just travels at maximal speed, right or left, depending only upon its position in the initial configuration. This is just what we mean by ‘a signal propagating at lightspeed’. In Subsection \[theqca\] we provide all extra information needed about Figure \[Uascircuit\] so that the behaviour of $U$ becomes fully-determined. We then state that ${| 1 \rangle}$ in the **program** subsystem is the control value required for the **+1 mod 3** to apply upon the **mode** subsystem. Hence the ‘Change colour’ signals are indeed implemented by setting some cells to have their subsystem $\textbf{program}$ initialized at ${| 1 \rangle}$ as in Figure \[ternary\], and the ‘Change colour’ signals indeed propagates at lightspeed, changing the value of the mode signals travelling at lightspeed in the opposite direction.\ Notice that later, when we will set the **data** subsystem to non-${| 0 \rangle}$ values in order to code for simulated cells, or use up the other possible values of the **program** subsystem in order to code for elementary universal gates to be applied upon the coded simulated cells, this ternary background pattern will remain unaffected. This is obvious from Figure \[Uascircuit\] and the fact that ${| 1 \rangle}$ is the only value of **program** which triggers the **+1 mod 3** gate.\ The required widths of the Light grey and Middle grey zones of the initial configuration vary depending upon the $V$-defined QCA we are seeking to simulate, in a way which we explain in Subsection \[hexa\] and \[coll\], respectively.\ Hexagonal data signals flow {#hexa} --------------------------- Second we seek to draw the *hexagonal data signals flow* of Figure \[dataflow\]. That is we want to implement data signals, and would like that the data signals remain stationary for a while, and then separate into a left moving and a right moving bunch of data signals, only to eventually rejoin their left and right counterparts in order to flow straight in time again, as was explained in Figure \[IntuitionWeak\]. As previously mentioned, the reason why we want to achieve this particular form of data signals flow is that it corresponds to the overall architecture of the QCA we are trying to simulate as in Figure \[UsimV\], with the interaction unitary $V$ taking its inputs as coming both from the left and the right, computing upon them, and spitting its outputs again both towards the left and the right for another run of $V$.\ In order to achieve this each cell must contain another $3$-dimensional subsystem to code for a data signal. This is really the purpose of subsystem **data**: when the **data** equals ${| 0 \rangle}$, ${| 1 \rangle}$ or ${| 2 \rangle}$ the cell carries no data, an encoded ${| 1 \rangle}$ or an encoded ${| 2 \rangle}$ respectively. We must then place the encoded data qubits in the Light grey coloured zones of the initial configuration, i.e. replacing the ${| 000 \rangle}$ cells by ${| 100 \rangle}$ cells in order to code for the presence of an encoded ${| 0 \rangle}$, and ${| 000 \rangle}$ by ${| 200 \rangle}$ in order to code for the presence of an encoded ${| 1 \rangle}$.\ Let us show that the scattering unitary $U$ which is given in Figure \[Uascircuit\] does the job of generating Figure \[dataflow\]. The intuitive explanation will of course be that the Grey levels of the ternary background pattern are here to tell the data whether it can move or not, with the Middle and Dark grey forcing it to remain stationary, and the Light grey allowing it to move freely until they are gathered by a Middle grey funnel again. The more formal explanation relies on looking at the scattering unitary matrix $U$ which is given in Figure \[Uascircuit\] in order to understand when the value of the left/right **data** subsystem is propagated to the right/left, and when it is just left sitting on the left/right. This is what determines whether a data signal is stationary or moving at lightspeed. In order to have a complete answer to this question one must look at the definition of the $S$ gate in Figure \[Uascircuit\], as provided in Subsection \[theqca\]. There we find that $S$ swaps the left and right **data** subsystems but only if one of them is ${| 0 \rangle}$ (i.e. a data signal moves right/left only if there is no data signal there) and if the values of both **mode** subsystems are ${| 0 \rangle}$ (i.e. the Light grey zones of the ternary background pattern). Initially the data signals are in a Light grey zones, but they are stuck upon another, so they cannot move. Not even the ones on the left and right ends can move – due to the surrounding Middle grey zones. Clearly this stationary situation will be maintained until the ones on the left and right ends become surrounded by Light grey zones. During this period the data signals are freed two by two, with the one on the left end going to the left, and the one on the right end going to the right. Finally the first right moving data signal meets up with the first left moving data signal, and so they are stuck by one another for one step. But that one step is enough so that the second right/left moving data signal rejoins the first, so that the second one also gets stuck against the first, and the first does not bounce back. When all of the right/left moving signals have done so we are back to the original situation. Hence this ability of the ternary background pattern to ‘mold’ the flow of data signals into the hexagonal shape of Figure \[dataflow\].\ Notice that the required width of the Light grey zone of Figure \[ternary\] has now become apparent. Since the purpose of this Light grey zone is to encode the state of two cells of the simulated QCA, it needs to be twice as big as the number of qubits that are needed to encode one cell of the simulated, $V$-defined QCA. Collision gates {#coll} --------------- Third we seek simulate the scattering unitary $V$, as in Figure \[circuitry\]. As was explained in Figure \[IntuitionWeak\], the key idea here is that during the time the data signals are stationary, they may be crossed by program signals incoming from both their left and their right, and sometimes these program signals even collide against one another upon the data signals. The value of the colliding program signals is what will specify *what* should happen to the data signals, through a numbering of a set elementary universal quantum gates. The relative positions of the program signals is what will specify *where* this should happen, by determining where they collide.\ In order to achieve this we must change some of the ${| 0 \rangle}$ values of the $\textbf{program}$ subsystems of the Middle grey zones of the initial configuration, and allow them to take extra values ${| 2 \rangle}$ and ${| 3 \rangle}$. I.e. we will change some ${| 001 \rangle}$ cells into ${| 021 \rangle}$ or ${| 031 \rangle}$ cells and generate program signals carrying value ${| 2 \rangle}$ or ${| 3 \rangle}$, which will then travel at lightspeed, collide and so implement gates upon the data signals.\ Let us show that the scattering unitary $U$ which is given in Figure \[Uascircuit\] does the job of generating Figure \[circuitry\]. We have already shown in Subsection \[tern\] that the program signals travel at lightspeed, unaffected. The only thing we need to explain is what happens when they collide with one another. Again from Figure \[Uascircuit\] and Subsection \[theqca\] we have that whenever two program signals (**xprogram** and **yprogram** are both ${| 2 \rangle}$ or ${| 3 \rangle}$) cross each other upon some data signals (**xdata** and **ydata** non-${| 0 \rangle}$) then some elementary quantum gate is applied upon **xdata**$\otimes$**ydata** as given in Table \[Mgate\]. The mechanism is illustrated in Figure \[circuitry\] but with essentially classical elementary gates – so that we may draw their effect.\ Notice that the required width of the Middle grey zone of Figure \[ternary\] is starting to become apparent. Since the purpose of this Middle grey zone is to hold pairs of program signals, each pair coding for one of the elementary universal quantum gate implementing $V$, its size will depend will depend upon the number of elementary universal quantum gates of the circuit-description of $V$. However the exact size of those Middle grey zones will be determined in Subsection \[weakresults\]. Fitting things together {#theqca} ----------------------- So overall our universal, $U$-defined, QCA consists of a repeated application of one scattering unitary $U$ as in Figure \[structure\]. Now follows the summarized description of the structure of the cells and the scattering unitary $U$. *Structure of the cells* In Figure \[structure\] each vertical line do not represent just one qubit but a $36$-dimensional quantum system made of the subsystems described in the following table. Name Size Function ------------- ------ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- **data** 3 To hold one qubit of data of the QCA being simulated (${| 0 \rangle}\equiv$Empty, ${| 1 \rangle}\equiv$Encoded ${| 0 \rangle}$, ${| 2 \rangle}\equiv$Encoded ${| 1 \rangle}$). **program** 4 To code for what quantum gate should be applied to the data (${| 0 \rangle}\equiv$Empty, ${| 1 \rangle}\equiv$Change colour, ${| 2 \rangle}/{| 3 \rangle}\equiv$cf. Table \[Mgate\]) **mode** 3 To synchronize the flow of the data signals (${| 0 \rangle}\equiv$Light grey ${| 1 \rangle}\equiv$Middle grey, ${| 2 \rangle}\equiv$Dark grey). : Subsystems of the copy band[]{data-label="bandsubsystems"} *Structure of the scattering unitary $U$.* In Figure \[structure\] the scattering unitary $U$ takes two inputs **x** and **y**, each of which decomposes into three subsystems **xdata**, **xprogram**, **xmode** and **ydata**, **yprogram**, **yprogram** respectively, as mentioned. Therefore it could be given as a $36^2\times 36^2$ matrix of complex numbers, yet fortunately it decomposes as in Figure \[Uascircuit\]. ![The scattering unitary $U$ as a quantum circuit.\[Uascircuit\] The three left/right lines represent the three subsystems making up the left/right cell. Each square represents an elementary quantum gate being applied of these subsystems, conditional to the value of the control systems designated by the horizontal lines. Time flows upwards. The whole thing represents the scattering unitary $U$, it takes two cells and yields back two cells.](Uascircuit.pdf) In this diagram horizontal lines are control lines. Let us explain what we mean by control lines in general, as this will provide us with the main step of the proof of the unitarity of $U$. Say that a box labelled $B$ applies upon systems $S_1,\ldots, S_p$ whilst having an horizontal line crossing systems $T_1, \ldots, T_q$, then its effect is to apply the unitary evolution $B_i$ whenever the control system is in state ${| i \rangle}$ – and linearly so. The following well-known lemma formalizes this construct.  \ Let $S_1, \ldots, S_p$ and $T_1, \ldots, T_q$ be quantum systems of finite dimensions $|S_1|, \ldots, |S_p|, |T_1|, \ldots, |T_q|$. Let $\{B_i\}$ be a set of $|T_1|\cdots|T_q|$ unitary matrices of dimension $|S_1|\cdots|S_p|\times|S_1|\cdots|S_p|$. Then the evolution defined by linear extension of\ ${| i \rangle}^{T_1\otimes\ldots\otimes T_q}{| \psi \rangle}^{S_1\otimes\ldots\otimes S_p}\mapsto {| i \rangle}^{T_1\otimes\ldots\otimes T_q}B_i{| \psi \rangle}^{S_1\otimes\ldots\otimes S_p}$ is itself unitary. Here the notation ${| \phi \rangle}^S$ means the quantum system $S$ in in state ${| \phi \rangle}$. **Proof.** (This has appeared many times before in the literature but we reproduce it for completeness.) Since the evolution is a square matrix it suffices to check that it has orthonormal columns.\ Let $\{{| i \rangle}^{T_1\otimes\ldots\otimes T_q}\}$ and $\{{| j \rangle}^{S_1\otimes\ldots\otimes S_p}\}$ denote the canonical orthonormal basis for $T_1\otimes\ldots\otimes T_q$ and $S_1\otimes\ldots\otimes S_p$ respectively. Then the inner product between column $ij$ and column $kl$ is given by\ $({\langle i |}\otimes{\langle j |}B_i^\dagger)({| k \rangle}\otimes B_k{| l \rangle})=\delta_{ik}{\langle j |}B_i^\dagger B_i{| l \rangle}=\delta_{ik}\delta_{jl}$.$\Box$\ This being said we are now in a position to describe the three gates used, in terms of whatever canonical basis state their control systems may take. Of course the control systems do not actually have to be in a basis state, by linear extension and the above lemma. We shall see that the first two gates are used really just to move the encoded data qubits around, whereas the last gate actually does perform the computation. - The **+1 mod 3** gate. If **xprogram**$\neq{| 0 \rangle}$ and **yprogram**$={| 4 \rangle}$, then **xmode** is incremented by one modulo three. Else this is the identity. (Symmetrically so swapping the roles of **x** and **y**.) This step’s role is to increment the modes, which in turn synchronizes the flow of the data signals. - The **S** gate. If **xprogram**$={| 0 \rangle}$, and **yprogram**$={| 0 \rangle}$, then permute ${| 01 \rangle}$ with ${| 10 \rangle}$ and ${| 02 \rangle}$ with ${| 20 \rangle}$. Else this is the identity. This means we are swapping **xdata** and **ydata** only if one of them is empty and the modes are ‘White’. This step’s role is to move the data when appropriate, in order to actually perform the flow of data. - The **M** gate. If **xdata**$\neq{| 0 \rangle}$, **ydata**$\neq{| 0 \rangle}$, **xprogram**$={| 2 \rangle}$ and **yprogram**$={| 2 \rangle}$, then the system **xdata**$\otimes$**ydata** undergoes an elementary quantum gate according to the state of the system **xmode**$\otimes$**ymode** as in Table \[Mgate\]. This step’s role is to apply a quantum gate upon two qubits of data, in order to perform the computation. $$\begin{aligned} &\textbf{xprogram}\otimes\textbf{yprogram}\qquad &\textbf{Action of M}\\ &{| 22 \rangle} &Swap &\\ &{| 23 \rangle} &\mathbb{I} \otimes H&\quad\textrm{Hadamard on the second qubit}\\ &{| 32 \rangle} &H \otimes \mathbb{I}&\quad\textrm{Hadamard on the first qubit}\\ &{| 33 \rangle} &cPhase&\\ &otherwise &\mathbb{I} \otimes \mathbb{I}\end{aligned}$$ where $cPhase$ stands for $\left(\begin{array}{cccc} 1 &0 &0 &0\\ 0 &1 &0 &0\\ 0 &0 &1 &0\\ 0 &0 &0 &e^{i\pi/8} \end{array}\right)$.\ In the next section we show how this all fits together to obtain the intrinsic universality. Beforehand however note that since we have described the evolution $U$ as a combination of smaller unitary matrices via tensors, composition and the control-construct, it is indeed unitary as required by Def. \[pqca\]. Results {#weakresults} ------- *Universal set of elementary quantum gates.* First let us show that we have a universal set of gates available in the QCA. The set of gates which the gate is able to perform upon the data qubits has been chosen to be universal in the traditional sense, i.e. any finite dimensional unitary evolution $V$ can be approximated by tensors and compositions of these gates. We have not chosen the standard set ($cNot$, $H$, $Phase$) so as to preserve the **xy** symmetry of the unitary evolution $U$ and yet keep the dimension of **program** to a minimum, but it is easy to see that we can recover the standard set since $$\begin{aligned} cNot{| \psi \rangle}&=(\mathbb{I}\otimes H)(cPhase)^8(\mathbb{I}\otimes H){| \psi \rangle}\\ {| 1 \rangle}\otimes Phase{| \psi \rangle}&=cPhase{| 1 \rangle}\otimes{| \psi \rangle}\end{aligned}$$ where the ancilla ${| 1 \rangle}$ can be brought over via applications of the $Swap$ gate.\ *Addressing each elementary quantum gate.* Second, let us show that we may combine our universal quantum gates in and arbitrary fashion so as to be able to implement any quantum circuit within the QCA. Consider some circuit $C$ made of $m$ elementary two qubit quantum gates $(g_i)_{i=0\ldots m-1}$ taken from our universal set of two-qubit gates ($Swap$, $\mathbb{I}\otimes H$, $H\otimes \mathbb{I}$, $cPhase$). Suppose that the circuit is $2n$ qubits wide, so that the positions $(p_i)_{i=0\ldots m-1}$ telling us where to apply those gates can be given as numbers in $-(n-1)\ldots (n-1)$ relative to the center. For instance if $g_3=Swap$ and $p_3=0$, this means the fourth gate in the circuit consists in swapping the qubits at the center. Or say if $g_0=\mathbb{I}\otimes H$ and $p_0=n-1$, this means that the first gate of the circuit consists in applying a Hadamard to the rightmost qubit. Hence a position actually refers to a pair of qubits, position $-n+1$ being the leftmost pair of qubits, $-n+2$ the second leftmost pair etc. This is coherent with the fact that we have only two qubits gates in our chose universal set of quantum gates.\ Then circuit as acting upon the initial state ${| \psi \rangle}$ can be encoded in our QCA via the state: $$\begin{aligned} {| \Gamma(C,\psi) \rangle}&= (\bigotimes_{i=m-1}^{0}l_i)\otimes F{| \psi \rangle} \otimes (\bigotimes_{i=0}^{m-1}r_i).\end{aligned}$$ Here the $F{| \psi \rangle}$ region holds the qubits of the circuit, encoded within the cells of the $QCA$. The way to do this is as explained in Subsection \[hexa\]: i.e. ${| 0 \rangle}/{| 1 \rangle}$ is encoded as ${| 100 \rangle}/{| 200 \rangle}$. (It is important that the presence of data ${| 1?? \rangle}/{| 2?? \rangle}$ be distinguished from the absence of data ${| 0?? \rangle}$.) Formally (with ${| \phi \rangle}^j$ the $j^{th}$ subsystem of ${| \phi \rangle}$): $$\begin{aligned} F{| \psi \rangle}&=\bigotimes_{j=0}^{n-1}Inc{| \psi \rangle}^j{| 00 \rangle}\\ Inc{| 0 \rangle}&={| 1 \rangle}\quad Inc{| 1 \rangle}={| 2 \rangle}.\end{aligned}$$ The $(\bigotimes_{i=m-1}^{0}l_i)$ and $(\bigotimes_{i=0}^{m-1}r_i)$ hold the description of the circuit, encoded within the cells of the QCA. Say we want to perform the gate $g_i$ at position $p_i$ of these $2n$ encoded qubits. We will surround the qubits (and the previous gate descriptions) with a description of $(g_i, p_i)$, as in: $$\begin{aligned} \textrm{with for $p_i\leq 0$, }\; l_i&= {| 011 \rangle}^{\otimes 4n-1-2|p_i|}\otimes{| 0l(g_i)1 \rangle}\otimes{| 001 \rangle}^{\otimes 2|p_i|}\\ r_i&= {| 0r(g_i)2 \rangle}\otimes{| 001 \rangle}^{\otimes 4n-1}\\ \textrm{and for $p_i\geq 0$, }\; l_i&= {| 001 \rangle}^{\otimes 4n-1}\otimes{| 0l(g_i)1 \rangle}\\ r_i&= {| 001 \rangle}^{\otimes 2|p_i|}\otimes{| 0r(g_i)1 \rangle}\otimes{| 001 \rangle}^{\otimes 4n-1-2|p_i|}\\ \textrm{moreover }\; l(Swap)&=r(Swap)=l(\mathbb{I}\otimes H)=2\\ l(cPhase)&=r(cPhase)=r(H\otimes \mathbb{I})=3\end{aligned}$$ This deserves some explanations. Say, as in the previous example, that we want for our fourth gate to have a $Swap$ happen at position $0$. We know from Table \[Mgate\] that in order for $U$ to perform a $Swap$ upon its **data** subsystems, we need both **program** registers to be in state ${| 2 \rangle}$. So we need to have a program signal coming from the left and holding value ${| 2 \rangle}={| l(Swap) \rangle}$, and the same from the right. Moreover we need them to originate equally far from the center so that they meet above the center, and we need to make sure that the next gate gets encoded far away enough so that its program signals do not intersect ours – thereby inducing some unwanted operations. Finally the data signals need to be sandwiched by Middle grey mode signals so that they remain stationary, very much as explained in Subsection \[hexa\]. Here $l_3={| 001 \rangle}^{\otimes 2n-1}\otimes{| 021 \rangle}$ and $r_3={| 021 \rangle}\otimes{| 001 \rangle}^{\otimes 2n-1}$ will do the job. The rest is only a matter of placing the program signals in a adequate manner so that the gate is performed in the right position. Note that in the left/right encoded circuit description region, only one cell out of two is potentially used, because the other cells can only move in the opposite direction which will not go over our data, see Figure \[encoding\]. ![Encoding a gate $g$.\[encoding\] Here the position of the gate relative to the center position is $k=-1$, meaning that the gate will be applied not on the center qubits but on the pair just left of it. And here the radius of the data signals bus is $n=3$, meaning that three qubits are needed to code for just one cell of the simulated QCA. The two black lines represent the right-moving and left-moving program signals that make up the encoding of the gate $g$. The fact that the right-moving program signal originates further apart from the stationary data signals than its left counterpart is what codes for the position $k=-1$. In the worst case scenario the position is $-n+1$. Clearly we must not send another left-moving signal until the worst case scenario right-moving signal has come out of the data signals, otherwise their collision may trigger the application of an unwanted gate. So each gate requires at least $2(n+k)$ cells on each side in order to be encoded. In the chosen encoding we systematically take $4n$ cells on each side in order to encode one gate.](encoding.pdf) More convoluted encodings can be used to save up space and time, but only marginally. Also they seem to make the size of the encoding of a gate $g_i$ dependent upon its position $p_i$, which makes it more difficult to explain. What we have here does formalize the way to address each elementary quantum gate with respect to the data qubits in our scheme. It shows that we are able to simulate any such circuit $C$ by using up $8mn$ cells during $4mn$ steps. *Intrinsic simulation of any partitioned one-dimensional QCA.* Third, let us show that this particular PQCA is intrinsically universal. In order to do this, consider a PQCA has scattering unitary $V$, and say we seek to approximate it just as illustrated in Figure \[UsimV\]. In order to do so we must consider the quantum circuit $C(V)$ which approximates $V$ in terms of the quantum gates $(g_i)_{i=0\ldots m-1}$ acting at positions $(p_i)_{i=-n\ldots n-1}$. The approximated QCA as acting upon the initial state ${| \phi \rangle}$, can be encoded in our QCA via repetitions of the state ${| \Gamma(C,\phi,k) \rangle}$: $$\begin{aligned} {| \Gamma(C,\phi,k) \rangle}&= {| 011 \rangle}\otimes w \otimes {| 011 \rangle}\otimes F{| \phi \rangle}^{2k,2k+1}\\ \textrm{with for $p_i\leq 0$, }\; l_i&= {| 011 \rangle}^{\otimes 2n-1-|p_i|}\otimes{| 0l(g_i)1 \rangle}\otimes{| 011 \rangle}^{\otimes |p_i|}\\ r_i&= {| 0r(g_i)1 \rangle}\otimes{| 011 \rangle}^{\otimes 2n-1}\\ \textrm{and for $p_i\geq 0$, }\; l_i&= {| 011 \rangle}^{\otimes 2n-1}\otimes{| 0l(g_i)1 \rangle}\\ r_i&= {| 011 \rangle}^{\otimes |p_i|}\otimes{| 0r(g_i)1 \rangle}\otimes{| 011 \rangle}^{\otimes 2n-1-|p_i|}\\ \textrm{moreover }\; r&=\bigotimes_{i=0}^{m-1} r_i,\;l=\bigotimes_{i=m-1}^{0} l_i\\ w&=\bigotimes_k l^k \otimes r^k\end{aligned}$$ Here the scheme is exactly that of Fig. \[circuitry\]. Compared to the previous coding here is what deserves explanations. The data is encoded exactly in the same way, namely each pair of cells ${| \phi \rangle}^{k,k+1}$ is encoded as $2n$ cells $F{| \phi \rangle}^{k,k+1}$. Then the description of the circuit $C$ is also encoded in the same way, in the surroundings of each encoding cells. There are a couple of differences however. Previously we were using only one out of two cells in the left/right encoded circuit description region, but now we can interleave the left and right encoded circuit descriptions to use them all, this is what $w$ does and the reason why the factor of $2$ was dropped in the $l_i/r_i$ descriptions. Also, we now need to handle mode and ‘Change colour’ signals appropriately so as to obtain the ternary background pattern and the hexagonal data signals flow. By inspection of Fig. \[dataflow\] we see that this is done by setting the left/right encoded circuit description regions to having subsystems **mode** at ‘Middle grey’ and separating them from the encoded pair of cells with a ‘Change colour’ signal (the $|011\rangle$ cells).\ Again this is not quite optimal, but it does entail the following result. There exists $G'$ a $U$-defined QCA which is intrinsically universal QCA in the following sense. Let $G$ be $V$-defined QCA such that $V$ can be expressed as a quantum circuit $C$ made of $m$ gates acting upon $2n$ qubits. Then $G'$ is able to intrinsically simulate $G$ with space expansion factor $s=4nm+2+2n$ and time expansion factor $t=(3/2)s$. **Proof.** Each cell of the $V$-defined QCA will be encoded into a string of cells of the $U$-defined QCA which we have described in this Section, according to the formula ${| \Gamma(C,\phi,k) \rangle}$ given in this Subsection (remember that that the two cells $2k$ and $2k+1$ of a PQCA are really just the $k$ cell of the corresponding QCA in the original definition, as we explained in Subsection \[subsecstruc\]). This constitutes our isometric encoding $E$. The size of ${| \Gamma(C,\phi,k) \rangle}$ can be seen to be $4nm+2+2n$, which explains the value of the space expansion factor $s$. The ratio between the time expansion factor and the space expansion factor is deduced by inspection of Figure \[circuitry\].\ $\Box$ Note that if the scattering unitary $V$ is only approximated with an error of $\epsilon=\max_{{| \psi \rangle}}||V{| \psi \rangle}-C(V){| \psi \rangle}||$ by the quantum circuit $C(V)$, then this entails we are able to intrinsically approximate the evolution of $s$ cells over $t$ steps with an error bounded by $st\epsilon$ – again using supercells of size $s=4nm+2+2n$ and a time expansion of factor $t=(3/2)s$. This is the general statement that errors in quantum circuits grow no more than proportional to time and space [@Nielsen], which stems from the fact that if $||U-U'||\leq\epsilon$ then $||U^{\otimes s}-U'^{\otimes s}||\leq s\epsilon$ and $||U^t-U'^t||\leq t\epsilon$. Strong intrinsic universality {#strong} ============================= The idea of tackling the problem of strong intrinsic universality was suggested to us by Torsten Franz, whom we would like to thank. Why do we need another flavour? ------------------------------- The QCA explained in subsection  \[theqca\] has one problem: one needs to be able to prepare an infinite initial configuration somehow. If we do not do that, then the quiescent cells surrounding the background pattern will slowly begin to mix with the pattern, thus rendering it unusable. As can be seen in Fig.  \[dispersingbg\], after certain amount of time, the whole ternary background pattern will be reduced to a single cell, and no more computation is possible. Thus, in order to perform an arbitrary long computation, we have to know the length of the computation beforehand, so that we can prepare an initial configuration of adequate width and protect the ternary background pattern long enough for the computation to finish. But in practice finding out how long the computation is going to last is a very difficult task, and may even be undecidable. In order to solve this problem, we need a second flavour to the notion of intrinsic universality, namely strong intrinsic universality, as was formally presented and defined in Subsection \[subsecsim\]. Hence we shall build a new QCA that has the ability to weave the ternary background pattern as the computation proceeds. This section will be organized as follows: first we give an intuitive idea about the mechanisms we use to solve this problem, then we explain the different components of the QCA in detail, and finally we explain how the whole system fits together. ![The ternary background pattern will shrink as the computation goes on and the surrounding quiescent cells enter the computation region.[]{data-label="dispersingbg"}](circuitry.pdf) Intuition --------- In John Conway’s ‘Game of Life’ CA, there are objects called ‘guns’ that can fire cells in different directions at different time intervals. Here we use a similar idea in order to fire the signals that make up the ternary background pattern. Essentially, we use what we call a ‘copy band’, i.e. a structure made out of two walls and a synchronization signal bouncing between them, as in Fig. \[copyband\]. The distance between the walls controls the frequency of duplication, and the bouncer signal decides when to fire a piece of signal constituting the ternary background. Recall that this background was essentially made out of two signals: the mode (the **mode** subsystem is described in subsection  \[theqca\]), and the program (each program signal carries half of the information needed to code for a universal quantum gate, refer back to Table  \[Mgate\] to see how each gate is encoded as the state of two **program** subsystems). The left/right half of the coding of a universal quantum gate will hence be held by a left/right copy band. Since the fired mode and program signals travel at maximum speed, they appear on the space-time diagram of the QCA as having a 45/135-degree angle with respect to the horizontal line. So the two program signals each encoding one half of the program will eventually meet in the middle, inducing the application of a quantum gate. The program signals are being duplicated at a constant rate, thus creating the ‘squares’ in Fig.  \[ternary\]. Part of the difficulty is to determine what level of grey should be sent (the value of the mode signal). This will be explained further down. The speed of the computation, manifested in Fig.  \[ternary\] as the size of the ‘squares’, can be controlled by adjusting the distance between the walls and hence the frequency of the bounces. The copy band ------------- We continue with the convention described in Subsection \[conventions\], i.e. the names of the subsystems will be given in bold (Examples of subsystems are the **mode** and **program** subsystems in Table \[bandsubsystems2\]), and names of the signals will not (An example of a signal is the bouncer signal in Fig. \[copyband\]). A copy band is shown in Fig.  \[copyband\]. It is composed of a left wall, a right wall and the bouncer signal traveling in between. The subsystems used in order to implement the copy band are summarized in Table  \[bandsubsystems2\]. ![A band. Both walls of the band contain a counter signal, and the left wall also contains a copy of the program signal which will be fired later in order to make up the ternary background pattern. The bouncer signal can be seen bouncing between the walls, with different grey level corresponding to different values of this signal. The duplicated program signal propagates to the right at light speed. Time flows upwards.[]{data-label="copyband"}](copyband.pdf) ![Two copy bands placed side by side. The band will not alter a program signal that is already traveling at light speed, so the program signals emitted by the band to the left just pass through the band to the right. Time flows upwards.[]{data-label="twocopybands"}](doubleband.pdf) Name Size Function ------------- ------ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- **counter** 4 The counter guarantees that the ternary background pattern grows at a constant rate. It increments by one at each step, and whenever it gets to 3, the copy band will start to shift left (or symmetrically, right ) to transform one quiescent cell into part of the ternary background pattern. **wall** 3 Walls are what mark the border of a ‘copy band’, separates it from the quiescent cells and the ternary background pattern. Cells that are not part of any copy band have value $|0\rangle$ (i.e. quiescent cells, the ternary background pattern and data cells). The walls that are closer to the half-line of quiescent cells have value $|1\rangle$, they are referred to from now on as ‘outer walls’. The right walls in the left half of the QCA have value $|2\rangle$, they are referred to as ‘inner walls’ from now on. **bouncer** 4 The bouncer is a signal that travels inside the *copy band*, and changes its value whenever it hits a wall of value $|1\rangle$. Every time the bouncer bounces on the outer wall, a copy operation is performed on the **copy** subsystem. The resulting copy is put in the **program** subsystem, which travels at maximum speed to construct gates in the ternary background pattern. Together with the **history** subsystem, it enables calculation of the correct **mode** value for the fired ternary background signal. **history** 3 To calculate the correct mode of the duplicated program bit, a ‘history’ of how many mode-change signals have been encountered by this copy band is indispensable. Together with the **bouncer** subsystem, the correct mode can then be inferred. : Subsystems of the copy band[]{data-label="bandsubsystems2"} The program signal that encodes half the information of a quantum gate is kept in the outer wall. This wall also stores the counter, which is used to slow down the speed of the walls by counting from $1$ to $3$ and then letting the walls move right/left only when the value of the counter signal is $3$, the $0$ meaning that there is no counter signal. The bouncer signal is placed next to this wall in the initial configuration. When the computation begins, the bouncer signal will travel towards the opposite wall at maximum speed. When it hits the outer wall, its value will change. The value of the bouncer signal is essentially another counter that cycles from $1$ to $3$, the $0$ meaning that there is no bouncer signal. Hence each change of value is just the sequence of operation $-1,\,(+1\mod{3}),\, +1$, applied only when the value differs from $0$. For simplicity we refer to this operation as $Inc_3$. Whenever the bouncer signal hits the outer wall, the program/mode signals that constitute the background pattern will be fired. Namely, a copy of the value of the **copy** subsystem (cf. Table  \[allsubsystems\]) will be placed in the **program** subsystem, and the value of the **mode** subsystem will be set appropriately. The way this is done involves the value of the **counter** and the **history** subsystems, and will be explained in detail in subsection  \[modesync\]. Since each copy band only holds half the information needed to code a quantum gate in the ternary background pattern, another band is needed in order to reconstruct the gate. By exploiting the symmetry of the QCA depicted in subsection \[theqca\], we use two symmetrically placed bands to ‘weave’ one gate in the ternary background pattern, as shown in Fig. \[gates\] ![Universal quantum gates encoded by two symmetrically placed copy bands. Gates are formed when two duplicated programs traveling in opposite directions intersect. Seven gates can be seen in the graph.[]{data-label="gates"}](gates.pdf) Mode synchronization {#modesync} -------------------- In subsection  \[theqca\], all gates start in the same mode at time 0, and the mode cycles each time a ‘change colour’ signal is encountered. Since program signals start and travel at the same speed towards each other, it takes them an equal amount of time to meet in the middle, meeting the same number of ‘change colour’ signals on the way, thus making mode synchronization automatic. In the QCA described in this section, however, gates start in different modes at different times, and although they still travel at the same speed towards each other, they may very well meet different numbers of ‘change colour’ signals on the way, so when two program signals meet, their mode may be out of sync, which may wrongly render this gate inactive/active. To address the problem, we used the information stored in the **history** subsystem and value of the **bouncer** subsystem in order to compute the correct value for this mode signal. The bouncer signal decides when to fire the program signal and in which mode. It has three possible values, each corresponding to a value of the fired mode signal. The bouncer signal cycles through these values each time it ‘touches’ the outer wall (i.e. wall of value $|1\rangle$), at this time the program stored in the outer wall will also be duplicated and fired as a signal, with the current value of the bouncer signal as its default mode, so to say. The **history** subsystem stores the number of ‘change colour’ signals that a program signal ‘would have encountered’ if the signal had been sent at time 0. The effects of these ‘hypothetical’ signals cannot be ignored, we must keep track of them modulo $3$. So actually the value of the fired mode will be that of the bouncer \*plus\* the offset kept by **history**. This mechanism is illustrated in Figure \[sync\]. ![Synchronization of the mode of the fired ternary background signal. The initial value of the **history** subsystem should be the number of ‘change colour’ signals the copy band would have been encountered had the ternary background pattern been complete (i.e. had the ‘hypothetical’ signals actually been sent). In the depicted situation, the value of the **history** subsystem of the outer copy bands would be $1$ because the outer bands would have encountered a ‘change colour’ signal if the region occupied by the copy bands had been part of the initial configuration. []{data-label="sync"}](sync.pdf) Fitting things together {#strongfitting} ----------------------- ### Components and circuit overview Table \[allsubsystems\] summarizes the subsystems that make up a cell. In addition to the ones already mentioned in Table \[bandsubsystems\] and  \[bandsubsystems2\], there is also the **copy** subsystem, which is used to store the value of the program signal to be fired. Name Size Function ------------- ------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- **counter** 3 cf. Table \[bandsubsystems2\]. **wall** 3 cf. Table \[bandsubsystems2\]. **bouncer** 4 cf. Table \[bandsubsystems2\]. **history** 3 cf. Table \[bandsubsystems2\]. **data** 3 cf. Table \[bandsubsystems\]. **copy** 4 The system that carries half the information needed to construct a gate. This information is permanently stored in a copy band, and will be duplicated when a ternary background signal needs be fired. It has the same size as the **program** subsystem of subsection \[theqca\]. **program** 4 cf. Table \[bandsubsystems\]. **mode** 3 cf. Table \[bandsubsystems\]. : Subsystems of the QCA[]{data-label="allsubsystems"} The circuit of the QCA has a similar structure to the circuit depicted in Fig. \[Uascircuit\], albeit being more complex. To help visualize the circuit, it is divided into several different ‘layers’, with each layer performing one or two basic functions. A schematic diagram of the different layers is given in Fig. \[layers\]. We will layout the circuit diagrams for the different layers and explain their functions in detail below. We follow the same conventions in the detailed diagrams as in Fig. \[Uascircuit\] (with horizontal lines representing control lines, and different subsystems represented by different vertical lines, time flowing upwards. The **x**s denote subsystems of the left cell, and the **y**s denote subsystems of the right cell.) ![The schematic layer diagram of the QCA’s circuit. The **X** and **Y** lines denote the left and right cells undergoing the transformation. Time flows upwards.[]{data-label="layers"}](layers.pdf) ### Layers of the QCA’s circuit - The Copy Band Expansion Layer (Fig. \[expansion\_layer\]): This layer ‘expands’ the ternary background two new cells at a time (one cell one the left, one cell on the right) by shifting the walls of the copy bands outwards, converting quiescent cells into part of the background pattern. Because the **copy** and the **history** subsystems are stored in the walls, they follow the expansion of **counter** and **wall** subsystems. The **counter** subsystem controls the swapping of **copy**, **history** and **wall** subsystems. That is, the swap gates are applied whenever the value of **xcounter** or **ycounter** is $|3\rangle$. The $SC$ gate performs the function of swapping the value of **xcounter** and **ycounter** whenever one of them is $|3\rangle$. It is unitary because it is essentially just a permutation of the different base vectors of the subsystems upon which it acts. Its permutation table is given in Table \[sc\]. This layer also lets the bouncer signal travel and bounce, i.e. the values of **xbouncer** and **ybouncer** are swapped whenever the value of **xwall** and **ywall** are both $|0\rangle$ (i.e. the bouncer signal is traveling between the walls). Depending on the position of the copy band (i.e. the band is at the left or right side of the ternary background), only one $Inc_3$ gate will be applied. This gate is to cycle the value of the bouncer signal when the value of **xwall** or **ywall** is $|1\rangle$. ![The copy band expansion layer.[]{data-label="expansion_layer"}](expansion_layer.pdf) [|c|c|p[10cm]{}|]{}\ Input&Output\ $|30\rangle$&$\langle03|$\ $|03\rangle$&$\langle30|$\ $|31\rangle$&$\langle13|$\ $|13\rangle$&$\langle31|$\ $|32\rangle$&$\langle23|$\ $|23\rangle$&$\langle32|$\ $|ij\rangle (i\neq3,j\neq3)$&$\langle ij| (i\neq3,j\neq3)$\ - The Mode Computation Layer (Fig. \[mode\_layer\]): this layer implements the rest of the synchronization mechanism discussed in subsection \[modesync\]. Here both the **counter** and the **history** subsystems must be used to determine the correct mode of the fired program signal. The **bouncer** subsystem contains the ‘base mode’ of the program signal. We call it ‘base mode’ because in order to compute the correct mode, we must ‘offset’ the base mode by the value of the **history** subsystem. The function of the $CM$ gate is to ‘offset’ the base mode by the value of the **history** subsystem, giving a new mode and leaving the **history** subsystem unchanged. To understand this one must know that even though the ternary background signal is fired at regular time intervals by the outer wall of a copy band, its mode is only worked out when it crosses the inner wall and leaves out the copy band. Hence the $CM$ gate is only applied when the value of **xwall** is $|0\rangle$, the value of **ywall** is $|2\rangle$, to make sure that the gate is applied at the inner wall, and when there is a non-empty **xprogram** (to make sure that a ternary background signal is crossing this inner wall). The $CM$ gate shown in the figure only acts on half of the copy bands, namely the copy bands that fire program signals to the *right*. Because the QCA is symmetric, the $CM$ gate that acts on the other half can be obtained by replacing every $x$ in the figure with $y$. Another function played by this layer is the incrementation of the history. Recall that the value of **history** is incremented when the inner wall that holds the **history** subsystem meets a program signal with value $|4\rangle$ (cf. the **program** subsystem in Table \[bandsubsystems\]). But the inner wall may also meet the program signal fired by the outer wall of the same copy band. To differentiate these two program signals, we can use the bouncer signal: the program signal fired by the outer wall always travel with the bouncer signal, which is non-zero, until it acquires its mode and leaves the copy band. On the other hand, the program signal coming from the ‘outside’ of the copy band does not have a bouncer signal traveling with it. Contrary to the $CM$ gate, the $+1\mod{3}$ gate shown in the figure acts on copy bands which fire program signals to the *left*. The $+1\mod{3}$ gate is applied when the value of **xprogram** is $|4\rangle$, and the value of **xbouncer** is not zero (to make sure the ‘change colour’ signal comes from the outside of the copy band), and the value of **ywall** is $|2\rangle$ (to make sure we are applying the gate to the **yhistory** subsystem in the inner wall). ![Left half of the mode computation layer.[]{data-label="mode_layer"}](mode_layer.pdf) - The Program Duplication Layer (Fig. \[program\_layer\]): the function of the $CP$ gate in this layer is to change the value of **program** subsystem to the value stored in **copy** subsystem whenever a bouncer signal meets the outer wall that stores the **copy** subsystem. It is also responsible for incrementing the value of the bouncer signal. This figure also shows only the left half of the actual diagram. The $CP$ gate is applied when the value of **xwall** is $|0\rangle$, the value of **ywall** is $|1\rangle$ (to make sure the gate is applied to the outer wall), and a non-zero **xbouncer** (to make sure that the bouncer signal has reached the outer wall). ![Left half of the program duplication layer.[]{data-label="program_layer"}](program_layer.pdf) - The Displacement and Interaction Layer: this layer is identical to the scattering unitary $U$ of the weak intrinsic universal QCA described in subsection \[theqca\] (cf. Fig. \[Uascircuit\]). It only acts on the three subsystems given in Table \[bandsubsystems\] while leaving the other subsystems unchanged. It is responsible for moving the duplicated program and data around, and performing the computation by having the reconstructed gates act on the data. ### Initial state preparation {#initprep} We now turn our attention towards the question of how to set up the initial state of out strong intrinsic QCA in order to simulate some $V$-defined QCA. This will be very much alike what we did in Subsection \[weak\]. But for the QCA described in this section we must also set up the copy bands, and special attention must be paid to the initialization of the mode synchronization mechanisms. *Determining the speed of computation.* When setting up our strong intrinsic universal QCA to simulate a QCA having scattering unitary $V$, the first task is to determine the size $s$ of the supercells used for the encoding. This then determines how many steps $t$ of the simulating QCA are needed to simulate one step of the simulated QCA. Refer back to Subsection \[subsecsim\] for a formal definition of this $s$ and $t$.\ This step is very much identical to that of Subsection \[weakresults\]. Again the key elements in order to work out $s$ are the sizes of the quantum circuit description $C$ of the scattering unitary $V$, namely $2n$ the number of qubits upon which $C$ acts and $m$ the number of elementary universal quantum gates which make up $C$. Because each simulated cell will be encoded into the simulating QCA via just the same isometric coding $E$ (with all of the extra subsystems left in ${| 0 \rangle}$), we have $s=4nm+2+2n$ and $t=(3/2)s$ once more.\ But then the number, width and the separation of the copy bands needs be worked out too. The number of copy bands is easily found: since two copy bands are necessary to encode one gate, there are $4mn+2$ ($4mn$ for program signals and $2$ for change colour signals) copy bands in total. These $4mn+2$ copy bands are placed symmetrically at the left and right side of the data, with each side containing $2mn+1$ copy bands.\ The rate at which the ternary background signals are fired is equal to the distance traveled by the bouncer signal from the inner wall to the outer wall (refer to Figure \[copyband\] for a visualization). We want this rate to be $s/2$, because there are $s/2$ ternary background signals to be fired on each side. Let us call $d$ the initial value of the width of each band (the distance between the inner wall and the outer wall, excluding walls). Clearly this is what determines the rate at which the ternary background signals are fired, yet this rate is not just $d$, because in traveling this distance, the bouncer signal has to ‘catch up’ with the outer wall which is moving at speed one third, and hence $$\begin{aligned} \frac{s}{2}&=d+\frac{1}{3}d+\frac{1}{9}d+\frac{1}{27}d+\ldots=\frac{1}{1-(1/3)}d=\frac{3}{2}d\\ \textrm{so}\quad d&=s/3=\frac{1}{3}(4mn+2+2n)\end{aligned}$$ Here the fact that $\frac{1}{3}s$ may not be an integer arises from the flexibility of the initial configuration, namely the flexibility in choosing initial values of the **counter** subsystem. As a result, it is possible to find two values of $d$, one equals to $\lfloor\frac{1}{3}s\rfloor$ and the other one equals to $\lceil\frac{1}{3}s\rceil$, by adjusting the initial values of the **counter** subsystem in the inner and outer walls, that produce the same result (i.e. correspond to the same value of $s$). Note however, the choice of $d$ and the initial values of the **counter** subsystem is not arbitrary, because the current neighborhood scheme of the QCA may render the walls moving in the wrong direction. If all the **counter**s start in the same value, we can fix the value of $d$ by round $\frac{1}{3}s$ to the closest integer. The separation between different copy bands just needs to be done by leaving one quiescent cell in between them, so that the proximity of each others wall do not interfere with their functioning.\ Now by multiplying the widths and separations of the bands together with the number of bands, we obtain the size of the $r$ parameter of the strong isometric coding, as was formally defined Subsection \[subsecsim\]. In our scheme we have: $$r=(\frac{1}{3}s+2+1)*(2mn+1)=(\frac{1}{3}(4mn+2+2n)+3)*(2nm+1)$$\ Notice a nice feature of our construction: our copy bands move apart at speed $1/3$, but the expansion rate of the simulated QCA is at most $s/2t$, which is also $1/3$. Hence the ternary background pattern, which is being constantly ‘weaved’ by our copy bands, grows fast enough so that the data it holds never ‘leak’ out of it. This point is illustrated in Figure \[seededqca\]. *Prepare the initial bouncer and history value.* The next step would be to figure out the initial **bouncer** value, and the corresponding **history** value. The best way to do this is to use a fixed **bouncer** value for all copy bands, and use the technique of subsection \[modesync\] to trace the ‘hypothetical’ change colour signals, and calculate the appropriate **history** value for each band by hand. A complete example of the strong intrinsic universal QCA is shown in Fig. \[seededqca\]. It has four program bits, hence four copy bands at each side. The bands are numbered 1 to 8 from left to right, and their initial configurations are shown in Table \[initconf\]. This figure does not show the actual interaction between gates and data, so data just ‘flow through’ the ternary background pattern. Band Configuration ------ ---------------------------------------------------------------------------------------------------- -- 1 $|30130000\rangle$, $|00001300\rangle$, $|00000000\rangle$, $|00000000\rangle$, $|30200001\rangle$ 2 $|30130000\rangle$, $|00001300\rangle$, $|00000000\rangle$, $|00000000\rangle$, $|30200421\rangle$ 3 $|30130000\rangle$, $|00001300\rangle$, $|00000000\rangle$, $|00000000\rangle$, $|30200330\rangle$ 4 $|30110420\rangle$, $|00003400\rangle$, $|00000000\rangle$, $|00000000\rangle$, $|30200322\rangle$ 5 $|30200322\rangle$, $|00000000\rangle$, $|00000000\rangle$, $|00003400\rangle$, $|30110410\rangle$ 6 $|30200310\rangle$, $|00000000\rangle$, $|00000000\rangle$, $|00001300\rangle$, $|30130000\rangle$ 7 $|30200411\rangle$, $|00000000\rangle$, $|00000000\rangle$, $|00001300\rangle$, $|30130000\rangle$ 8 $|30200001\rangle$, $|00000000\rangle$, $|00000000\rangle$, $|00001300\rangle$, $|30130000\rangle$ : Initial configurations of each copy band. The subsystems in each cell is ordered by **$|counter|data|wall|copy|bouncer|program|mode|history\rangle$**[]{data-label="initconf"} ![A complete example of the QCA. The ternary background pattern is produced by the four pairs of copy bands. Two data bits are shown as two white lines propagating through the ternary background pattern. In this example no quantum gates act on the data bits. Time flows upwards.[]{data-label="seededqca"}](seededqca.pdf) Formal results -------------- The strong intrinsic universal QCA shares many formal properties with the weak one. The way that universal quantum gates are constructed and addressed, as described in subsection \[weakresults\], can be applied directly to the strong intrinsic universal QCA. The additional structures, namely the copy bands, correspond to the $L$ and $R$ isometries in Definition \[strongisomcode\]. With the construction details described in this section, we have shown the following theorem: There exists $G'$ a $U$-defined QCA which is strongly intrinsically universal QCA in the following sense. Let $G$ be $V$-defined QCA such that $V$ can be expressed as a quantum circuit $C$ made of $m$ gates acting upon $2n$ qubits. Then $G'$ is able to strongly intrinsically simulate $G$ with space expansion factor $s=4nm+2+2n$, time expansion factor $t=(3/2)s$. Moreover if the initial configuration we want to simulate is of size $x$, then the initial configuration of the simulating configuration is of size $sx+2r$ with $r=(\frac{1}{3}(4mn+2+2n)+3)*(2nm+1)$. **Proof.** The $E$ encoding is the same as the one in Subsection \[weakresults\], and the $L$ and $R$ are the constructions of the let and right ‘copy bands’ described in this section. Their different sizes have been worked out previously.\ $\Box$ Conclusion ========== *Main claim and future work.* We have formalized the notion of a QCA capable of simulating all others with linear overhead, exactly if the scattering unitaries they are made of decompose into a circuit of elementary quantum gates, and approximately otherwise. We have constructed such an intrinsic universal QCA, which turns out to be a Partitioned QCA (Figure \[structure\]) of cell-dimension $36$ and whose scattering unitary we have given explicitly (Figure \[Uascircuit\] and Subsection \[theqca\]). If we insist that the initial configuration of the simulating QCA be finite (and not just periodic) we get to a stronger notion of universality, which was also formalized here. We have also constructed such a strong intrinsic universal QCA, which turns out to be a Partitioned QCA (Figure \[structure\]) of cell-dimension $15 552$ and whose scattering unitary we have given explicitly (Figures \[layers\] to \[program\_layer\] and Subsection \[strongfitting\]).\ Clearly the main challenge we now face is to find an intrinsically universal QCA in $n>1$ dimensions. The construction described in this paper is unlikely to be useful then, as it seems to blow up in complexity. But surprising as it may seem, in the classical reversible cellular automata literature intrinsically universal $n$-dimensional reversible cellular automata turn out have appeared before the intrinsically one-dimensional reversible cellular automata [@Ollinger2]. The common belief in the classical CA community seems to be that the $n$-dimensional constructions are actually simpler, due to the ability to use up the more than one-dimensional space to draw up circuits and sort out wire cross-overs. Intuitively at least, in $n$-dimensions intrinsic universality is very much the same as circuit universality. This is the direction which we plan to take in the close future. Acknowledgements {#acknowledgements .unnumbered} ================ P.J.A would like to thank Torsten Franz, Miguel Lezama and Philippe Jorrand for their support. [99]{} J. Albert, K. Culik, *A simple universal cellular automaton and its one-way totalistic version*, Complex Systems, $\mathbf{1}$, 1-16, (1987). P. Arrighi, *An algebraic study of one-dimensional quantum cellular automata*, Proceedings of MFCS 2006, LNCS 4162, 122–133, (2006). P. Arrighi, V. Nesme, R. Werner, [*Quantum cellular automata upon finite, unbounded configurations*]{}, LATA’08, Proceedings to be published in LNCS. Pre-print arXiv:0711.3517. Pablo Arrighi, Vincent Nesme, Reinhard Werner, [*N-dimensional quantum cellular automata*]{}, Pre-print arXiv:0711.3975. R. B. Banks, *Universality in cellular automata*, Symposium on Switching and Automata Theory, (Santa Monica, California, 1970), IEEE, 194–215, (1970). P. Boykin, T. Mor, M. Pulver, V. Roychowdhury, F. Vatan, *On universal and fault-tolerant quantum computing*, arxiv:quant-ph/9906054 D. Cheung, C. A. Perez-Delgado, *Local Unitary Quantum Cellular Automata*, arXiv:/0709.0006. M. Delorme, *An introduction to cellular automata*, in ‘Cellular Automata: A Parallel Model’, Kluwer, (1999). J. Durand-Lose, *Reversible Cellular Automaton Able to Simulate Any Other Reversible One Using Partitioning Automata*, Proc. of LATIN ’95, LNCS **911**, 230-244, (1995). J. Durand-Lose, *Intrinsic universality of a 1-dimensional reversible cellular automaton*, STACS 97, LNCS $\mathbf{1200}$, 439–450, (1997). J. Durand-Lose, *Representing Reversible Cellular Automata with Reversible Block Cellular Automata*, DM-CCG’01, Discrete Mathematics and Theoretical Computer Science Proceedings, **AA**, 145-154, (2001). C. Dürr, H. LêThanh, M. Santha, *A decision procedure for well formed quantum cellular automata*, Random Structures and Algorithms, $\mathbf{11}$, 381–394, (1997). C. Dürr, M. Santha, *A decision procedure for unitary quantum linear cellular automata*, SIAM J. of Computing, $\mathbf{31}(4)$, 1076–1089, (2002). R. P. Feynman, *Quantum mechanical computers*, Found. Phys. $\mathbf{16}$, 507-531, (1986). D. Meyer, *Unitarity in one dimensional nonlinear quantum cellular automata*, arXiv:quant-ph/9605023, (1995). M.A. Nielsen, I.L. Chuang, *Quantum computation and quantum information*, Cambridge University Press, 194-195, (2000). N. Ollinger, *Automates cellulaires: structures.*, Ph.D. Thesis at ENS-Lyon, (2002). N. Ollinger, *Universalities in Cellular Automata: A short survey.*, JAC 2008. J. Kari, *Representation of Reversible Cellular Automata with Block Permutations*, Theory of Computing Systems, **29**, 47-61, (1996). J. Kari, *On the Circuit Depth of Structurally Reversible Cellular Automata*, Fundamenta Informaticae, **38**(1-2), 93-107, (1999). R. Raussendorf, *Quantum cellular automaton for universal quantum computation*, Phys. Rev. A, $\mathbf{72}$, 022301, (2005). B. Schumacher, R. F. Werner, *Reversible quantum cellular automata*, arXiv:quant-ph/0405174. D. J. Shepherd, T. Franz, R. F. Werner, *Universally programmable quantum cellular automata*, Phys. Rev. Lett., $\mathbf{97}$, 020502 (2006). T. Toffoli, *Computation and construction universality of reversible cellular automata*, J. of Computer and System Sciences, $\mathbf{15}(2)$, (1977). T. Toffoli, N. Margolus, *Cellular Automata Machine – A New Environment for Modelling*, MIT Press, Cambridge MA, (1987). T. Toffoli, N. Margolus, *Invertible Cellular Automata : A Review*, Physica D, **45**, 229-253, (1990). W. Van Dam, *Quantum Cellular Automata*, Master thesis, Department of Mathematics and Computer Science, University of Nijmegen, The Netherlands, (1996). J. Watrous, *On one-dimensional quantum cellular automata*, Complex Systems $\mathbf{5}(1)$, 19–30, (1991).
{ "pile_set_name": "ArXiv" }
--- author: - | [^1]\ $^1$JIVE, Dwingeloo, Netherlands\ $^2$MTA Research Group for Physical Geodesy and Geodynamics, Penc, Hungary\ E-mail: - | Sándor Frey\ $^1$FÖMI Satellite Geodetic Observatory\ $^2$MTA Research Group for Physical Geodesy and Geodynamics, Penc, Hungary\ E-mail: - | Bob Campbell\ JIVE, Dwingeloo, Netherlands\ E-mail: - | Attila Moór\ MTA Konkoly Observatory, Budapest, Hungary\ E-mail: title: 'A suspected Dark Lens revealed with the e-EVN' --- A Dark Lens candidate ===================== In a campaign to study the optical properties of faint VLA FIRST (Faint Images of the Radio Sky at Twenty-Centimeters) sources with the Hubble Space Telescope, Russell et al. \[\] identified an optical arc 4 arcseconds away from FIRST J121839.7 +295325 (hereafter J1218+2953). There was no optical identification of the radio source itself. The possible relation between the two objects was further investigated by Ryan et al. \[\], who proposed that the arc may be a gravitationally lensed image and the radio source may belong to the lensing object. There are two major difficulties with this interpretation. First of all the lensing galaxy is not seen. Moreover the redshifts of the radio source and the optical arc are not known. Ryan et al. \[\] estimated a redshift range of $0.8<z<1.5$ for the radio source on the basis that the galaxy is not seen in the optical ($z>0.8$), but it has a relatively bright radio flux density ($z<1.5$). They carried out photometric redshift measurements of the optical arc with the SAO Multi-Mirror Telescope. A weak spectral break was seen at around 4300 Å, which could be due to the Balmer/4000 Å break at $z\sim0.13$, or due to the Lyman-break at $z\sim2.5$. Probability analysis showed that the Lyman-break is much more likely, which suggests that the arc is located at a high redhsift. Obviously, if the spectral break is related to the Balmer series then the redshift is much smaller and the two objects are unrelated. Ryan et al. \[\] carried out gravitational lens modelling and found that the mass distribution must be elliptical to produce such a long arc. However the observed sub-structure within the arc, and especially a bright knot at the end, cannot be easily explained by gravitational lensing. They predict a secondary image which appears close to a pointlike source in the HST image, but this is likely due to a “hot pixel”. According to the model, the enclosed mass within the Einstein-radius of 1.3 arcseconds is $10^{12\pm0.5}M_{\odot}$. It is not clear how such a massive object may remain hidden. Even if there is strong obscuration by dust, there should be an infrared counterpart detected. IR imaging with the SAO Wide-field camera showed no galaxy with limiting magnitudes of $J=22.0$mag and $H=20.7$mag \[\]. Ryan et al. conclude that either there is an early-type galaxy with significant amount of dark matter, or this could be a massive system with an AGN that is completely obscured, with dynamic mass-to-light ratio exceeding 100 $M_{\odot}L_{\odot}^{-1}$. VLBI imaging of FIRST J1218+2953 ================================ ![image](SHORT.ps){width="7cm"} ![image](UVPLOT.ps){width="7cm"} Short e-EVN observations ------------------------ We applied for short e-EVN observations in December 2008, to test the AGN hypothesis for J1218+2953. The total flux density of the source at 1.6 GHz is 33 mJy, which can be easily detected with limited EVN resources if it is compact. The observations took place on 23 January 2009 at a data rate of 512 Mbps with 2-bit sampling, dual polarization, and lasted for 2 hours. The array consisted of Cambridge and Knockin (MERLIN telescopes, limited to 128 Mbps), Jodrell Bank MkII, Medicina, Onsala, Torun and the Westerbork Synthesis Radio Telescope (WSRT). The target was phase-referenced to the nearby calibrator J1217+3007. The data were pipelined and then imaged in Difmap \[\]. The source was detected and was resolved to two components, separated by about 500 mas (see Fig. \[short\]). This result opened up new possibilities for the interpretation of the radio source. Although it was not predicted from the “best-fit” gravitational lens model of Ryan et al., one may speculate that these two components are gravitationally lensed images of the same background source that gives rise to the optical arc images, or perhaps a completely unrelated background source. Alternatively, we may see a core-jet system or a medium-size symmetric object (MSO). To distinguish between these scenarios we put in an observing proposal by the 1 February 2009 deadline (just 8 days following the e-EVN observations), for full-track e-EVN observations at 5 and 1.6 GHz. Follow-up experiments --------------------- J1218+2953 was observed at 5 GHz on 24-25 March for 8 hours. The array this time included the 100m Effelsberg telescope as well. Four telescopes (Ef, On, Tr, Wb) sent data to the correlator at 1024 Mbps rate, the rest at 512 Mbps or lower. The 1.6 GHz observations were carried out on 21-22 April 2009 at 512 Mbps data rate. In the array Knockin was replaced by Darnhall, the 76m Lovell Telescope was used instead of the MkII in Jodrell Bank, and Arecibo joined as well (for 2 hours and 20 minutes). These observations and the data processing were similar to the short project described above. In addition, we reduced the synthesis array data from the WSRT that was obtained during the VLBI run. The 5 GHz image resolves the South-East component into an elongated, slightly curved jet-like structure, which does not point towards the North-West component (see Fig. \[fulltracks\]). There is no very compact component that could be firmly identified as a core. The total cleaned flux density is only 3 mJy compared to the total WSRT flux density of 9 mJy, indicating that most of the source is resolved out in this image. The 1.6 GHz image reveals an even more complex, but more continuous structure. The total cleaned flux density was about 20 mJy, close to the total WSRT flux density of 27 mJy. Interpretation of the results ============================= These preliminary e-EVN results show that the radio source near the optical arc has a complex structure. The spectrum of the components is steep; that of the faint component near the phase centre is somewhat flatter. Because of the apparent quasi-continuous structure, the various components are likely not gravitationally lensed images of an unrelated background source. Comparing the total cleaned flux density of about 20 mJy to the WSRT flux density of 27 mJy, it is evident that most of the flux density is recovered at 1.6 GHz with the EVN. This indicates that the radio source and the optical arc cannot be lensed image pairs of the same background object, because in that case most of the radio flux density would be present near the optical arc since that image is strongly magnified (if the arc is indeed a lensed image). Further, gravitational lensing should be achromatic; thus were the pair of components to the SE and NW in the 1.6 GHz image (Fig. \[fulltracks\], bottom panel) lensed images, the flux density ratio of the inner:outer components of each should be similar, a condition that is clearly violated. The most likely scenario is that the images show a single compact steep-spectrum (CSS) source (projected linear size &lt; 20 $h^{-1}$ kpc), that might be categorized as a medium-size symmetric object (MSO, projected linear size &gt; 1 $h^{-1}$ kpc) as well \[\]. There is thus evidence for AGN activity in the radio source. Note that if the optical arc is lensed by an object related to this AGN, then the mass-centre of the lens should be at the position of the AGN, which constrains further modelling of the system. Finally we note that this research benefited strongly from two aspects of the (e-)EVN: the additional short MERLIN spacings were of great use in recovering flux density on several-hundred mas scales, and the simultaneous recording of WSRT synthesis array data was very important for the intepretation of our results. ![image](5GHz_map.ps){width="9.5cm"} ![image](1.6GHz_map.ps){width="9.5cm"} [99]{} J. Russell, R.E. Ryan Jr., S.H. Cohen, R.A. Windhorst, I. Waddington, *Optical morphologies of millijansky radio galaxies observed by the Hubble Space Telescope and in the Very Large Array FIRST survey*, *Astrophys. J. Suppl.* [**179**]{} (2008) 306 E. Ryan Jr., S.H. Cohen, R.A. Windhorst, C.R. Keeton, T.J. Veach, *Is the optically unidentified radio source FIRST J121839.7+295325 a Dark Lens?*, *Astrophys. J.* [**688**]{} (2008) 43 M.C. Shepherd, T.J. Pearson, G.B. Taylor, *DIFMAP: an interactive program for synthesis imaging*, *BAAS* [**26**]{} (2008) 987 C. Fanti, R. Fanti, D. Dallacasa, R.T. Schilizzi, R.E. Spencer, C. Stanghellini, *Are compact steep-spectrum sources young?*, *Astron. Astrophys.* [**302**]{} (2008) 317 [^1]: e-VLBI developments in Europe are supported by the EC DG-INFSO funded Communication Network Developments project ’EXPReS’, Contract No. 02662. The European VLBI Network is a joint facility of European, Chinese, South African and other radio astronomy institutes funded by their national research councils. The Westerbork Synthesis Radio Telescope is operated by the ASTRON (Netherlands Institute for Radio Astronomy) with support from the Netherlands Foundation for Scientific Research (NWO). SF and ZP acknowledge partial support from the Hungarian Scientific Research Fund (OTKA, grant no. K72515).
{ "pile_set_name": "ArXiv" }
--- abstract: 'The package CosmoLib is a combination of a cosmological Boltzmann code and a simulation toolkit to forecast the constraints on cosmological parameters from future observations. In this paper we describe the released linear-order part of the package. We discuss the stability and performance of the Boltzmann code. This is written in Newtonian gauge and including dark energy perturbations. In CosmoLib the integrator that computes the CMB angular power spectrum is optimized for a $\ell$-by-$\ell$ brute-force integration, which is useful for studying inflationary models predicting sharp features in the primordial power spectrum of metric fluctuations. As an application, CosmoLib is used to study the axion monodromy inflation model that predicts cosine oscillations in the primordial power spectrum. In contrast to the previous studies by Aich [*et al*]{} and Meerburg [*et al*]{}, we found no detection or hint of the osicllations. We pointed out that the CAMB code modified by Aich [*et al*]{} does not have sufficient numerical accuracy. CosmoLib and its documentation are available at <http://www.cita.utoronto.ca/~zqhuang/CosmoLib>.' author: - Zhiqi Huang title: 'A Cosmology Forecast Toolkit – CosmoLib' --- Introduction ============ The hot big bang model and the cosmological perturbation theory, where the physical metric is perturbed around the spatially homogeneous and isotropic Friedmann-Robertson-Walker (FRW) metric [@Peebles:1980; @Peacock:1999; @Mukhanov:2005; @Weinberg:2008a], have led to a remarkable success in interpreting the plethora of observational data of the last two decades [@Riess/etal:2011; @Amanullah/etal:2010; @Sullivan/etal:2011; @Reid/etal:2010; @Percival/etal:2010; @Komatsu/etal:2011]. Observations of the temperature anisotropy in the Cosmic Microwave Background (CMB) have been playing an essential role in building the standard cosmological model and measuring its parameters [@Komatsu/etal:2011; @Fixsen/etal:1996]. In order to maximize the usage of the observational data, one would like to compute the theoretical prediction on the CMB anisotropy for a given model as accurately as possible, with tolerable time consumption. Computation tools developed over the years such as CMBFAST [@Seljak/Zaldarriaga:1996], CAMB [@Lewis/etal:2000], CMBEASY [@Doran:2005] and CLASS [@Lesgourgues:2011; @Blas/etal:2011] are capable of computing a CMB angular power spectrum to percent-level accuracy within a few seconds on a modern desktop personal computer. The crucial technique used in all the fast CMB codes to date is the line-of-sight integration approach [@Seljak/Zaldarriaga:1996; @Hu/White:1997; @Hu/etal:1998] and an assumption that the primordial power spectrum of metric perturbations is smooth. (Here for readability we focus on the comoving curvature perturbations and temperature anisotropies, although the same arguments can be as well applied to the tensor perturbations and CMB polarization.) A CMB code first computes the radiation transfer function $\Delta_\ell^k$ by solving the linear-order Boltzmann equations and using the line-of-sight integration method, then convolves $|\Delta_\ell^k|^2$ with the primordial power spectrum $\mathcal{P}(k)$ to obtain the CMB angular power spectrum $C_\ell$. The smoothness assumption allow us to compute only a few tens of multipoles spanning from $\ell_{\min} = 2$ to $\ell_{\max}\sim$ a few thousands and interpolate the remaining $C_\ell$’s. Furthermore, since $\mathcal{P}(k)$ is assumed to a smooth function, sparse sampling of the radiation transfer function has been implemented for the integration of each $C_\ell$. A smooth primordial power spectrum is a prediction of the simplest single-field slow-roll inflation models [@Abbott/Wise:1984; @Lucchin/Matarrese:1985; @Lucchin/Matarrese:1985a; @Lyth/Stewart:1992; @Stewart/Lyth:1993]. However, local signatures in the primordial power spectrum that makes it deviate from smoothness can arise in various alternative models, for instance, when the inflaton potential has sharp features [@Starobinsky:1992; @Chen/etal:2007], when there is a transition between different stages in the inflaton evolution [@Contaldi/etal:2003; @Cline/etal:2003], when more than one field is present [@Polarski/Starobinsky:1992; @Langlois/Vernizzi:2005], from particle production during inflation [@Barnaby/etal:2009a; @Barnaby/Huang:2009], modulated preheating  [@Chambers/Rajantie:2008; @Bond/etal:2009], or, more recently, in models motivated by monodromy in the extra dimensions [@Silverstein/Westphal:2008] (see also [@Bean/etal:2008]). These features represent an important window on new physics because they are often related to UV scale phenomena inaccessible to experiments in the laboratory. For these models, the CMB angular power spectrum is not necessarily smooth, and therefore needs to be computed at each multipole without interpolation. This increases the computing time by a factor of a few tens. Moreover, for the numerical-integration of each $C_\ell$, the sampling frequency in the wavenumber $k$ often needs to be increased, again, by a factor of a few tens. The required sampling frequency in $k$ is model-dependent. It is determined by the larger between the minimum width of the features in the primordial power spectrum and the minimum width of the oscillations in the radiation transfer function. To keep track of the features in the primordial power spectrum, one can modify standard CMB codes by naively doing an $\ell$-by-$\ell$ brute-force calculation with increased integration sampling frequency in $k$. However, in the case where the features in the primordial power spectrum are really sharp ($\delta\ln k \lesssim 0.01$), this naive modification increases the computing time by a factor of $\sim 10^3$ (a few tens in $\ell$ sampling times a few tens in $k$ sampling). Moreover, the memory that is required to store all the transfer functions and tables of spherical Bessel functions can be too large for most desktop personal computers. One of the purposes of this paper is to introduce a more optimized algorithm to treat these problems. In fact, apart from increasing the sampling frequency, that cannot be avoided, all the other problems can be significantly alleviated by using the recurrence relation of spherical Bessel functions. An optimized algorithm, which we detail in Section \[sec:cmb\], is $\lesssim 10^2$ times slower than the standard algorithm for the smooth-$\mathcal{P}(k)$ case. This new algorithm has been implemented in the CosmoLib package, a self-contained package that we developed to compute cosmological perturbations, CMB angular power spectra, and the forecast constraints on cosmological parameters from future cosmological surveys using Fisher matrix analysis and Monte Carlo Markov Chain (MCMC) calculation. In particular, the cosmological surveys that we consider are CMB, large scale structure (LSS) and supernovae (SN). In addition to the enhanced CMB integrator, CosmoLib has a few other features that are complementary to the other publicly available Boltzmann/CMB/MCMC codes. For instance, the MCMC engine in CosmoLib has a modified rejection rule that allows the proposal density (the probability of random-walking to a new point in the parameter space) to periodically depend some parameters. This is useful when one considers a likelihood that depends on some periodic parameter. This happens, for instance, in the context of inflation from axion monodromy [@Silverstein/Westphal:2008; @Chen/etal:2008; @McAllister/etal:2010; @Flauger/etal:2010], where the oscillations in the predicted power spectrum depend on a free phase. Moreover, CosmoLib treats the dark energy equation of state (EOS) $w(a)$ and the primordial scalar and tensor power spectra $\mathcal{P}_s(k)$ and $\mathcal{P}_t(k)$, as free functions, which can be either chosen from a list of build-in models or defined by the user. This makes CosmoLib a convenient tool to study non-standard parametrizations of dark energy EOS and primordial power spectra. Finally, CosmoLib is written in Newtonian gauge (also called Poisson gauge) [@Mukhanov/etal:1992; @Ma/Bertschinger:1995; @Bertschinger:1996], while many other codes are mainly developed in synchronous gauge (see e.g. [@Peacock:1999]). This is a plus-and-minus point. We found that our Newtonian-gauge Boltzmann code is slightly slower than the codes written in synchronous gauge. However, many theoretical works in the literature have derived equations in Newtonian gauge. For instance, second-order Boltzmann equations have been derived in this gauge [@Bruni/etal:1997; @Bartolo/etal:2006; @Bartolo/etal:2007; @Nakamura:2007; @Enqvist/etal:2007; @Nakamura:2009; @Malik/Wands:2009; @Boubekeur/etal:2009; @Pitrou:2009; @Nitta/etal:2009; @Senatore/etal:2009; @Nakamura:2010; @Beneke/Fidler:2010; @Bernardeau/etal:2011]. Implementing these equations in a code already in Newtonian gauge would be much easier. To conclude the discussion, we list the differences between CosmoLib and other publicly available CMB codes in Table \[tbl:codes\]. ----------------------------- ----------- ------------------ ----------------- ------------- --------------- CAMB CLASS CMBEASY CMBquick CosmoLib [^2] Language F90 C C++ Mathematica F90[^3] gauge [^4] syn. syn./Newt. [^5] syn./gauge-inv. Newt. Newt. open/close universe Yes No No No No massive neutrinos Yes Yes Yes Yes No tensor perturb. Yes Yes Yes Yes Yes CDM isocurvature mode Yes Yes Yes Yes Yes dark energy perturb. Yes Yes Yes No Yes nonzero $c^2_{s,b}$ Yes Yes Yes No Yes dark energy EOS. constant   $w_0+w_a(1-a)$   arbitrary -1 arbitrary non-smooth primordial power No No No No Yes MCMC driver Yes No Yes No Yes periodic proposal density No No No No Yes data simulation No No No No Yes second-order perturb. [^6] No No No Yes No [^7] ----------------------------- ----------- ------------------ ----------------- ------------- --------------- : Comparison between CMB Codes [^1] \[tbl:codes\] As an application, CosmoLib is used to study the “hints” of cosine osicllations in the primordial power spectrum that was recently found in Refs. [@Aich/etal:2011; @Meerburg/etal:2012]. In an accompanying paper [@Huang/etal:2012], CosmoLib is applied to forecast the constraining power of future CMB and galaxy survey data on the primordial power spectrum from inflation, with an emphasis on models generating features in the power spectrum. This paper is organized as follows. In Section \[sec:perturb\] we introduce the Boltzmann code in Newtonian gauge and discuss its stability and performance. Section \[sec:cmb\] details the algorithm used in the enhanced CMB integrator. In Section \[sec:forecast\] we introduce the forecast technique and parameter estimation methods. Section \[sec:conclusion\] concludes. Throughout this paper, unless otherwise specified, repeated indices are summed over. Greek indices run from 0 to 3. Latin indices run from 1 to 3, that is only over spatial dimensions. We use natural units $c=\hbar = 1$ and the reduced Planck Mass $M_p\equiv 1/\sqrt{8\pi G_N} = 2.43 \times 10^{18} {\rm GeV}$. CosmoLib in Newtonian Gauge \[sec:perturb\] =========================================== The Background Solutions ------------------------ Let us start discussing the background solutions. We consider a flat FRW metric $ds^2 = a^2(\tau)(-d\tau^2 + dx^idx^i) $, where $a$ is the scale factor and $\tau$ is the conformal time. The normalization of $a$ is arbitrary. We normalize it such that $a=1$ today. CosmoLib uses the e-fold number $N \equiv \ln a$ as the time variable. The physical Hubble expansion rate is defined as $H \equiv \frac{da/d\tau}{a^2}$. Its present value is denoted by $H_0 \equiv 100 h {\,\rm km\,} {\rm s}^{-1} {\,\rm Mpc}^{-1}$. We assume a universe with cold dark matter (labeled with a subscript $c$), dark energy (labeled with a subscript $\Lambda$), baryons (labeled with a subscriber $b$), radiation (labeled with a subscript $\gamma$), and 3 species of massless neutrinos (labeled with a subscript $\nu$). For a component $X$ ($X= b, c, \gamma, \nu, \Lambda$) the background density is denoted as $\rho_X$, and the background pressure $p_X$. The present-day fractional energy density is written as $\Omega_{X0}$ . Dark energy is assumed to be a perfect fluid with known equation of state (pressure to density ratio) $w(a)$ and a constant sound speed $c_{s,\Lambda}^2$ in its rest frame. The users can either choose $w(a)$ from a list of build-in models or define their own $w(a)$ functions. The build-in models of $w(a)$ include the cosmological constant model $w(a)=-1$ [@Einstein:1917], a constant EOS $w(a)=w_0$, a linear function $w(a)=w_0+w_a(1-a)$ [@Chevallier/Polarski:2001; @Linder:2003], and a general three-parameter parametrization for the minimally coupled quintessence/phantom models [@Huang/etal:2011]. For a given $w(a)$ the background solutions are $$\begin{split} a &= e^N, \\ \rho_c &= 3 H_0^2M_p^2 \Omega_{c0} a^{-3}\;,\; p_c = 0\;, \\ \rho_b &= 3 H_0^2M_p^2 \Omega_{b0} a^{-3}\;,\; p_b = 0\;, \\ \rho_\gamma &= 3 H_0^2M_p^2 \Omega_{\gamma 0} a^{-4}\;,\; p_\gamma = \frac{1}{3} \rho_\gamma \; , \\ \rho_\nu &= 3 H_0^2M_p^2 \Omega_{\nu 0} a^{-4}\;,\; p_\nu = \frac{1}{3} \rho_\nu\;, \\ \rho_\Lambda &= 3 H_0^2M_p^2 \Omega_{\Lambda 0} a^{-3} \exp{\left[3\int_N^0 w(a) dN\right]}\;,\; p_\Lambda = w(a) \rho_\Lambda\; ,\\ H &= \frac{1}{M_p}\sqrt{\frac{\rho_c+\rho_b+\rho_\gamma +\rho_\nu+\rho_\Lambda }{3}}\; . \end{split}$$ We will also use the derived quantities $\Omega_X(a) \equiv \rho_X/(3H^2M_p^2)$ ($X= b, c, \gamma, \nu, \Lambda$), $R\equiv (3\rho_b)/(4\rho_\gamma)$, and $$\epsilon = -\frac{d\ln H}{dN} = \frac{3}{2}\left[1+\frac{p_\Lambda + p_\gamma + p_\nu}{\rho_c+\rho_b+\rho_\Lambda+\rho_\gamma + \rho_\nu}\right]\; .$$ The conformal time $\tau$ can be related to the scale factor $a=e^N$ via $$\tau = \int_0^a\frac{da}{Ha^2}.$$ The electron number density $n_e(a)$ is obtained using RecFast version 1.5 [@Seager/etal:1999; @Wong/etal:2008], which has been incorporated into CosmoLib. We denote the differential optical depth (increment of optical depth per $dN$) as $$\kappa_N \equiv \frac{d\kappa}{dN} = \frac{n_e \sigma_T}{H}\; ,$$ where $\sigma_T = 6.652\times 10^{-25}{\rm cm}^2$ is the Thomson scattering cross section. The baryon sound speed $c_{s,b}^2(a)$ is obtained by solving the differential equations (68-69) in Ref. [@Ma/Bertschinger:1995]. With these background solutions in hand, now we can write down the governing equations for scalar perturbations. Scalar Perturbations -------------------- The metric in the (generalized) Newtonian gauge can be written as $$ds^2 = a^2(\tau) \left\{-(1 + 2\Phi)d\tau^2 + \omega_i dx^id\tau + \left[(1-2\Psi)\delta_{ij} + h_{ij}\right]dx^idx^j \right\} \;,$$ where $\partial_i\omega_i = 0$, $h_{ii}=0$ and $\partial_i h_{ij} = 0$. The vector perturbation $\omega_i$ decays in an expanding universe and hence it is set to zero in CosmoLib. The tensor perturbation $h_{ij}$ is gauge-invariant and its governing equations are identical in all gauges. Thus, we will only focus on the scalar perturbation equations that in CosmoLib differ from those in many other Boltzmann codes. The linear-order relative density perturbation of $X$ is denoted by $\delta_X \equiv \delta\rho_X/\rho_X$, and the linear-order velocity $\vel{X}$. Unless otherwise specified, $\delta_X$ and $\vel{X}$ are all defined in Fourier space, that are functions of $\tau$ and the wave vector $\veck$. The radiation relative temperature fluctuation $\Delta T/T$ from direction $\vecn$ seen by an observer at position $\vecx$ is expanded as [@Hu/White:1997] $$\frac{\Delta T}{T} \left(\vecx, \vecn, \tau \right)= \int \frac{d^3\veck}{(2\pi)^3}\sum_{\ell=0}^{\infty}\sum_{m=-2}^{2} \Theta_\gamma(\ell,m) (-i)^\ell \sqrt{\frac{\pi}{4(2\ell +1)}}Y_\ell^m(\vecn) e^{i\veck\cdot\vecx}\; ,$$ where $Y_{\ell}^m$ are the spherical harmonic functions. Note that the moments $\Theta_\gamma(\ell,m)$ are functions of the wavenumber $\veck$ and the conformal time $\tau$. The energy density fluctuation and velocity of photons are related to the moments $\ell = 0$ and $\ell = 1$ via $$\begin{aligned} \delta_\gamma = \Theta_\gamma(0,0);\ \; \vel{\gamma} = \frac{1}{4}\Theta_\gamma(1,0)\; .\end{aligned}$$ The neutrino moments $\Theta_\nu(\ell,m)$ are defined in the same way, by replacing the subscript $\gamma$ with $\nu$. For the polarization of radiation, the Stokes parameters $Q, U$ are expanded using the spin-$2$ harmonics $_{\pm 2}Y_\ell^m$ [@Hu/White:1997] $$\left(Q\pm i U\right)\left(\vecx, \vecn ,\tau\right) = \int \frac{d^3\veck}{(2\pi)^3}\sum_{\ell=0}^{\infty}\sum_{m=-2}^{2} \left[E(\ell,m) \pm B(\ell,m) \right] (-i)^\ell \sqrt{\frac{\pi}{4(2\ell +1)}}\left[_{\pm 2}Y_\ell^m(\vecn)\right] e^{i\veck\cdot\vecx}\; ,$$ where $E(\ell, m)$ and $B(\ell, m)$ are functions of the wave vector $\veck$ and conformal time. The linear-order Fourier modes are decoupled. The Fourier-space variables to be evolved are $\Psi$, $\Psi_N \equiv d\Psi/dN$, $\delta_b$, $\vel{b}$, $\delta_c$, $\vel{c}$, $\delta_\Lambda$, $\theta_\Lambda\equiv (1+w)\vel{\Lambda}$, $\Theta_\gamma(\ell, 0)$ ($\ell$ = $0$, $1$, $2$, ..., $\ell_{\max, \gamma}$), $\Theta_\nu(\ell, 0)$ ($\ell$ = $0$, $1$, $2$, ..., $\ell_{\max, \nu}$), $E(\ell, 0)$ ($\ell$ = $2$, ..., $\ell_{\max, E}$). The truncations $\ell_{\max, \gamma}$, $\ell_{\max, \nu}$ and $\ell_{\max, E}$ are adjustable integers. In CosmoLib their default values are taken to be $14$, $12$, $14$, respectively. Without loss of generality we choose the azimuthal direction (the $z$-axis direction that is used to define $Y_{\ell,m}(\vecn)$) to be parallel to $\veck$. The gravitational potential $\Phi$ can be obtained from the Einstein equations [@Bond/Efstathiou:1984; @Ma/Bertschinger:1995; @Hu/White:1997] $$\Phi = \Psi - \frac{3}{5k_H^2} \left[\Omega_\gamma \Theta_\gamma(2,0) + \Omega_\nu \Theta_\nu(2,0)\right]\; ,$$ where we have introduced the reduced wavenumber $$k_H \equiv \frac{k}{aH}\; .$$ Note that $k_H, \Omega_\gamma, \Omega_\nu$ are functions of time. We do not treat $\Phi$ as an independent-variable. Instead we view it as a function of the variables $\Psi$, $\Theta_\gamma(2,0)$ and $\Theta_\nu(2,0)$. The close set of first-order differential equations including all the truncation schemes is: [$$\begin{aligned} \frac{d\Psi}{dN} &= \Psi_N\; ,\\ \frac{d\delta_c}{dN} &= - k_H \vel{c} + 3 \Psi_N \;,\\ \frac{d \vel{c}}{dN} &= - \vel{c} + k_H \Phi \; , \\ \frac{d\delta_b}{dN} &= - k_H \vel{b} + 3 \Psi_N \;, \\ \frac{d\vel{b}}{dN} &= - \vel{b} + k_H\left(\Phi + c_{s,b}^2\delta_b\right) - \frac{\kappa_N}{R} \left[\vel{b} - \frac{1}{4}\Theta_{\gamma}(1,0)\right]\;,\\ \frac{d\delta_\Lambda}{dN} &= - 3 \left(c_{s,\Lambda}^2 - w\right) \delta_\Lambda - 9\left[c_{\Lambda,s}^2 - \left(w-\frac{dw/dN}{3\left(1+w\right)}\right)\right] \frac{\theta_\Lambda}{k_H} - k_H \theta_\Lambda + 3 (1 + w) \Psi_N \;, \\ \frac{d\theta_\Lambda}{dN} &= 3 \left[w + c_{s,\Lambda}^2 - \left(w-\frac{dw/dN}{3\left(1+w\right)}\right) - \frac{1}{3}\right] \theta_\Lambda + k_H\left[c_{s,\Lambda}^2 \delta_\Lambda + (1+w) \Phi\right]\;,\\ \frac{d \Theta_\gamma(0,0)}{dN} &= - \frac{1}{3} k_H \Theta_\gamma(1,0) + 4 \Psi_N\;, \label{eq:hiestart}\\ \frac{d \Theta_\gamma(1, 0)}{dN} &= k_H\left[ \Theta_\gamma(0,0) - \frac{2}{5} \Theta_\gamma(2,0) + 4 \Phi\right] + \kappa_N \left[4\vel{b} - \Theta_\gamma(1,0)\right]\;,\\ \frac{d \Theta_\gamma(2, 0)}{dN} &= k_H\left[ \frac{2}{3} \Theta_\gamma(1,0) - \frac{3}{7} \Theta_\gamma(3,0) \right] - \kappa_N \left[\frac{9}{10}\Theta_\gamma(2,0)+\frac{\sqrt{6}}{10}E(2,0)\right]\;,\\ \frac{d \Theta_\gamma(\ell, 0)}{dN} &= k_H \left[\frac{\ell}{2\ell - 1} \Theta_\gamma(\ell - 1, 0) - \frac{\ell+1}{2\ell + 3} \Theta_\gamma(\ell+1, 0)\right] - \kappa_N \Theta_\gamma(\ell , 0)\; \;(2<\ell<\ell_{\max,\gamma})\;,\\ \frac{d\Theta_\gamma(\ell_{\max,\gamma}, 0)}{dN} &= \frac{2\ell_{\max,\gamma}+1}{2\ell_{\max,\gamma}-1} k_H\Theta_\gamma(\ell_{\max,\gamma}-1, 0) - \left(\kappa_N + \frac{\ell_{\max,\gamma}+1}{aH\tau}\right) \Theta_\gamma(\ell_{\max,\gamma},0)\;,\\ \frac{d \Theta_\nu(0,0)}{dN} &= - \frac{1}{3} k_H \Theta_\nu(1,0) + 4 \Psi_N\;, \\ \frac{d \Theta_\nu(1, 0)}{dN} &= k_H\left[ \Theta_\nu(0,0) - \frac{2}{5} \Theta_\nu(2,0) + 4 \Phi\right]\;,\\ \frac{d \Theta_\nu(\ell, 0)}{dN} &= k_H \left[\frac{\ell}{2\ell - 1} \Theta_\nu(\ell - 1, 0) - \frac{\ell+1}{2\ell + 3} \Theta_\nu(\ell+1, 0)\right]\; \;(2\le\ell<\ell_{\max,\nu})\;,\\ \frac{d\Theta_\gamma(\ell_{\max,\nu}, 0)}{dN} &= \frac{2\ell_{\max,\nu}+1}{2\ell_{\max,\nu}-1} k_H\Theta_\nu(\ell_{\max,\nu}-1, 0) - \frac{\ell_{\max,\nu}+1}{aH\tau} \Theta_\nu(\ell_{\max,\nu},0)\;,\\ \frac{d E(2,0)}{dN} &= -k_H \frac{K_{3,0,2}}{7}E(3,0) - \kappa_N\left[\frac{2}{5} E(2,0) + \frac{\sqrt{6}}{10} \Theta_\gamma(2,0)\right]\; ,\\ \frac{dE(\ell,0)}{dN} &= k_H \left[\frac{K_{\ell,0,2}}{2\ell-1} E(\ell-1,0) - \frac{K_{\ell+1, 0, 2}}{2\ell+3} E(\ell+1,0)\right] - \kappa_N E(\ell,0) \;\;(2<\ell<\ell_{\max,E})\;, \\ \frac{dE(\ell_{\max,E},0)}{dN} &= \frac{2\ell_{\max,E}+1}{2\ell_{\max,E}-1}k_H E(\ell_{\max,E}-1,0) - \left(\kappa_N + \frac{\ell_{\max,E}+1}{aH\tau}\right)E(\ell_{\max,E},0)\;, \label{eq:hieend}\\ \frac{d\Psi_N}{dN} &= \frac{1}{2}\left\{(3c_{s,b}^2-1)\delta_b\Omega_b -\delta_c\Omega_c + \left[(3c_{s,\Lambda}^2-1) \delta_\Lambda + 9\left(c_{s,\Lambda}^2 - w+\frac{dw/dN}{3(1+w)}\right)\frac{\theta_\Lambda}{k_H}\right]\Omega_\Lambda\right\} - 2\Psi \nonumber \\ & - 2(1-\epsilon)\Phi - \frac{k_H^2}{3}(2\Psi - \Phi) - (5-\epsilon)\Psi_N + \frac{3}{5k_H^2}\left(\Omega_\gamma\frac{d\Theta_\gamma(2,0)}{dN} + \Omega_\nu\frac{d\Theta_\nu(2,0)}{dN}\right) \;. \label{eq:PsiNN}\end{aligned}$$]{} In the radiation and neutrino hierarchy equations (\[eq:hiestart\]-\[eq:hieend\]) we have used the Clebsch-Gordan coefficients $K_{l,m,s}$, which are defined as [@Hu/White:1997] $$K_{\ell,m,s} = \left\{ \begin{array}{ll} \frac{\sqrt{(\ell^2-m^2)(\ell^2-s^2)}}{\ell} & \text{, if\ } \ell \ge \max\{{|m|,|s|, 1}\}\; ; \\ 0 & \text{, otherwise.} \end{array} \right.$$ These equations are already written in the form that can be directly implemented into a generic first-order ordinary-differential-equation (ODE) solver. For a derivation of these equations, see Refs. [@Bond/Efstathiou:1984; @Ma/Bertschinger:1995; @Hu/White:1997; @Fang/etal:2008]. (The change of time variable from $\tau$ to $N$ can be done straightforwardly using $d/d\tau = aH d/dN$.) The initial conditions can be found in Ref. [@Ma/Bertschinger:1995]. For the tight-coupling approximation we follow Ref. [@Doran:2005a], where the obvious typos in Eqs. (15-16) has been fixed. Since these treatments are identical to the original source, we do not repeat the discussion here. The interested readers are referred to these references for the governing equations and their derivation. CosmoLib allows the user-input $w(a)$ to be a phantom-crossing function, that is a function crossing the line $w=-1$. In this case we force $d\delta_\Lambda/dN$ and $d \theta_\Lambda/dN$ to be zero around the proximity of the phantom crossing. This is an approximation. Exact treatment requires input of at least one more degree of freedom [@Vikman:2005; @Hu:2005; @Caldwell/Doran:2005], which cannot be implemented in a generic code. In Ref. [@Fang/etal:2008] the reader can find an alternative treatment that works better for multiple scalar field models. Equation  is the key equation that guarantees the numerical stability of the code (even for isocurvature initial conditions). It is obtained by subtracting the $ii$ components of the perturbed Einstein equations (pressure perturbations) from the $00$ component (density perturbations). This particular combination of the Einstein equations has been applied in the numerical code CMBquick [@Pitrou:2009; @Pitrou/etal:2010], which assumes that dark energy is a cosmological constant and ignores the baryon sound speed. Eq.  is a generalized version that includes dark energy perturbations and a nonzero baryon sound speed. We can use the energy constraint ($00$ component of the perturbed Einstein equations, that is $\delta G_{00} = \delta T_{00}$) and the momentum constraint ($0i$-component of the perturbed Einstein equations, that is $\delta G_{0i} = \delta T_{0i}$) to estimate the numerical error of the code. As shown in shown in Figure \[fig:eins\], the relative errors are $\lesssim 10^{-4}$ for a wide range of scales and different initial conditions. ![Testing the energy constraint ($00$-component of the perturbed Einstein equation) and momentum constraint ($0i$-component of the perturbed Einstein equation). The cosmological parameters used here are $\Omega_{b0} h^2 = 0.022$, $\Omega_{c0}h^2 = 0.1128$, $h = 0.72$. In the lower-right panel the CDM-isocurvature initial conditions are used, while in the other panels we have used adiabatic initial conditions. \[fig:eins\]](eins-1.eps "fig:"){width="\halffigurewidth"}![Testing the energy constraint ($00$-component of the perturbed Einstein equation) and momentum constraint ($0i$-component of the perturbed Einstein equation). The cosmological parameters used here are $\Omega_{b0} h^2 = 0.022$, $\Omega_{c0}h^2 = 0.1128$, $h = 0.72$. In the lower-right panel the CDM-isocurvature initial conditions are used, while in the other panels we have used adiabatic initial conditions. \[fig:eins\]](eins-2.eps "fig:"){width="\halffigurewidth"} ![Testing the energy constraint ($00$-component of the perturbed Einstein equation) and momentum constraint ($0i$-component of the perturbed Einstein equation). The cosmological parameters used here are $\Omega_{b0} h^2 = 0.022$, $\Omega_{c0}h^2 = 0.1128$, $h = 0.72$. In the lower-right panel the CDM-isocurvature initial conditions are used, while in the other panels we have used adiabatic initial conditions. \[fig:eins\]](eins-4.eps "fig:"){width="\halffigurewidth"}![Testing the energy constraint ($00$-component of the perturbed Einstein equation) and momentum constraint ($0i$-component of the perturbed Einstein equation). The cosmological parameters used here are $\Omega_{b0} h^2 = 0.022$, $\Omega_{c0}h^2 = 0.1128$, $h = 0.72$. In the lower-right panel the CDM-isocurvature initial conditions are used, while in the other panels we have used adiabatic initial conditions. \[fig:eins\]](eins-2-iso.eps "fig:"){width="\halffigurewidth"} CMB Angular Power Spectra \[sec:cmb\] ===================================== Algorithm --------- Optionally CosmoLib can compute the CMB angular power spectrum for each multipole $\ell$ by brute force, i.e., without interpolation. The angular spectrum for the temperature anisotropies is given by $$C_\ell = \int |\Delta_\ell^k |^2 \primsca(k) d\ln k \;, \label{eq:clint}$$ where $\Delta_\ell^k$ is the temperature transfer function given by the line-of-sight integration $$\Delta_\ell^k = \int_0^{\tau_0} S(k,\tau) j_\ell\left[k(\tau_0-\tau)\right] d\tau \; , \label{eq:trans}$$ where $j_\ell$ is the spherical Bessel function and $\tau_0$ is $\tau$ at redshift zero. The source $S(k,\tau)$ can be computed from the perturbations $\Psi$, $\Phi$, $\delta_X$, $\vel{X}$ ($X=c$, $b$, $\Lambda$, $\gamma$, $\nu$), $\Theta_\gamma(2, 0)$ and $\Theta_\nu(2,0)$ [@Seljak/Zaldarriaga:1996; @Hu/White:1997]. In Ref. [@Hu/White:1997] the line-of-sight integration involves the functions $j_\ell$, $j'_\ell$ and $j_\ell''$. As shown in Ref. [@Seljak/Zaldarriaga:1996], however, the dependence on $j_\ell'$ and $j_\ell''$ can be eliminated by integrating by part. (We have corrected the obvious typos in eq. (12b) in Ref. [@Seljak/Zaldarriaga:1996].) Since $\Delta_\ell^k$ is evaluated numerically and it typically oscillates quickly, its sampling is time consuming. Indeed, in modern fast CMB codes – such as CAMB, CLASS, CMBEASY – the integral  is computed by sampling $\Delta_\ell^k$ using a step size in $k$ that can be typically much larger than the oscillation period in $\Delta_\ell^k$. For instance, in Fig. \[fig:trans\] we show an example of $\Delta_\ell^k$ for a fixed $\ell = 300$. A typical sampling scheme is shown by the red solid triangles in the upper-right panel, which zooms-in part of the figure. According to Parseval’s theorem, if $\primsca (\ln k)$ is a smooth function, such sparse sampling of $\Delta_\ell^k$ is enough. ![The temperature transfer function $\Delta_\ell^k$ for a fixed $\ell = 300$. A typical sampling scheme is shown by the red solid triangles in the upper-right panel, which zooms-in part of the figure. \[fig:trans\]](TransExample.eps){width="\figurewidth"} However, when $\primsca(k)$ has local sharp features, the minimum sampling distance should be determined by the larger between the typical relative width (the width measured in $\ln k$) of the oscillations in $\Delta_\ell^k$ and $\delta\ln k$, the typical relative width of the features in $\primsca(k)$. The former is about $10^{-4}-10^{-3}$, while the latter is model-dependent. For instance, if our goal is to sample features with width $\delta \ln k \sim 10^{-3}$, the required sampling frequency is typically $\sim 20 $ to $100$ times higher than that used for a smooth $\primsca(k)$. Furthermore, as we wish to compute the $C_\ell$’s for each $\ell$ rather than interpolating it over few tens of $\ell$’s, the total time consumption will be again multiplied by a factor of $\sim 10 - 50$. The naively estimated total time consumption is hence $\sim 10^3$ times more than that in the smooth-$\primsca(k)$ case. A final complication is due to the fact that, if all the transfer functions and the precomputed $j_\ell(x)$ tables are to be stored, one has also to face a memory barrier that cannot be easily bypassed. For these reasons, simply increasing the $\ell$ and $k$ resolution in standard codes such as CAMB, CLASS or CMBEASY, will not be efficient enough for the purpose of scanning the whole parameter space. The algorithm can be significantly improved, however, if we notice that the output $S(k,\tau)$ from the Boltzmann code is a 2D matrix in $k$-$\tau$ space. If $j_\ell\left[k(\tau_0-\tau)\right]$ is also a precomputed 2D matrix with the same structure, the integration  can be obtained by taking the inner product of the two matrices. Modern Fortran90 compilers can optimize such operation and make the computation much faster. The difficulty, however, is that the $j_\ell\left[k(\tau_0-\tau)\right]$ matrices for all $\ell$’s will occupy too much memory (can be up to a few tens of Giga bytes in the worst scenario). Our solution is then to only store the matrices for two neighboring $\ell$’s and update them using the recurrence relation of spherical Bessel functions. Let us describe our strategy. We first compute two neighboring $C_\ell$’s by brute force. Two matrices of spherical Bessel functions $j_{\ell+1}[k(\tau_0 -\tau)]$ and $j_\ell[k(\tau_0-\tau)]$ are stored in the memory for each $(k, \tau)$ indices. Then we compute $C_{\ell-1}$. To do that, we update the $j_{\ell+1}$ matrix to the $j_{\ell -1}$ matrix using the recurrence relation $$j_{\ell -1}(x) = \frac{2\ell +1}{x} j_\ell(x) - j_{\ell +1} (x) \ .\label{eq:jlrecur}$$ Again, using $j_\ell$ and $j_{\ell-1}$ we then calculate $j_{\ell -2}$ and hence $C_{\ell -2}$. This downward iteration is very stable for a few tens of steps, after which we need to recompute another couple of neighboring $C_\ell$’s and iterate downward again. The initial neighboring $j_{\ell}$’s are calculated using precomputed 25-th order Chebyshev fitting formulas. (For the rapidly oscillating part at $x\gg l$, the phase and amplitude of oscillations are fitted using Chebyshev polynomials.) Chebyshev fitting is slightly slower than the cubic-spline fitting used in other publicly available CMB codes, but it is more memory-efficient and more accurate – it has an accuracy of $\sim 10^{-8}$ – and allows more downward iterative steps. Finally, note that the algorithm proposed here is more efficient both CPU-wise and memory-wise, enhancing the speed of $\ell$-by-$\ell$ computation of $C_\ell$’s by a factor of $\sim 10$ to $30$. For CMB lensing we use the power spectrum approach as described in Refs. [@Seljak:1996; @Zaldarriaga/Seljak:1998]. Testing the Code ---------------- The trivial comparison between CosmoLib and CAMB for smooth-$\primsca$ models can be found in the online documentation at <http://www.cita.utoronto.ca/~zqhuang/CosmoLib>. Here we focus on the enhanced CMB integrator that does not assume the smoothness in $\primsca(k)$. Since this feature is not available in other CMB codes, direct numerical comparison is not possible when there is very sharp features in $\primsca$. Thus, we need to study a model in which we have some theoretical insights. An ideal candidate is the axion monodromy inflation model, where the primordial power spectrum displays sinusoidal oscillations superimposed to a smooth power spectrum. It can be written as [@Flauger/etal:2010] $$\primsca(k) = A_s \left(\frac{k}{\kpiv}\right)^{n_s-1} \left[1 + \delta n_s \cos \left(\frac{\ln (k/\kpiv)}{\delta\ln k} + \varphi \right)\right]\; , \label{eq:monoPs}$$ where $A_s$ and $n_s$ are the amplitude and tilt, respectively. The parameter $\delta\ln k$ describes the width of the oscillations in $\primsca(k)$, while $\delta n_s$ gives their relative amplitude. The pivot scale $\kpiv$ is chosen to be $0.05{\rm Mpc}^{-1}$ in our computation. We compute the CMB temperature power spectrum using the enhanced CMB integrator in CosmoLib and compare the results to the smooth-$\primsca(k)$ case. The relative difference between the non-smooth (for a series of $\delta\ln k$) and the smooth model is shown in Figure \[fig:pkcls\]. For $\delta \ln k =0.1$ and $\delta \ln k = 0.03$ we compare the results to CAMB output (both with lensing) and find good agreement. The CAMB outputs are obtained by a straightforward modification of CAMB, i.e., increasing the $\ell$ sampling frequency in the input file and increasing the $k$ sampling frequency in the source code. For $\delta \ln k \lesssim 0.01$ the simple modification of CAMB fails due to insufficient memory to store the transfer functions. For $\delta\ln k\ll 1$, the amplitude of oscillations in the CMB angular power spectrum (right-hand panels) is smaller than that in $\primsca(k)$ (left-hand panels). This suppression is generic when a 3D power spectrum is projected to a 2D one, even though in the CMB case it is further complicated by the finite duration of recombination and the recombination physics [@Adshead/etal:2011]. As shown in [@Huang/etal:2012], when the frequency of oscillations is constant in $\ln k$, such as in eq. , the relative suppression is given by $\sim \sqrt{\delta\ln k}$, as confirmed by the examples shown in Figure \[fig:pkcls\]. Moreover, for $\delta\ln k \lesssim 0.01$, in addition to the projection effect, CMB lensing also significantly smears out the oscillations in $C_\ell$ at high $\ell \gtrsim 2000$. While for $\delta\ln k = 0.1$, the lensing smearing effect is almost negligible. See [@Adshead/etal:2011; @Lewis/Challinor:2006] for more detailed discussions about the lensing smearing effect. Finally, note that, although the oscillations in $C_\ell$ are damped, they maintain the same relative width of those of the left-hand panels, i.e., $\delta\ln \ell = \delta\ln k$ where $\ell \gtrsim (\delta\ln k)^{-1}$. At low $\ell$ where $\ell \lesssim 1/\delta\ln k$ the oscillations in $k$ space disappear in $\ell$ space due to the discreteness of $\ell$. ![The differences in $\ln \primsca$ (left panels) or $\ln C^{TT}_\ell$ (right panels) between a fiducial axion monodromy model with $\ln \left(10^{10}A_s\right)=3.027$, $n_s = 0.975$, amplitude of cosine modulation $\delta n_s = 0.01$, phase $\varphi=0$ and a smooth power-law spectrum with the same $A_s$ and $n_s$. For the top to bottom a series of $\delta\ln k = 0.1$, $0.03$, $0.01$ are used, respectively. The $\tau_{\rm recomb}$ in the $x$-axis legend of left panels is the conformal time at recombination ($z\approx 1100$). For $\delta \ln k=0.1$ and $0.03$ the results are compared to CAMB outputs. For $\delta\ln k=0.01$ a simple modification of CAMB cannot be applied as too much memory is required to store the transfer functions for all ($\ell$, $k$) pairs. \[fig:pkcls\]](Cldiff0.1.eps "fig:"){width="\figurewidth"} ![The differences in $\ln \primsca$ (left panels) or $\ln C^{TT}_\ell$ (right panels) between a fiducial axion monodromy model with $\ln \left(10^{10}A_s\right)=3.027$, $n_s = 0.975$, amplitude of cosine modulation $\delta n_s = 0.01$, phase $\varphi=0$ and a smooth power-law spectrum with the same $A_s$ and $n_s$. For the top to bottom a series of $\delta\ln k = 0.1$, $0.03$, $0.01$ are used, respectively. The $\tau_{\rm recomb}$ in the $x$-axis legend of left panels is the conformal time at recombination ($z\approx 1100$). For $\delta \ln k=0.1$ and $0.03$ the results are compared to CAMB outputs. For $\delta\ln k=0.01$ a simple modification of CAMB cannot be applied as too much memory is required to store the transfer functions for all ($\ell$, $k$) pairs. \[fig:pkcls\]](Cldiff0.03.eps "fig:"){width="\figurewidth"} ![The differences in $\ln \primsca$ (left panels) or $\ln C^{TT}_\ell$ (right panels) between a fiducial axion monodromy model with $\ln \left(10^{10}A_s\right)=3.027$, $n_s = 0.975$, amplitude of cosine modulation $\delta n_s = 0.01$, phase $\varphi=0$ and a smooth power-law spectrum with the same $A_s$ and $n_s$. For the top to bottom a series of $\delta\ln k = 0.1$, $0.03$, $0.01$ are used, respectively. The $\tau_{\rm recomb}$ in the $x$-axis legend of left panels is the conformal time at recombination ($z\approx 1100$). For $\delta \ln k=0.1$ and $0.03$ the results are compared to CAMB outputs. For $\delta\ln k=0.01$ a simple modification of CAMB cannot be applied as too much memory is required to store the transfer functions for all ($\ell$, $k$) pairs. \[fig:pkcls\]](Cldiff0.01.eps "fig:"){width="\figurewidth"} This discussion shows that the enhanced CMB integrator can accurately compute the oscillations in CMB to $\Delta C_\ell/C_\ell \lesssim 10^{-3}$. This does not mean, however, that the total $C_\ell$ is accurate to $10^{-3}$. The $C_\ell$ power spectrum can be systematically biased at subpercent level due to, e.g., recombination uncertainties [@Rubino-Mart/etal:2010]. Understanding and eliminating these theoretical errors is important if we want to extract generic features in $\primsca(k)$ with $10^{-3}$ accuracy. On the other hand, if we are only interested in a model predicting a specific feature in $C_\ell$ that cannot be mimicked by other effects, we can focus only on the relative difference in $C_\ell$. The forecast techniques \[sec:forecast\] ======================================== CosmoLib uses Fisher matrix analysis and MCMC method to forecast the constraints on cosmological parameters for future CMB, LSS and SN experiments. In this section we discuss the modeling of the likelihoods and the parameter estimation methods. The likelihoods --------------- ### CMB simulation Given a likelihood function ${\cal L}$, we define $\chi^2 \equiv -2 \ln {\cal L}$. For a nearly full-sky CMB experiment $\chi^2$ can be approximated by [@Verde/etal:2006; @Baumann/etal:2009] [$$\begin{aligned} \label{chisq_CMB} \chi^2 =& \sum_{\ell=\ell_{\rm min}}^{\ell_{\rm max}} (2\ell+1)f_{\rm sky}\, \left[ \frac{\hat{{\cal C}}_\ell^{BB}}{{\cal C}_\ell^{BB}} - 3 + \ln\left(\frac{{\cal C}_\ell^{BB}}{\hat{{\cal C}}_\ell^{BB}}\right) + \frac{\hat{{\cal C}}_\ell^{TT}{\cal C}_\ell^{EE} + \hat{{\cal C}}_\ell^{EE}{\cal C}_\ell^{TT} - 2\hat{{\cal C}}_\ell^{TE}{\cal C}_\ell^{TE}}{{\cal C}_\ell^{TT}{\cal C}_\ell^{EE}-({\cal C}_\ell^{TE})^2} + \ln{\left(\frac{{\cal C}_\ell^{TT}{\cal C}_\ell^{EE}-({\cal C}_\ell^{TE})^2}{\hat{{\cal C}}_\ell^{TT}\hat{{\cal C}}_\ell^{EE}-(\hat{{\cal C}}_\ell^{TE})^2}\right)}\right] \ , \end{aligned}$$]{} where $\ell_{\rm min}$ and $\ell_{\rm max}$ are suitable cutoffs that are determined by the observed fraction of sky $f_{\rm sky}$ and the beam resolution of the experiment. In this formula, ${\cal C}^{XY}_\ell$ are the model-dependent theoretical angular power spectra (including the noise contributions) for the temperature, $E$ and $B$ polarizations and their cross-correlations, with $X,Y=\{T,E,B \}$. We compute the noise contribution $N_\ell$ assuming Gaussian beams. The mock data $\hat{{\cal C}}^{XY}_\ell$ are ${\cal C}^{XY}_\ell$ for the fiducial model. We use the model introduced in [@Verde/etal:2006] (and later updated in [@Baumann/etal:2009]) to propagate the effect of polarization foreground residuals into the estimated uncertainties on the cosmological parameters. For simplicity, in our simulation we consider only the dominant components in the frequency bands that we are using, i.e., the synchrotron and dust signals. We assume that foreground subtraction can be done correctly down to a level of 5%. (This parameter is adjustable by the user.) ### SN simulation For the SN simulation, we use the model given by the Dark Energy Task Force (DETF) forecast [@Albrecht/etal:2006]. In this case $$\chi^2= \sum_i\left(\frac{m_i - \hat{m}_i}{\delta m_i}\right)^2\; ,$$ with $i$ going over the SN samples. More specifically, here $m_i$ and $\hat{m}_i$ are the theoretical expectation and observed magnitude of the $i$-th supernova, respectively. The uncertainty $\delta m_i$ is computed by quadratically adding a peculiar velocity (a user-defined constant) to the intrinsic uncertainty in the supernova absolute magnitude (another user-specified constant). The apparent magnitude of SN is modeled as $$\begin{aligned} m &= M -\mu^Lz-\mu^Qz^2 +5\log_{10}\left(\frac{d_L}{\text{Mpc}}\right)+25 - \mu^S\delta_{\text{near}}\ . \label{snm}\end{aligned}$$ The first three terms model the redshift evolution of the absolute magnitude of the supernova peak luminosity. In particular, $M$ is a free parameter with a flat prior over $-\infty<M<+\infty$; for $\mu^L$ and $\mu^Q$, Gaussian priors are applied. The widths of the Gaussian priors are user-input parameters. Finally, given that the nearby samples are likely to be a collection from many other experiments, an offset $-\mu^S\delta_{\text{near}}$, where $\delta_{\text{near}}$ is unity for the nearby samples ($z<z_{\rm near}$) and zero otherwise, is added to account for the systematics. For $\mu^S$ we also apply a Gaussian prior with a user-specified width. The threshold redshift $z_{\rm near}$ is also user-defined. In conclusion, in this model there are four nuisance parameters $(M, \mu^L, \mu^Q, \mu^S)$, which we marginalized analytically. ### LSS likelihood We model the galaxy power spectrum in redshift space as (e.g., [@Kaiser:1987; @Peacock:1992; @Peacock/Dodds:1994]) $$P_g(k,\mu; z) = \left(b+ f \mu^2\right)^2 D^2(z)P_m(k) \exp\left(-k^2\mu^2\sigma_r^2 \right), \label{eq:Pg}$$ where $\mu$ is the cosine of the angle between the wave vector $\mathbf{k}$ and the line of sight, $D(z)$ is the linear growth factor, $f \equiv {d\ln D}/{d\ln a}$ is the linear growth rate, $P_m(k)$ is the matter power spectrum today (at $z=0$) and $\sigma_r$ parameterizes the effect of small scales velocity dispersion and redshift errors as explained below. The matter power spectrum $P_m(k)$ is computed using Poisson’s equation, that is, $P_m(k) = 4k^4|\Phi_k|^2/(9H^4\Omega^2_m)$. The term $f \mu^2$ accounts for the redshift distortions due to the large-scale peculiar velocity field [@Kaiser:1987], which is correlated with the matter density field. The exponential factor on the right-hand side accounts for the radial smearing due to the redshift distortions that are uncorrelated with the LSS. In particular, we consider two contributions. The first is due to the redshift uncertainty of the spectroscopic galaxy samples which is estimated to be $\sigma_z=0.001(1+z)$ [@Laureijs:2009]. (In CosmoLib the user is allowed to change this value.) The second comes from the Doppler shift due to the virialized motion of galaxies within clusters, which typically has a pairwise velocity dispersion $\sigma_g$ of the order of few hundred $\text{km/s}$. This can be parameterized as $\frac{\sigma_g}{\sqrt{2}} (1+z)$ [@Peacock/Dodds:1994]. The two contributions are quadratically added together $$\sigma_r^2 = \frac{(1+z)^2}{H^2(z)} \left(\sigma_z^2 + {\sigma_g^2}/{2} \right)\;, \label{eq:sigma2}$$ where $H(z)$ is the Hubble parameter. Practically, neither the redshift measurement nor the virialized motion of galaxies can be precisely modeled. In particular, the radial smearing due to peculiar velocity is not necessarily close to Gaussian. Thus, eq. (\[eq:Pg\]) should not be used for wavenumbers $k>\frac{H(z)}{\sigma_g (1+z)}$, where the radial smearing effect is important. We introduce a UV cutoff $k_{\max}$ as the smallest value between $\frac{H}{\sigma_g(1+z)}$ and $\frac{\pi}{2R}$, where $R$ is chosen such that the r.m.s. linear density fluctuation of the matter field in a sphere with radius $R$ is $0.5$. The survey volume is split into $n_z$ redshift bins from $z_{\min}$ to $z_{\max}$, with all these parameters to be specified by the user. The number density of galaxies that can be used is $\bar{n}=\varepsilon \bar{n}_{\rm obs}$, where $\varepsilon$ is the fraction of galaxies with measured redshift to be specified by the user. Due to the high accuracy of the spectroscopic redshift and the width of the bins, we ignore the bin-to-bin correlations and write $\chi^2$ as $$\chi^2 = \sum_{k,\mu,z\ \rm bins} \left(\frac{P_{g, \rm model} - P_{g, \rm fiducial}}{\Delta P_{g,\rm fiducial}}\right)^2\ .$$ As on large scales the matter density field has, to a very good approximation, Gaussian statistics and uncorrelated Fourier modes, the band-power uncertainty is given by [@Tegmark/etal:1998] $$\Delta P_g = \left[ \frac{2 (2\pi)^3}{(2\pi k^2dk d\mu) (4\pi r^2f_{\rm sky} dr)}\right]^{1/2}\left(P_g+\frac{1}{\bar{n}}\right), \label{eq:dPg}$$ where $r$ is the comoving distance given, for a flat FRW universe, by $r(z)=\int_0^z cdz'/H(z')$. The second term in the parenthesis is due to shot noise, under the assumption that the positions of the observed galaxies are generated by a random Poisson point process. In practice $\bar{n}$ is not known a priori and is calibrated by galaxies themselves. The imperfect knowledge of $\bar{n}$ can bias $P_g$ on the scale of the survey [@Tegmark/etal:1998]. This is taken into account by using an IR cutoff $k_{\min}\sim {\rm Gpc}^{-1}$. This is chosen such that $k_{\min}^{(i)} = 2\pi/V^{1/3}_i$, where $V_i$ is the comoving volume of the $i$-th ($i = 1, \ldots, n_z$) redshift slice. Finally, the user has to specify the binning scheme for $k$ and $\mu$. For $k$ we allow uniform binning in $\ln k$ or in $k$. For $\mu$ only uniform binning in $\mu$ is allowed. In the special case where $P_m(k)$ has sharp features, we must consider the smearing effect due to the fact that we are only observing a finite volume. This effect is approximated by replacing $P_m(k)$ in with its convolution with a Gaussian window, where the width of the Gaussian window $\sigma_W$ has been chosen to be $$\sigma_W = \frac{\sqrt{2 \ln 2}}{2 \pi} \left(\frac{4 \pi}{3} \right)^{1/3} k_{\rm min} \simeq 0.302 \; k_{\rm min}\;.$$ In such a way, the real-space representation of the window, if cut off at its half-height, contains the same volume as that of the redshift bin. The fact that $\sigma_W$ is smaller than $k_{\rm min}$ allows us to neglect the overlap between window functions centered around neighboring values of $k$. Parameter Estimation -------------------- ### Fisher Matrix Analysis In general, the likelihood can be written as $$\ln \mathcal{L} (\mathbf{p}; \mathbf{p}_{\rm fid}) = -\frac{1}{2} \left[\mathbf{d}(\mathbf{p}) - \mathbf{d}(\mathbf{p}_{\rm fid})\right]^T C^{-1}(\mathbf{p}; \mathbf{p}_{\rm fid}) \left[\mathbf{d}(\mathbf{p}) - \mathbf{d}(\mathbf{p}_{\rm fid})\right]\;,$$ where $\mathbf{d}$ is the data vector, $\mathbf{p}_{\rm fid}$ the fiducial parameter vector, $\mathbf{p}$ the parameter vector for which one wants to evaluate the likelihood, and $C^{-1}(\mathbf{p}; \mathbf{p}_{\rm fid})$ the covariance matrix. The fisher matrix for $p_i$, $p_j$ (two components of $\mathbf{p}$) is then $$F_{ij} \equiv -\left.\frac{\partial^2 \ln \mathcal{L}}{\partial p_i\partial p_j} \right\vert_{\mathbf{p} =\mathbf{p}_{\rm fid}} =\left. \frac{\partial \mathbf{d}(\mathbf{p})}{\partial p_i} C^{-1}(\mathbf{p}; \mathbf{p}_{\rm fid}) \frac{\partial \mathbf{d}(\mathbf{p})}{\partial p_j} \right\vert_{\mathbf{p} =\mathbf{p}_{\rm fid}}\;,$$ where the partial derivatives $ \frac{\partial \mathbf{d}(\mathbf{p})}{\partial p_i}$ can be evaluated numerically: $$\frac{\partial \mathbf{d}(\mathbf{p})}{\partial p_i} = \frac{1}{2\Delta p_i} \left[\mathbf{d}(p_1, p_2, \ldots,p_i+\Delta p_i, \ldots, p_n)- \mathbf{d}(p_1, p_2, \ldots, p_i-\Delta p_i, \ldots, p_n)\right]\;.$$ The stepsize $\Delta p_i$ is initially supplied by the user, and then adjusted by the software in such a way that the variation of $\chi^2$ is of $O(1)$ when $p_i$ is varied by $\Delta p_i$. By doing this, we have assumed that the likelihood is approximately Gaussian in the proximity of $\mathbf{p}_{\rm fid}$ where the variation of $\chi^2$ is $\lesssim O(1)$. If the likelihood is highly non-Gaussian, Fisher matrix analysis does not give reliable estimations of the error bars of parameters. In this case, one should use the MCMC method to fully explore the structure of the likelihood. ### MCMC method CosmoLib has an independent MCMC engine using the Metropolis-Hastings algorithm. The traditional approach is to define the proposal density $Q(\vecx; \vecx')$ (the probability of walking from $\vecx$ to $\vecx '$ in the parameter space) using a roughly estimated covariance matrix $C_e$ $$Q(\vecx; \vecx') \propto \exp{\left[-\frac{1}{2}(\vecx-\vecx')^TC_e^{-1} (\vecx-\vecx')\right]} \;. \label{eq:prop}$$ Convergence can be achieved quickly if $C_e$ is close the posterior covariance matrix of $\vecx$. However, sometimes we need to treat models where the likelihood periodically depends on some phase parameters. Here we take the axion monodromy inflation model for example. The likelihood $\mathcal{L}$ is a periodic function of the axion phase parameter $\varphi$, $$\mathcal{L}(P, \varphi) = \mathcal{L}(P, \varphi + 2\pi)\;,$$ where we have used $P$ to represent the collection of other parameters. If $\varphi$ is not well constrained, we will obtain multi-branches in the posterior, i.e., for a fixed value of $\varphi$ and a chosen confidence level, the allowed values of $P$ locate in well separated regions in the parameter space. Intuitively the separated regions can be more efficiently explored by restricting the range of $\varphi$ to one period and proposing with wrap-around or, in a more rigorous language, by using a periodic proposal density. For $\vecx = (P, \varphi)$ and $\vecx' = (P', \varphi')$, we use $$Q(P, \varphi;P',\varphi' )\propto \sum_{n=-\infty}^\infty\exp{\left[-\frac{1}{2}(\vecx-\vecx'_n)^TC_e^{-1} (\vecx-\vecx'_n)\right]} \; , \label{eq:propupdate}$$ where $x'_n \equiv (P', \varphi'+2n\pi)$. The estimation of covariance matrix, $C_e$, is practically computed with a trial run that is terminated before the multi-branches of the posterior are explored by the random walk. We find that the periodic proposal density leads to significant improvement of the convergence. For the axion monodromy model, it takes about 5-10 times longer to achieve convergence using than using . The output chains in CosmoLib have the same format as those in CosmoMC [@Lewis/Bridle:2002]. The chains can hence be directly analyzed using the GetDist tool in CosmoMC. For completeness, an independent postprocessing tool is supplied in CosmoLib to analyze and visualize the marginalized posterior of parameters. In the online documentation the reader can find the instructions on how to use this tool. ### Oscillations in the Current CMB Data? Recently a hint of the axion monodromy cosine oscillations (see eq. ) in WMAP-7yr [@Komatsu/etal:2011; @Larson/etal:2011] and ACT CMB data [@Dunkley/etal:2011] has been claimed in Ref. [@Aich/etal:2011]. Ref. [@Meerburg/etal:2012] confirms the finding that $\chi^2$ can be significantly improved in some regions of parameter space where oscillations in the primordial power spectrum are assumed. In this section we use CosmoLib to constrain the axion monodromy model with the same data sets. We find that when the CMB power spectrum is accurately computed and rigorous statistical method is used, there is [*no*]{} detectable axion monodromy oscillations in the CMB data. In Refs. [@Aich/etal:2011] the authors used their modified CAMB to compute the CMB power spectrum. As discussed in previous sections, such a modification is not trivial for $\delta\ln k \lesssim 10^{-2}$. Since the best-fit $\delta\ln k$ found in Ref. [@Aich/etal:2011] is small – $\delta\ln k \approx 0.005$ (derived from Table III of Ref. [@Aich/etal:2011] and equation 51 in Ref. [@Huang/etal:2012]), it is necessary to exam the numerical accuracy of the modified CAMB used in [@Aich/etal:2011]. For $\delta\ln k\approx 0.005$, the modulation period in $\ln k$ is $T_{\ln k} = 2\pi \delta\ln k\approx 0.03$. In the CMB power spectrum one should see same modulation period in $\ln \ell$, i.e., $T_{\ln \ell} = T_{\ln k} \approx 0.03$. Thus, from $\ell = 1000$ to $\ell = 1200$ there should be about $7$ oscillations in $C_\ell$. However, in Figure 5 of Ref. [@Aich/etal:2011] the number of oscillations in $C_\ell$ between $\ell = 1000$ and $\ell = 1200$ are much more than $7$. This implies that the “modulations” in $C_\ell$ shown in Ref. [@Aich/etal:2011] may just be numerical noises. In Figure \[fig:am\_Cls\] we show the CMB temperature angular power spectrum computed with CosmoLib, where the parameters are chosen to be close to the ones used in Figure 5 of Ref. [@Aich/etal:2011]. Qualitative difference can be seen between the two figures. The $C_\ell$ spectrum computed using CosmoLib presents clear modulations that agrees with the $\delta\ln k$ value, while the modified CAMB used in Ref. [@Aich/etal:2011] failed to produce the expected modulations. ![The CMB angular power spectrum for axion monodromy model with $\delta\ln k=0.005$, $\delta n_s = 0.18$. The other cosmological parameters are $\Omega_{b0}h^2 = 0.0223$, $\Omega_{c0}h^2 = 0.1119$, $\theta = 1.041$, $\tau_{\rm re} = 0.0884$, $n_s = 0.975$, $\ln (10^{10}A_s) = 3.04$. Modulation of $C_\ell$ is uniform in $\ln \ell$ and is almost invisible at high-$\ell$ due to lensing smearing. This should be compared to Fig. 5 in Ref. [@Aich/etal:2011], where the random fluctuations in $C_\ell$ implies insufficient numerical accuracy of the modified CAMB used therein. \[fig:am\_Cls\]](am_Cls.eps){width="\halffigurewidth"} In Ref. [@Huang/etal:2012] we pointed out that, a significant improvement of $\chi^2$ does not necessarily imply a detection of models with periodic oscillations, which typically has a spiky likelihood that is highly non-Gaussian. A rigorous treatment is to compute the marginalized probability of the amplitude of oscillations $\delta n_s$. The marginalization should be done in such a way that all the other cosmological and nuisance parameters are allowed to vary. A detection of monodromy oscillations should not be claimed unless $\delta n_s=0$ is excluded by the data. We did the full marginalization using MCMC method. The CMB power spectra are computed using the accurate integrator in CosmoLib. The marginalized 68.3% and 95.4% confidence level posterior contours are shown in Figure \[fig:deltansdeltalnk\]. ![The marginalized 68.3% and 95.4% confidence level contours of $\delta n_s$ and $\delta\ln k$ for axion monodromy model. WMAP-7yr and ACT data are used. CMB angular power spectrum are computed up to $\ell = 4000$ with CMB lensing effect included. Uniform priors $0.003\le \delta\ln k\le 0.2$ and $0\le n_s\le 0.2$ are used. No detection of axion monodromy oscillations are found since zero amplitude of oscillations ($\delta n_s = 0$) is consistent with the data. \[fig:deltansdeltalnk\]](mono_wmapact_p13_p12_2D.eps){width="\halffigurewidth"} Conclusions \[sec:conclusion\] ============================== We introduced the numerical package CosmoLib and focused on its features that are complementary to other numerical codes. The major advantage of CosmoLib is that it can accurately compute CMB angular power spectrum for inflationary models that predict sharp features in the primordial power spectrum of metric perturbations. This is not available in any other publicly available CMB codes. CosmoLib can calculate the relative fluctuations in $C_\ell$ to accuracy $\sim 10^{-3}$. Because of cosmic variance, we cannot measure $C_\ell$ to this accuracy [*if all $C_\ell$ are treated independently*]{}. However, our purpose is to use CosmoLib to study specific models where the degrees of freedom in the $C_\ell$ spectrum is small. In other words, if we assume a specific model (such as the axion monodromy model), the relative error in $C_\ell$ can be constrained to a level that is well below cosmic variance. In the naive limit where the $C_\ell$ spectrum is controlled by a single scaling parameter $s$, that is, $C_\ell = s C_{\ell, \rm fiducial}$, we can constrain $C_\ell$ to a relative accuracy $1/\sqrt{N} \approx 1/\ell_{\max}$, where $N= \sum_\ell (2\ell +1) \approx \ell_{\max}^2$ is the total number of measured spherical harmonic modes. For a future experiment that measures $C_\ell$ to cosmic variance for $\ell$ up to a few thousands [@PlanckScienceTeam:2009], the aforementioned $10^{-3}$ relative accuracy is necessary. While a straightforward (but not optimized) modification of CAMB and CLASS to use non-smooth $P(k)$ seems to be trivial, in practice it is often limited by the available memory and tolerable computation time. We pointed out that the modified CAMB in Ref. [@Aich/etal:2011] produces numerical noises instead of the expected modulation in $C_\ell$ spectrum. Repeating the computation in Ref. [@Aich/etal:2011] using CosmoLib and the same data sets (WMAP + ACT), we found no detection or hint of axion monodromy model in the current CMB data. This forecast toolkit contains a fisher matrix calculator, a MCMC engine, a postprocessing tool for chain analysis, and likelihoods for future CMB, galaxy survey, and supernova observations. The MCMC engine has an option of using a periodic proposal density, which can significantly accelerate the convergence of the chains in the case where the likelihood is a periodic function of some parameters. Although the likelihood models in CosmoLib are likely to be too simple for real experiments with complicated specifications, they provide a quick [*estimation*]{} of the performance of future CMB/LSS/SN experiments, for which the details of specifications are not yet known. We are planning to include more likelihoods for, e.g., weak lensing experiments in future releases. I thank Licia Verde, Filippo Vernizzi, Cyril Pitrou, Julien Lesgourgues and Emiliano Sefusatti for useful advice and discussions. [10]{} P. J. E. [Peebles]{}, in [*Research supported by the National Science Foundation. Princeton, N.J., Princeton University Press, 1980. 435 p.*]{}, edited by [Peebles, P. J. E.]{} (PUBLISHER, 41 William Street, Princeton, New Jersey, USA, 08540-5237, 1980). J. A. [Peacock]{}, [*Cosmological Physics*]{} (Cambridge University Press, 100 Brook Hill Drive, West Nyack, NY 10994-2133, 1999). V. F. [Mukhanov]{}, [*Physical Foundations of Cosmology*]{} (Cambridge University Press, 100 Brook Hill Drive, West Nyack, NY 10994-2133, 2005). S. [Weinberg]{}, [*Cosmology*]{} (Cambridge University Press, 100 Brook Hill Drive, West Nyack, NY 10994-2133, 2008). A. G. [Riess]{} [*et al.*]{}, Astrophys. J. [**732**]{}, 129 (2011). R. [Amanullah]{} [*et al.*]{}, Astrophys. J. [**716**]{}, 712 (2010). M. [Sullivan]{} [*et al.*]{}, Astrophys. J. [**737**]{}, 102 (2011). B. A. [Reid]{} [*et al.*]{}, Mon. Not. R. Astron. Soc. [**404**]{}, 60 (2010). W. J. [Percival]{} [*et al.*]{}, Mon. Not. R. Astron. Soc. [**401**]{}, 2148 (2010). E. [Komatsu]{} [*et al.*]{}, Astrophys. J., Suppl. [**192**]{}, 18 (2011). D. J. [Fixsen]{} [*et al.*]{}, Astrophys. J. [**473**]{}, 576 (1996). U. [Seljak]{} and M. [Zaldarriaga]{}, Astrophys. J. [**469**]{}, 437 (1996). A. [Lewis]{}, A. [Challinor]{}, and A. [Lasenby]{}, Astrophys. J. [**538**]{}, 473 (2000). M. [Doran]{}, Journal of Cosmology and Astro-Particle Physics [**10**]{}, 11 (2005). J. [Lesgourgues]{}, ArXiv:1104.2932 (2011). D. [Blas]{}, J. [Lesgourgues]{}, and T. [Tram]{}, Journal of Cosmology and Astro-Particle Physics [**7**]{}, 34 (2011). W. [Hu]{} and M. [White]{}, Phys. Rev. D [**56**]{}, 596 (1997). W. [Hu]{}, U. [Seljak]{}, M. [White]{}, and M. [Zaldarriaga]{}, Phys. Rev. D [**57**]{}, 3290 (1998). L. F. [Abbott]{} and M. B. [Wise]{}, Nuclear Physics B [**244**]{}, 541 (1984). F. [Lucchin]{} and S. [Matarrese]{}, Phys. Rev. D [**32**]{}, 1316 (1985). F. [Lucchin]{} and S. [Matarrese]{}, Physics Letters B [**164**]{}, 282 (1985). D. H. [Lyth]{} and E. D. [Stewart]{}, Physics Letters B [**274**]{}, 168 (1992). E. D. [Stewart]{} and D. H. [Lyth]{}, Physics Letters B [**302**]{}, 171 (1993). A. A. [Starobinsky]{}, Pis ma Zhurnal Eksperimental noi i Teoreticheskoi Fiziki [**55**]{}, 477 (1992). X. [Chen]{}, R. [Easther]{}, and E. A. [Lim]{}, Journal of Cosmology and Astro-Particle Physics [**6**]{}, 23 (2007). C. R. [Contaldi]{}, M. [Peloso]{}, L. [Kofman]{}, and A. [Linde]{}, Journal of Cosmology and Astro-Particle Physics [**7**]{}, 2 (2003). J. M. [Cline]{}, P. [Crotty]{}, and J. [Lesgourgues]{}, Journal of Cosmology and Astro-Particle Physics [**9**]{}, 10 (2003). D. [Polarski]{} and A. A. [Starobinsky]{}, Nuclear Physics B [**385**]{}, 623 (1992). D. [Langlois]{} and F. [Vernizzi]{}, Journal of Cosmology and Astro-Particle Physics [**1**]{}, 2 (2005). N. [Barnaby]{}, Z. [Huang]{}, L. [Kofman]{}, and D. [Pogosyan]{}, Phys. Rev. D [ **80**]{}, 043501 (2009). N. Barnaby and Z. Huang, Phys. Rev. D [**80**]{}, 126018 (2009). A. [Chambers]{} and A. [Rajantie]{}, Physical Review Letters [**100**]{}, 041302 (2008). J. R. [Bond]{}, A. V. [Frolov]{}, Z. [Huang]{}, and L. [Kofman]{}, Physical Review Letters [**103**]{}, 071301 (2009). E. [Silverstein]{} and A. [Westphal]{}, Phys. Rev. D [**78**]{}, 106003 (2008). R. [Bean]{} [*et al.*]{}, Journal of Cosmology and Astro-Particle Physics [ **3**]{}, 26 (2008). X. [Chen]{}, R. [Easther]{}, and E. A. [Lim]{}, Journal of Cosmology and Astro-Particle Physics [**4**]{}, 10 (2008). L. [McAllister]{}, E. [Silverstein]{}, and A. [Westphal]{}, Phys. Rev. D [**82**]{}, 046003 (2010). R. [Flauger]{} [*et al.*]{}, Journal of Cosmology and Astro-Particle Physics [ **6**]{}, 9 (2010). V. F. [Mukhanov]{}, H. A. [Feldman]{}, and R. H. [Brandenberger]{}, Phys. Rep. [ **215**]{}, 203 (1992). C.-P. [Ma]{} and E. [Bertschinger]{}, Astrophys. J. [**455**]{}, 7 (1995). E. [Bertschinger]{}, in [*Cosmology and Large Scale Structure*]{}, edited by [R. Schaeffer, J. Silk, M. Spiro, & J. Zinn-Justin]{} (PUBLISHER, ADDRESS, 1996), p. 273. M. [Bruni]{}, S. [Matarrese]{}, S. [Mollerach]{}, and S. [Sonego]{}, Classical and Quantum Gravity [**14**]{}, 2585 (1997). N. [Bartolo]{}, S. [Matarrese]{}, and A. [Riotto]{}, Journal of Cosmology and Astro-Particle Physics [**6**]{}, 24 (2006). N. [Bartolo]{}, S. [Matarrese]{}, and A. [Riotto]{}, Journal of Cosmology and Astro-Particle Physics [**1**]{}, 19 (2007). K. [Nakamura]{}, Progress of Theoretical Physics [**117**]{}, 17 (2007). K. [Enqvist]{}, J. [H[ö]{}gdahl]{}, S. [Nurmi]{}, and F. [Vernizzi]{}, Phys. Rev. D [**75**]{}, 023515 (2007). K. [Nakamura]{}, Phys. Rev. D [**80**]{}, 124021 (2009). K. A. [Malik]{} and D. [Wands]{}, Phys. Rep. [**475**]{}, 1 (2009). L. [Boubekeur]{} [*et al.*]{}, Journal of Cosmology and Astro-Particle Physics [**8**]{}, 29 (2009). C. [Pitrou]{}, Classical and Quantum Gravity [**26**]{}, 065006 (2009). D. [Nitta]{} [*et al.*]{}, Journal of Cosmology and Astro-Particle Physics [ **5**]{}, 14 (2009). L. [Senatore]{}, S. [Tassev]{}, and M. [Zaldarriaga]{}, Journal of Cosmology and Astro-Particle Physics [**8**]{}, 31 (2009). K. [Nakamura]{}, Advances in Astronomy [**2010**]{}, (2010). M. [Beneke]{} and C. [Fidler]{}, Phys. Rev. D [**82**]{}, 063509 (2010). F. [Bernardeau]{}, C. [Pitrou]{}, and J.-P. [Uzan]{}, Journal of Cosmology and Astro-Particle Physics [**2**]{}, 15 (2011). M. [Aich]{}, D. K. [Hazra]{}, L. [Sriramkumar]{}, and T. [Souradeep]{}, ArXiv:1106.2798 (2011). P. D. [Meerburg]{}, R. A. M. J. [Wijers]{}, and J. P. [van der Schaar]{}, Mon. Not. R. Astron. Soc. 2297 (2012). Z. [Huang]{}, L. [Verde]{}, and F. [Vernizzi]{}, Journal of Cosmology and Astro-Particle Physics [**4**]{}, 5 (2012). A. [Einstein]{}, Sitzungsberichte der K[ö]{}niglich Preu[ß]{}ischen Akademie der Wissenschaften (Berlin), Seite 142-152. 142 (1917). M. [Chevallier]{} and D. [Polarski]{}, International Journal of Modern Physics D [**10**]{}, 213 (2001). E. V. [Linder]{}, Physical Review Letters [**90**]{}, 091301 (2003). Z. [Huang]{}, J. R. [Bond]{}, and L. [Kofman]{}, Astrophys. J. [**726**]{}, 64 (2011). S. [Seager]{}, D. D. [Sasselov]{}, and D. [Scott]{}, Astrophys. J., Lett. [**523**]{}, L1 (1999). W. Y. [Wong]{}, A. [Moss]{}, and D. [Scott]{}, Mon. Not. R. Astron. Soc. [**386**]{}, 1023 (2008). J. R. [Bond]{} and G. [Efstathiou]{}, Astrophys. J., Lett. [**285**]{}, L45 (1984). W. [Fang]{}, W. [Hu]{}, and A. [Lewis]{}, Phys. Rev. D [**78**]{}, 087303 (2008). M. [Doran]{}, Journal of Cosmology and Astro-Particle Physics [**6**]{}, 11 (2005). A. [Vikman]{}, Phys. Rev. D [**71**]{}, 023515 (2005). W. [Hu]{}, Phys. Rev. D [**71**]{}, 047301 (2005). R. R. [Caldwell]{} and M. [Doran]{}, Phys. Rev. D [**72**]{}, 043527 (2005). C. [Pitrou]{}, J.-P. [Uzan]{}, and F. [Bernardeau]{}, Journal of Cosmology and Astro-Particle Physics [**7**]{}, 3 (2010). U. [Seljak]{}, , 1 (1996). M. [Zaldarriaga]{} and U. [Seljak]{}, , 023003 (1998). P. [Adshead]{}, R. [Easther]{}, J. [Pritchard]{}, and A. [Loeb]{}, Journal of Cosmology and Astro-Particle Physics [**2**]{}, 21 (2011). A. [Lewis]{} and A. [Challinor]{}, Phys. Rep. [**429**]{}, 1 (2006). J. A. [Rubi[ñ]{}o-Mart[í]{}n]{}, J. [Chluba]{}, W. A. [Fendt]{}, and B. D. [Wandelt]{}, Mon. Not. R. Astron. Soc. [**403**]{}, 439 (2010). L. [Verde]{}, H. V. [Peiris]{}, and R. [Jimenez]{}, Journal of Cosmology and Astro-Particle Physics [**1**]{}, 19 (2006). D. [Baumann]{} [*et al.*]{}, in [*American Institute of Physics Conference Series*]{}, Vol. 1141 of [*American Institute of Physics Conference Series*]{}, edited by [S. Dodelson, D. Baumann, A. Cooray, J. Dunkley, A. Fraisse, M. G. Jackson, A. Kogut, L. Krauss, M. Zaldarriaga, & K. Smith ]{} (PUBLISHER, ADDRESS, 2009), pp. 10–120. A. [Albrecht]{} [*et al.*]{}, ArXiv:astro-ph/0609591 (2006). N. [Kaiser]{}, Mon. Not. R. Astron. Soc. [**227**]{}, 1 (1987). J. A. [Peacock]{}, in [*New Insights into the Universe*]{}, Vol. 408 of [ *Lecture Notes in Physics, Berlin Springer Verlag*]{}, edited by [V. J. Martinez, M. Portilla, & D. Saez]{} (PUBLISHER, ADDRESS, 1992), pp. 1–+. J. A. [Peacock]{} and S. J. [Dodds]{}, Mon. Not. R. Astron. Soc. [**267**]{}, 1020 (1994). R. [Laureijs]{}, ArXiv:0912.0914 (2009). M. [Tegmark]{} [*et al.*]{}, Astrophys. J. [**499**]{}, 555 (1998). A. [Lewis]{} and S. [Bridle]{}, Phys. Rev. D [**66**]{}, 103511 (2002). D. [Larson]{} [*et al.*]{}, Astrophys. J., Suppl. [**192**]{}, 16 (2011). J. [Dunkley]{} [*et al.*]{}, Astrophys. J. [**739**]{}, 52 (2011). , [*Planck: The Scientific Programme (Blue Book). ESA-SCI (2005)-1. Version 2.*]{} (European Space Agency., ADDRESS, 2009). [^1]: Here we do not include CMBFast, which is no longer supported by its authors or available for download. [^2]: This refers to CosmoLib Version 0.2. [^3]: CosmoLib is a mixture of Fortran and C codes. The main part is written in Fortran. [^4]: syn.: synchronous gauge; Newt.: Newtonian gauge; gauge-inv.: gauge-invariant variables. [^5]: Newtonian gauge is implemented in CLASS version 1.3. [^6]: A second-order perturbation code is used to study the CMB non-Gaussianity. [^7]: The second-order part of CosmoLib is not released with this paper.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We discuss the origin of charge density wave ( CDW ) and spin density wave ( SDW ) in p-wave, d-wave and f-wave superconductors. To describe the low-energy quasiparticle excitation of p-wave case, we introduce a two- ( one for time and one for space ) dimensional massless Dirac model. After the non-Abelian bosonization is performed, the charge and spin density waves emerge from the model. By using this scheme, we try to explain the characteristic aspect of phase diagrams of various compounds, oxides and organic superconductors. The purpose of this paper is to make an argument that the dimensionality of the nodal excitation in superconductors plays an important role in the determination of the structure of the phase diagram.' address: 'Research Center for Nuclear Physics, Osaka University, Ibaraki, Osaka, Japan.' author: - Tadafumi Ohsaku title: 'Chiral Symmetry and Collective Excitations in p-wave, d-wave and f-wave Superconductors' --- Motivated by recent experimental discoveries of the coexistence of antiferromagnetic and superconducting orders \[1\], Franz and Tesanovic, and independently Herbut found that the chiral symmetry and its dynamical breaking ( dynamical chiral symmetry breaking, ${\rm D}\chi{\rm SB}$ ) is realized in d-wave copper oxide superconductors \[2\]. They introduced a four-component Dirac field $\Psi$ to describe the nodal excitation of quasiparticles in d-wave superconductors, and derived a low energy effective theory. They considered the coupling between the quasiparticles and fluctuating vortices of the system by a gauge interaction. Then the low energy effective theory becomes the three- ( one for time and two for space ) dimensional two-flavor massless quantum electrodymanics ( ${\rm QED}_{3}$ ). The massless ${\rm QED}_{3}$ has a chiral symmetry; The Lagrangian is invariant under $\Psi\to e^{i\gamma_{5}\theta}\Psi$. It is a famous fact that the four-component ${\rm QED}_{3}$ dymanically generates a parity-conserving Dirac mass \[3\]. Consulting on the field-theoretical result of ${\rm QED}_{3}$, they discussed the chiral symmetry breaking in their low-energy effective theory, and they observed that the chiral condensate $m_{dyn}\langle\bar{\Psi}\Psi\rangle$ ( $m_{dyn}$; the dynamical mass ) is an alternating spin density wave ( SDW ). Based on the result, they argued that the system ( d-wave superconductor with fluctuating vortices ) has an antiferromagnetic instability as the dynamical origin of ${\rm QED}_{3}$ model, and explained the reason of the existence of the antiferromagnetic order in the phase diagram of copper oxide superconductors. The essential part of their discussions and conclusions, especially about the phenomenon of ${\rm D}\chi{\rm SB}$ can also be obtained by the following Lagrangian: $$\begin{aligned} \sum^{2}_{n=1}\bigl(\bar{\Psi}_{n}i\gamma^{\mu}\partial_{\mu}\Psi_{n}+G^{(3)}_{0}[(\bar{\Psi}_{n}\Psi_{n})^{2}+(\bar{\Psi}_{n}i\gamma_{5}\Psi_{n})^{2}]\bigr).\end{aligned}$$ Here, we take the same definition of $\Psi_{n}$ as that of the ${\rm QED}_{3}$ model given by Franz-Tesanovic and Herbut. This model is simple, and at least for studying the ${\rm D}\chi{\rm SB}$, the calculation is easier ( though we have to introduce a cutoff ) than the gauge theory, ${\rm QED}_{3}$. It is clear from their logic, the ${\rm QED}_{3}$ model can be applied to [*all*]{} d-wave superconductors ( not only to copper oxide but also to d-wave organic superconductors ). Our four-fermi model (1) can also be applied to [*all*]{} d-wave superconductors. By introducing the local one-particle density matrix $Q(x)=-\langle\Psi(x)\bar{\Psi}(x)\rangle$, we proceed to perform the group-theoretical classification for the order parameter developed from our theory \[4$\sim$7\]. $Q(x)$ is a 4$\times$4 matrix, then we can expand it by 16-dimensional complete set of gamma matrices: $$\begin{aligned} Q &=& Q^{s}\hat{1}+Q^{V}_{\mu}\gamma^{\mu}+Q^{T}_{\mu\nu}\sigma^{\mu\nu}+Q^{A}_{\mu}\gamma_{5}\gamma^{\mu}+Q^{P}i\gamma_{5}.\end{aligned}$$ Here $S$, $V$, $T$, $A$ and $P$ denote the scalar, vector, tensor, axial vector and pseudoscalar, respectively. In fact, the dynamical mass discussed by Franz-Tesanovic and Herbut corresponds to the scalar density $Q^{S}$. If we examine each component of the matrix $Q$ more in detail, we can discuss the possibility of the appearances of other types of order. Now we study the problem, and intend to publish our results elsewhere. On the other hand, a low-energy effective theory for point-like-node p-wave superconductors ( similar to the case of the ABM ( Anderson-Brinkman-Morel ) state ) becomes a two-dimensional ( one for time and one for space ) massless Dirac fermion model: The system has two Fermi points in a specific direction in momentum space, and quasiparticles are easily excited near the Fermi points. If we describe the low-energy long-wavelength excitation by $\psi_{R\sigma}(z)e^{ik_{F}z}+\psi_{L\sigma}(z)e^{-ik_{F}z}$, ( here, $R$ denotes a right mover, $L$ denotes a left mover and $\sigma$ denotes a spin quantum number ) we will obtain a two-flavor massless Dirac fermion model: $$\begin{aligned} \sum_{\sigma}\bar{\psi}_{\sigma}i\gamma^{\mu}\partial_{\mu}\psi_{\sigma}.\end{aligned}$$ Here, we take the definition of the two-component Dirac field as $\psi_{\sigma}=(\psi_{R\sigma}(z),\psi_{L\sigma}(z))$. The gamma matrices are given by $\gamma_{0}=\sigma_{1}$, $\gamma_{1}=-i\sigma_{2}$, $\gamma_{5}=\gamma_{0}\gamma_{1}=\sigma_{3}$. The effective Lagrangian (3) also has the chiral symmtery. If we consider a chiral invariant four-body contact interaction, its mathematical form is severely restricted. Add a chiral invariant interaction to the Dirac kinetic term (3), we get $$\begin{aligned} \sum^{2}_{n=1}\bigl(\bar{\psi}_{n}i\gamma^{\mu}\partial_{\mu}\psi_{n}+G^{(2)}_{0}[(\bar{\psi}_{n}\psi_{n})^{2}+(\bar{\psi}_{n}i\gamma_{5}\psi_{n})^{2}]\bigr).\end{aligned}$$ Any continuous symmetry in one-dimension cannot be spontaneously broken \[8\]. In the one-dimensional case, the non-Abelian bosonization procedure \[9$\sim$14\] should be employed. After incorporate the band multiplicity in our model, our Hamiltonian will be decoupled to three sectors: $U(1)$ ( charge ), $SU(2)$ ( spin ) and $SU(N)$ ( orbital or band multiplicity ). Then we can write down the bosonized Hamiltonian in the Sugawara form \[15\]: $$\begin{aligned} H &=& H_{U(1)}+H_{SU(2)}+H_{SU(N)}, \\ H_{U(1)} &=& 2\pi v_{charge}\int dx(:J_{R}J_{R}:+:J_{L}J_{L}:+G:J_{R}J_{L}:), \\ H_{SU(2)} &=& \frac{2\pi}{2+N}v_{spin}\sum^{3}_{a=1}\int dx(:J^{a}_{R}J^{a}_{R}:+:J^{a}_{L}J^{a}_{L}:-G:J^{a}_{R}J^{a}_{L}:), \\ H_{SU(N)} &=& \frac{2\pi}{2+N}v_{orb}\sum^{N^{2}-1}_{A=1}\int dx(:J^{A}_{R}J^{A}_{R}:+:J^{A}_{L}J^{A}_{L}:), \end{aligned}$$ In the expression given above, the spin-charge-orbital separation was occured. By using the conformal field theoretical techniques with renormalization group approach \[12,13\], we can predict that the excitation in each sector becomes gapless ( massless ) or gapful ( massive ) \[13\]. Then we determine what kind of order ( CDW, SDW and “orbital wave” ) will emerge. For example, when the spectrum of the charge sector is massless, CDW will arise, while it is massive, CDW will not appear. It is clear from our discussion, this model can be applied to [*all*]{} systems which have point-like p-wave nodes. To examine the physics of CDW, Sakita et al. used the same Lagrangian with (4). They discussed the chiral symmetry of the Lagrangian to study the CDW \[16\]. The most important point in our discussion is in the following logic: The excitation of p-wave superconductors will be described by the chiral invariant model, and when a kind of perturbation ( interaction between particles ) is applied, CDW or SDW may appear/disappear. Because of the dimensionality of the nodal excitation in p-wave systems, CDW and/or SDW can appear. Our context in this paper is different from that of Su and Sakita. It should be emphasized that our theory, combined (1) with (4), can explain SDW, CDW and other possible phases, while a phenomenological Landau-Ginzburg-type $SO(5)$ model introduced by Zhang \[17\] can only explain antiferromagnetic phase, superconducting phase and coexistense of them. Let us consider various superconducting systems of real substances. Recently, some experiments found the existence and/or coexistence of CDW, SDW and other ordered state in some superconductors. For example, the coexistences of CDW and SDW in $({\rm TMTTF})_{2}{\rm Br}$, $({\rm TMTSF})_{2}{\rm PF}_{6}$ ( p- or f-wave superconductor ) and $\alpha -({\rm BEDT-TTF})_{2}{\rm MHg}({\rm SCN})_{4}$ ( non-pure s-wave ) were observed. The phese diagram of $({\rm BEDT-TTF})_{3}{\rm Cl}_{2}({\rm H_{2}O})_{2}$ has a CDW phase neighbor a superconducting phase. The importance of charge fluctuation with ferromagnetic spin fluctuation in ${\rm Sr_{2}RuO_{4}}$ ( p- or f-wave superconductor ) was pointed out by Takimoto \[18\]. Kuroki et al. performed a theoretical investigation about the effect of the coexistence of CDW and SDW in $({\rm TMTSF})_{2}{\rm PF}_{6}$ \[19\]. Neighbor the superconducting phase of ${\rm UGe_{2}}$ ( p- or f-wave superconductor ), there is a CDW/SDW coexistent phase. We recognize almost all of these substances are p- or f-wave superconductors. We speculate that the CDW phase or CDW/SDW coexistent phase may emerge by the mechanism of the generation of chiral mass in two-dimensional system, or by collective excitations of charge and spin in one-dimensinal system. We suppose the pairing symmetry of the superconducting phase in $({\rm BEDT-TTF})_{3}{\rm Cl}_{2}({\rm H_{2}O})_{2}$ is a p-wave type ( though there is no experimental report about it ). We would like to make an argument that p-wave, d-wave and f-wave superconductors generally have the SDW/CDW instability. To the contrary, s-wave superconductors do not have such kind of instability. Usually, the phase diagrams of p-wave, d-wave and f-wave superconductors have several ordered phases, while the phase diagram of s-wave should become a simple one. The chiral symmetry arises from the nodal structure of superconducting gap, and play the key-role in the coexistence/competition of various phases in phase diagrams of superconductors. Finally, we wish to make a comment on the confinememt-deconfinement transition ( CDT ) in superconductors. In quantum chromodynamics ( QCD ), quark confinement occurs at low-energy low-density state, and the ${\rm D}\chi{\rm SB}$ is realized, dynamical mass is generated. At high-density state, the quark-deconfinement occurs and the color-superconductivity will be realized \[20,21\]. Similar to this case, the confinement wil be realized at low-energy in ${\rm QED}_{3}$, while deconfinement will occur at high-density state \[22\]. There is a similarity between the phase diagram of copper oxide and that of QCD: We speculate SDW corresponds to ${\rm D}\chi{\rm SB}$ phase, while superconductivity corresponds to color-superconductivity. It is case that there are several similarities between ${\rm QED}_{3}$ and ${\rm QCD}_{4}$. Therefore, there is a possibility to understand the phase diagram of copper oxide by the concept of CDT. We suppose both the ${\rm D}\chi{\rm SB}$ and CDT are universal phenomena in various condensed matter. For example, Gabovich et al. Phys. Rep. [**367**]{}, 583 (2002). M. Franz and Z. Tesanovic, Phys. Rev. Lett. [**87**]{}, 257003 (2001), I. F. Herbut, Phys. Rev. Lett. [**88**]{}, 047006 (2002), Z. Tesanovic, O. Vafek and M. Franz, Phys. Rev. [**B65**]{}, 180511 (2002), M. Franz, D. E. Sheehy and Z. Tesanovic, Phys. Rev. Lett. [**88**]{}, 257005 (2002), O. Vafek, Z. Tesanovic and M. Franz, cond-mat/0203047, M. Franz, Z, Tesanovic and O. Vafek, Phys. Rev. [**B66**]{}, 054535 (2002), I. F. Herbut, Phys. Rev. [**B66**]{}, 094504 (2002), D. J. Lee and I. F. Herbut, Phys. Rev. [**B66**]{}, 094512 (2002), B. H. Seradjeh and I. F. Herbut, cond-mat/027221, G.-Z. Liu and G. Cheng, Phys. Rev. [**B66**]{}, 100505 (2002). T. Perag-Barnea and M. Franz, cond-mat/0209301. R. D. Pisarski, Phys. Rev. [**D29**]{}, 2423 (1984), T. Appelquist, M. J. Bowick, E. Cohler and L. C. R. Wijewardhana, Phys. Rev. Lett. [**55**]{}, 1715 (1985), T. W. Appelquist, M. Bowick, D. Karabali and L. C. R. Wijewardhana, Phys. Rev. [**D33**]{}, 3704 (1986). T. Ohsaku, Thesis, Department of Physics, Osaka University (2000). T. Ohsaku, Phys. Rev. [**B65**]{}, 024512 (2002). T. Ohsaku, Phys. Rev. [**B66**]{}, 054518 (2002). T. Ohsaku, cond-mat/0209352, submitted for publication. S. Coleman, Commun. Math. Phys. [**31**]{}, 259 (1973). E. Witten, Commun. Math. Phys. [**92**]{}, 455 (1984). A. M. Polyakov and P. B. Wiegmann, Phys. Lett. [**131B**]{}, 121 (1983), Phys. Lett. [**141B**]{}, 223 (1984). I. Affleck, Nucl. Phys. [**B265**]{}, 409, 448 (1986). V. G. Knizhnik and A. B. Zamolodchikov, Nucl. Phys. [**B247**]{}, 83 (1984), A. B. Zamolodchikov and V. A. Fateev, Sov. J. Nucl. Phys. [**43**]{}, 657 (1986). A. M. Tsvelik, [*Quantum Field Theory in Condensed Matter Physics*]{} (Cambridge University Press, New York, 1995). P. B. Wiegmann, Phys. Rev. [**B59**]{}, 15705 (1999). H. Sugawara, Phys. Rev. [**170**]{}, 1659 (1968). Z.-b. Su and B. Sakita, Phys. Rev. Lett. [**56**]{}, 780 (1986), Phys. Rev [**B38**]{}, 7421 (1988), B. Sakita and K. Shizuya, Phys. Rev. [**B42**]{}, 5586 (1990). S.-C. Zhang, Science, [**275**]{}, 1089 (1997). T. Takimoto, Phys. Rev. [**B62**]{}, R14641 (2000). K. Kuroki et al., cond-mat/0006218. M. Iwasaki and T. Iwado, Phys. Lett. [**B350**]{}, 163 (1995). M. Alford, K. Rajagopal and F. Wilczek, Phys. Lett. [**B422**]{}, 247 (1998), R. Rapp, T. Schaefer, E. V. Shuryak and M. Velkovsky, Phys. Rev. Lett. [**81**]{}, 53 (1998). P. Maris, Phys. Rev. [**D52**]{}, 6087 (1995).
{ "pile_set_name": "ArXiv" }
--- abstract: 'Under the assumption that the quantum parameter $q$ is an $l$th primitive root of unity with $l$ odd in a field $F$ of characteristic 0 and $m+n\geq r$, we obtained a complete classification of irreducible modules of the $q$-Schur superalgebra $\sS_F(m|n,r)$ introduced in [@DR].' address: - 'J.D., School of Mathematics and Statistics, University of New South Wales, Sydney NSW 2052, Australia' - 'H.G. & J.W., Department of Mathematics, East China Normal University, Shanghai, China' author: - 'Jie Du, Haixia Gu and Jianpan Wang' title: 'Irreducible representations of $q$-Schur superalgebras at a root of unity' --- [^1] Introduction ============ The investigation on Schur superalgebras and their quantum analogue has achieved significant progress in the last decade; see, e.g., [@BKv; @BK; @D; @M; @DR; @TK]. This includes an establishment of a super analogue of the Schur–Weyl reciprocity by Mitsuhashi, a Kazhdan–Lusztig type cell approach to the representation theory of $q$-Schur superalgebras by Rui and the first author, and, recently, a presentation of $q$-Schur superalgebras by Turkey and Kujawa. As part of a super version of the quantum Schur–Weyl theory, these developments are all in the generic case where the quantum parameter is generic or not a root of unity. This paper attempts to investigate the structure and representations of $q$-Schur superalgebra at a root of unity. As is seen from [@DR] and [@TK], $q$-Schur superalgebras share many of the properties of $q$-Schur algebras. For example, both have several definitions in terms of Hecke algebras via (super) $q$-permutation modules, quantum enveloping algebras via tensor (super)spaces, and quantum coordinate (super)algebras via the dual of homogenous components; both have integral versions and the base change property; both have Drinfeld-Jimbo type presentations; etc. However, the structure of $q$-Schur superalgebras is fundamentally different. For example, no quasihereditary structure or cellular structure is seen naturally. Thus, classifying irreducible modules requires an approach that is different from the usual quasihereditary or cellular approach. In this paper, we will generalise the approach used in [@DU; @DU3] for $q$-Schur algebras to $q$-Schur superalgebras. This approach was developed as a $q$-deformation of the modular permutation representation theory established by L. Scott in 1973 [@S]. One important feature of this approach is that the resulting classification of irreducible representations is analogous to Alperin’s weight conjecture [@A] for modular representations of a finite group. More precisely, we consider a $q$-analogue of an endomorphism algebra of a signed permutation module instead of a group algebra, consider $l$-parabolic subgroups instead of $p$-subgroups, and consider representations of a quotient algebra defined by an $l$-parabolic subgroup as “local” representations. Thus, the isomorphism types of irreducible modules are determined by an $l$-parabolic subgroup $P$ together with a projective indecomposable module of the quotient algebra associated with $P$. The main ingredients in this approach are the notion of relative norms for representations of Hecke algebras of type $A$, which was introduced by P. Hoefsmit and L. Scott in 1977, and its corresponding properties which include transitivity, Mackey decomposition, Frobenius reciprocity, Nakayama relation, Higman criterion; see [@J]. By describing a basis for a $q$-Schur superalgebra in relative norms, we will attach a defect group, which is an $l$-parabolic subgroup, to a basis element and introduce a filtration of ideals in terms of defect groups. Thus, every primitive idempotent $e$ lies in a minimal ideal in the sequence. We then attach to $e$ a defect group $D(e)$. This defect group is the vertex of the corresponding indecomposable module for the Hecke algebra. If $D(e)$ is trivial, then the indecomposable module is projective. Hence, it is determined by an $l$-regular partition. If $D(e)$ is nontrivial, then the image of $e$ in the quotient algebra by the ideal right below $e$ has the trivial defect group. Thus, once we show that the quotient algebra is again an endomorphism algebra part of which is again a $q$-Schur superalgebra, we may determine all primitive idempotents with the trivial defect group in the quotient algebra. In this way, we completely determine the isomorphism types of irreducible modules for a $q$-Schur superalgebras $\sS_F(m|n,r)$ with $m+n\geq r$ by the index set $\mathcal P_r$ (Theorem \[ML4\]). We point out that classification of irreducible supermodules for Schur superalgebras was first completed by Donkin [@D] under the condition $m,n\geq r$. Brundan and Kujawa extended the result to arbitrary $m,n,r$ ([@BK Th. 6.5]) in their beautiful paper for a proof of the Mullineux conjecture. Their approach is quite different from ours, involving a certain category equivalence, Mullineux conjugation function, a process of removing nodes from $p$-rims of a partition, and a couple of other relevant functions. See Appendix II for more details, where we will also make a comparison between the two classification index sets $\La^{++}(m|n,r)$ and $\mathcal P_r$ when $m+n\geq r$ and $l=p$. Like Alperin’s weight conjecture, our approach is not constructive for irreducible modules. However, it indicates a certain tensor product structure. We plan to address this issue in a different paper; see Remark \[Alperin\]. One possible approach is to quantise the work of Brundan–Kujawa. However, it would be interesting to find an explicit construction compatible with the index set $\mathcal P_r$. We also point out that we only consider the $q$-Schur superalgebra as an algebra and representations as algebra representation in this paper. The contents of the paper are organised as follows: 1. Introduction 2. Quantum Schur superalgebras 3. Relative norms: the first properties 4. Some vanishing properties of relative norms on $V_R(m|n)^{\otimes r}$ 5. Bases for $\sS_R(m|n,r)$ in relative norms 6. A filtration of ideals of $\sS_R(m|n,r)$ 7. Alternative characterisation of the ideals $I_F(P,r)$ 8. Quantum matrix superalgebras 9. Frobenius morphism and Brauer homomorphisms 10. Shrinking defect groups via Brauer homomorphisms 11. Classification of irreducible $\sS_F(m|n,r)$-modules 12. Appendix I: Brauer homomorphisms without Frobenius 13. Appendix II: A comparison with the classification of Brundan-Kujawa Throughout the paper, let $\sZ=\mathbb{Z}[\up,\up^{-1}]$ be the Laurent polynomial ring in indeterminate $\up$ and let $R$ be a domain which is a $\sZ$-module via a ring homomorphism $\sZ\to R$. We further assume that the image $q$ of $\up$ in $R$ is a primitive $l$th root of 1 with $l$ odd. We will always assume the characteristic char$(R)\neq2$. From §7 onwards, $R=F$ is a field of characteristic 0. For $a,b\in\mathbb N$ with $a<b$, we often write $$[a,b]=\{a,a+1,\ldots,b\}.$$ Quantum Schur superalgebras =========================== We first recall the definitions of $q$-Schur superalgebras in terms of signed $q$-permutation modules and tensor superspaces. Let $(W,S)$ be the symmetric group on $r$ letters where $W=\fS_r=\fS_{\{1,2,\ldots,r\}}$ and $S=\{s_k\mid 1\leq k<r\}$ is the set of basic transpositions $s_k=(k,k+1),$ and let $\ell:W\to\mathbb{N}$ be the length function with respect to $S$. An $n$-tuple $\lambda=(\lambda_1,\lambda_2,\cdots,\lambda_n)\in{\mathbb N}^n$ is called a composition of $r$ into $n$ parts if $|\la|:=\sum_i\la_i=r$. A composition $\la$ of $r$ is called a partition of $r$ if $\la_1\geq\la_2\geq\ldots$. Let $\La(n,r)\subset \mathbb{N}^n$ be the set of all compositions of $r$ into $n$ parts and let $\La^+(n,r)$ be the subset of partitions in $\La(n,r)$. In particular, let $\La^+(r)=\La^+(r,r)$ be the set of all partitions of $r$. We sometimes use the notation $\la\models r$ or $\la\vdash r$ for a composition or partition of $r$. The parabolic (or the standard Young) subgroup $W_{\lambda}$ of $W$ associated with a composition $\la$ consists of the permutations of $\{1,2,\cdots,r\}$ which leave invariant the following sets of integers $$\{1,2,\cdots,\lambda_1\},\{\lambda_1+1,\lambda_1+2,\cdots,\lambda_1+\lambda_2\},\{\lambda_1+\lambda_2+1,\lambda_1+\lambda_2+2,\cdots\},\cdots.$$ We will frequently use the following notation: if $W'$ is a subgroup of $W$ and $W_\mu$ is another parabolic subgroup, the notation $W_\mu=_WW_\la$ means $W_\mu^x:=x^{-1}W_\mu x=W_\la$ for some $x\in W$, while the notation $W_\mu\leq_WW'$ means that a parabolic conjugate of $W_\mu$ is a subgroup of $W'$, i.e.,$W_\mu^x\leq W'$ and $W_\mu^x$ is also parabolic for some $x\in W$. We will also denote by $\sD_\la:=\mathcal{D}_{W_\la}$ the set of all distinguished (or shortest) coset representatives of the right cosets of $W_\la$ in $W$. Let $\mathcal{D}_{\lambda\mu}=\mathcal{D}_\lambda\cap\mathcal{D}^{-1}_{\mu}$. Then $\mathcal{D}_{\lambda\mu}$ is the set of distinguished $W_\lambda$-$W_\mu$ coset representatives. For $d\in\mathcal{D}_{\la\mu}$, the subgroup $W_\la^d\cap W_\mu=d^{-1}W_\la d\cap W_\mu$ is a parabolic subgroup associated with a composition which is denoted by $\la d\cap\mu$. In other words, we define $$\label{ladmu} W_{\la d\cap\mu}=W_\la^d\cap W_\mu.$$ We will often regard a pair $(\lambda^{(0)},\lambda^{(1)})\in \Lambda(m,r_1)\times \Lambda(n,r_2)$ of compositions of $m$ parts and $n$ parts as a composition $\la$ of $m+n$ parts and write $$\la=(\lambda^{(0)}|\lambda^{(1)})=(\lambda^{(0)}_1,\lambda^{(0)}_2,\cdots,\lambda^{(0)}_m|\lambda^{(1)}_1, \lambda^{(1)}_2,\cdots,\lambda^{(1)}_n)$$ to indicate the“even” and “odd” parts of $\la$. Let $$\begin{aligned} \Lambda(m|n,r)&=\{\lambda=(\lambda^{(0)}|\lambda^{(1)}) \mid\la\in\La(m+n,r)\}\\ &=\bigcup_{r_1+r_2=r}(\La(m,r_1)\times\La(n,r_2))\\ \end{aligned}$$ By identifying $\Lambda(m|n,r)$ with $\Lambda(m+n,r)$, the notations $W_\la$, $\sD_\la$, $\sD_{\la\mu}$, etc., are all well defined for all $\la,\mu\in\La(m|n,r)$. \[xxx\] For $\rho=(\rho_1,\rho_2,\ldots,\rho_t)\models r$, we will sometimes write the parabolic subgroups $$W_\rho=W_{\rho_1}\times W_{\rho_2}\times\cdots \times W_{\rho_t}$$ where $W_{\rho_1}=\fS_{\{1,\ldots,\rho_1\}}$, $W_{\rho_2}=\fS_{\{\rho_1+1,\ldots,\rho_1+\rho_2\}}$, and so on. For $\la_{(i)},\mu_{(i)}\models\rho_i$, the notation $\sD_{\la_{(i)}\mu_{(i)}}$, etc., are defined relative to $W_{\rho_i}$. In particular, for $\lambda=(\lambda^{(0)}|\lambda^{(1)})\in\Lambda(m|n,r)$, we make the following notational convention: $$\label{notation} W_\la= W_{\la^{(0)}}W_{\la^{(1)}}=W_{\la^{(0)}}\times W_{\la^{(1)}},$$ where $W_{\lambda^{(0)}}\leq\fS_{\{1,2,\ldots,m\}}$ and $W_{\lambda^{(1)}}\leq\fS_{\{m+1,\ldots,m+n\}}$ are the even and odd parts of $W_\la$, respectively. \[00-11\] For $\lambda=(\lambda^{(0)}\mid\lambda^{(1)}),\mu=(\mu^{(0)}\mid\mu^{(1)})\in\Lambda(m|n,r)$, $d\in\mathcal{D}_{\la}$, and $d'\in\mathcal{D}_{\mu}$ let $W_{\la d\cap\mu d'}:=W_\la^d\cap W_\mu^{d'}$ and $W_{\la d\cap\mu d'}^{ij}=W_{\la^{(i)}}^d\cap W_{\mu^{(j)}}^{d'}$ for all $i,j\in\{0,1\}$. Then $$W_{\la d\cap\mu d'}=W_{\la d\cap\mu d'}^{00} \times W_{\la d\cap\mu d'}^{11}\times W_{\la d\cap\mu d'}^{01}\times W_{\la d\cap\mu d'}^{10}.$$ Moreover, if $s\in S\cap W_{\la d\cap\mu d'}$, then $s\in W_{\la d\cap\mu d'}^{ij}$ for some $i,j\in\{0,1\}$. In particular, any parabolic subgroup $W_\eta$ of $W_{\la d\cap\mu d'}$ is a product of parabolic subgroups $W_{\eta^{ij}}=W_\eta\cap W_{\la d\cap\mu d'}^{ij}$ with $i,j\in\{0,1\}$. The assertion about direct product is clear. For $s\in S$, if $s\in W_\la^d$, then $s=d^{-1}s'd$ or $ds=s'd$ for some $s'\in W_\la$. Thus, $1+\ell(d)\leq\ell(s')+\ell(d)=\ell(ds)\leq 1+\ell(d)$ forces $\ell(s')=1$. Hence, $s'\in S$ and $s'\in W_{\la^{(0)}}$ or $s'\in W_{\la^{(1)}}$ and so, $s\in W_{\la^{(0)}}^d$ or $s\in W_{\la^{(1)}}^d$. Similarly, one proves that $s\in W_{\mu^{(0)}}^{d'}$ or $s\in W_{\mu^{(1)}}^{d'}$. Hence, $s\in W_{\la d\cap\mu d'}^{ij}$ for some $i,j\in\{0,1\}$. The rest of the proof is clear. For $d\in\mathcal{D}_{\la\mu}$, $W_{\la d\cap \mu}$ itself is a parabolic subgroup which is decomposed into parabolic subgroups $$\label{ladmu00-11} W_{\la d\cap\mu}=W_{\la d\cap\mu}^{00} \times W_{\la d\cap\mu}^{11}\times W_{\la d\cap\mu}^{01}\times W_{\la d\cap\mu}^{10}.$$ In this case, the composition $\nu=\la d\cap \mu$ has the form $\nu=(\nu^1,\ldots,\nu^{m+n})$ with $\nu^i\in\La(m+n,\mu_i)$. We say that $d$ satisfies the [*even-odd trivial intersection property*]{} if $W_{\la d\cap\mu}^{01}=1=W_{\la d\cap\mu}^{10}$. For $\lambda,\mu\in\Lambda(m|n,r)$, let $$\label{Dcirc} \mathcal{D}^\circ_{\lambda\mu}=\{d\in\mathcal{D}_{\lambda\mu}\mid W^d_{\lambda^{(0)}}\cap W_{\mu^{(1)}}=1,W^d_{\lambda^{(1)}}\cap W_{\mu^{(0)}}=1\}.$$ This set is the super version of the usual $\sD_{\la\mu}$. For $d\in\mathcal{D}^\circ_{\lambda\mu}$, if we put $W_{\nu^{(0)}}=W^d_{\lambda^{(0)}}\cap W_{\mu^{(0)}},W_{\nu^{(1)}}=W^d_{\lambda^{(1)}}\cap W_{\mu^{(1)}},\nu=(\nu^{(0)}|\nu^{(1)})$, then $W_\nu=W_{\nu^{(0)}}\times W_{\nu^{(1)}}$. We have, in general, $\nu\in\La(m'|n',r)$ where $m'=m(m+n)$ and $n'=n(m+n)$. The Hecke algebra $\mathcal{H}_R=\sH_R(W)$ corresponding to $W=\fS_r$ is a free $R$-module with basis $\{\mathcal{T}_w; w\in W\}$ and the multiplication is defined by the rules: for $s\in S$, $$\mathcal{T}_w\mathcal{T}_s=\left\{\begin{aligned} &\mathcal{T}_{ws}, &\mbox{if } \ell(ws)>\ell(w);\\ &(q-q^{-1})\mathcal{T}_w+\mathcal{T}_{ws}, &\mbox{otherwise}. \end{aligned} \right.$$ Note that $\{T_w=q^{\ell(w)}\sT_w\}_{w\in W}$ is the usual defining basis satisfying relations $${T}_w{T}_s=\left\{\begin{aligned} &{T}_{ws}, &\mbox{if } \ell(ws)>\ell(w);\\ &(q^2-1)T_w+q^2{T}_{ws}, &\mbox{otherwise}. \end{aligned} \right.$$ If $W'$ is a parabolic subgroup of $W$, then the $R$-module $\sum_{w\in W'}R\mathcal{T}_w$ is a subalgebra of $\mathcal{H}_R$, which is called a [*parabolic subalgebra*]{} of $\mathcal{H}_R$, denoted by $\mathcal{H}_{W'}$. We will use the abbreviation $\mathcal{H}_\lambda$ instead of $\mathcal{H}_{W_\lambda}$. Let $W'$ be a parabolic subgroup of $W$ and let $d_{W'}$ denote the [*Poincaré polynomial*]{} of $W'$, i.e., $$d_{W'}=d_{W'}(\bsu)=\sum_{w\in W'} \bsu^{\ell(w)}.$$ In particular, the Pioncaré polynomial of $W$ has the form $d_W=[\![r]\!]^{!}:=[\![1]\!][\![2]\!]\cdots[\![r]\!]$, where $[\![i+1]\!]=1+\bsu+\bsu^2+\cdots+\bsu^i.$ For later use in §9 define, for $0\leq t\leq s$, the Gaussian polynomials by $$\left[\!\!\left[t\atop s\right]\!\!\right]=\frac{[\![s]\!]^!}{[\![t]\!]^![\![s-t]\!]^!}.$$ Thus, $d_W(q^2)=0=[\![l]\!]_{q^2}$ if $r\geq l$, where $q$ is a primitive $l$th root of 1 with $l$ odd. For $\lambda=(\lambda^{(0)}\mid\lambda^{(1)})\in\Lambda(m|n,r)$, define (see ) $$x_{\lambda^{(0)}}=\sum_{w\in W_{\lambda^{(0)}}}T_w,\qquad y_{\lambda^{(1)}}=\sum_{w\in W_{\lambda^{(1)}}}(-q^2)^{-\ell(w)}T_w,$$ where $T_w=q^{\ell(w)}\mathcal{T}_w$. We call the module $x_{\lambda^{(0)}}y_{\lambda^{(1)}}\mathcal{H}_{R}$ a [*signed $q$-permutation module*]{} (cf. [@D]). \[S(m|n,r)\] Let $\fT_R(m|n,r)=\bigoplus_{\lambda\in\Lambda(m| n,r)}x_{\lambda^{(0)}}y_{\lambda^{(1)}}\mathcal{H}_{R}.$ The algebra $$\sS_R(m|n,r):= \End_{\mathcal{H}_R}(\fT_R(m|n,r))$$ is called a $q$-[*Schur superalgebra*]{} over $R$ on which the $\mathbb Z_2$-graded structure is induced from the $\mathbb Z_2$-graded structure on $\fT_R(m|n,r)$ with $$\fT_R(m|n,r)_i=\bigoplus_{{\lambda\in\Lambda(m| n,r)}\atop {|\la^{(1)}|\equiv i(\text{mod}2)}}x_{\lambda^{(0)}}y_{\lambda^{(1)}}\mathcal{H}_{R} \;\;(i=0,1).$$ If $R=\sZ:={\mathbb Z}[\up,\up^{-1}]$, we simply write $\sS(m|n,r)$, $\fT(m|n,r)$, etc. for $\sS_\sZ(m|n,r)$, $\fT_\sZ(m|n,r)$, etc. Note that it is proved in [@DR] that $\sS_R(m|n,r)\cong\sS(m|n,r)\otimes_\sZ R$. Following [@DR], define, for $\lambda,\mu\in\Lambda(m|n,r)$ and $d\in\mathcal{D}^\circ_{\lambda\mu}$, $$T_{W_\lambda d W_\mu}:=\sum_{\substack{w_0w_1 \in W_\mu\cap \mathcal{D}_\nu,\\ w_0\in W_{\mu^{(0)}},w_1\in W_{\mu^{(1)}}}}(-q^2)^{-\ell(w_1)}x_{\lambda^{(0)}}y_{\lambda^{(1)}}T_dT_{w_0}T_{w_1}.$$ There exists $\mathcal{H}_{R}$-homomorphism $\psi^d_{\lambda\mu}$ such that $$(x_{\alpha^{(0)}}y_{\alpha^{(1)}}h)\psi_{\mu,\la}^d=\delta_{\mu,\alpha}T_{W_\lambda d W_\mu}h, \forall \alpha\in\Lambda(m| n,r),h\in\mathcal{H}_R.$$ Note that we changed the left hand notation $\phi^d_{\lambda\mu}$ used in [@DR (5.7.1)] to the right hand notation $\psi^d_{\mu,\la}$ here for notational simplicity later on.[^2] The following result is given in [[@DR 5.8]]{}. We will provide a different proof below in §5 by using relative norms. \[DR5.8\] The set $\{\psi^d_{\mu\lambda}\mid \lambda,\mu\in\Lambda(m|n, r), d\in\mathcal{D}^\circ_{\lambda\mu}\}$ forms a $R$-basis for $\sS_R(m|n,r )$. \[m’|n’ case\] For $m\leq m'$ and $n\leq n'$, we may embed $\La(m|n,r)$ into $\La(m'|n',r)$ by adding zeros at the end of each of the two sequences $\la^{(0)}$ and $\la^{(1)}$ for every $\la\in\La(m|n,r)$. Let $\vep=\sum_{\la\in\La(m|n,r)}\psi_{\la\la}^1$. Then $\sS(m|n,r)\cong \vep\sS(m'|n',r)\vep$. Thus, we simply regard $\sS(m|n,r)$ as a centralizer subalgebra of $\sS(m'|n',r)$. \[DR8.1\] Let $R=\mathbb{Q}(\up)$. \(1) The non-isomorphic irreducible $\sS_R(m|n,r)$-modules are indexed by the set $$\La^+(r)_{m|n}=\{(\la_1,\la_2,\ldots)\in\La^+(r)\mid \la_{m+1}\leq n\}.$$ \(2) Assume $m+n\geq r$. The $\sS_R(m|n,r )$-$\mathcal{H}_R$ bimodule structure $\fT_R(m|n,r )$ satisfies the following double centralizer property $$\sS_R(m|n,r )=\End_{\mathcal{H}_R}(\fT_R(m| n,r ))\mbox{ and } \mathcal{H}_R\cong \End_{\sS_R(m| n,r )}(\fT_R(m|n,r )).$$ Moreover, there is a category equivalence $$\Hom_{\mathcal{H}_R}(-,\fT_R(m| n,r)):\modh\mathcal{H}_R\longrightarrow \sS_R(m| n,r )\hmod.$$ We now describe $\sS_R(m|n,r )$ in terms of tensor superspaces. Let $V_R(m|n)$ be a free $R$-module of rank $m+n$ with basis $v_1,v_2,\cdots,v_{m+n}$. The [*parity map*]{}, by setting $\hat{i}={0}$ if $1\leq i\leq m$, and $\hat{i}={1}$ otherwise, gives $V_R(m|n)$ a $\mathbb{Z}_2$-graded structure: $V_R(m|n)=V_0\oplus V_1$, where $V_0$ is spanned by $v_1,\cdots,v_m$, and $V_1$ is spanned by $v_{m+1},\cdots,v_{m+n}$. Thus, $V_R(m|n) $ becomes a “superspace”. Let $$I(m|n,r)=\{\bsi=(i_1,i_2,\cdots,i_r)\in \mathbb{N}^r\mid 1\leq i_j\leq m+n,\forall j\}.$$ For each $\lambda=(\lambda^{(0)}\mid\lambda^{(1)} )\in \Lambda(m| n,r)$, define $\bsi_\lambda\in I(m|n,r)$ by $$\bsi_\lambda=(\underbrace{1,\ldots,1}_{\la^{(0)}_1},\ldots,\underbrace{m,\ldots,m}_{\la^{(0)}_m},\underbrace{m+1,\ldots,m+1}_{\la^{(1)}_1}\ldots,\underbrace{m+n,\ldots,m+n}_{\la^{(1)}_n}).$$ For convenience, denote $v_\lambda:=v_{\bsi_\lambda}$. The symmetric group $W=\fS_r$ acts on $I(m|n, r)$ by place permutation. For $w\in W, \bsi\in I(m|n,r)$ $$\bsi w=(i_{w(1)},i_{w(2)},\cdots,i_{w(r)}).$$ For each $\bsi\in I(m|n,r)$, define $\lambda\in \Lambda (m|n,r)$ to be the weight $wt(\bsi)$ of $v_\bsi$ be setting $\lambda_k=\#\{k\mid i_j=k,1\leq j\leq r\},\lambda^{(0)}=(\lambda_1,\lambda_2,\cdots,\lambda_m),\lambda^{(1)}=(\lambda_{m+1},\lambda_{m+2},\cdots,\lambda_{m+n}).$ For $\bsi=(i_1,i_2,\cdots,i_r)\in I(m| n,r)$, let $$v_\bsi=v_{i_1}\otimes v_{i_2}\otimes\cdots \otimes v_{i_r}=v_{i_1}v_{i_2}\cdots v_{i_r}.$$ Clearly, the set $\{v_\bsi\}_{\bsi\in I(m|n,r)}$ form a basis of $V_R(m| n)^{\otimes r}$. Following [@M], $V_R(m|n,r)^{\otimes r}$ is a right $\sH_R$-module with the following action: $$\label{action2} v_\bsi\mathcal{T}_{s_k}= \begin{cases} (-1)^{\widehat{i_k}\widehat{i_{k+1}}} v_{\bsi s_k}, &\mbox{ if } i_k<i_{k+1};\\ q v_\bsi, &\mbox{ if } i_k=i_{k+1}\leq m;\\ (-q^{-1}) v_\bsi, &\mbox{ if } i_k=i_{k+1}\geq m+1;\\ (-1)^{\widehat{i_k}\widehat{i_{k+1}}} v_{\bsi s_k}+(q-q^{-1}) v_\bsi,&\mbox{ if } i_k>i_{k+1}. \end{cases}$$ where $s_k=(k,k+1)$. Note that when $n=0$, this action coincides with the action on the usual tensor space as given in [@DDPW (14.6.4)], commuting with the action of quantum $\mathfrak{gl}_n$. \[dhat\] For $\lambda\in\Lambda(m|n,r), d\in \mathcal{D}_\lambda$, if $\bsi_\lambda d=(i_1,i_2,\cdots,i_r)$, define $$(\la,d)^\wedge =\sum_{k=1}^{r-1}\sum_{k<l,i_k>i_l}\widehat{i_k}\widehat{i_l}.$$ When $\la$ is clear from the context, we write $\whd=(\la,d)^\wedge$. The following relation will be repeatedly used in the sequel. $$\label{diaoyudao} (-1)^{\whd}(-1)^{\widehat{i_k}\widehat{i_{k+1}}}=(-1)^{\widehat{ds_k}}\text{ for all } d\in\mathcal{D}_\lambda, s_k\in S\text{ with }ds_k\in\mathcal{D}_\lambda.$$ \[iso f\] The right $\mathcal{H}_R$-module $V_R(m|n)^{\otimes r}$ is isomorphic to the $\mathcal{H}_R$-module $\fT_R(m|n,r)$ defined in Definition \[S(m|n,r)\] under the map: $$f:V_R(m| n)^{\otimes r}\longrightarrow \bigoplus_{\lambda\in\Lambda(m| n,r)}x_{\lambda^{(0)}}y_{\lambda^{(1)}}\mathcal{H}_R,\quad(-1)^{\hat{d}}v_{\bsi_\lambda d}\longmapsto x_{\lambda^{(0)}}y_{\lambda^{(1)}}\mathcal{T}_d.$$ This isomorphism induces a superalgebra isomorphism $$\sS_R(m| n,r)\cong \End_{\mathcal{H}_R}(V_R(m| n)^{\otimes r}),$$ sending $g\in\sS_R(m|n,r)$ to $fgf^{-1}$ for all $g\in\sS_R(m|n,r)$. The action on a signed $q$-permutation module is given by $$x_{\lambda^{(0)}}y_{\lambda^{(1)}}\mathcal{T}_d\mathcal{T}_{s_k}= \begin{cases} x_{\lambda^{(0)}}y_{\lambda^{(1)}}\mathcal{T}_{ds_k},&\mbox{ if }ds_k\in\mathcal{D}_\lambda;\\ qx_{\lambda^{(0)}}y_{\lambda^{(1)}}\mathcal{T}_d, &\mbox{ if } ds_k=s_ld, s_l\in W_{\lambda^{(0)}};\\ -q^{-1}x_{\lambda^{(0)}}y_{\lambda^{(1)}}\mathcal{T}_d, &\mbox{ if } ds_k=s_ld, s_l\in W_{\lambda^{(1)}};\\ x_{\lambda^{(0)}}y_{\lambda^{(1)}}\mathcal{T}_{ds_k}+(q-q^{-1})x_{\lambda^{(0)}}y_{\lambda^{(1)}}\mathcal{T}_d,&\mbox{ if } ds_k<d. \end{cases}$$ Now, the relation $(-1)^{\hat{d}}(-1)^{\widehat{i_k}\widehat{i_{k+1}}}=(-1)^{\widehat{ds_k}}$ implies that $f$ is a right $\mathcal{H}_R$-module homomorphism. From now on, we will identify $\sS_R(m| n,r)$ with $\End_{\mathcal{H}_R}(V_R(m| n)^{\otimes r})$. Relative norms: the first properties ==================================== In 1977, P. Hoefsmit and L. Scott introduced the notion of relative norms, which is the $q$-analogue of relative traces in group representations (see, e.g., [@Lan]) and use it to investigate the representation theory of Hecke algebras. The following material is taken from their unpublished manuscript. A proof can now be found in [@J]. \[rel norm\] (1) Let $\lambda,\mu$ be the compositions of $r$ such that $W_\lambda\leq W_\mu$. Let $M$ be an $\mathcal{H}_\mu$-$\mathcal{H}_\mu$-bimodule and $b\in M$. Define the [*relative norm*]{} $$N_{W_\mu,W_\lambda}(b)=\sum_{w\in\mathcal{D}_\lambda\cap W_\mu} \mathcal{T}_{w^{-1}}b\mathcal{T}_w.$$ \(2) For an $\mathcal{H}_\lambda$-$\mathcal{H}_\lambda$-bimodule $M$, we define $$Z_M(\mathcal{H}_\lambda)=\{m\in M\mid hm=mh,\forall h\in \mathcal{H}_\lambda\}.$$ It is easy to see that $Z_{\mathcal{H}_\lambda}(\mathcal{H}_\lambda)$ is the center of $\mathcal{H}_\lambda$ and, for $M=\Hom_R(N,N)$ where $N$ is a right $\mathcal{H}_\lambda$-module, $Z_M(\mathcal{H}_\lambda)=\Hom_{\mathcal{H}_\lambda}(N,N)$. In particular, if $M=\End_R(V_R(m|n)^{\otimes r})$, then $Z_M(\mathcal{H}_R)=\sS_R(m|n,r)$. We first list some properties of relative norms. \[RNT\] Let $M$ be an $\mathcal{H}_R$-$\mathcal{H}_R$-bimodule and let $W_\lambda$ and $W_\mu$ be parabolic subgroups of $W$. - (Transitivity) If $W_\mu\leq W_\lambda$ and $b\in M$, then $$N_{W,W_\mu}(N_{W_\mu,W_\lambda}(b))=N_{W,W_\lambda}(b).$$ - $N_{W,W_\lambda}(Z_M(\mathcal{H}_\lambda))\subseteq Z_M(\mathcal{H}_{R}).$ Next, we list the $q$-analogues of four useful results known as Mackey decompsition, Frobenius reciprocity, Nakayama relation, and Higman criterion for relative projectivity. See [@Lan] or [@Feit] for their classical version for groups. \[Mackey\] If $N$ is an $\mathcal{H}_{R}$-$\mathcal{H}_\lambda$-bimodule, then $$(N\otimes_{\mathcal{H}_\lambda}\mathcal{H}_R)|_{\mathcal{H}_\mu}\cong \bigoplus_{d\in \mathcal{D}_{\lambda\mu}}(N\otimes_{\mathcal{H}_\lambda}\mathcal{T}_d)\otimes_{\mathcal{H}_{\nu}}\mathcal{H}_\mu,$$ where $\nu$ is defined by $W_{\nu}=W^d_\lambda\cap W_\mu$ for all $d\in \mathcal{D}_{\lambda\mu}$. In particular, if $M$ is an $\mathcal{H}_R$-$\mathcal{H}_R$-bimodule and $b\in Z_M(\mathcal{H}_\mathcal{\lambda})$, then $$N_{W, W_\lambda}(b)=\sum_{d\in \mathcal{D}_{\lambda\mu}}N_{W_\mu, W_{\la d\cap\mu}}(\mathcal{T}_{d^{-1}}b\mathcal{T}_d).$$ The Frobenius reciprocity simply follows from the fact that induction is a left adjoint functor to restriction. \[Frobenius\] Let $M$ by an $\sH_R$-module and $N$ be an $\sH_\mu$-module. Then there is an $R$-module isomorphism $$\varphi:\Hom_{\sH_\mu}(N, M|_{\sH_\mu})\overset\sim\longrightarrow \Hom_{\sH_R}(N\otimes_{\sH_\mu}\sH_R,M).$$ In particular, for $\la,\mu\in\La(m|n,r)$ and tensors $v_{\nu}=v_{\bsi_\nu}\in V_R(m|n)^{\otimes r}$ ($\nu=\la$ or $\mu$), the $R$-module isomorphism $$\varphi:\Hom_{\mathcal{H}_\mu}(Rv_\mu,v_\lambda\mathcal{H}_R)\longrightarrow \Hom_{\mathcal{H}_R}(v_\mu\mathcal{H}_R,v_\lambda\mathcal{H}_R)$$ induced from the Frobenius reciprocity is the restriction of the relative norm map $N_{W,W_\mu}(-)$, and so $$N_{W,W_\mu}(\Hom_{\mathcal{H}_\mu}(Rv_\mu,v_\lambda\mathcal{H}_R))= \Hom_{\mathcal{H}_R}(v_\mu\mathcal{H}_R,v_\lambda\mathcal{H}_R).$$ The proof of the last assertion is almost identical to that in [@DU Lem. 2.5] and is omitted. \[Nakayama\]Let $M$ be an $\mathcal{H}_R$-$\mathcal{H}_R$-bimodule. If $N$ is an $\mathcal{H}_{R}$-$\mathcal{H}_\lambda$-submodule of $M$ such that $M\cong N\otimes_{\mathcal{H}_\lambda}\mathcal{H}_{R}$. Then $$Z_M(\mathcal{H}_{R})=N_{W,W_\lambda}(Z_N(\mathcal{H}_\lambda)).$$ Moreover, if $W_\mu=W^d_\lambda$ for some $d\in \mathcal{D}_{\lambda\mu}$, then there exists an $\mathcal{H}_{R}$-$\mathcal{H}_\mu$-submodule $N'\cong N\otimes_{\mathcal{H}_\lambda}\mathcal{T}_d$ of $M$ such that $$N_{W, W_\lambda}(Z_N(\mathcal{H}_\lambda))=N_{W, W_\mu}(Z_{N'}(\mathcal{H}_\mu)).$$ If $X$ is an $\sH_R$-module, $Y$ is an $\sH_\la$-module, $$M=\Hom_R(X,Y\otimes_{\sH_\la}\sH_R),\;\text{ and }\; N=\Hom(X_{\sH_\la},Y),$$ then the usual Nakayama relation (see, e.g., [@DJ 2.6]) now becomes $$N_{W, W_\lambda}(\Hom_{\sH_\la}(X|_{\sH_\la},Y))=\Hom_{\sH_R}(X,Y\otimes_{\sH_\la}\sH_R).$$ A right $\mathcal{H}_R$-module $M$ is [*projective relative*]{} to $\mathcal{H}_\lambda$ or simply $\mathcal{H}_\lambda$-projective if for every pair of right $\mathcal{H}_R$-modules $M',M{''}$ the exact sequence $$0\rightarrow M'\rightarrow M{''}\rightarrow M\rightarrow 0$$ split provided it is a split exact sequence as $\mathcal{H}_\lambda$-modules. In the following result, the notation $X\mid Y$ means that $X$ is isomorphic to a direct summand of $Y$. \[Higman\] Let $M$ be a right $\mathcal{H}_R$-module. Then the following are equivalent: - $M$ is $\mathcal{H}_\lambda$-projective; - $M\mid M\otimes_{\mathcal{H}_\lambda} \mathcal{H}_R$; - $M\mid U\otimes_{\mathcal{H}_\lambda} \mathcal{H}_R$ for some right $\mathcal{H}_\lambda$-module $U$; - $N_{W,W_\lambda}(\Hom_{\mathcal{H}_\lambda}(M,M))=\Hom_{\mathcal{H}_R}(M,M)$. Let $M$ be a finitely generated indecomposable right $\mathcal{H}_R$-module. Then by [@J 3.35], there exists a parabolic subgroup $W_\lambda$ of $W$ unique up to conjugation such that $M$ is $\mathcal{H}_\lambda$-projective and such that $W_\lambda$ is $W$-conjugate to a parabolic subgroup of any parabolic subgroup $W_\mu$ of $W$ for which $M$ is $\mathcal{H}_\mu$-projective. We call $W_\lambda$ a [*vertex*]{} of $M$ which is unique up to conjugation. This notion is a generalisation of the vertex theory in the representation theory of finite groups. Motivated from the fact that a vertex must be a $p$-subgroup, we need the notion of $l$-parabolic subgroups. Let $l$ be a positive odd number and $l\leq r$. A parabolic subgroup $W_\lambda$ is called $l$-[*parabolic*]{} if all parts of $\lambda$ are 0, $1$, or $l$. Write $r=sl+t$ with $0\leq t<l$ and let $P_{(r)}$ be the parabolic subgroup of $W=\fS_r$ associated with the composition $(l^s,1^t)$. This is called in [@DU] ‘the’ maximal $l$-parabolic subgroup of $W$. For any composition $\la=(\la_1,\ldots,\la_a)$ of $r$, let the maximal $l$-parabolic subgroup of $W_\la$ be the parabolic subgroup[^3] $$\label{max l-parabolic} P_\la=P_{(\la_1)}\times\cdots\times P_{(\la_a)}.$$ A maximal $l$-parabolic subgroup of $W_\lambda$ is a parabolic subgroup $P$ of $W_\la$ such that $P=_{W_\la}P_\la$ (meaning $P^x:=x^{-1}Px=P_\la$ for some $x\in W_\la$). Let $\Phi_l=\Phi_l(\bsu)$ denote the $l$th cyclotomic polynomial. The Pioncaré polynomial of an $l$-parabolic subgroup $W_\la$ has the form $d_{W_\la}=(d_l)^s$, where $s$ is the number of parts $l$ in $\la$ and $$d_l=\prod^{l-1}_{i=1}(1+\bsu+\bsu^2+\cdots+\bsu^i).$$ Here are a few simple facts. \[DU1.1\] Let $\lambda$ be a composition of $r$ and $x\in W$. Then \(a) If $W^x_\lambda$ is parabolic, then $d_{W^x_\lambda}=d_{W_\lambda}$, \(b) $\Phi_l\nmid (d_{W_\lambda}/d_{P_\lambda} )$. \(c) Let $W_\lambda,W_\mu,W_\theta$ be parabolic subgroups of $W$ such that $$W_\theta\leq W^x_\mu, W_\mu\leq W^y_\lambda$$ where $W^x_\mu,W^y_\lambda$ are parabolic and $x,y\in W$. Assume $d_{P_\theta}\neq d_{P_\lambda}$. Then $\Phi_l\mid (d_{W_\lambda}/d_{W_\theta})$. The following result is a $q$-analogue of the fact for group representations that a vertex must be a $p$-group and will be repeatedly used later on. Let $F$ be a field of characteristic 0 in which $q$ is a primitive $l$th root of 1 (with $l$ odd). If $M$ is an indecomposable $\sH_F$-module, then the vertex of $M$ is an $l$-parabolic subgroup., We will describe the vertices of indecomposable direct summand of the $\sH_F$-module $V_F(m|n)^{\otimes r}$ in §10 in terms of the defect group of the corresponding primitive idempotent and apply this to classify all simple $\sS_F(m|n,r)$-modules when $m+n\geq r$. Some vanishing properties of relative norms on $V_R(m|n)^{\otimes r}$ ===================================================================== As seen from the remarks right after Definition \[rel norm\], the $q$-Schur superalgebra $\sS_R(m|n,r)= Z_M(\sH_R)$ where $M=\End_R(V_R(m|n)^{\otimes r})$. Thus, by Lemma \[RNT\](b), we can construct elements in $\sS_R(m|n,r)$ by applying relative norms to the matrix units in $M$. We first show in this section that some of these elements are simply the 0 transformation. We will construct a basis from this type of elements in the next section. For $\bsi,\bsj\in I(m| n,r)$, we define $e_{\bsi,\bsj}\in \End_R(V_R(m| n)^{\otimes r})$ to be the linear map $$(v_{\bsi'})e_{\bsi,\bsj}=\left\{ \begin{aligned} &v_\bsj,\mbox{ if }\bsi'=\bsi,\\ &0,\mbox{otherwise}. \end{aligned} \right.$$ If $(\bsi,\bsj)=(\bsi_\mu,\bsi_\lambda d)$ with $d\in \mathcal{D}_{\lambda\mu}$, we use the abbreviation $e_{\mu,\lambda d}$ instead of $e_{\bsi_\mu, \bsi_\lambda d}$. The following result is obvious from the definition, but will be useful later on. \[SRNP\] For $b\in V_R(m|n)^{\otimes r}=\bigoplus_{\la\in\La(m|n,r)}v_\la\sH_R$, if the projection of $b$ on $v_\mu\mathcal{H}_R$ is $0$ for some $\mu\in\La(m|n,r)$, then $(b)N_{W,W_{\la d\cap \mu}}(f)=0$ for all $\lambda\in\Lambda(m|n,r)$, $d\in\mathcal{D}^\circ_{\lambda\mu},$ and $f\in \Hom_{\mathcal{H}_{\la d\cap\mu}}(v_\mu\mathcal{H}_R,v_\lambda\mathcal{H}_R)$, extended (by sending other $v_\nu\sH_R$ to 0 for all $\nu\neq \mu$) to an element in $\End_{\sH_{\la d\cap\mu}}(V_R(m|n)^{\otimes r})$. Moreover, $N_{W,W_\mu}(e_{\mu,\mu})$ is the identity map on $v_\mu\mathcal{H}_R$ and $0$ elsewhere. By , $v_\bsi\mathcal{H}_R=\mathrm{span}\{v_{\bsi w}\mid w\in W\}$ for any $\bsi\in I(m|n,r)$. Hence, if the projection of $b$ on $v_\mu\mathcal{H}_R$ is $0$, then $(b)N_{W,W_{\la d\cap\mu}}(f)=0$. Note that if $x\in\mathcal{D}_\mu$ and $x\neq 1$, then $x^{-1}\notin W_\mu$. Hence $$\begin{aligned} (v_\mu)N_{W,W_\mu}(e_{\mu\mu}) &=(v_\mu)\sum_{w\in\mathcal{D}_\mu}\mathcal{T}_{w^{-1}}e_{\mu\mu}\mathcal{T}_w\\ &=(v_\mu)\sum_{w\in\mathcal{D}_\mu\cap W_{\mu}}\mathcal{T}_{w^{-1}}e_{\mu\mu}\mathcal{T}_w=v_\mu. \end{aligned}$$ Hence, $N_{W,W_\mu}(e_{\mu\mu})$ is the identity on $v_\mu\mathcal{H}_R$. By the proof above, it is $0$ elsewhere. The place permutation action of $W$ on $I(m|n,r)$ induces an action on $I(m|n,r)^2$: $(\bsi,\bsj)w=(\bsi w,\bsj w)$ for all $\bsi,\bsj\in I(m|n,r)$ and $w\in W$. Clearly, if $\bsi=\bsi_\la d$ and $\bsj=\bsi_\mu d'$ for some $d\in\sD_\la$ and $d'\in \sD_\mu$, then $$\text{Stab}_W(\bsi,\bsj):=\{w\in W\mid (\bsi,\bsj)w=(\bsi,\bsj)\} =W_\la^d\cap W_\mu^{d'}.$$ The following result is a super version of [@DU Lem. 2.2]. \[super2.2\] Let $\bsi=\bsi_\la d$ and $\bsj=\bsi_\mu d'$, where $d\in\sD_\la$ and $d'\in \sD_\mu$, and let $s\in W_{\la d\cap\mu d'}^{ij}\cap S$, where $i,j\in\{0,1\}$. Then $\sT_se_{\bsi,\bsj}=e_{\bsi,\bsj}\sT_s$ if and only if $(i,j)=(0,0)$ or $(1,1)$. Suppose $s=(a,a+1)$. Since $\bsi s=\bsi$ and $\bsj s=\bsj$, by the definition of the action , $(v_{\bsi'})\sT_se_{\bsi,\bsj}\neq 0$ if and only if $\bsi'=\bsi $ or $\bsi'=\bsi s$, or equivalently, $\bsi'=\bsi$. Now, write $s=d^{-1}s'd$ or $d'^{-1}s''d'$ for some $s'\in W_\la\cap S$ or $s''\in W_\mu\cap S$. If $(i,j)=(0,0)$, then $s'\in W_{\la^{(0)}}$, $s''\in W_{\mu^{(0)}}$, and $$(v_\bsi)\sT_se_{\bsi,\bsj}=(-1)^{\widehat d}v_{\bsi_\la}\sT_d\sT_se_{\bsi,\bsj}=(-1)^{\widehat d}v_{\bsi_\la}\sT_{s'}\sT_de_{\bsi,\bsj}=qv_\bsj=v_\bsj\sT_{s}=(v_\bsi) e_{\bsi,\bsj}\sT_s.$$ Similarly, if $(i,j)=(1,1)$, then $s'\in W_{\la^{(1)}}$, $s''\in W_{\mu^{(1)}}$, and $$(v_\bsi)\sT_se_{\bsi,\bsj}=(-q^{-1})v_\bsj=v_\bsj\sT_{s}=(v_\bsi) e_{\bsi,\bsj}\sT_s.$$ However, if $(i,j)=(0,1)$, then $s'\in W_{\la^{(0)}}$, $s''\in W_{\mu^{(1)}}$ and $$(v_\bsi)\sT_se_{\bsi,\bsj}=qv_\bsj\text{ and }(v_\bsi) e_{\bsi,\bsj}\sT_s=-q^{-1}v_\bsj.$$ Hence, $\sT_se_{\bsi,\bsj}\neq e_{\bsi,\bsj}\sT_s$. The proof for the (1,0) case is similar. \[SPC\] Let $\lambda,\mu\in \Lambda(m|n,r),d\in\mathcal{D}_{\lambda\mu}$, and $W_{\nu}=W^d_\lambda\cap W_\mu$. - If $d\in\mathcal{D}_{\lambda\mu}^\circ$, then $e_{\mu,\lambda d}\in \End_{\mathcal{H}_{\nu}}(V_R(m|n)^{\otimes r})$. In particular, we have $$N_{W,W_{\nu}}(e_{\mu,\lambda d})\in \End_{\mathcal{H}_R}(V_R(m|n,r)^{\otimes r}).$$ - If $d\in\mathcal{D}_{\lambda\mu}\setminus\mathcal{D}^\circ_{\lambda\mu}$, then $e_{\mu,\lambda d}\notin \End_{\mathcal{H}_{\nu}}(V_R(m|n)^{\otimes r})$. If $d\in \mathcal{D}^\circ_{\lambda\mu}$, we have $W^d_\lambda\cap W_\mu=W_{\nu}=W_{\nu^{(0)}}\times W_{\nu^{(1)}} $ where $W_{\nu^{(0)}}=W^d_{\lambda^{(0)}}\cap W_{\mu^{(0)}},W_{\nu^{(1)}}=W^d_{\lambda^{(1)}}\cap W_{\mu^{(1)}}$. By the lemma above, for $w=w_0w_1\in W_{\nu}$ with $w_i\in W_{\nu^{(i)}},$ $$\begin{aligned} (v_\mu\mathcal{T}_w)e_{\mu,\lambda d}&=q^{\ell(w_0)}(-q^{-1})^{\ell(w_1)} (v_\mu)e_{\mu,\lambda d}\\ &= v_{\bsi_\lambda d}\sT_{w_0}\sT_{w_1}\\ &=(v_\mu)e_{\mu,\la d}\sT_{w_0}\sT_{w_1}. \end{aligned}$$ Hence, $\mathcal{T}_we_{\mu,\lambda d}= e_{\mu,\lambda d}\mathcal{T}_w,$ proving $e_{\mu,\lambda d}\in \End_{\mathcal{H}_{\nu}}(V_R(m|n)^{\otimes r})$. The last assertion in (1) follows from Lemma \[RNT\](b). If $d\in\mathcal{D}_{\lambda\mu}\backslash\mathcal{D}^\circ_{\lambda\mu}$, by definition of $\mathcal{D}^\circ_{\lambda\mu}$, we have $$W^d_{\lambda^{(0)}}\cap W_{\mu^{(1)}}\neq 1\mbox{ or } W^d_{\lambda^{(1)}}\cap W_{\mu^{(0)}}\neq 1.$$ Hence, by the lemma above, there exists $s\in W_\nu$ such that $\sT_se_{\bsi,\bsj}\neq e_{\bsi,\bsj}\sT_s$. The following vanishing property is somewhat surprising. Recall the notation introduced in . \[SRTC\] For $\lambda,\mu \in \Lambda(m|n,r)$ and $d\in\mathcal{D}_{\lambda\mu}\setminus \mathcal{D}^\circ_{\lambda\mu}$, if a parabolic subgroup $W_\eta\leq W_{\la d\cap\mu}^{00}\times W_{\la d\cap\mu}^{11}$, then $N_{W,W_\eta}(e_{\mu,\lambda d})=0$. By the hypothesis and Lemma \[super2.2\], $e_{\mu,\la d}\in\End_{\sH_\eta}(V_R(m|n)^{\otimes r})$. Hence, by Lemma \[RNT\](b), $N_{W,W_{\eta}}(e_{\mu,\lambda d})\in \End_{\mathcal{H}_R}(V_R(m|n)^{\otimes r})$. From the definition of $e_{\mu,\lambda d}$, it is enough by Lemma \[SRNP\] to consider the action of $N_{W,W_\eta}(e_{\mu,\lambda d})$ on $v_\mu$. Thus, we have $$\begin{aligned} (v_\mu)N_{W,W_\eta}(e_{\mu,\lambda d})&=(v_\mu)\sum_{w\in \sD_\eta}\mathcal{T}_{w^{-1}} e_{\mu,\lambda d}\mathcal{T}_w\\ &=\sum_{w_0w_1\in \sD_\eta\cap(W_{\mu^{(0)}}\times W_{\mu^{(1)}})}(v_\mu) \mathcal{T}_{(w_0w_1)^{-1}} e_{\mu,\lambda d}\mathcal{T}_{w_0w_1}\\ &=\sum_{w_0w_1\in \sD_\eta\cap(W_{\mu^{(0)}}\times W_{\mu^{(1)}})}q^{\ell(w_0)}(-q^{-1})^{\ell(w_1)} v_{\bsi_\lambda d}\mathcal{T}_{w_0w_1} \end{aligned}$$ Since $d\in\mathcal{D}_{\lambda\mu}\setminus \mathcal{D}^\circ_{\lambda\mu}$, we have $$W_{\nu}:=W^d_\lambda\cap W_\mu=W_\nu^{00}\times W_\nu^{01}\times W_\nu^{10}\times W_\nu^{11},$$ where $W_\nu^{ij}=W^d_{\lambda^{(i)}}\cap W_{\mu^{(j)}},$ for all $i,j=0,1$, and $W_\mu=W_{\nu}(\mathcal{D}_{\nu}\cap W_\mu)$. For $w_0\in \sD_\eta\cap W_{\mu^{(0)}}$, there exist $x_0x_1\in (\sD_\eta\cap W^{00}_\nu)\times W^{10}_\nu$ and $d_0\in \mathcal{D}_{\nu}\cap W_{\mu^{(0)}}$ such that $w_0=x_0x_1d_0$. For $w_1\in W_{\mu^{(1)}} $, there are $y_0y_1\in W^{01}_\nu \times(\sD_\eta\cap W^{11}_\nu)$ and $d_1\in \mathcal{D}_{\nu}\cap W_{\mu^{(1)}}$ such that $w_1=y_0y_1 d_1$. Therefore, $w_0w_1=x_0x_1y_0y_1d_0d_1$ and $\ell(w_0w_1)=\ell(x_0)+\ell(x_1)+\ell(y_0)+\ell(y_1)+\ell(d_0)+\ell(d_1)$. Consequently, we have $$\begin{aligned} &(v_\mu)N_{W,W_\eta}(e_{\mu,\lambda d})\\ &=\sum_{\substack{x_0x_1\in(\sD_\eta\cap W^{00}_\nu) \times W^{10}_\nu\\ y_0y_1\in W^{01}_\nu\times(\sD_\eta\cap W^{11}_\nu)\\ d_0\in\mathcal{D}_{\nu(d)}\cap W_{\mu^{(0)}},d_1\in \mathcal{D}_{\nu(d)}\cap W_{\mu^{(1)}}} } q^{\ell(x_0x_1d_0)}(-q^{-1})^{\ell(y_0y_1d_1)}v_{\bsi_\lambda d}\mathcal{T}_{x_0x_1}\mathcal{T}_{y_0y_1}\mathcal{T}_{d_0}\mathcal{T}_{d_1}\\ &=\sum_{\substack{x_0x_1\in(\sD_\eta\cap W^{00}_\nu) \times W^{10}_\nu\\ y_0y_1\in W^{01}_\nu\times(\sD_\eta\cap W^{11}_\nu)\\ d_0\in\mathcal{D}_{\nu(d)}\cap W_{\mu^{(0)}},d_1\in \mathcal{D}_{\nu(d)}\cap W_{\mu^{(1)}} } } (-1)^{\widehat{dd_0d_1}-\widehat{d}}(-1)^{\ell(x_1)+\ell(y_0)}q^{2\ell(x_0)-2\ell(y_1)}v_{\bsi_\lambda dd_0d_1}\\ &=\sum_{\substack{d_0\in\mathcal{D}_{\nu(d)}\cap W_{\mu^{(0)}},\\ d_1\in \mathcal{D}_{\nu(d)}\cap W_{\mu^{(1)}} }}(-1)^{\widehat{dd_0d_1}-\widehat{d}}\sum_{\substack{x_0\in\sD_\eta\cap W_\nu^{00}\\ y_1\in\sD_\eta\cap W^{11}_\nu } }q^{2\ell(x_0)-2\ell(y_1)}\sum_{\substack{x_1\in W^{10}_\nu\\ y_0\in W^{01}_\nu}}(-1)^{\ell(x_1)+\ell(y_0)}v_{\bsi_\lambda dd_0d_1}. \end{aligned}$$ Now, the sum $\sum_{(x_1,y_0)\in W^{10}_\nu \times W^{01}_\nu }(-1)^{\ell(x_1)+\ell(y_0)}$ is the value at $-1$ of a Poincaré polynomial of a product of symmetric groups. Obviously, it is zero if and only if $ W^{10}_\nu \times W^{01}_\nu \neq 1$. However, the hypothesis $d\in \mathcal{D}_{\lambda\mu}\setminus \mathcal{D}^\circ_{\lambda\mu}$ implies $ W^{10}_\nu \times W^{01}_\nu \neq 1$. Hence, we have $N_{W,W_\eta}(e_{\mu,\lambda d})=0$. \[MC\] For $\lambda,\mu\in\Lambda(m|n,r),y\in \mathcal{D}_\lambda$, if $W^y_{\lambda^{(0)}}\cap W_{\mu^{(1)}}\neq 1$ or $ W^y_{\lambda^{(1)}}\cap W_{\mu^{(0)}}\neq1$, then $N_{W,1}(e_{\mu,\lambda y})=0$ in $\sS_R(m|n,r)$. If, in addition, a parabolic subgroup $W_\eta$ is a subgroup of $(W_{\la^{(0)}}^y\cap W_{\mu^{(0)}})\times(W_{\la^{(1)}}^y\cap W_{\mu^{(1)}}) $, then we have $N_{W,W_\eta}(e_{\mu,\lambda y})=0$. For the double coset $W_\lambda y W_\mu$ of $W$, there is a unique distinguish element $d\in \mathcal{D}_{\lambda\mu}$ such that $W_\lambda y W_\mu=W_\lambda d W_\mu$. Since $W_\lambda y W_\mu=W_\lambda\times\{d\}\times(\mathcal{D}_{\nu}\cap W_\mu)$ (as sets), there is $x\in\mathcal{D}_{\nu}\cap W_\mu$ such that $y=dx$ where $W_{\nu}=W^d_{\lambda}\cap W_{\mu}$. Without loss of generality, we assume $W^y_{\lambda^{(0)}}\cap W_{\mu^{(1)}}\neq 1$. Since $W^y_{\lambda^{(0)}}\cap W_{\mu^{(1)}}=W^{dx}_{\lambda^{(0)}}\cap W_{\mu^{(1)}}=(W^d_{\lambda^{(0)}}\cap W_{\mu^{(1)}})^x\neq 1$, $W^{d}_{\lambda^{(0)}}\cap W_{\mu^{(1)}}\neq 1$. Thus, $d\in\mathcal{D}_{\lambda\mu}\setminus\mathcal{D}^\circ_{\lambda\mu}$ in this case. In order to get the claim, it is enough as above to consider the action of $N_{W,1}(e_{\mu,\lambda y})$ over $v_\mu$. That is, $$\begin{aligned} &(v_\mu)N_{W,1}(e_{\mu,\lambda y})\\ &=\sum_{w\in W}(v_\mu\mathcal{T}_{w^{-1}}) e_{\mu,\lambda y}\mathcal{T}_w\\ &=\sum_{w\in W_\mu}(v_\mu\mathcal{T}_{w^{-1}}) e_{\mu,\lambda y}\mathcal{T}_w\\ &=\sum_{w_0w_1\in W_{\mu^{(0)}}\times W_{\mu^{(1)}}}(v_\mu \mathcal{T}_{(w_0w_1)^{-1}}) e_{\mu,\lambda y}\mathcal{T}_{w_0w_1}\\ &=\sum_{w_0w_1\in W_{\mu^{(0)}}\times W_{\mu^{(1)}}}q^{\ell(w_0)}(-q^{-1})^{\ell(w_1)}v_{\bsi_\lambda y}\mathcal{T}_{w_0}\mathcal{T}_{w_1}\\ &=\sum_{w_0w_1\in W_{\mu^{(0)}}\times W_{\mu^{(1)}}}(-1)^{\hat{y}-\hat{d}}q^{\ell(w_0)}(-q^{-1})^{\ell(w_1)}v_{\bsi_\lambda d}\mathcal{T}_x\mathcal{T}_{w_0}\mathcal{T}_{w_1}\\ &=(-1)^{\hat{y}-\hat{d}}v_{\bsi_\lambda d}\mathcal{T}_x\sum_{w_0w_1\in W_{\mu^{(0)}}\times W_{\mu^{(1)}}}q^{\ell(w_0)}(-q^{-1})^{\ell(w_1)}\mathcal{T}_{w_0}\mathcal{T}_{w_1}. \end{aligned}$$ Since $$\sum_{w_0w_1\in W_{\mu^{(0)}}\times W_{\mu^{(1)}}}q^{\ell(w_0)}(-q^{-1})^{\ell(w_1)}\mathcal{T}_{w_0}\mathcal{T}_{w_1}=x_{\mu^{(0)}}y_{\mu^{(1)}},$$ it follows that $$\mathcal{T}_xx_{\mu^{(0)}}y_{\mu^{(1)}}= q^{-\ell(x)}q^{2\ell(x_0)}(-1)^{\ell(x_1)}x_{\mu^{(0)}}y_{\mu^{(1)}}=q^{\ell(x_0)}(-q^{-1})^{\ell(x_1)}x_{\mu^{(0)}}y_{\mu^{(1)}},$$ where $x=x_0x_1$ with $x_0\in W_{\mu^{(0)}}$ and $x_1\in W_{\mu^{(1)}}$ (so $\ell(x)=\ell(x_0)+\ell(x_1)$). Hence, by Theorem \[SRTC\], $$(v_\mu)N_{W,1}(e_{\mu,\lambda y})=(-1)^{\hat{y}-\hat{d}}q^{\ell(x_0)}(-q^{-1})^{\ell(x_1)}(v_\mu)N_{W,1}(e_{\mu,\lambda d})=0,$$ proving the first assertion. For last assertion, the argument does not carry over. However, we first give a proof for the $\up$-Schur super algebra $\sS(m|n,r)$ (i.e., the $R=\sZ$ case). By Lemmas \[00-11\] and \[super2.2\], for any $s\in S$, $\mathcal{T}_se_{\mu,\la y}=e_{\mu,\la y}\mathcal{T}_s$ if and only if $s\in (W^y_{\lambda^{(0)}}\cap W_{\mu^{(0)}})\times (W^y_{\lambda^{(1)}}\cap W_{\mu^{(1)}})$. Now, if $W_\eta=W_{\eta^{(0)}}\times W_{\eta^{(1)}}$ with $W_{\eta^{(i)}}\leq W_{\la y\cap\mu}^{ii}$ ($i=0,1$), then, by Lemma \[super2.2\], $$\sT_ze_{\mu,\la y}=e_{\mu,\la y}\sT_z=\up^{\ell(z_0)}(-\up^{-1})^{\ell(z_1)} e_{\mu,\la y}$$ where $z=z_0z_1$ with $z_i\in W_{\eta^{(i)}}$ ($i=0,1$). Thus, $$\aligned 0=N_{W,1}(e_{\mu,\la y})&=N_{W,W_\eta}(N_{W_\eta,1}(e_{\mu,\la y}))\\ &=d_{W_{\eta^{(0)}}}(\bsu)d_{W_{\eta^{(1)}}}(\bsu^{-1})N_{W,W_\eta}(e_{\mu,\la y}). \endaligned$$ Hence, $N_{W,W_\eta}(e_{\mu,\la y})=0$ in $\sS(m|n,r)$. The general case follows from base change, noting $\sS_R(m|n,r)\cong\sS(m|n,r)\otimes_\sZ R$. Bases for $\sS_R(m|n,r)$ in relative norms ========================================== By Lemma \[iso f\], we will always identify $\sS_R(m|n,r)$ with $\End_{\sH_R}(V_R(m|n)^{\otimes r})$. For $\lambda,\mu\in\Lambda(m|n,r)$ and $d\in\mathcal{D}^\circ_{\lambda\mu}$, let $$N_{\mu\la}^d:=N_{W,W_{\la d\cap\mu}}(e_{\mu,\lambda d}).$$ Then, by Corollary \[SPC\](1), $N_{\mu\la}^d\in\sS_R(m|n,r )$. \[RNB\] The set $$\mathcal{B}=\{N_{\mu\la}^d\mid \lambda,\mu\in\Lambda(m|n,r),d\in\mathcal{D}^\circ_{\lambda\mu}\}$$ forms a basis of the $q$-Schur superalgebra $\sS_R(m|n,r )$. Consider the $\mathcal{H}_\mu$-$\mathcal{H}_\mu$-bimodule $$M=\Hom_R(Rv_\mu,Rv_{\bsi_\lambda d}\otimes_{\mathcal{H}_{\nu}}\mathcal{H}_\mu),$$ where $d\in\mathcal{D}_{\lambda\mu}$ and $\nu=\la d\cap \mu$. Now, $M$ contains an $\mathcal{H}_\mu$-$\mathcal{H}_\nu$-submodule $N=\Hom_R(Rv_\mu,Rv_{\bsi_\lambda d})$ and $M\cong N\otimes_{\mathcal{H}_{\nu}}\mathcal{H}_\mu$. Thus, by Nakayama relation in Lemma \[Nakayama\], $$N_{W_\mu,W_{\nu}}(\Hom_{\mathcal{H}_{\nu}}(Rv_\mu,Rv_{\bsi_\lambda d}))=\Hom_{\mathcal{H}_\mu}(Rv_\mu,Rv_{\bsi_\lambda d}\otimes_{\mathcal{H}_{\nu}}\mathcal{H}_\mu).$$ This together with an application of Mackey decomposition yields $$\begin{aligned} \Hom_{\mathcal{H}_\mu}(Rv_\mu,v_\lambda\mathcal{H}_R)&=\bigoplus_{d\in\mathcal{D}_{\lambda\mu}}\Hom_{\mathcal{H}_\mu}(Rv_\mu,Rv_\lambda\otimes \mathcal{T}_d\otimes_{\mathcal{H}_{\nu}}\mathcal{H}_\mu)\\ &=\bigoplus_{d\in\mathcal{D}_{\lambda\mu}}\Hom_{\mathcal{H}_\mu}(Rv_\mu,Rv_{\bsi_\lambda d}\otimes_{\mathcal{H}_{\nu}}\mathcal{H}_\mu)\\ &=\bigoplus_{d\in\mathcal{D}_{\lambda\mu}}N_{W_\mu,W_{\nu}}(\Hom_{\mathcal{H}_{\nu}}(Rv_\mu,Rv_{\bsi_\lambda d})). \end{aligned}$$ We claim that, for $d\in\mathcal{D}_{\lambda\mu}\setminus\mathcal{D}^\circ_{\lambda\mu}$, $\Hom_{\mathcal{H}_{\nu}}(Rv_\mu,Rv_{\bsi_\lambda d})=0$. Indeed, in this case, there exists $w\in(W^d_{\lambda^{(0)}}\cap W_{\mu^{(1)}})\times (W^d_{\lambda^{(1)}}\cap W_{\mu^{(0)}})$ such that $\ell(w)=1$. Assume $w\in W^d_{\lambda^{(0)}}\cap W_{\mu^{(1)}}$. For $f\in \Hom_{\mathcal{H}_{\nu}}(Rv_\mu,Rv_{\bsi_\lambda d})$, $(v_\mu)f=av_{\bsi_\lambda d}$ for some $a\in R$. Since $(v_\mu\mathcal{T}_s)f=(v_\mu)f\mathcal{T}_s$ for all $s\in W_\nu\cap S$, applying this to $s=w$ yields $aqv_{\bsi_\lambda d}=a(-q^{-1})v_{\bsi_\lambda d}$. Because $R$ is a domain and $q\neq (-q^{-1})$, $a=0$ and $f=0$. Thus, by the claim, $$\Hom_{\mathcal{H}_\mu}(Rv_\mu,v_\lambda\mathcal{H}_R) =\bigoplus_{d\in\mathcal{D}^\circ_{\lambda\mu}}N_{W_\mu,W_{\nu(d)}}(Rv_\mu,Rv_{\bsi_\lambda d}).$$ By applying $N_{W,W_\mu}(-)$ to both sides, Lemma \[RNT\](a) and Frobenius reciprocity (Lemma \[Frobenius\]) imply $$\begin{aligned} \Hom_{\mathcal{H}_R}(v_\mu\mathcal{H}_R,v_\lambda\mathcal{H}_R) &=N_{W,W_\mu}(\Hom_{\mathcal{H}_\mu}(Rv_\mu,v_\lambda\mathcal{H}_R))\\ &=\bigoplus_{d\in \mathcal{D}^\circ_{\lambda\mu}}N_{W,W_{\nu}}(\Hom_{\mathcal{H}_{\nu}}(Rv_\mu,Rv_{\bsi_\lambda d})). \end{aligned}$$ Therefore, $\{N_{W,W_{\nu}}(e_{\mu,\lambda d})\mid d\in\mathcal{D}^\circ_{\mu\lambda}\}$ forms a basis of $\Hom_{\mathcal{H}_R}(v_\mu\mathcal{H}_R,v_\lambda\mathcal{H}_R)$. Hence, $\mathcal{B}$ is a basis of $\sS_R(m|n,r )$. Now we describe a basis of $\End_{\mathcal{H}_\rho}(V_R(m|n)^{\otimes r})$ for $\rho=(\rho_1,\rho_2,\cdots,\rho_t)\models r$, which will be used in §7 and §10. For $\lambda,\mu\in\Lambda(m|n,r)$, let $d\in\mathcal{D}_{\la\rho},d'\in\mathcal{D}_{\mu\rho}$, and $$W_\alpha=W^d_\la\cap W_\rho\;\;\text{ and }\;\; W_\beta=W^{d'}_{\mu}\cap W_\rho.$$ Then we have $\alpha=(\alpha_{(0)},\alpha_{(1)},\cdots,\alpha_{(t)})$ and $\beta=(\beta_{(0)},\beta_{(1)},\cdots,\beta_{(t)})$ where $\alpha_{(i)},\beta_{(i)}\in\Lambda{(m|n,\rho_i)}$ for $i=1,2,\cdots,t$. Let $$\mathcal{D}^\circ_{\alpha\beta}\cap W_\rho=\{d=d_1d_2\cdots d_t\in \mathcal{D}_{\alpha\beta}\cap W_\rho\mid d_i\in W_{\rho_i}\cap\mathcal{D}^\circ_{\alpha_{(i)}\beta_{(i)}},0\leq i\leq t\}.$$ See Notation \[xxx\]. Since $V_R(m|n)^{\otimes r}=V_R(m|n)^{\otimes \rho_1}\otimes V_R(m|n)^{\otimes \rho_2}\otimes \cdots \otimes V_R(m|n)^{\otimes \rho_t}$, then $$\label{Schur algebra product} \sS_R(m|n,\rho):=\End_{\mathcal{H}_\rho}(V_R(m|n)^{\otimes r})\cong \sS_R(m|n,\rho_1 )\otimes \sS_R(m|n,\rho_2 )\otimes\cdots\otimes \sS_R(m|n,\rho_t ).$$ Set $$\mathcal{B}(\lambda,\mu,d,d')=\{N_{W_\rho,W^y_\alpha\cap W_\beta}(e_{\mu d',\lambda dy})\mid y\in \mathcal{D}_{\alpha\beta}^\circ\cap W_\rho\}.$$ The set $\mathcal{B}(\lambda,\mu,d,d')$ forms a basis for the $R$-module $$\Hom_{\mathcal{H}_\rho}(v_{\bsi_\mu d'}\otimes_{\mathcal{H}_\beta} \mathcal{H}_\rho, v_{\bsi_\lambda d}\otimes_{\mathcal{H}_\alpha}\mathcal{H}_\rho).$$ By Lemmas \[Frobenius\], \[Mackey\], and \[Nakayama\], $$\aligned \Hom_{\mathcal{H}_\rho}(v_{\bsi_\mu d'}\otimes_{\mathcal{H}_\beta} \mathcal{H}_\rho,& v_{\bsi_\lambda d}\otimes_{\mathcal{H}_\alpha}\mathcal{H}_\rho)=N_{W_\rho,W_\beta}(\Hom_{\sH_\beta}(v_{\bsi_\mu d'},v_{\bsi_\la d}\otimes_{\sH_\alpha}\sH_\rho|_{\sH_\beta})\\ &=N_{W_\rho,W_\beta}(\Hom_{\sH_\beta}(v_{\bsi_\mu d'},\bigoplus_{y\in \mathcal{D}_{\alpha\beta}\cap W_\rho}v_{\bsi_\la d}\otimes\sT_y\otimes_{\sH_{\alpha y\cap \beta}}{\sH_\beta})\\ &=\bigoplus_{y\in \mathcal{D}_{\alpha\beta}\cap W_\rho}N_{W_\rho,W_{\alpha y\cap\beta}}(\Hom_{\mathcal{H}_{\alpha y\cap\beta}}(Rv_{\bsi_\mu d'},Rv_{\bsi_\lambda dy})).\endaligned$$ Since, for $y\in (\mathcal{D}_{\alpha\beta}\cap W_\rho)\setminus(\mathcal{D}^\circ_{\alpha\beta}\cap W_\rho)$, $\Hom_{\mathcal{H}_{\alpha y\cap \beta}}(Rv_{\bsi_\mu d'},Rv_{\bsi_\lambda dy})=0$, it follows that $$\Hom_{\mathcal{H}_\rho}(v_{\bsi_\mu d'}\otimes_{\mathcal{H}_\beta} \mathcal{H}_\rho, v_{\bsi_\lambda d}\otimes_{\mathcal{H}_\alpha}\mathcal{H}_\rho)=\bigoplus_{y\in \mathcal{D}^\circ_{\alpha\beta}\cap W_\rho}N_{W_\rho,W_{\alpha y\cap\beta}}(\Hom_{\mathcal{H}_{\alpha y\cap\beta}}(Rv_{\bsi_\mu d'},Rv_{\bsi_\lambda dy})),$$ Hence, $\mathcal{B}(\lambda,\mu,d,d')$ forms a basis. Let $$\mathcal{B}(\rho)=\bigcup_{\substack{\lambda,\mu\in \Lambda{(m|n,r)}\\ d\in\mathcal{D}_{\la\rho},d'\in\mathcal{D}_{\mu\rho}}}\mathcal{B}(\lambda,\mu,d,d').$$ \[SRNB\] The set $\mathcal{B}(\rho)$ is a basis of $\sS_R(m|n,\rho)=\End_{\mathcal{H}_\rho}(V_R(m|n)^{\otimes r})$. By Lemma \[iso f\] and [[@J 2.22]]{}, we have $$\begin{aligned} \End_{\mathcal{H}_\rho}(V_R(m|n)^{\otimes r})&=\bigoplus_{\lambda,\mu\in\Lambda{(m|n,r)}}\Hom_{\mathcal{H}_\rho}(v_\mu\mathcal{H}_R,v_\lambda\mathcal{H}_R)\\ &=\bigoplus_{\substack{\lambda,\mu\in\Lambda{(m|n,r)}\\ d\in\mathcal{D}_{\la\rho},d'\in\mathcal{D}_{\mu\rho}}}\Hom_{\mathcal{H}_\rho}(v_{\bsi_\mu d'}\otimes_{\mathcal{H}_{\beta}}\mathcal{H}_\rho,v_{\bsi_\lambda d}\otimes_{\mathcal{H}_{\alpha}}\mathcal{H}_\rho) \end{aligned}$$ The theorem follows from the above lemma. However, by , we may use the bases for $\sS_R(m|n,\rho_i)$ described in Theorem \[RNB\] to describe a basis for $\sS_R(m|n,\rho)$. We now show that this basis coincides with the basis above. Let $$\label{vecLa} \vec{\La}(m|n,\rho)=\Lambda{(m|n,\rho_1)}\times\cdots\times \Lambda{(m|n,\rho_t)}.$$ Then, for $\vec{\lambda}\in \vec{\La}(m|n,\rho)$, there exist $\lambda_{(i)}\in \Lambda{(m|n,\rho_i)} $ such that $\vec{\lambda}=(\lambda_{(1)},\ldots,\lambda_{(t)})$. Let $W_{\vec{\lambda}}$ be the corresponding parabolic subgroup of $W_\rho$ and, for $\vec{\lambda},\vec{\mu}\in \vec{\La}(m|n,\rho)$, define the set $\mathcal{D}_{\vec{\lambda}\vec{\mu};\rho}$ of distinguished double coset representatives by $$\label{scrD} \mathcal{D}^\circ_{\vec{\lambda}\vec{\mu};\rho}= \{d\in\mathcal{D}_{\vec{\lambda}\vec{\mu}}\mid d=d_1\cdots d_t, d_i\in W_{\rho_i}\cap \mathcal{D}^\circ_{\lambda_{(i)}\mu_{(i)}},1\leq i\leq t\}\subseteq W_\rho.$$ Here, again, see Notation \[xxx\] for the notational convention. Putting $$N_{\vec\mu\vec\la}^d=\bigotimes_{i=1}^tN_{W_{\rho_i},W_{\la_{(i)}}^{d_i}\cap W_{\mu_{(i)}}}(e_{\mu_{(i)},\la_{(i)}d_i}),$$ then the set $\{N^d_{\vec\mu\vec{\lambda}}\mid \vec{\lambda},\vec{\mu}\in \vec{\La}(m|n,\rho), d\in \mathcal{D}^\circ_{\vec{\lambda}\vec\mu;\rho}\}$ forms a basis of $\sS_R(m|n,\rho)$. If we define the multi-index $\bsi_{\vec{\mu}}\in I(m|n,r)$ in an obvious way so that $$v_{\vec{\mu}}=v_{\bsi_{\vec{\mu}}}=v_{\mu_{(1)}}\otimes v_{\mu_{(2)}}\otimes\cdots\otimes v_{\mu_{(t)}}\in V_R(m|n)^{\otimes r}$$ and set $e_{\vec{\mu},\vec{\lambda}d}$ to be the $R$-linear map by sending $v_{\vec{\mu}}$ to $v_{\bsi_{\vec{\lambda}}d}$ and the other basis vectors of $V_R(m|n)^{\otimes r}$ to $0$, then we may simply write $N^d_{\vec\mu\vec{\lambda}}=N_{W_\rho,W_{\vec\la}^d\cap W_{\vec\mu}}(e_{\vec\mu,\vec\la d})$. Here we regard $\vec\nu$ as a composition of $r$ by concatenation so that $W_{\vec\nu}$ is well defined. The coincidence of this basis and the one in Theorem \[SRNB\] can be seen as follows. For $\vec{\lambda}=(\lambda_{(1)},\ldots,\lambda_{(t)})\in\vec{\La}(m|n,\rho)$, let $\lambda=\lambda_{(1)}+\cdots+\lambda_{(t)} \in \Lambda{(m|n,r)}$. Then there is a unique $x\in\mathcal{D}_{\lambda\rho}$ such that $W_{\vec{\lambda}}=W^x_\lambda\cap W_\rho$. Similarly, for $\vec\mu\in\vec{\La}(m|n,\rho)$, there is a unique $y\in\mathcal{D}_{\mu\rho}$ such that $W_{\vec{\mu}}=W^y_\mu\cap W_\rho$. Thus, for any $z\in\mathcal{D}_{\vec\la\vec\mu;\rho}^\circ$, putting $W_{xyz}=(W_\la^x\cap W_\rho)^z\cap(W_\mu^y\cap W_\rho)$, we have $$\label{same}N_{W_\rho,W_{\vec\la}^d\cap W_{\vec\mu}}(e_{\vec\mu,\vec\la z})= N^z_{\vec\mu\vec{\lambda}}=N_{W_\rho,W_{xyz}}(e_{\mu y,\la xz}).$$ We end this section with a comparison of the basis given in Theorem \[RNB\] and the one in Lemma \[DR5.8\]. \[RTSB\] If we identify $\fT_R(m|n,r)=\bigoplus_{\lambda\in\Lambda{(m|n,r)}}x_{\lambda^{(0)}}y_{\lambda^{(1)}}\mathcal{H}_R$ with $V_R(m|n)^{\otimes r}$ by the isomorphism $f$ given in Lemma \[iso f\], then, for any $\lambda,\mu\in \Lambda{(m|n,r)},d\in\mathcal{D}^\circ_{\lambda\mu}$, and $\nu=\la d\cap\mu$, $$\label{psi=N} N_{W,W_{\nu}}(e_{\mu,\lambda d}) =(q^{-1})^{\ell(d)}(-1)^{\hat{d}} \psi^d_{\mu\lambda}.$$ We need to prove that $f\psi_{\mu\la}^df^{-1}$ equals the right hand side. Since $f$ is an $\mathcal{H}_R$-module isomorphism, it is enough to consider the actions on $v_\mu$. We have on the one hand, $$\begin{aligned} (v_\mu)f\psi^d_{\mu\lambda}f^{-1}&=(x_{\mu^{(0)}}y_{\mu^{(1)}})\psi^d_{\mu\lambda}f^{-1}\\ &=(T_{W_\lambda d W_\mu})f^{-1}\\ &=(\sum_{w_0w_1\in (W_{\mu^{(0)}}\times W_{\mu^{(1)}})\cap \mathcal{D}_{\nu}}(-q^2)^{-\ell(w_1)}x_{\mu^{(0)}}y_{\mu^{(1)}}T_dT_{w_0}T_{w_1})f^{-1}\\ &=(q^{\ell(d)}\sum_{w_0w_1\in (W_{\mu^{(0)}}\times W_{\mu^{(1)}})\cap \mathcal{D}_{\nu}}q^{\ell(w_0)}(-q^{-1})^{\ell(w_1)}x_{\mu^{(0)}}y_{\mu^{(1)}}\mathcal{T}_{dw_0w_1})f^{-1}\\ &=q^{\ell(d)}\sum_{w_0w_1\in (W_{\mu^{(0)}}\times W_{\mu^{(1)}})\cap \mathcal{D}_{\nu}}(-1)^{\widehat{dw_0w_1}}q^{\ell(w_0)}(-q^{-1})^{\ell(w_1)}v_{\bsi_\lambda dw_0w_1}. \end{aligned}$$ On the other hand, $$\begin{aligned} (v_\mu)N_{W,W_{\nu}}(e_{\mu,\lambda d})&=(v_\mu)\sum_{w\in\mathcal{D}_{\nu}}\mathcal{T}_{w^{-1}}e_{\mu,\lambda d}\mathcal{T}_w\\ &=\sum_{w_0w_1\in( W_{\mu^{(0)}}\times W_{\mu^{(1)}})\cap \mathcal{D}_{\nu}}(v_\mu\mathcal{T}_{w_0^{-1}}\mathcal{T}_{w_1^{-1}})e_{\mu,\lambda d}\mathcal{T}_{w_0}\mathcal{T}_{w_1}\\ &=\sum_{w_0w_1\in (W_{\mu^{(0)}}\times W_{\mu^{(1)}})\cap \mathcal{D}_{\nu}}q^{\ell(w_0)}(-q^{-1})^{\ell(w_1)}v_{\bsi_\lambda d}\mathcal{T}_{w_0}\mathcal{T}_{w_1}\\ &=(-1)^{\widehat{d}}\sum_{w_0w_1\in (W_{\mu^{(0)}}\times W_{\mu^{(1)}})\cap \mathcal{D}_{\nu}}(-1)^{\widehat{dw_0w_1}}q^{\ell(w_0)}(-q^{-1})^{\ell(w_1)}v_{\bsi_\lambda dw_0w_1}. \end{aligned}$$ Therefore, follows. In particular, by Theorem \[RNB\], the set $$\{\psi^d_{\mu\lambda}\mid\lambda,\mu\in\Lambda{(m|n,r)},d\in\mathcal{D}^\circ_{\lambda\mu}\}$$ forms a basis for $\sS_R(m|n,r)$, proving Lemma \[DR5.8\]. A filtration of ideals of $\sS_R(m|n,r)$ ======================================== The purpose of this section is to construct a filtration of ideals of $\sS_R(m|n,r)$ labelled by $l$-parabolic subgroups. The existence of the filtration is based on a nice property of the structure constants associated with the defect groups of the relative norm basis elements in $\mathcal{B}$ (see Theorem \[TI\]). We need certain technical results Lemmas \[DD\]–\[ML\] for the proof of the theorem. \[RTRN\] Let $\lambda,\mu\in\Lambda{(m|n,r)}$ and $d\in\mathcal{D}^\circ_{\lambda\mu}$. Then $$(-1)^{\hat{d}}N_{W,W^d_\lambda\cap W_\mu}(e_{\mu,\lambda d})=(-1)^{\widehat{d^{-1}}}N_{W,W_\lambda\cap W^{d^{-1}}_\mu}(e_{\mu d^{-1},\lambda}).$$ Note by Definition \[dhat\] that $\widehat d=(\la,d)^\wedge$ while $\widehat{d^{-1}}=(\mu,d^{-1})^\wedge$. Let $W_\nu=W^d_\lambda\cap W_\mu$ and $W_{\nu'}=W_\lambda\cap W^{d^{-1}}_\mu$. By Lemma \[SRNP\], it is sufficient to check that $$(-1)^{\hat{d}}(v_\mu)N_{W,W^d_\lambda\cap W_\mu}(e_{\mu,\lambda d})=(-1)^{\widehat{d^{-1}}}(v_\mu)N_{W,W_\lambda\cap W^{d^{-1}}_\mu}(e_{\mu d^{-1},\lambda}).$$ Now, $$\begin{aligned} \text{LHS}&=%(-1)^{\hat{d}}\sum_{w\in\mathcal{D}_\nu}(v_\mu\mathcal{T}_{x^{-1}})e_{\mu,\lambda d}\mathcal{T}_x= (-1)^{\hat{d}}\sum_{w\in\mathcal{D}_\nu\cap W_\mu}(v_\mu\mathcal{T}_{x^{-1}})e_{\mu,\lambda d}\mathcal{T}_x\\ &=(-1)^{\hat{d}}\sum_{x_0x_1\in\mathcal{D}_\nu\cap (W_{\mu^{(0)}}\times W_{\mu^{(1)}})} (v_\mu\mathcal{T}_{x_0^{-1}}\mathcal{T}_{x_1^{-1}})e_{\mu,\lambda d}\mathcal{T}_{x_0}\mathcal{T}_{x_1}\\ &=(-1)^{\hat{d}}\sum_{x_0x_1\in\mathcal{D}_ \nu\cap (W_{\mu^{(0)}}\times W_{\mu^{(1)}})}q^{\ell(x_0)}(-q^{-1})^{\ell(x_1)}v_{\bsi_\lambda d}\mathcal{T}_{x_0}\mathcal{T}_{x_1}\\ &=(-1)^{\hat{d}}\sum_{x_0x_1\in\mathcal{D}_ \nu\cap (W_{\mu^{(0)}}\times W_{\mu^{(1)}})}(-1)^{\widehat{dx_0x_1}-\hat{d}}q^{\ell(x_0)}(-q^{-1})^{\ell(x_1)}v_{\bsi_\lambda dx_0x_1}\\ &=\sum_{x_0x_1\in\mathcal{D}_ \nu\cap (W_{\mu^{(0)}}\times W_{\mu^{(1)}})}(-1)^{\widehat{dx_0x_1}}q^{\ell(x_0)}(-q^{-1})^{\ell(x_1)}v_{\bsi_\lambda dx_0x_1}. \end{aligned}$$ By [[@DU 3.1]]{}, $\mathcal{D}_{\nu'}\cap d W_\mu=d(\mathcal{D}_\nu\cap W_\mu)$. Thus, $$\begin{aligned} \text{RHS}&=%(-1)^{\widehat{d^{-1}}}\sum_{y\in\mathcal{D}_{\nu'}\cap %dW_\mu}(v_\mu\mathcal{T}_{y^{-1}})e_{\mu d^{-1},\lambda}\mathcal{T}_y= (-1)^{\widehat{d^{-1}}}\sum_{y\in d(\mathcal{D}_{\nu}\cap W_\mu)}(v_\mu\mathcal{T}_{y^{-1}})e_{\mu d^{-1},\lambda}\mathcal{T}_y\\ &=(-1)^{\widehat{d^{-1}}}\sum_{dy_0y_1\in d(\mathcal{D}_\nu\cap(W_{\mu^{(0)}}\times W_{\mu^{(1)}}))}(v_\mu\mathcal{T}_{y_0^{-1}}\mathcal{T}_{y_1^{-1}}\mathcal{T}_{d^{-1}})e_{\mu d^{-1},\lambda}\mathcal{T}_{d}\mathcal{T}_{y_0}\mathcal{T}_{y_1}\\ &=(-1)^{\widehat{d^{-1}}}\sum_{dy_0y_1\in d(\mathcal{D}_\nu\cap(W_{\mu^{(0)}}\times W_{\mu^{(1)}}))}q^{\ell(y_0)}(-q^{-1})^{\ell(y_1)}(-1)^{\widehat{d^{-1}}}v_\lambda\mathcal{T}_{dy_0y_1}\\ &=(-1)^{\widehat{d^{-1}}}\sum_{y_0y_1\in \mathcal{D}_\nu\cap(W_{\mu^{(0)}}\times W_{\mu^{(1)}})}q^{\ell(y_0)}(-q^{-1})^{\ell(y_1)}(-1)^{\widehat{d^{-1}}}(-1)^{\widehat{dy_0y_1}}v_{\bsi_\lambda dy_0y_1}\\ &=\sum_{y_0y_1\in \mathcal{D}_\nu\cap(W_{\mu^{(0)}}\times W_{\mu^{(1)}})}q^{\ell(y_0)}(-q^{-1})^{\ell(y_1)}(-1)^{\widehat{dy_0y_1}}v_{\bsi_\lambda dy_0y_1}. \end{aligned}$$ So LHS=RHS, proving the lemma. [*For the next three lemmas, we fix the following notations:*]{} $$\label{DD0a} \aligned \rho,\lambda,\mu\in&\Lambda{(m|n,r)},d'\in\mathcal{D}_{\rho\lambda}, d\in\mathcal{D}_{\lambda\mu},\text{ and } y\in\mathcal{D}_{\rho\mu},\text{ and define }\nu,\nu',\tau\text{ by }\\ W_{\nu}&=W^d_\lambda\cap W_\mu,\quad W_{\nu'}=W_\rho\cap W^{d'^{-1}}_\lambda,\;\text{ and }\;W_\tau=W^y_\rho\cap W_\mu. \endaligned$$ Then, $\nu,\nu'\in\La(m'|n',r)$ for some $m'\geq m,n'\geq n$; see remarks right after . \[DD\] If $d\in\mathcal{D}^\circ_{\lambda\mu},d' \in\mathcal{D}^\circ_{\rho\lambda}$, and $y\in\mathcal{D}^\circ_{\rho\mu}$, then $$W_\rho y W_\mu\cap\mathcal{D}_{\nu'\nu}=\{hyk\mid yk\in\mathcal{D}^\circ_{\rho\nu},k\in\mathcal{D}^\circ_{\tau\nu}\cap W_\mu, h\in \mathcal{D}_{\nu'}\cap W_\rho\}$$ and it is a subset of $\mathcal{D}^\circ_{\nu'\nu}$. Since $y\in\mathcal{D}_{\rho\mu}$, by [[@DU 3.2]]{}, we have $$\label{DD0b} W_\rho y W_\mu\cap\mathcal{D}_{\nu'\nu}=\{hyk\mid yk\in\mathcal{D}_{\rho\nu},k\in\mathcal{D}_{\tau\nu}\cap W_\mu,h\in \mathcal{D}_{\nu'}\cap W_\rho\}.$$ Since $y\in\mathcal{D}^\circ_{\rho\mu},d\in\mathcal{D}^\circ_{\lambda\mu}$, by definition, $$\aligned W^y_{\rho^{(0)}}\cap W_{\mu^{(1)}}=1,\quad&W^y_{\rho^{(1)}}\cap W_{\mu^{(0)}}=1, \\W_{\nu^{(0)}}=W^d_{\lambda^{(0)}}\cap W_{\mu^{(0)}},\;&\text{ and }\; W_{\nu^{(1)}}=W^d_{\lambda^{(1)}}\cap W_{\mu^{(1)}}. \endaligned$$ Thus, noting that $k\in W_\mu$ implies $W_{\mu^{(1)}}^k=W_{\mu^{(1)}}$, $$W^{yk}_{\rho^{(0)}}\cap W_{\nu^{(1)}}=W^{yk}_{\rho^{(0)}}\cap(W^d_{\lambda^{(1)}}\cap W_{\mu^{(1)}})=(W^y_{\rho^{(0)}}\cap W_{\mu^{(1)}})^{k}\cap W^d_{\lambda^{(1)}}=1.$$ Similarly, one shows $W^{yk}_{\rho^{(1)}}\cap W_{\nu^{(0)}}=1$. Hence, $yk\in\mathcal{D}^\circ_{\rho\nu}$. To see $k\in\mathcal{D}^\circ_{\tau\nu}\cap W_\mu$, we have $$W^{k}_{\tau^{(0)}}\cap W_{\nu^{(1)}}= (W^y_{\rho^{(0)}}\cap W_{\mu^{(0)}})^{k}\cap W^d_{\lambda^{(1)}}\cap W_{\mu^{(1)}}=W^{yk}_{\rho^{(0)}}\cap W_{\mu^{(0)}}\cap W^d_{\lambda^{(1)}}\cap W_{\mu^{(1)}}=1$$ since $W^{yk}_{\rho^{(0)}}\cap W_{\mu^{(1)}}=1$. By a similar proof, we obtain $W^{k}_{\tau^{(1)}}\cap W_{\nu^{(0)}}=1$, proving $k\in\mathcal{D}^\circ_{\tau\nu}$. Finally, we prove $W_\rho y W_\mu\cap \mathcal{D}_{\nu'\nu}\subseteq \mathcal{D}_{\nu'\nu}^\circ$. For $hyk\in W_\rho y W_\mu\cap \mathcal{D}_{\nu'\nu}$, we have $$W^{hyk}_{\nu'^{(0)}}\cap W_{\nu^{(1)}}=(W_{\rho^{(0)}}\cap W^{d'^{-1}}_{\lambda^{(0)}})^{hyk}\cap W^d_{\lambda^{(1)}}\cap W_{\mu^{(1)}}=W_{\rho^{(0)}}^{hyk}\cap W^{d'^{-1}hyk}_{\lambda^{(0)}}\cap W^d_{\lambda^{(1)}}\cap W_{\mu^{(1)}}.$$ Since $h\in\mathcal{D}_{\nu'}\cap W_\rho$, $W_{\rho^{(0)}}^{hyk}=W_{\rho^{(0)}}^{yk}$. Since $k\in W_\mu$, $W_{\rho^{(0)}}^{hyk}\cap W_{\mu^{(1)}}=(W^y_{\rho^{(0)}}\cap W_{\mu^{(1)}})^{k}$. Now $W^y_{\rho^{(0)}}\cap W_{\mu^{(1)}}=1$, implies $W^{hyk}_{\nu'^{(0)}}\cap W_{\nu^{(1)}}=1$. Similarly, one proves $W^{hyk}_{\nu'^{(1)}}\cap W_{\nu^{(0)}}=1$. Hence, $hyk\in \mathcal{D}^\circ_{\nu'\nu}$. The following is the super version of a modified [@DU Lem. 3.3]. \[DD1\] Let $\rho,\mu\in\La(m|n,r)$ and let $W_\alpha$ be a parabolic subgroup of $W_\mu$. For $y\in \mathcal{D}_{\rho\mu}^\circ$ and $k\in\mathcal{D}_{\rho y\cap\mu}\cap W_\mu\cap\sD_\alpha^{-1}$, let $W_\theta=W^{yk}_\rho\cap W_\alpha$ and $W_{\theta'}=W^y_\rho\cap W^{k^{-1}}_\alpha=W_\theta^{k^{-1}}$. Then there exists $c\in R$ such that $$N_{W,W_\theta}(e_{\mu,\rho yk})=cN_{W,W_{\theta'}}(e_{\mu,\rho y})$$ Let $z=yk$ and $\tau=\rho y\cap \mu$. The hypothesis $y\in\sD_{\rho\mu}^\circ$ and $W_\alpha\leq W_\mu$ implies $yk\in\sD_{\rho\alpha}^\circ$. Hence, $W_\theta$ is parabolic. Since $k\in\mathcal{D}_{\tau\alpha}\cap W_\mu$, it follow that $W_{\theta'}=W_\alpha^{k^{-1}}\cap W_\tau$ is parabolic. Moreover, both $W_\theta$ and $W_{\theta'}$ are parabolic subgroups of (parabolic) $W_\rho^y\cap W_\mu$ and of (possibly non-parabolic) $W_\rho^z\cap W_\mu$. By the transitivity of relative norm Lemma \[RNT\](a), it is enough to show that, for some $c\in R$, $$N_{W_\mu,W_\theta}(e_{\mu,\rho z})=cN_{W_\mu,W_{\theta'}}(e_{\mu,\rho y}).$$ Now consider the $\mathcal{H}_\mu$-$\mathcal{H}_\mu$-bimodule $$M=\Hom_R(Rv_\mu,Rv_{\bsi_\rho y}\otimes_{\mathcal{H}_{\theta'}}\mathcal{H}_\mu).$$ It is clear that there exists an $\mathcal{H}_\mu$-$\mathcal{H}_{\theta'}$-bimodule $N=\Hom_R(Rv_\mu,Rv_{\bsi_\rho y})$ and an $\mathcal{H}_\mu$-$\mathcal{H}_{\theta}$-bimodule $N'=\Hom_R(Rv_\mu,Rv_{\bsi_\rho z})$ such that $M\cong N\otimes_{\mathcal{H}_{\theta'}}\mathcal{H}_\mu$ and $M\cong N'\otimes_{\mathcal{H}_\theta}\mathcal{H}_\mu$ as $\mathcal{H}_\mu$-$\mathcal{H}_\mu$-bimodules. Applying Lemma \[Nakayama\] yields $$N_{W_\mu,W_{\theta'}}(\Hom_{\mathcal{H}_{\theta'}}(Rv_\mu,Rv_{\bsi_\rho y}))=Z_M(\sH_\mu)=N_{W_\mu,W_\theta}(\Hom_{\mathcal{H}_\theta}(Rv_\mu,Rv_{\bsi_\rho z})).$$ Hence, there exists $c\in R$ such that $N_{W,W_\theta}(e_{\mu,\rho z})=N_{W,W_{\theta'}}(ce_{\mu,\rho y})$. Note that, in the applications below of this lemma, we will take $W_\alpha=W_\nu$ as defined in or $W_\alpha=1$. \[ML\]Maintain the notations in and Lemma \[DD\]. For $hyk\in \mathcal{D}^\circ_{\nu'\nu}$, let $$\aligned W_\xi&=W^{yk}_\rho\cap W_\nu\leq W_\tau^k,\quad W_\eta=W^{hyk}_{\nu'}\cap W_\nu\leq W_\xi,\\ f(\bsu)&=\frac{d_{W_{\tau^{(0)}}}(\bsu)d_{W_{\tau^{(1)}}}(\bsu^{-1})}{d_{W_{\xi^{(0)}}}(\bsu)d_{W_{\xi^{(1)}}}(\bsu^{-1})},\text{ and }g(\bsu)=\frac{d_{W_{\xi^{(0)}}}(\bsu)d_{W_{\xi^{(1)}}}(\bsu^{-1})}{d_{W_{\eta^{(0)}}}(\bsu)d_{W_{\eta^{(1)}}}(\bsu^{-1})}.\endaligned$$ Then $$N_{W,W^{hyk}_{\nu'}\cap W_\nu}(e_{\mu,\rho yk})=cf(q^2)g(q^2)N_{W,W^y_\rho\cap W_\mu}(e_{\mu,\rho y})$$ for some $c\in R$. Moreover, if $q\in R$ is an $l$th primitive root of 1 with $l$ odd and $P_{\tau}\nleq_WP_\nu$ or $P_{\tau}\nleq_WP_{\nu'}$, then $f(q^2)g(q^2)=0$. By the transitivity of relative norms (Lemma \[RNT\](a)), we have $$N_{W,W^{hyk}_{\nu'}\cap W_\nu}(e_{\mu,\rho yk})=N_{W,W^{yk}_\rho\cap W_\nu}(N_{W^{yk}_\rho\cap W_\nu,W^{hyk}_{\nu'}\cap W_\nu}(e_{\mu,\rho yk})).$$ By Lemma \[DD\], $W_\xi=W_{\rho yk\cap\nu}^{00}\times W_{\rho yk\cap\nu}^{11}\leq W_{\rho yk\cap\mu}^{00}\times W_{\rho yk\cap\mu}^{11}$. Thus, we have $e_{\mu,\rho yk}\in \Hom_{\mathcal{H}_\xi}(V_R(m|n)^{\otimes r})$ by Lemma \[super2.2\]. Since $$\begin{aligned} (v_\mu)N_{W^{yk}_\rho\cap W_\nu,W^{hyk}_{\nu'}\cap W_\nu}(e_{\mu,\rho yk})&=(v_\mu)N_{W_\xi,W_\eta}(e_{\mu,\rho yk})\\ &=\sum_{w\in\mathcal{D}_\eta\cap W_\xi}(v_\mu\mathcal{T}_{w^{-1}})e_{\mu,\rho yk} \mathcal{T}_w\\ &=\sum_{w_0w_1\in\mathcal{D}_\eta\cap(W_{\xi^{(0)}}\times W_{\xi^{(1)}})}(v_\mu\mathcal{T}_{w_0^{-1}}\mathcal{T}_{w_1^{-1}})e_{\mu,\rho yk} \mathcal{T}_{w_0}\mathcal{T}_{w_1}\\ &=\sum_{w_0w_1\in\mathcal{D}_\eta\cap(W_{\xi^{(0)}}\times W_{\xi^{(1)}})}q^{\ell(w_0)}(-q^{-1})^{\ell(w_1)}v_{\bsi_\rho yk}\mathcal{T}_{w_0}\mathcal{T}_{w_1}\\ &=\sum_{w_0w_1\in\mathcal{D}_\eta\cap(W_{\xi^{(0)}}\times W_{\xi^{(1)}})}q^{2\ell(w_0)}(-q^{-1})^{2\ell(w_1)}v_{\bsi_\rho yk} \end{aligned}$$ According to the definition of Poincaré polynomial, we have $$(v_\mu)N_{W_\xi,W_\eta}(e_{\mu,\rho yk}) =\frac{d_{W_{\xi^{(0)}}}(\bsu)d_{W_{\xi^{(1)}}}(\bsu^{-1})}{d_{W_{\eta^{(0)}}}(\bsu)d_{W_{\eta^{(1)}}}(\bsu^{-1})}\bigg|_{\bsu=q^2}v_{\bsi_\rho yk} =g(q^2)v_{\bsi_\rho yk}.$$ So $$N_{W,W^{hyk}_{\nu'}\cap W_\nu}(e_{\mu,\rho yk})=g(q^2)N_{W,W^{yk}_\rho\cap W_\nu}(e_{\mu,\rho yk}).$$ By Lemma \[DD1\], there exists $c\in R$ such that $$N_{W,W^{yk}_\rho\cap W_\nu}(e_{\mu,\rho yk})=cN_{W,W^y_\rho\cap W^{k^{-1}}_\nu}(e_{\mu,\rho y}).$$ Thus, $$\label{first half} N_{W,W^{hyk}_{\nu'}\cap W_\nu}(e_{\mu,\rho yk})=cg(q^2)N_{W,W^y_\rho\cap W^{k^{-1}}_\nu}(e_{\mu,\rho y}).$$ By Corollary \[SPC\], $e_{\mu,\rho y}\in \Hom_{\mathcal{H}_\tau}(V_R(m|n)^{\otimes r})$. Since $W^y_\rho\cap W^{k^{-1}}_\nu\leq W_\tau, $ we have by the transitivity of relative norms $$N_{W,W^y_\rho\cap W^{k^{-1}}_\nu}(e_{\mu,\rho y})=N_{W,W_\tau}(N_{W_\tau,W^y_\rho\cap W^{k^{-1}}_\nu}(e_{\mu,\rho y})).$$ Since $(W^{yk}_\rho\cap W_\nu)^{k^{-1}}=W^y_\rho\cap W_\mu\cap W^{k^{-1}}_\nu=W^{k^{-1}}_\xi$ and $k\in\mathcal{D}^\circ_{\tau\nu}\cap W_\mu$, $W^{k^{-1}}_\xi$ is a parabolic subgroup of $W$. By Lemma \[DU1.1\](a), we have $d_{W^y_\rho\cap W^{k^{-1}}_\nu}=d_{W_\xi}$. Hence, $$N_{W_\tau,W^y_\rho\cap W^{k^{-1}}_\nu}(e_{\mu,\rho y})=\frac{d_{W_{\tau^{(0)}}}(\bsu)d_{W_{\tau^{(1)}}}(\bsu^{-1})}{d_{W_{\xi^{(0)}}}(\bsu)d_{W_{\xi^{(1)}}}(\bsu^{-1})}\bigg|_{\bsu=q^2}e_{\mu,\rho y}=f(q^2)e_{\mu,\rho y}.$$ Applying $N_{W,W_\tau}(\ )$ to both sides and combining it with give $$N_{W,W^{hyk}_{\nu'}\cap W_\nu}(e_{\mu,\rho yk})=cf(q^2)g(q^2)N_{W,W^y_\rho\cap W_\mu}(e_{\mu,\rho y}),$$ proving the first assertion. To prove the last assertion, we first assume that $P_\tau\nleq_W P_\nu$, equivalently, $d_{P_\nu}\mid d_{P_\tau}$ and $d_{P_\nu}\neq d_{P_\tau}$. Since $W_\xi\leq W_\nu$, it follows that $d_{P_{\xi^{(0)}}}\mid d_{P_{\nu^{(0)}}}$ and $d_{P_{\xi^{(1)}}}\mid d_{P_{\nu^{(1)}}}$. Thus, $d_{P_\nu}\neq d_{P_\tau}$ implies $d_{P_{\xi^{(0)}}}\mid d_{P_{\tau^{(0)}}}$ and $d_{P_{\xi^{(1)}}}\mid d_{P_{\tau^{(1)}}}$, one of which is not equal. This implies $f(q^2)=0$, and hence, $f(q^2)g(q^2)=0$. From the argument, we see that, if $P_\xi<_WP_\tau$, then $f(q^2)=0$. We now assume that $P_\tau\leq_W P_\nu$ but $P_\tau\nleq_W P_{\nu'}$. As seen above, we may further assume $P_\xi=P_\tau$. Then, $d_{P_{\nu'}}|d_{P_\tau}$ and $d_{P_{\nu'}}\neq d_{P_\tau}$. Since $W_\eta\leq_WW_{\nu'}$ and $W_\eta\leq W_\xi$, $P_\xi=P_\tau$ implies that one of the relations $d_{P_{\eta^{(0)}}}|d_{P_{\xi^{(0)}}}$ and $d_{P_{\eta^{(1)}}}|d_{P_{\xi^{(1)}}}$ is not an equality. Hence, $g(q^2)=0$, and hence, $f(q^2)g(q^2)=0$, proving the last assertion. Let $N^d_{\mu\lambda}=N_{W,W_{\la d\cap\mu}}(e_{\mu,\lambda d})$ be a standard basis element as described in Theorem \[RNB\]. We define the [*defect group*]{} of $N^d_{\mu\lambda}$ to be the maximal $l$-parabolic subgroup $P_{\la d\cap\mu}$ of $W_{\la d\cap\mu}=W_\la^d\cap W_\mu$. We now have the following result about the coefficients of the product of two standard basis elements. Recall that $R$ is a domain and $q\in R$ is an $l$-th primitive root of 1 with $l$ odd. \[TI\] For basis elements $N^d_{\mu\lambda},N^{d'}_{\lambda\rho}\in\mathcal{B}$ of $\sS_R(m|n,r)$, assume that $$N^d_{\mu\lambda}N^{d'}_{\lambda\rho}=\sum_{y\in\mathcal{D}^\circ_{\rho\mu}}a_y N^y_{\mu\rho}$$ where $a_y\in R$. If $a_y\neq 0$, then $$P_{\rho y\cap\mu}\leq_W P_{\lambda d\cap\mu}\mbox{ and }P_{\rho y\cap\mu}\leq_W P_{\rho d'\cap\la}.$$ We first compute in the $\up$-Schur superalgebra $\sS(m|n,r)$ over $\sZ=\mathbb{Z}[\up,\up^{-1}]$. By \[RTRN\] and Lemma \[Frobenius\] and noting $\nu=\la d\cap\mu$ and $\nu'=\la d^{\prime -1}\cap\rho$, we have $$\begin{aligned} N^d_{\mu\lambda}N^{d'}_{\lambda\rho}&=(-1)^{\widehat{d'}}(-1)^{\widehat{d'^{-1}}}N_{W,W_\nu}(e_{\mu,\lambda d})N_{W,W_{\nu'}}(e_{\lambda d'^{-1},\rho})\\ &=(-1)^{\widehat{d'}+\widehat{d'^{-1}}}N_{W,W_\nu}(e_{\mu,\lambda d}N_{W,W_{\nu'}}(e_{\lambda d'^{-1},\rho}))\\ &=(-1)^{\widehat{d'}+\widehat{d'^{-1}}}N_{W,W_\nu}(e_{\mu,\lambda d}\sum_{x\in\mathcal{D}_{\nu'\nu}}N_{W_\nu,W^x_{\nu'}\cap W_\nu}(\mathcal{T}_{x^{-1}}e_{\lambda d'^{-1},\rho}\mathcal{T}_x))\\ &=(-1)^{\widehat{d'}+\widehat{d'^{-1}}}\sum_{x\in\mathcal{D}_{\nu'\nu}}N_{W,W^x_{\nu'}\cap W_\nu}(e_{\mu,\lambda d}\mathcal{T}_{x^{-1}}e_{\lambda d'^{-1},\rho}\mathcal{T}_x) \end{aligned}$$ Now, by [@DU 3.2] (or ), $$\mathcal{D}_{\nu'\nu}=\{hyk\mid y\in\mathcal{D}_{\rho\la},yk\in\mathcal{D}_{\rho\nu},k\in\mathcal{D}_{\tau\nu}\cap W_\mu,h\in \mathcal{D}_{\nu'}\cap W_\rho\}.$$ For $x\in \mathcal{D}_{\nu'\nu}$ there exist $y\in\mathcal{D}_{\rho\mu}$ and $h,k$ such that $x=hyk$ and $yk\in\mathcal{D}_{\rho\nu},h\in\mathcal{D}_{\nu'}\cap W_\rho$. By a direct computation, we have for some $b_x\in \sZ$. $$e_{\mu,\lambda d}\mathcal{T}_{x^{-1}}e_{\lambda d'^{-1},\rho}\mathcal{T}_x=b_x e_{\mu,\rho yk}.$$ We claim that if $b_x\neq0$ $$\label{gu1} \mathcal {T}_ze_{\mu,\rho yk}=e_{\mu,\rho yk}\mathcal{T}_z\text{ for all }z\in W^x_{\nu'}\cap W_\nu.$$ Indeed, $$\aligned b_x\cdot(\text{LHS})&=\mathcal{T}_z (e_{\mu,\lambda d}\mathcal{T}_{x^{-1}}e_{\lambda d'^{-1},\rho}\mathcal{T}_x)\\ &=e_{\mu,\lambda d}\mathcal{T}_z\mathcal{T}_{x^{-1}}e_{\lambda d'^{-1},\rho}\mathcal{T}_x\quad\text{(by Lemma \ref{SPC})}\\ &=e_{\mu,\lambda d}\mathcal{T}_{x^{-1}} \mathcal{T}_{z'} e_{\lambda d'^{-1},\rho}\mathcal{T}_x\quad(z'\in W_{\nu'}\text{ with }zx^{-1}=x^{-1}z')\\ &=e_{\mu,\lambda d}\mathcal{T}_{x^{-1}} e_{\lambda d'^{-1},\rho} \mathcal{T}_{z'} \mathcal{T}_x\\ &= (e_{\mu,\lambda d}\mathcal{T}_{x^{-1}}e_{\lambda d'^{-1},\rho}\mathcal{T}_x)\mathcal{T}_z \\ &=b_x\cdot(\text{RHS}) \endaligned$$ Since $W^x_{\nu'}\cap W_\nu\leq W_\rho^{yk}\cap W_\nu\leq W_\rho^{yk}\cap W_\mu$ and $W_{\nu^{(i)}}=W_{\la d\cap\mu}^{ii}\leq W_{\mu^{(i)}}$ for $i=0,1$, it follows that $W^x_{\nu'}\cap W_\nu=\prod_{i,j=0,1}W_{\nu' x\cap\nu}^{ij}$ is a product of parabolic subgroups and each $W_{\nu'x\cap\nu}^{ij}\leq W_{\rho^{(i)}}^{yk}\cap W_{\mu^{(j)}}$. Thus, if $b_x\neq0$, we conclude that $W^x_{\nu'}\cap W_\nu=W_{\nu'x\cap\nu}^{00}\times W_{\nu'x\cap\nu}^{11}\leq(W_{\rho^{(0)}}^{yk}\cap W_{\mu^{(0)}})\times( W_{\rho^{(1)}}^{yk}\cap W_{\mu^{(1)}})$, since any element in $ W_{\nu'x\cap\nu}^{01}\leq W_{\rho^{(0)}}^{yk}\cap W_{\mu^{(1)}} $ or in $W_{\nu'x\cap\nu}^{10}\leq W_{\rho^{(1)}}^{yk}\cap W_{\mu^{(0)}} $ does not satisfy ; see Lemma \[super2.2\]. Now, if $y\in\mathcal{D}_{\rho\mu}\setminus \mathcal{D}^\circ_{\rho\mu}$, then $ W_{\rho^{(0)}}^{yk}\cap W_{\mu^{(1)}}=(W_{\rho^{(0)}}^{y}\cap W_{\mu^{(1)}} )^k\neq1$ or $ W_{\rho^{(1)}}^{yk}\cap W_{\mu^{(0)}}=(W_{\rho^{(1)}}^{y}\cap W_{\mu^{(0)}} )^k\neq1$. Thus, by Corollary \[MC\], $N_{W,W^x_{\nu'}\cap W_{\nu}}(e_{\mu,\rho yk})=0$. Thus, for $y\in \mathcal{D}^\circ_{\rho\mu}$, if we write by Lemma \[DD\] $$W_\rho y W_\mu\cap\mathcal{D}_{\nu'\nu}=\{h_iyk_j\mid 1\leq i\leq m_y, 1\leq j\leq n_y\},$$ then $$\begin{aligned} N^d_{\mu\lambda}N^{d'}_{\lambda\rho} &=(-1)^{\widehat{d'}+\widehat{d'^{-1}}}\sum_{hyk\in\mathcal{D}^\circ_{\nu'\nu}}N_{W,W^x_{\nu'}\cap W_\nu}(b_{hyk}e_{\mu,\rho yk})\\ &=(-1)^{\widehat{d'}+\widehat{d'^{-1}}}\sum_{y\in\mathcal{D}^\circ_{\rho\mu}}\sum_{i,j}N_{W,W^{h_iyk_j}_{\nu'}\cap W_\nu}(b_{h_iyk_j}e_{\mu,\rho yk_j}). \end{aligned}$$ By Lemma \[ML\], there are Laurent polynomials $f_{y,j}(\bsu)$ associated with $yk_j$ and $g_{i,y,j}(\bsu)$ associated with $h_iyk_j$ such that in $\sS(m|n,r)$, $$N^d_{\mu\lambda}N^{d'}_{\lambda\rho}=(-1)^{\widehat{d'}+\widehat{d'^{-1}}}\sum_{y\in\mathcal{D}^\circ_{\rho\mu}}(\sum_{i,j}b'_{h_iyk_j}f_{y,j}(\bsu)g_{i,y,j}(\bsu))N_{W,W^y_\rho\cap W_\mu}(e_{\mu,\rho y}) .$$ We now look at the product in $\sS_R(m|n,r)=\sS(m|n,r)\otimes_\sZ R$ by specializing $\up$ to $q\in R$ and obtain $N^d_{\mu\lambda}N^{d'}_{\lambda\rho}=\sum_{y\in\mathcal{D}^\circ_{\rho\mu}}a_yN_{\mu\rho}^y$ in $\sS_R(m|n,r)$, where $$a_y=(-1)^{\widehat{d'}+\widehat{d'^{-1}}}\sum_{i=1}^{m_y}\sum_{j=1}^{n_y}b'_{i,y,j}f_{y,j}(q^2)g_{i,y,j}(q^2).$$ By Lemma \[ML\] again, if $P_{\rho y\cap\mu}\nleq_WP_{\lambda d\cap\mu}$ or $P_{\rho y\cap\mu}\nleq_W P_{\rho d'\cap\la}$, then all $f_{y,j}(q^2)g_{i,y,j}(q^2)=0$. Hence, $a_y=0$. Let $P$ be an $l$-parabolic subgroup of $W$. We define $$\label{ideal I} I_R(P,r)=\text{span}\{N^d_{\mu\lambda}\mid \la,\mu\in\La(m|n,r),d\in\sD_{\la\mu}^\circ,P_{\lambda d\cap\mu}\leq_W P\}.$$ We also write $I_R(P,r)=I_R(P,r)_{m|n}$ if $m|n$ needs to be mentioned. By the above theorem, we have \[MC2\] The space $I_R(P,r)$ is an ideal of $\sS_R(m|n,r)$. Let $r=sl+t$, where $0\leq t<l$. For each $k$, $0\leq k\leq s$, let $P_k$ be the $l$-parabolic subgroup of $W$ associated with the composition $(1^{t+(s-k)l},l^k)$. Then we have a chain of ideals in $\sS_R(m|n,r)$ $$\label{filtration} 0\subseteq I_R(P_0,r)\subseteq I_R(P_1,r)\subseteq\cdots\subseteq I_R(P_s,r)=\sS_R(m|n,r).$$ In Definition \[defect group\], we will use the sequence to define the defect group of a primitive idempotent and discuss the effect of Brauer homomorphisms on the defect groups. Alternative characterisation of the ideals $I_F(P, r)$ ====================================================== [*For the rest of the paper, we assume that $R=F$ is a field of characteristic 0 and $q\in F$ is a primitive $l$th root of 1 with $l$ odd.* ]{} \[P=1\] We have $$I_F(\{1\},r)=N_{W,1}(\End_F(V_F(m|n)^{\otimes r})).$$ Clearly, $$\{e_{\mu d,\lambda d'}\mid \lambda,\mu\in\Lambda{(m|n,r)},d\in\mathcal{D}_\mu,d'\in\mathcal{D}_\lambda\}$$ is a basis of $\End_F(V_F(m|n)^{\otimes r})$. By Lemmas \[RNT\], \[Mackey\], and \[SRNP\], we have $$\begin{aligned} N_{W,1}(e_{\mu, \lambda d'})&=N_{W,W_\mu}(e_{\mu\mu})N_{W,1}(e_{\mu d,\lambda d'})\\ &=N_{W,W_\mu}(e_{\mu\mu}N_{W,1}(e_{\mu d,\lambda d'}))\\ &=N_{W,W_\mu}(e_{\mu\mu}\sum_{x\in\mathcal{D}^{-1}_\mu}N_{W_\mu,1}(\mathcal{T}_{x^{-1}}e_{\mu d,\lambda d'}\mathcal{T}_x))\\ &=N_{W,W_\mu}(\sum_{x\in\mathcal{D}^{-1}_\mu}N_{W_\mu,1}(e_{\mu\mu}\mathcal{T}_{x^{-1}}e_{\mu d,\lambda d'}\mathcal{T}_x))\\ &=\sum_{x\in\mathcal{D}^{-1}_\mu\cap d^{-1}W_\mu}N_{W,1}(e_{\mu\mu}\mathcal{T}_{x^{-1}}e_{\mu d,\lambda d'}\mathcal{T}_x)\\ &=N_{W,1}(e_{\mu\mu}\mathcal{T}_d e_{\mu d,\lambda d'}\mathcal{T}_{d^{-1}}). \end{aligned}$$ Write $e_{\mu\mu}\mathcal{T}_d e_{\mu d,\lambda d'}\mathcal{T}_{d^{-1}}=\sum_{y\in \mathcal{D}_\la}a_y e_{\mu,\lambda y} $ where $a_y\in F$. Then $$N_{W,1}(e_{\mu d,\lambda d'})=\sum_{y\in \mathcal{D}_\lambda}a_yN_{W,1}(e_{\mu,\lambda y}).$$ For $y\in \mathcal{D}_\lambda$, there exists $x\in\mathcal{D}_{\lambda\mu},k \in \sD_{\la x\cap\mu}\cap W_\mu$ such that $y=xk$. By Corollary \[MC\], if $x\in\mathcal{D}_{\lambda\mu}\setminus\mathcal{D}^\circ_{\lambda\mu}$, then $N_{W,1}(e_{\mu,\lambda y})=0$. If $x\in\mathcal{D}^\circ_{\lambda\mu}$, then we apply Lemma \[DD1\] with $W_\alpha=1$ to give, for some $c\in F$, $$N_{W,1}(e_{\mu,\lambda y})=cN_{W,1}(e_{\mu,\la x})=cd_{W_{\la x\cap\mu}}(q^2)N_{W,W_{\la x\cap \mu}}(e_{\mu,\la x})\in I_F(\{1\},r),$$ since $d_{W_{\la x\cap\mu}}(q^2)=0$ if $P_{\la x\cap\mu}\neq1$. The result is proved. We now generalise this result to an arbitrary $l$-parabolic subgroup. [*For the rest of the section, fix a non-negative integer $k$ and let* ]{} $$\theta=(\underbrace{l,\cdots,l}_{k},1,\cdots,1).$$ Then $W_\theta=_WP_k$ is an $l$-parabolic subgroup of $W$. For $\lambda\in\Lambda{(m|n,r)},d\in\mathcal{D}_{\mu\theta}$, if $W_\theta\leq W_\la^d$ then, by Lemma \[00-11\], $W_\theta=(W_{\la^{(0)}}^d\cap W_\theta)\times (W_{\la^{(1)}}^d\cap W_\theta)$, and both $W_{\la^{(0)}}^d\cap W_\theta$ and $W_{\la^{(1)}}^d\cap W_\theta$ are $l$-parabolic. \[key\] Let $\theta\models r$ be given as above. If $\lambda\in\Lambda{(m|n,r)}, d\in\mathcal{D}_{\lambda\theta}$ and $W_\theta\leq W^d_\lambda$, then there exist $\theta'\in\Lambda{(m'|n',r)}$ for some $m'\geq m$ and $n'\geq n$ and $w\in N_W(W_\theta)\cap \mathcal{D}_{\theta\theta}$ such that $W_{\theta'}=W_\theta$, $dw^{-1}\in\sD_{\la\theta}^\circ$, and $$e_{\lambda d, \theta' w}\in \End_{\mathcal{H}_\theta}(V_F(m'|n')^{\otimes r}).$$ Moreover, $N_{W,W_\theta}(e_{\lambda d,\theta'w})\in I_F(W_\theta,r)_{m'|n'}.$ Since $d\in \mathcal{D}_{\lambda\theta}$ and $W_\theta\leq W^d_\lambda$, by [@DU2 1.3], we have $$\bsi_\lambda d=(\underbrace{i_1,\cdots,i_1}_l,\underbrace{i_2,\cdots,i_2}_l,\cdots,\underbrace{i_{k},\cdots,i_{k}}_l,j_1,\cdots,j_{t}).$$ The sequence $( \widehat{i_1},\widehat{i_2},\ldots, \widehat{i_k})$ is called [*the parity of the $k$ blocks of length $l$ in $\bsi_\la d$*]{}. Let $a=\#\{j\mid \widehat{i_j}=0, 1\leq j\leq k\}$ and $b= \#\{j\mid \widehat{i_j}=1, 1\leq j\leq k\}$. We will call $a$ (resp., $b$) the [*number of even*]{} (resp., [*odd*]{}) [*blocks of length $l$*]{}. Then $a+b=k$ and $$\bsi_{\la^{(0)}}d=\bsi_\la d|_{[1,m]}\text{ and } \bsi_{\la^{(1)}}d=\bsi_\la d|_{[m+1, m+n]}.$$ Consider 0-1 sequence $( \widehat{i_1},\widehat{i_2},\ldots, \widehat{i_k})$. There is a shortest $x\in\fS_k$ such that $$( \widehat{i_1},\widehat{i_2},\ldots, \widehat{i_k})x=(0^a,1^b).$$ Then $g=\ell(x)$ is the number of inversions in the sequence (see proof of [@DU3 Lem. 1.3]). Every such inversion $\widehat{i_c}>\widehat{i_d}$ with $c<d$ determines $l^2$ inversions in the sequence $\underbrace{i_c,\cdots,i_c}_l,\underbrace{i_d,\cdots,i_d}_l$. Let $\theta'^{(0)}=(l^{a})$ and $\theta'^{(1)}=(l^b,1^{t})$. Then, by adding some 0’s at the end of $\theta'^{(i)}$ if necessary, we may assume $\theta'=(\theta'^{(0)}|\theta'^{(1)})\in\La(m'|n',r)$ for some $m'\geq m,n'\geq n$. Clearly, $W_\theta=W_{\theta'}$. Define $w\in\sD_{\theta\theta}$ such that $$\bsi_{\theta'}w=(\underbrace{i'_1,\cdots,i'_1}_l,\underbrace{i'_2,\cdots,i'_2}_l,\cdots,\underbrace{i'_{k},\cdots,i'_{k}}_l,j'_1,\cdots,j'_{t})$$ where $\widehat{i'_s}=\widehat{i_s}$ for $1\leq s\leq k$ and $(j)w=j$ for all $kl<j\leq r$. Since the first $k$ blocks in $\bsi_{\theta'}$ contains $gl^2$ inversion, we have $\ell(w)=gl^2$ and $w\in N_W(W_\theta)\cap\mathcal{D}_{\theta\theta}$. Moreover, $$\bsi_\lambda dw^{-1}=((i_1,\cdots,i_1,i_2,\cdots,i_2,\ldots,i_{k},\cdots,i_{k})_{\text{sorted}},j_1,\cdots,j_{t}).$$ Now the other inversions in $\bsi_\la d$ is unchanged during this sorting process. We conclude that $\ell(d)=\ell(dw^{-1})+\ell(w)$. Hence, $dw^{-1}\in\sD_{\la\theta'}$. Also, since Stab$_W(\bsi_{\la^{(0)}}d,\bsi_{\theta'^{(0)}}w)=_WP_a$ and Stab$_W(\bsi_{\la^{(1)}}d,\bsi_{\theta'^{(1)}}w)=_WP_b$, it follows that $W^d_{\lambda^{(0)}}\cap W^w_{\theta'^{(1)}}=1 \mbox{ and } W^d_{\lambda^{(1)}}\cap W^w_{\theta'^{(0)}}=1$. Hence, $dw^{-1}\in\mathcal{D}_{\lambda\theta'}^\circ$. By Lemma \[super2.2\], $e_{\lambda d, \theta'w}\in \End_{\mathcal{H}_\theta}(V_F(m|n)^{\otimes r})$. So the transitivity of relative norms (Lemma \[RNT\](a)) gives $$N_{W,W_\theta}(e_{\lambda d, \theta'w}) \in \End_{\mathcal{H}_F}(V_F(m'|n')^{\otimes r}).$$ The last assertion is clear since $N_{W,W_\theta}(e_{\lambda d, \theta'w})$ is a linear combination of basis elements of the form $N_{W,W_\la^x\cap W_{\theta'}}(e_{\la x,\theta'})$ each of which has a defect group $\leq_WW_{\theta'}$ and hence, is in $I_F(W_\theta,r)_{m'|n'}$. \[MC3\] Let $\theta=(l^k,1^t)$ be as above. For $\lambda,\mu\in \Lambda(m|n,r), d\in \mathcal{D}_{\la\theta},d'\in\mathcal{D}_{\mu\theta}$, assume $W_\theta\leq (W^{d}_{\lambda^{(0)}}\cap W^{d'}_{\mu^{(0)}})\times (W^{d}_{\lambda^{(1)}}\cap W^{d'}_{\mu^{(1)}}) $. Then there is $\theta'\in \Lambda(m'|n',r)$ for some $m'\geq m,n'\geq n$ and $w\in\mathcal{D}_{\theta\theta}$ such that $W_\theta=W_{\theta'}$ and $$N_{W,W_\theta}(e_{\mu d,\lambda d'})=N_{W,W_\theta}(e_{\mu d',\theta'w})N_{W,W_\theta}(e_{\theta'w,\lambda d}).$$ We first prove that $\bsi_\lambda d=(i_1,\cdots,i_r)$ and $\bsi_\mu d'=(i'_1,\cdots,i'_r)$ have the same parity sequence for the $k$ blocks of length $l$. Since $W_\theta\leq W^{00}_{\lambda d\cap\mu d'}\times W^{11}_{\lambda d\cap\mu d'}$, by Lemma \[super2.2\], $\mathcal{T}_s e_{\mu d',\lambda d}= e_{\mu d',\lambda d}\mathcal{T}_s$ for $s=(j,j+1)\in W_\theta\cap S$. Applying this equality to $v_{\bsi_\mu d'}$ gives $(-1)^{\widehat{i'_j}}q^{(-1)^{\widehat{i'_j}}} v_{\bsi_\lambda d}=(-1)^{\widehat{i_j}}q^{(-1)^{\widehat{i_j}}} v_{\bsi_\lambda d}$. Hence, $\widehat{i_j}=\widehat{i'_j}$ for $1\leq j\leq kl$. By the proof of Lemma \[key\], there are $\theta'\in\Lambda(m'|n',r)$ and a common $w\in N_W(W_\theta)\cap\mathcal{D}_{\theta\theta}$ such that $e_{\mu d',\theta'w}, e_{\theta'w,\lambda d}\in \End_{\mathcal{H}_\theta}(V_F(m|n)^{\otimes r})$. Then $$N_{W,W_\theta}(e_{\mu d',\theta'w}), N_{W,W_\theta}( e_{\theta'w,\lambda d})\in \End_{\mathcal{H}_F}(V_F(m'|n')^{\otimes r}).$$ Also, from the construction of $w$, we see that Stab$_W(\bsi_{\theta'} w)=W_{\theta'}$. Thus, $v_{\bsi_{\theta'}w}\sT_xe_{\theta'w,\la d}\neq0$ if and only if $x\in W_{\theta'}$. Hence, $$\begin{aligned} &N_{W,W_\theta}(e_{\mu d',\theta'w})N_{W,W_\theta}( e_{\theta'w,\lambda d})\\ &=N_{W,W_\theta}(e_{\mu d',\theta'w}N_{W,W_\theta}( e_{\theta'w,\lambda d}))\\ &=N_{W,W_\theta}(e_{\mu d',\theta'w}\sum_{x\in\sD_{\theta\theta}}N_{W_\theta,W_\theta^x\cap W_\theta}(\sT_{x^{-1}} e_{\theta'w,\lambda d}\sT_x))\\ &=\sum_{x\in\mathcal{D}_{\theta\theta}\cap W_{\theta'}}N_{W,W^x_\theta\cap W_\theta}(e_{\mu d',\theta'w}\mathcal{T}_{x^{-1}}e_{\theta'w,\lambda d}\mathcal{T}_x)\\ &= N_{W,W_\theta}(e_{\mu d',\lambda d}),\end{aligned}$$ as desired. We now establish the main result of this section and leave its application to §10 where we will use this result to compute the vertices of indecomposable summands of $V_F(m|n)^{\otimes r}$. \[IT\] Let $W_\theta=_WP_k$ be the $l$-parabolic subgroup of $W$ associated with composition $(l^k,1^{r-kl})$. Then $$I_F(W_\theta,r)=N_{W,W_\theta}(\End_{\mathcal{H}_\theta}(V_F(m|n)^{\otimes r})).$$ We first show that $$N_{W,W_\theta}(\End_{\mathcal{H}_\theta}(V_F(m|n)^{\otimes r}))\subseteq I_F(W_\theta,r).$$ By \[SRNB\], it is enough to prove that $$N_{W,W^y_\alpha\cap W_\beta}(e_{\mu d',\lambda dy})\in I_F(W_\theta,r)$$ for all $\mu,\lambda\in\Lambda{(m|n,r)},d'\in\mathcal{D}_{\mu\theta},d\in\mathcal{D}_{\lambda\theta},W_\alpha=W^d_\la\cap W_\theta,W_\beta=W^{d'}_\mu\cap W_\theta,y\in\mathcal{D}^\circ_{\alpha\beta}\cap W_\theta$. As in [[@DU 4.7]]{}, we proceed induction on $k$. If $k=0$, i.e $W_{\theta}=1,W^y_\alpha\cap W_\beta=1$, then the assertion follows from Lemma \[P=1\]. Assume that $W_\theta\neq 1$. Let $W_\gamma=W^y_\alpha\cap W_\beta\leq W_\theta$. If the maximal parabolic subgroup $P_\gamma$ of $W_\gamma$ is conjugate a proper parabolic subgroup of $W_\theta$, i.e., $P_\gamma<_{W} W_\theta$, then, by induction, $$N_{W,W_\gamma}(e_{\mu d,\lambda d'y})=f(q^2)^{-1}N_{W,P_\gamma}(e_{\mu d,\lambda d'y})\in I_F(P_\gamma,r),$$ where $f(\bsu)=\frac{d_{W_{\gamma^{(0)}}}(\bsu)d_{W_{\gamma^{(1)}}}(\bsu^{-1})}{d_{P_{\gamma^{(0)}}}(\bsu)d_{P_{\gamma^{(1)}}}(\bsu^{-1})}$. Hence, $N_{W,W_\gamma}(e_{\mu d,\lambda d'y})\in I_F(W_\theta,r)$. Thus, it remains to look at the case where $W_\gamma=W_\theta$. That is, $$W_\theta=(W^{d}_\lambda\cap W_\theta)^y\cap(W^{d'}_\mu\cap W_\theta)=W^{dy}_\lambda\cap W^{d'}_\mu\cap W_\theta.$$ This forces that $W_\alpha=W_\beta=W_\theta$ and $y=1$ as $y\in\mathcal{D}_{\alpha\beta}^\circ\cap W_\theta$. In particular, $W_\theta\leq W_{\la d\cap\mu d'}^{00}\times W_{\la d\cap\mu d'}^{11}$. Hence, by Lemma \[key\] and Corollaries \[MC3\] and \[MC2\], there exist $m'\geq m,n'\geq n$, $\theta'\in\La(m'|n',r)$, and $w\in\sD_{\theta\theta}\cap N_W(W_\theta)$ such that $$N_{W,W_\theta}(e_{\mu d',\lambda d})=N_{W,W_\theta}(e_{\mu d',\theta'w})N_{W,W_{\theta}}(e_{\theta'w,\lambda d})\in \vep I_F(W_\theta,r)_{m'|n'}\vep,$$ where $\vep$ is defined in Remark \[m’|n’ case\]. Hence, by identifying $I_F(W_\theta,r)$ with $\vep I_F(W_\theta,r)_{m'|n'}\vep $, we proved that all $N_{W,W^y_\beta\cap W_\alpha}(e_{\mu d,\lambda d'y})\in I_F(W_\theta,r)$. We next prove that $$I_F(W_\theta,r)\subseteq N_{W,W_\theta}(\End_{\mathcal{H}_\theta}(V_F(m|n)^{\otimes r})).$$ Equivalently, we want to prove every basis element $N_{W,W_\nu}(e_{\mu,\lambda d})\in I_F(W_\theta,r)\cap \mathcal{B}$, where $W_{\nu}=W^d_\lambda\cap W_\mu,d\in\mathcal{D}^\circ_{\lambda\mu}$ satisfying $P_\nu\leq_W W_\theta$, is in the R.H.S. Let $z$ be a distinguished representative of the $P_\nu-W_\theta$ double coset such that $P^z_\nu\leq W_\theta$ and $P^z_\nu$ is also parabolic (of course $l$-parabolic). $z$ is also a distinguished representative of the $P_\nu-P^z_\nu$ double coset. Let $\tau,\tau'$ be the decompositions of $r$ such that $W_\tau=P_\nu,W_{\tau'}=P^z_\nu$. We consider $\mathcal{H}_F$-$\mathcal{H}_F$-bimodule $$M=\Hom_F(v_\mu\mathcal{H}_F,v_{\bsi_\lambda d}\otimes_{\mathcal{H}_\tau}\mathcal{H}_F)$$ which has an $\mathcal{H}_F$-$\mathcal{H}_\tau$-bisubmodule $N=\Hom_F(v_\mu\mathcal{H}_F,Rv_{\bsi_\lambda d})$ and an $\mathcal{H}_F$-$\mathcal{H}_{\tau'}$-bisubmodule $N'=\Hom_F(v_\mu\mathcal{H}_F,Rv_{\bsi_\lambda d}\otimes_{\mathcal{H}_\tau}\mathcal{T}_z)$. Since $$\begin{aligned} M&\cong\bigoplus_{w\in\mathcal{D}_\tau}\Hom_F(v_\mu\mathcal{H}_F,v_{\bsi_\lambda d}\otimes \mathcal{T}_w)\\ &\cong\bigoplus_{w\in\mathcal{D}_\tau}\Hom_F(v_\mu\mathcal{H}_F,v_{\bsi_\lambda d})\otimes \mathcal{T}_w\\ &\cong N\otimes_{\mathcal{H}_\tau}\mathcal{H}_F \end{aligned}$$ as $\mathcal{H}_F$-$\mathcal{H}_F$-bimodules and, similarly, $M\cong N'\otimes_{\mathcal{H}_{\tau'}}\mathcal{H}_F$, it follows from Lemma \[RNT\] that $$N_{W,W_\tau}(\Hom_{\mathcal{H}_\tau}(v_\mu\mathcal{H}_F,v_{\bsi_\lambda d}))=N_{W,W_{\tau'}}(\Hom_{\mathcal{H}_{\tau'}}(v_\mu\mathcal{H}_F,v_{\bsi_\lambda d}\otimes_{\mathcal{H}_\tau}\mathcal{T}_z)).$$ Thus, there exists $h\in \Hom_{\mathcal{H}_{\tau'}}(v_\mu\mathcal{H}_F,v_{\bsi_\lambda d}\otimes_{\mathcal{H}_\tau}\mathcal{T}_z)$ such that $$N_{W,P_\nu}(e_{\mu,\lambda d})=N_{W,P^z_\nu}(h).$$ Hence, putting $g(\bsu)=\frac{d_{W_{\nu^{(0)}}}(\bsu)d_{W_{\nu^{(1)}}}(\bsu^{-1})}{d_{P_{\nu^{(0)}}}(\bsu)d_{P_{\nu^{(1)}}}(\bsu^{-1})}$, $$\begin{aligned} N_{W,W_\nu}(e_{\mu,\lambda d})&=g(q^2)^{-1}N_{W,P_\nu}(e_{\mu,\lambda d})\\ &=N_{W,P^z_\nu}(g(q^2)^{-1}h)\\ &=N_{W,W_\theta}(N_{W_\theta,P^z_\nu}(g(q^2)^{-1}h))\in \text{RHS} %N_{W,W_\theta}(\End_{\mathcal{H}_\theta}(V(m|n)^{\otimes r})). \end{aligned}$$ proving the theorem. Quantum matrix superalgebras ============================ We follow [@M] to introduce the quantum matrix superalgebras. \[DefSA\] Let $\mathcal{A}_q(m|n)$ be the associated superalgebra over $F$ generated by $x_{ij},1\leq i,j\leq m+n$ subject to the following relations 1. $x^2_{ij}=0$ for $\hat{i}+\hat{j}=1$; 2. $x_{ij}x_{ik}=(-1)^{(\hat{i}+\hat{j})(\hat{i}+\hat{k})}q^{(-1)^{\hat{i}+1}}x_{ik}x_{ij}$ for $j<k$; 3. $x_{ij}x_{kj}=(-1)^{(\hat{i}+\hat{j})(\hat{k}+\hat{j})}q^{(-1)^{\hat{j}+1}}x_{kj}x_{ij}$ for $i<k$; 4. $x_{ij}x_{kl}=(-1)^{(\hat{i}+\hat{j})(\hat{k}+\hat{l})}x_{kl}x_{ij}$ for $i<k$ and $j>l$; 5. $x_{ij}x_{kl}=(-1)^{(\hat{i}+\hat{j})(\hat{k}+\hat{l})}x_{kl}x_{ij}+(-1)^{\hat{k}\hat{j}+\hat{k}\hat{l}+\hat{j}\hat{l}}(q^{-1}-q)x_{il}x_{kj}$ for $i<k$ and $j<l$. Note that if all indices $i,j,k,l$ are taken from $[1,m]$ or $[m+1,m+n]$ then the relations coincide with those for the quantum matrix algebra; see, e.g., [@PW (3.5a)]. Manin proved that $\sA_q(m|n)$ has also the (usual) coalgebra structure with comultiplication $\Delta:\sA_q(m|n)\rightarrow \sA_q(m|n)\otimes \sA_q(m|n)$ and counit $\varepsilon:\sA_q(m|n)\rightarrow F$ defined by $$\Delta(x_{ij})=\sum_{k=1}^{m+n}x_{ik}\otimes x_{kj} \mbox{ and } \varepsilon(x_{ij})=\delta_{ij},\forall 1\leq i,j\leq m+n.$$ Further, the $\mathbb{Z}_2$ grading degree of $x_{ij}$ is $\hat{i}+\hat{j}\in\mathbb{Z}_2$. Hence, $\sA_q(m|n)$ is a super bialgebra. We now describe a basis for $\sA_q(m|n)$. For $\bsi=(i_1,i_2,\ldots,i_r),\bsj=(j_1,j_2,\ldots,j_r)\in I(m|n,r)$, let $$x_{\bsi,\bsj}=x_{i_1j_1}x_{i_2j_2}\cdots x_{i_rj_r}.$$ In particular, for $\lambda,\mu\in\Lambda{(m|n,r)},d\in\mathcal{D}^\circ_{\lambda\mu}$, define $x_{\mu,\lambda d}:=x_{\bsi_\mu,\bsi_\lambda d}.$ The elements $x_{\mu,\la d}$ are described in [@DR] by matrices which follows Manin [@M]. Let $M_{m+n}(\mathbb N)$ be the $(m+n)\times(m+n)$ matrix semigroup over $\mathbb N$ and let $$\aligned M(m|n)&=\{(a_{ij})\in M_{m+n}({\mathbb N})\mid a_{ij}=0,1 \text{ if }\hat i+\hat j=1\}\\ M(m|n,r)&=\{A\in M(m|n)\mid r=|A|\}, \endaligned$$ where $|A|$ is the sum of the entries of $A$. Then, by [@DR (3.2.1)], every triple $\la,\mu\in\La(m|n,r)$, and $d\in\sD_{\la\mu}^\circ$ defines a unique matrix $\jmath(\la,d,\mu)^t$, the transpose of $\jmath(\la,d,\mu)$, whose concatenation of row 1, row 2, and so on is the composition $\la d\cap\mu$ and whose row sum (resp., column) vector is $\mu$ (resp., $\la$). We will write $x^A=x_{\mu,\la d}$ if $A=\jmath(\la,d,\mu)^t$. . \[DR9.7\] (1) ([[@M],[@DR 9.3]]{}) The set $\{x^A\}_{A\in M(m|n)}$ forms a basis for $\sA_q(m|n)$. \(2) ([@DR 9.7]) Let $\sA_q(m|n,r)$ be the subspace spanned by $$\mathcal{B}^\vee=\{x_{\mu,\la d}\mid \lambda,\mu\in\Lambda{(m|n,r)},d\in\mathcal{D}^\circ_{\lambda\mu}\}= \{x^A\}_{A\in M(m|n,r)}.$$ Then $\sA_q(m|n,r)$ is a subcoalgebras with basis $\mathcal{B}^\vee$ and the quantum Schur superalgebra $\sS_F(m|n,r)$ is isomorphic to the dual algebra $\mathcal{A}_q(m|n,r)^*$. Recall from [@DR 9.8] that the isomorphism $\sS_F(m|n,r)\cong \sA_q(m|n,r)^*$ is obtained from an isomorphism $$\mathcal{A}_q(m|n,r)^*\cong \End_{\mathcal{H}_F}(V_F(m|n)^{\otimes r})$$ which is induced by an $\mathcal{A}_q(m|n,r)$-comodule structure on $V_F(m|n)^{\otimes r}$. This structure is the restriction of the comodule structure $\delta:V_F(m|n)^{\otimes r}\to \sA_q(m|n)\otimes V_F(m|n)^{\otimes r}$ defined by, for any $\bsi\in I(m|n,r)$, $$\label{comodule map} \delta(v_\bsi)=\sum_{\bsj\in I(m|n,r)}(-1)^{\sum_{1\leq k<l\leq r}\widehat{j_k}(\widehat{j_l}+\widehat{i_l})}x_{\bsi,\bsj}\otimes v_\bsj.$$ Now, the (left) $\mathcal{A}_q(m|n,r)$-comodule $V_F(m|n)^{\otimes r}$ turns into a right $\mathcal{A}_q(m|n,r)^*$-module with the action given by $$\label{module map} v\cdot f=(f\otimes\text{id}_{V_F(m|n)^{\otimes r}})\delta(v),\forall f\in \mathcal{A}_q(m|n,r)^*,v\in V_F(m|n)^{\otimes r}.$$ This action commutes with the Hecke algebra action as shown in [@DR 9.7] and results in the isomorphism above. We now make a comparison between the basis $\{\psi^d_{\mu\lambda}\mid \lambda,\mu\in\Lambda{(m|n,r)},d\in\mathcal{D}^\circ_{\lambda\mu}\}$ (or the relative norm basis $\mathcal{B}$) for $\sS_F(m|n,r)$ and the dual basis $$\{x^*_{\mu,\lambda d}\mid \lambda,\mu\in\Lambda{(m|n,r)},d\in\mathcal{D}^\circ_{\lambda\mu}\}.$$ First we observe $v_{\bsi_\nu}\cdot x^*_{\mu,\la d}=0$ if $\mu\neq \nu$. \[N-x basis\] For $\lambda,\mu\in\Lambda{(m|n,r)},d\in\mathcal{D}^\circ_{\lambda\mu}$, and $\nu=\la d\cap\mu$, we have $$N_{W,W_\nu}(e_{\mu,\lambda d})=(-1)^{\sum_{1\leq k<l\leq r}\widehat{(\bsi_\lambda d)_k}(\widehat{(\bsi_\lambda d)_l}+\widehat{(\bsi_\mu)_l})}x^*_{\mu,\lambda d}.$$ For any $\lambda,\mu\in\Lambda{(m|n,r)},d\in\mathcal{D}^\circ_{\lambda\mu}$. Since $N_{W,W_\nu}(e_{\mu,\lambda d})$ and $x^*_{\mu, \la d}$ are in $\End_{\mathcal{H}_F}(V_F(m|n)^{\otimes r})$, it is enough to consider their actions on $v_\mu$. By the proof of Theorem \[RTSB\], we know $$(v_\mu)N_{W,W_\nu}(e_{\mu,\la d})=(-1)^{\widehat{d}}\sum_{w_0w_1\in(W_{\mu^{(0)}}\times W_{\mu^{(1)}})\cap\mathcal{D}_{\nu}} (-1)^{\widehat{dw_0w_1}}q^{\ell(w_0)}(-q^{-1})^{\ell(w_1)}v_{\bsi_\lambda dw_0w_1}$$ where $W_{\nu}=W^d_\lambda\cap W_\mu$. On the other hand, $$\begin{aligned} v_\mu\cdot x^*_{\mu,\la d}&=( x^*_{\mu,\la d}\otimes \text{id}_{V_F(m|n)^{\otimes r}})\delta(v_\mu)\\ &=\sum_{\bsj\in I(m|n,r)}(-1)^{\sum_{1\leq k<l\leq r}\widehat{\bsj_k}(\widehat{\bsj_l}+\widehat{(\bsi_\mu)_l})} x^*_{\mu,\la d}(x_{\bsi_\mu, \bsj})v_\bsj. \end{aligned}$$ If $x^*_{\mu,\la d}(x_{\bsi_\mu, \bsj})\neq 0$, then there exists $w^{-1}\in W_\mu$ such that $(\bsi_\mu,\bsi_\lambda d)=(\bsi_\mu,\bsj w^{-1})$. So $\bsj=\bsi_\la dw\in (\bsi_\la d)\cdot W_\mu$, the orbit of $\bsi_\la d$. Then $$\begin{aligned} v_\mu\cdot x^*_{\mu,\la d}&=\sum_{\bsj\in(\bsi_\la d)\cdot W_\mu}(-1)^{\sum_{1\leq k<l\leq r}\widehat{\bsj_k}(\widehat{\bsj_l}+\widehat{(\bsi_\mu)_l})}x^*_{\mu,\la d}(x_{\bsi_\mu, \bsj})v_\bsj\\ &=\sum_{w\in W_\mu\cap \mathcal{D}_{\nu},\bsj=\bsi_\lambda dw}(-1)^{\sum_{1\leq k<l\leq r}\widehat{\bsj_k}(\widehat{\bsj_l}+\widehat{(\bsi_\mu)_l})}x^*_{\mu,\la d}(x_{\bsi_\mu, \bsj})v_\bsj. \end{aligned}$$ For $w\in \mathcal{D}_{\nu}\cap W_\mu$, there are $w_0\in W_{\mu^{(0)}},w_1\in W_{\mu^{(1)}}$ such that $w=w_0w_1$. Let $\bsi=\bsi_\la d=(i_1,\ldots,i_r)$ and $\bsj=\bsi_\la dw=(j_1,\dots,j_r)$. Then $j_k=i_{w(k)}$. Let $$\aligned \widetilde\sJ&=\{(k,l)\mid 1\leq k<l\leq r, j_k>j_l\},\quad\widetilde \sJ^*=\{(k,l)\mid 1\leq k<l\leq r, j_k<j_l\},\text{ and } \\ \sJ&=\sJ_{[1,m+n]}=\{(k,l)\in\widetilde \sJ\mid a_{i-1}+1\leq k<l\leq a_{i-1}+\mu_i, i\in[1,m+n]\},\endaligned$$ where $a_i=\sum_{j=1}^{i-1}\mu_j$ for $1\leq i\leq m+n$ and $a_0=0$. Define $\widetilde \sI$, $\widetilde \sI^*$, $\sI$ etc. similarly with respect to $\bsi$. We also define $\sJ^{(0)}=\sJ_{[1,m]}$ and $\sJ^{(1)}=\sJ_{[m+1,m+n]} $ similarly (so that $\sJ=\sJ^{(0)}\cup \sJ^{(1)}$). Then $\ell(w)=|\sJ|$, $\ell(w_i)=|\sJ^{(i)}|$ ($i=0,1$) and $$\sJ=\sJ^{(0)}\cup \sJ_{00}^{(1)}\cup\sJ_{10}^{(1)}\cup\sJ_{11}^{(1)},$$ where $\sJ_{ij}^{(1)}=\{(k,l)\in\sJ^{(1)}\mid \widehat j_k=i,\widehat j_l=j\}$ for all $ij\in\{00,10,11\}$. Note that $\sI=\emptyset$, $$\{(j_k,j_l)\mid(k,l)\in\widetilde\sJ\setminus\sJ\}=\{(i_k,i_l)\mid(k,l)\in\widetilde\sI\},$$ and $$A:=\{(j_k,j_l)\mid(k,l)\in\widetilde\sJ^*\cup\sJ\}=\{(i_k,i_l)\mid(k,l)\in\widetilde\sI^*\}=:B.$$ Hence, $(-1)^{\sum_{(k,l)\in\widetilde\sJ\setminus\sJ}\widehat{j_k}(\widehat{j_l}+\widehat{(\bsi_\mu)_l})}= (-1)^{\sum_{(k,l)\in\widetilde\sI}\widehat{i_k}(\widehat{i_l}+\widehat{(\bsi_\mu)_l})}$. On the other hand, if $(j_k,j_l)\in A$ with $(k,l)\in\sJ$, then $(j_l,j_k)\in B$. However, for $(k,l)\in\sJ^{(0)}$, $\widehat{(\bsi_\mu)_l}=0$, while for $(k,l)\in\sJ^{(1)}$, $\widehat{(\bsi_\mu)_l}=1$. Hence, $(-1)^{\widehat{j_k}(\widehat{j_l}+\widehat{(\bsi_\mu)_l)}}=(-1)^{\widehat{j_l}(\widehat{j_k}+\widehat{(\bsi_\mu)_l)}}$ for $(k,l)\in\sJ^{(0)}\cup\sJ^{(1)}_{00}\cup\sJ^{(1)}_{11}$, while $(-1)^{\widehat{j_k}(\widehat{j_l}+\widehat{(\bsi_\mu)_l)}}=(-1)(-1)^{\widehat{j_l}(\widehat{j_k}+\widehat{(\bsi_\mu)_l)}}$ for $(k,l)\in\sJ^{(1)}_{10}$. Hence, $$\aligned (-1)^{\sum_{1\leq k<l\leq r}\widehat{j_k}(\widehat{j_l}+\widehat{(\bsi_\mu)_l})}&=(-1)^{\sum_{(k,l)\in\widetilde\sJ}\widehat{j_k}(\widehat{j_l}+\widehat{(\bsi_\mu)_l})}\times(-1)^{\sum_{(k,l)\in\widetilde\sJ^*}\widehat{j_k}(\widehat{j_l}+\widehat{(\bsi_\mu)_l})}\\ &=(-1)^{\sum_{(k,l)\in\widetilde\sJ\setminus\sJ}\widehat{j_k}(\widehat{j_l}+\widehat{(\bsi_\mu)_l})}\times(-1)^{\sum_{(k,l)\in\widetilde\sJ^*\cup\sJ}\widehat{j_k}(\widehat{j_l}+\widehat{(\bsi_\mu)_l})}\\ &=(-1)^{\sum_{(k,l)\in\widetilde\sI}\widehat{i_k}(\widehat{i_l}+\widehat{(\bsi_\mu)_l})}\times(-1)^{\sum_{(k,l)\in\widetilde\sJ^*\cup\sJ}\widehat{i_k}(\widehat{i_l}+\widehat{(\bsi_\mu)_l})+|\sJ_{10}^{(1)}|}\\ &=(-1)^{\sum_{1\leq k<l\leq r}\widehat{i_k}(\widehat{i_l}+\widehat{(\bsi_\mu)_l})}\times (-1)^{|\sJ_{10}^{(1)}|}. \endaligned$$ Let $\bsj'=\bsi_\la dw_0$ and define $\sJ'$ as $\sJ$ above with $\bsj$ replaced by $\bsj'$. Then, by Definition \[DefSA\] and ([@DU2 1.3]), $$x_{\bsi_\mu,\bsi_\la d w_0}=(-1)^{\sum_{(k,l)\in\sJ'}\widehat{j'_k}\widehat{j'_l}}q^{\ell(w_0)}x_{\mu,\la d}=(-1)^{\widehat{dw_0}-\widehat d}q^{\ell(w_0)}x_{\mu,\la d},$$ and $$x_{\bsi_\mu,\bsi_\la d w_0w_1}=(-1)^{\star}q^{-\ell(w_1)}x_{\mu,\la dw_0}=(-1)^\ell(-1)^{\widehat{dw_0w_1}-\widehat {dw_0}}q^{-\ell(w_1)}x_{\mu,\la dw_0},$$ where $$\aligned \star&=\sum_{(k,l)\in\sJ^{(1)}}(1+\widehat{j_k})(1+\widehat{j_l})\\ &=\sum_{(k,l)\in\sJ^{(1)}}(1+\widehat{j_k}+\widehat{j_l})+\sum_{(k,l)\in\sJ^{(1)}}\widehat{j_k}\widehat{j_l}\\ &=\ell+(\widehat{dw_0w_1}-\widehat {dw_0}),\endaligned$$ where $\ell=\sum_{(k,l)\in\sJ^{(1)}}(1+\widehat{j_k}+\widehat{j_l})$ and so $(-1)^\ell=(-1)^{|\sJ_{00}^{(1)}|+|\sJ_{11}^{(1)}|}$. Hence, $$(-1)^\ell\times (-1)^{|\sJ_{10}^{(1)}|}=(-1)^{\ell(w_1)}$$ and substituting gives $$x_{\bsi_\mu, \bsi_\lambda dw_0w_1}=(-1)^{|\sJ_{10}^{(1)}|}(-1)^{\widehat{dw_0w_1}-\widehat{d}}q^{\ell(w_0)}(-q^{-1})^{\ell(w_1)}x_{\mu,\lambda d}.$$ Therefore, $$\begin{aligned} v_\mu\cdot x^*_{\mu,\la d}=&(-1)^{\sum_{1\leq k<l\leq r}\widehat{i_k}(\widehat{i_l}+\widehat{(\bsi_\mu)_l})}(-1)^{\widehat{d}}\\ &\sum_{w_0w_1\in(W_{\mu^{(0)}}\times W_{\mu^{(1)}})\cap\mathcal{D}_{\nu(d)}}(-1)^{\widehat{dw_0w_1}}q^{\ell(w_0)}(-q^{-1})^{\ell(w_1)}v_{\bsi_\lambda dw_0w_1}\\ &=(-1)^{\sum_{1\leq k<l\leq r}\widehat{(\bsi_\lambda d)_k}(\widehat{(\bsi_\lambda d)_l}+\widehat{(\bsi_\mu)_l})}(v_\mu)N_{W,W_\nu}(e_{\mu,\la d})\end{aligned}$$ proving the theorem. By Theorem \[RTSB\], we have the following. For $\lambda,\mu\in\Lambda{(m|n,r)},d\in\mathcal{D}^\circ_{\mu\la}$, we have $$\psi_{\mu,\lambda}^ d=q^{\ell(d)}(-1)^{\widehat d}(-1)^{\sum_{1\leq k<l\leq r}\widehat{(\bsi_\lambda d)_k}(\widehat{(\bsi_\lambda d)_l}+\widehat{(\bsi_\mu)_l})} x^*_{\mu,\la d}.$$ Frobenius morphisms and Brauer homomorphisms ============================================ Let $\sA_q(m)$ (resp., $\sA_q(n)$) be the subalgebras of $\sA_q(m|n)$ generated by $x_{ij}$ for $i,j\in[1,m]$ (resp., $i,j\in[m+1,m+n]$). Then $\sA_q(m)\otimes\sA_q(n)$ is a subalgebra, but not a subcoalgebra. We now show that $\sA_q(m)\otimes\sA_q(n)$ contains a central subbialgebra of $\sA_q(m|n)$. \[SAC\] For $1\leq i,j\leq m$ or $m+1\leq i,j\leq m+n$, the elements $x^l_{ij}$ are in the center of $\sA_q(m|n)$. Since $\hat i=\hat j$, the (first) signs on the right hand side of the relations in Definition \[DefSA\] are all $+1$. Note also that $q^l=(q^{-1})^l=1,[\![l]\!]_q=0$. For $j<k$, by \[DefSA\](2) $$x^l_{ij}x_{ik}=q^{l(-1)^{\hat{i}+1}}x_{ik}x^l_{ij}=x_{ik}x^l_{ij}.$$ For $i<k$, by \[DefSA\](3) $$x^l_{ij}x_{kj}=q^{l(-1)^{\hat{j}+1}}x_{kj}x^l_{ij}=x_{kj}x^l_{ij}.$$ For $i<k$ and $j>l$, by \[DefSA\](4) $$x^l_{ij}x_{kl}=x_{kl}x^l_{ij}.$$ Finally, for $i<k$ and $j<l$, we claim $$[x^s_{ij},x_{kl}]=(-1)^{\hat{k}\hat{j}+\hat{k}\hat{l}+\hat{j}\hat{l}}(q^{-1}-q)[\![s]\!]_{q^{2(-1)^{\hat{i}+1}}}x^{s-1}_{ij} x_{il}x_{kj}.$$ Indeed, this is clear, by \[DefSA\](5), for $s=1$. In general, we apply induction to $[x^{s+1}_{ij},x_{kl}]=x_{ij}[x_{ij}^s,x_{kl}]+[x_{ij},x_{kl}]x_{ij}^s$ to prove the claim. Now taking $s=l$ and noting ${[\![l]\!]}_{q^{\pm2}}=0$ give $x^l_{ij}x_{kl}=x_{kl}x^l_{ij}$ in this case. \[CCA\] For $1\leq i,j\leq m$ or $m+1\leq i,j\leq m+n$, we have $$\begin{aligned} &\Delta(x_{ij}^l)=\sum_{k=1}^mx^l_{ik}\otimes x^l_{kj}\mbox{ for }1\leq i,j\leq m;\\ &\Delta(x_{ij}^l)=\sum_{k=m+1}^{m+n}x^l_{ik}\otimes x^l_{kj}\mbox{ for }m+1\leq i,j\leq m+n. \end{aligned}$$ Since $\mathcal{A}_q(m|n)$ is a bialgebra, $$\Delta(x^l_{ij})=(\Delta(x_{ij}))^l=(\sum^{m+n}_{k=1}x_{ik}\otimes x_{kj})^l=(A+B)^l,$$ where $A=\sum^{m}_{k=1}x_{ik}\otimes x_{kj}$ and $B=\sum^{m+n}_{k=m+1}x_{ik}\otimes x_{kj}$. Putting $u=q^{-2(-1)^{\hat{i}+1}}$ (and so $u^l=1$), we have by the relations in Definition \[DefSA\], $AB=uBA$. Thus, by the quantum binomial theorem [@PW (7.1.a)], $$\Delta(x^l_{ij})=\sum^l_{h=0}\left[\!\!\left[l\atop h\right]\!\!\right]_uA^hB^{l-h}.$$ Since $\left[\!\!\left[l\atop h\right]\!\!\right]_u=0$ for $0<h<l$, it turns out $\Delta(x^l_{ij})=A^l+B^l$. Now, by the quantum multinomial theorem (see, e.g., [@DDPW Ex. 0.14]), \[DefSA\](1) implies $B^l=0$ for $1\leq i,j\leq m$, and $A^l=0$ for $m+1\leq i,j\leq m+n$. Hence, we have reduced the proof to the non-super case and the assertion follows from [@PW 7.2.2]. The subalgebra generated by $x_{ij}^l$ with $i,j\in[1,m]$ or $i,j\in[m+1,m+n]$ is also a subcoalgebra in the center of $\sA_q(m|n)$, and hence, is a central subbialgebra. The existence of the subbialgebra is the key to the definition of the Frobenius morphism. Let $A(m,n)$ be the polynomial algebra over $F$ generated by $t_{ij}$ for $1\leq i,j\leq m$ or $m+1\leq i,j\leq m+n$. If we also define polynomial algebras $A(m)=F[t_{i,j}]_{1\leq i,j\leq m}$ and $A(n)=F[t_{i,j}]_{m+1\leq i,j\leq m+n}$, then $A(m,n)\cong A(m)\otimes_FA(n)$. We will identify the two polynomial algebras in the sequel. Then, by the lemma above, the map $$\begin{aligned} \mathcal{F}:A(m,n)&\longrightarrow \mathcal{A}_q(m|n)\\ t_{ij}&\longmapsto x^l_{ij} \end{aligned}$$ is a bialgebra monomorphism. This is called the [*Frobenius morphism*]{}. For the fixed $r\in\mathbb{N}$, set $$\label{R_r} \mathcal{R}_r=\sR_r(l)=\{(r_{-1},r_0)\mid r_{-1},r_0\in\mathbb N, r_{-1}+lr_0=r\}.$$ For $\vecr=(r_{-1},r_0)\in\mathcal{R}_r$, we put $$\mathcal{A}_q(m|n,\vecr)=\mathcal{A}_q(m|n,r_{-1})\otimes A(m,n)_{r_0}$$ where $A(m,n)_{r_0}=\sum_{s+t=r_0}A(m)_s\otimes A(n)_t$ is the $r_0$th homogenous component of $A(m,n)$. It is easy to show that $\mathcal{A}_q(m|n,\vecr)$ is a subcoaglebra of $\mathcal{A}_q(m|n)\otimes A(m,n)$. Let $$\label{Schur vecr} \sS_F(m|n,\vecr):=\sA_q(m|n,\vecr)^*\cong\sS_F(m|n,r_{-1})\otimes S(m,n)_{r_0},$$ where $$\label{S(m,n)} S(m,n)_{r_0}=A(m,n)_{r_0}^*=\bigoplus_{i=0}^{r_0}(S(m,i)\otimes S(n,r_0-i)),$$ which is a sum of tensor products of certain classical Schur algebras. Consider the map $$\mu\circ(1\otimes \mathcal{F}):\mathcal{A}_q(m|n)\otimes A(m,n)\rightarrow \mathcal{A}_q(m|n)$$ where $\mu$ is the multiplication in $\mathcal{A}_q(m|n)$, then $\mu\circ(1\otimes \mathcal{F})$ is a coalgebra map and $$\mu\circ(1\otimes \mathcal{F})(\mathcal{A}_q(m|n,\vecr))\subseteq \mathcal{A}_q(m|n,r).$$ Thus, upon restriction, we obtain a coalgebra map $$F_{\vecr}:\mathcal{A}_q(m|n,\vecr)\rightarrow \mathcal{A}_q(m|n,r).$$ If we identify $\sS_F(m|n,r)$ with $\sA_q(m|n,r)^*$ under Lemma \[DR9.7\](2), then the [*Brauer homomorphism*]{} associated with $\vecr\in\mathcal{R}_r$ is the [*surjective*]{} map $$\phi_\vecr:\sS_F(m|n,r)\rightarrow \sS_F(m|n,\vecr):=(\mathcal{A}_q(m|n,\vecr))^*$$ dual to the coalgebra map $F_{\vecr}:\mathcal{A}_q(m|n,\vecr)\rightarrow \mathcal{A}_q(m|n,r)$. We now determine the kernel of $\phi_\vecr$. First, we determine the image of $F_{\vecr}$. For $\vecr=(r_{-1},r_0)\in\mathcal{R}_r$, define the $l$-parabolic subgroup $P_{\vecr}$ of $W$ to be the parabolic subgroup associated with $(1^{r_{-1}},l^{r_0})\in\Lambda{(m|n,r)}$. Thus, with the notation used in , $P_\vecr=P_{r_0}$. For $\bsi=(i_1,i_2,\cdots,i_r)\in I(m|n,r)$, let $$l\bsi=(\underbrace{i_1,\cdots,i_1}_l,\underbrace{i_2,\cdots,i_2}_l,\cdots,\underbrace{i_r,\cdots,i_r}_l)\in I(m|n,rl).$$ \[image\] The image of $F_{\vecr}$ is the subspace spanned by $$\{x_{\mu,\lambda d}\mid \lambda,\mu\in\Lambda{(m|n,r)},d\in\mathcal{D}^\circ_{\lambda\mu},P_{\vecr}\leq_W W^d_\lambda\cap W_\mu\}.$$ We first observe that the set $$\{x_{\bsi_{-1},\bsj_{-1}}\otimes t_{\bsi_0,\bsj_0}\mid \bsi_{-1},\bsj_{-1}\in I(m|n,r_{-1}),\bsi_0,\bsj_0\in I(m,n;r_0)\}$$ spans $\mathcal{A}_q(m|n,\vecr)$, where $$I(m,n;r_0)=\bigcup_{s+t=r_0}I(m,s)\times(\mathbf{m}+ I(n,t))\text{ with }\mathbf{m}=(\underbrace{m,\ldots,m}_n).$$ To get a basis, by Lemma \[DR9.7\](2), we simply take the $\bsi,\bsj$ involved to satisfy the conditions that $\bsi$ weakly increasing and $j_a\leq j_b$ whenever $i_a=i_b$. So we may assume that $x_{\bsi_{-1},\bsj_{-1}}=x_{\mu_{(-1)},\lambda_{(-1)}d_{-1}},t_{\bsi_0,\bsj_0}=t_{\mu_{(0)},\lambda_{(0)}d_0}$ for $\lambda_{(i)},\mu_{(i)}$ into $m+n$ parts and $d_i\in\mathcal{D}^\circ_{\lambda_{(i)},\mu_{(i)}},i=-1,0$. Now, $F_{\vecr}(x_{\bsi_{-1},\bsj_{-1}}\otimes t_{\bsi_0,\bsj_0})=x_{\bsi_{-1},\bsj_{-1}}x_{l\bsi_0,l\bsj_0}$. By Proposition \[CCA\], every factor $x_{i_kj_k}^l$ of $x_{l\bsi_0,l\bsj_0}$, where $i_k,j_k\in[1,m]$ or $[m+1,m+n]$, is in the center of $\mathcal{A}_q(m|n)$. Thus, we may move them around so that the product $x_{\bsi_{-1},\bsj_{-1}}x_{l\bsi_0,l\bsj_0}$ can be written as $x_{\bsi,\bsj}$ with $\bsi$ weakly increasing and $\bsj$ satisfying $j_a\leq j_b$ whenever $i_a=i_b$ and, if $i_a=i_{a+1}$ and $j_a=j_{a+1}$, then $i_a,j_a\in[1,m]$ or $i_a,j_a\in[m+1,m+n]$. This means that we may find $\lambda,\mu\in\Lambda{(m|n,r)}$ and $d\in\mathcal{D}^\circ_{\lambda\mu}$ such that $F_{\vecr}(x_{\bsi_{-1},\bsj_{-1}}\otimes t_{\bsi_0,\bsj_0})=x_{\bsi_\mu,\bsi_\lambda d}$. Clearly, by the construction, $P_{\vecr}\leq_W W_\la^d\cap W_\mu$. Conversely, if $\lambda,\mu\in\Lambda{(m|n,r)}$ and $d\in\mathcal{D}^\circ_{\lambda\mu}$ such that $P_{\vecr}\leq_W W^d_\lambda\cap W_\mu$, then both $W^d_{\la^{(0)}}\cap W_{\mu^{(0)}}$ and $W^d_{\la^{(1)}}\cap W_{\mu^{(1)}}$ contain $l$-parabolic subgroups. Thus, if we write $\bsi=(i_1,i_2,\cdots,i_r)=\bsi_\mu$ and $\bsj=(j_1,j_2,\cdots,j_r)=\bsi_\la d$, then $x_{\mu,\lambda d}=x_{\bsi,\bsj}$ has factors of the form $x_{i_kj_k}^l$, where $i_k,j_k\in[1,m]$ or $[m+1,m+n]$. Moving all such factors to the right, we may rewrite $x_{\bsi\bsj}=x_{\bsi_{-1},\bsj_{-1}}x_{l\bsi_0,l\bsj_0}$. Clearly, $(\bsi_{-1},\bsj_{-1})$ satisfies the even-odd trivial intersection property. Hence, we see that $x_{\bsi,\bsj}=F_{\vecr}(x)$ for some $x\in\mathcal{A}_q(m|n,\vecr)$. As an application of Theorem \[N-x basis\], we now determine the kernel of the Brauer homomorphism. \[kernel\] If $\{x^*_{\mu,\lambda d}\mid \lambda,\mu\in\Lambda{(m|n,r)},d\in\mathcal{D}^\circ_{\lambda\mu}\}$ is the dual basis of $\mathcal{B}^\vee$ in Lemma \[DR9.7\](2), then the kernel of $\phi_\vecr$, for $\vecr=(r_{-1},r_0)\in\mathcal{R}_r$, is spanned by $$\{x^*_{\mu,\lambda d}\mid\lambda,\mu\in\Lambda{(m|n,r)},d\in\mathcal{D}^\circ_{\lambda\mu},P_{\vecr}\nleq_W W^d_\lambda\cap W_\mu\}.$$ In particular, $\ker \phi_\vecr=I_F(P_{r_0-1},r).$ Suppose that $\phi_\vecr(x^*_{\mu,\lambda d})=0$. Thus we have $x^*_{\mu,\lambda d}(F_{\vecr}(a))=0$ for all $a\in \mathcal{A}_q(m|n,\vecr)$. The lemma above implies that $P_{\vecr}\nleq_W W^d_\lambda\cap W_\mu$. Now, the largest $l$-parabolic subgroup $P$ satisfying $P_{\vecr}\not\leq_W P$ is $P_{r_0-1}$, and by Theorem \[N-x basis\], $I_F(P,r)$ is spanned by $$\{x^*_{\mu,\la d}\mid \la,\mu\in\La(m|n,r), d\in\sD_{\la\mu}^\circ,P\leq_W W^d_\lambda\cap W_\mu\}.$$ The last assertion follows from the filtration and Theorem \[N-x basis\]. Shrinking defect groups by Brauer homomorphisms =============================================== In this section we continue to assume that $R=F$ is a field. We shall prove that the Brauer homomorphism $\phi_\vecr$ sends certain primitive idempotents in $\sS_F(m|n,r)$ to primitive idempotents in $\sS_F{(m|n,\vecr)}$ with the trivial defect group. Using the filtration , we make the following definition. \[defect group\] For a primitive idempotent $e\in \sS_F(m|n,r)$, there is a number $k=k(e)$ such that $e\in I_F(P_k,r),e\not\in I_F(P_{k-1},r)$. We set $D(e)=P_{k}$ and call $D(e)$ the [*defect group*]{} of $e$. Every primitive idempotent $e$ of $\sS_F(m|n,r)$ defines an indecomposable $\mathcal{H}_F$-submodule $V_F(m|n)^{\otimes r}e$ of $V_F(m|n)^{\otimes r}$. We now determine its defect group. The following result is a super version of [[@DU Th. 4.8]]{}. For completeness, we include a proof. \[VAP\] If $e\in \sS_F(m|n,r)$ is a primitive idempotent, then the defect group $D(e)$ of $e$ is the vertex of the indecomposable $\sH_F$-module $V_F(m|n)^{\otimes r}e$. Let $T=V_F(m|n)^{\otimes r}$ and suppose $W_\theta$ is the vertex of $Te$. Then $Te$ is $\sH_\theta$-projective and so $$\aligned e\sS_F(m|n,r)e&=\End_{\sH_F}(Te)=N_{W,W_\theta}(\End_{\sH_\theta}(Te))= N_{W,W_\theta}(e\End_{\sH_\theta}(T)e)\\ &= eN_{W,W_\theta}(\End_{\sH_\theta}(T))e =eI_F(W_\theta,r)e,\;\text{by Theorem \ref{IT}} \endaligned$$ Hence, $e\in I_F(W_\theta,r)$ and $D(e)\leq_WW_\theta$. However, by Theorem \[IT\] again, the equalities above continue to hold with $W_\theta$ replaced by $D(e)$. Hence, $Te$ is $\sH_{D(e)}$-projective which implies $W_\theta\leq_WD(e)$. Hence, $W_\theta=_WD(e)$. \[BHP\] Let $e\in \sS_F(m|n,r) $ be a primitive idempotent and let $ \phi_{\vecr} $ be the Brauer homomorphism associated with $\vecr\in \mathcal R_r$ . Then $\phi_{\vecr}(e)\neq 0$ if and only if $P_{\vecr}$ is conjugate to a parabolic subgroup of $D(e)$. For $\vecr=(r_{-1},r_0)$, by Theorem \[kernel\], $\phi_{\vecr}(e)\neq 0$ if and only if $e\not\in\text{ker}(\phi_\vecr)=I_F(P_{r_0-1},r)$. This is equivalent to $P_\vecr\leq_W D(e)$. Choose $\vecr\in\mathcal{R}_r$ satisfying $D(e)=P_\vecr$. Then we will prove that $\phi_\vecr(e)$ is primitive and determine its defect group. This is the key to the classification theorem in next section. For $\vecr\in\mathcal R_r$, let $$\label{rhotheta} \begin{aligned} \rho&=\rho_{\vecr}=(\rho_{-1},\rho_0)=(r_{-1},lr_0)\\ \theta&=\theta_{\vecr}=(\underbrace{1,\cdots,1}_{r_{-1}},\underbrace{l,\cdots,l}_{r_0}) \end{aligned}$$ Since $W_\rho\leq W$, $V_F(m|n)^{\otimes r}$ is a $\mathcal{H}_F$-module, and $$V_F(m|n)^{\otimes r}=V_F(m|n)^{\otimes \rho_{-1}}\otimes V_F(m|n)^{\otimes \rho_0},$$ restriction makes $V_F(m|n)^{\otimes r}$ into an $\mathcal{H}_\rho$-module. Let $\sS_F(m|n,\rho)=\End_{\mathcal{H}_\rho}(V_F(m|n)^{\otimes r})$, then (see ) $$\label{yyy} \sS_F(m|n,\rho)\cong \sS_F(m|n,\rho_{-1})\otimes \sS_F(m|n,\rho_0).$$ Recall from and that $$\vec{\La}(m|n,\rho)=\Lambda{(m|n,\rho_{-1})}\times \Lambda{(m|n,\rho_0)},$$ and, for $\vec{\lambda},\vec{\mu}\in \vec{\La}(m|n,\rho)$, $$\mathcal{D}^\circ_{\vec{\lambda}\vec{\mu};\rho}= \{d\in\mathcal{D}_{\vec{\lambda}\vec{\mu}}\mid d=d_1d_2\in W_{\rho_{-1}}\times W_{\rho_0},d_1\in \mathcal{D}^\circ_{\la_{(-1)}\mu_{(-1)}},d_2\in \mathcal{D}^\circ_{\lambda_{(0)}\mu_{(0)}}\}\subseteq W_\rho.$$ So $\sS_F(m|n,\rho)$ has a basis (see ) $$\{N_{W_\rho,W_{\vec\la}^d\cap W_{\vec\mu}}(e_{\vec{\mu},\vec{\lambda}d})\mid \vec{\lambda},\vec{\mu}\in \vec{\La}(m|n,\rho), d\in \mathcal{D}^\circ_{\vec{\lambda}\vec{\mu};\rho}\}.$$ Let $P$ be an $l$-parabolic subgroup of $W_\rho$ and define $I_F(P,\rho)$ to be the subspace of $\sS_F(m|n,\rho)$ spanned by $N^d_{\vec\mu\vec{\lambda}}$ where $\vec{\lambda},\vec{\mu}\in \vec{\La}(m|n,\rho),d\in\mathcal{D}^\circ_{\vec{\lambda}\vec{\mu};\rho}$ and the maximal $l$-parabolic subgroup of $W_{\vec\la}^d\cap W_{\vec\mu}$ is conjugate in $W_\rho$ to parabolic subgroup of $P$. By , the first two items of the following results are the $W_\rho$ version of Corollary \[MC2\] and Theorem \[IT\]. \[ISS\] Let $P=W_\theta$ be an arbitrary $l$-parabolic subgroup of $W_\rho$. - The space $I_F(P,\rho)$ is an ideal of $\sS_F(m|n,\rho)$. - $I_F(W_\theta,\rho)=N_{W_\rho,W_\theta}(\End_{\mathcal{H}_\theta}(V_F(m|n)^{\otimes r})).$ - If $\xi$ is a composition of $r$ such that $W_\xi\leq W_\rho$ and $P_\xi=W_\theta$ is the maximal l-parabolic subgroup of $W_\xi$, then $$N_{W_\rho,W_\theta}(\End_{\mathcal{H}_\theta}(V_F(m|n)^{\otimes r}))=N_{W_\rho,W_\xi}(\End_{\mathcal{H}_\xi}(V_F(m|n)^{\otimes r})).$$ \(3) The RHS is clearly contained in the LHS. The rest of the proof follows from a similar argument as in the proof of [@DD 4.3]. The Pioncaré polynomials involved in the proof there have to be replaced by the product of the Pioncaré polynomials in $q^2$ for the even parts and the Pioncaré polynomials in $q^{-2}$ for the odd parts. We fix an element $\vecr=(r_{-1},r_0)\in\mathcal R_r$, and define $\rho=\rho_{\vecr}$ and $\theta=\theta_{\vecr}$ as in . Let $$\mathcal{A}_q(m|n,\rho)=\mathcal{A}_q(m|n,r_{-1})\otimes \mathcal{A}_q(m|n,lr_0).$$ Recall the definition of the Frobenius morphism $\mathcal{F}$. Define the coalgebra homomorphism $$F_\rho=\text{id}\otimes \mathcal{F}:\mathcal{A}_q(m|n,\vecr)\rightarrow \mathcal{A}_q(m|n,\rho).$$ Consider the dual of $F_\rho$: $$\psi_\rho=F_\rho^*:\sS_F(m|n,\rho)\rightarrow \sS_F(m|n,\vecr).$$ Since $F_\rho$ is injective, $\psi_\rho$ is a surjective algebra homomorphism. On the other hand, the multiplication map $$\mu: \mathcal{A}_q(m|n,\rho)\rightarrow \mathcal{A}_q(m|n,r),$$ induces an algebra monomorphism $$\iota_\rho=\mu^*:\sS_F(m|n,r)\hookrightarrow\sS_F(m|n,\rho),$$ defined by $\iota(f)=f\circ\mu$ for all $f\in\sS_F(m|n,r)$. Thus, taking the dual of the relation $\mu\circ\mathcal{F}_\rho=\mathcal{F}_\vecr$ gives the following commutative diagram (recall $\rho=\rho_\vecr$): $$\label{comm diag} {\unitlength=1cm \begin{picture}(5,3.2) \put(-.5,2.5){$\sS_F(m|n,\rho)$} \put(4.1,2.5){$\sS_F(m|n,r)$} \put(-0.5,0.5){$\sS_F(m|n,\vecr)$} \put(3.9,2.6){\vector(-1,0){2.3}} \put(4.2,2.4){\vector(-3,-2){2.7}} \put(0.5,2.3){\vector(0,-1){1.35}} \put(2.5,2.8){$\iota_\rho$} \put(2.5,0.95){$\phi_\vecr$} \put(0.05,1.6){$\psi_\rho$} \end{picture}}$$ We claim that $\iota_\rho$ coincides with the inclusion $\End_{\mathcal{H}_R}(V_F(m|n)^{\otimes r})\subseteq \End_{\mathcal{H}_\rho}(V_F(m|n)^{\otimes r})$. Indeed, for $i=-1,0$, let $ \delta_{i}:V_F(m|n)^{\otimes \rho_{i}}\rightarrow \mathcal{A}_q(m|n,\rho_i)\otimes V_F(m|n)^{\otimes \rho_i}$ be the comodule structure map as defined in and let $\delta_\rho=(23)(\delta_{-1}\otimes\delta_0):V_F(m|n)^{\otimes r}\to\sA_q(m|n,\rho)\otimes V_F(m|n)^{\otimes r}$. Then it is direct to check that $(\mu\otimes1)\circ \delta_\rho=\delta:V_F(m|n)^{\otimes r}\to\sA_q(m|n,r)\otimes V_F(m|n)^{\otimes r}$. (The signs involved in both sides are the same!) Thus, for any $v\in V_F(m|n)^{\otimes r}$ and $f\in\sS_F(m|n,r)$, by , $$\aligned v\cdot f&=(f\otimes 1)\delta(v)=(f\otimes 1)(\mu\otimes 1)\delta_\rho(v)\\ &=(f\mu\otimes 1)\delta_\rho(v)=v\cdot\iota_\rho(f), \endaligned$$ proving the claim. By the definition of $\psi_\rho$ and Theorem \[kernel\], we have $$\label{kernel psi} \ker (\psi_\rho)=\sS_F(m|n,r_{-1})\otimes I_F(P_{r_0-1},r_0l)=\sum_{P_{r_0}\nleq_{W_{\rho}}P}I_F(P,\rho).$$ Consider the ideal of the form $I_F(P_{k},r_{-1})\otimes S(m,n)_{r_0}$ in $\sS_F(m|n,\vecr)$ and the ideal $$I_F(P_k,\vecr):=\phi_\vecr(I_F(P_{r_0+k},r))=I_F(P_{r_0+k},r)/I_F(P_{r_0-1},r).$$ We now show that both ideals are the same. \[new\] For $\vecr=(r_{-1},r_0)\in\sR_r$ and $r_0\leq k\leq s$, where $r=sl+t$ with $0\leq t<l$, the restriction $\phi_\vecr:I_F(P_k,r)\to I_F(P_{k-r_0},r_{-1})\otimes S(m,n)_{r_0}$ is surjective. In other words, we have $$I_F(P_k,\vecr)=I_F(P_{k-r_0},r_{-1})\otimes S(m,n)_{r_0}.$$ By Theorem \[N-x basis\], $I_F(P_k,r)$ is spanned by all $x_{\mu,\la d}^*$ satisfying $P_{\la d\cap\mu}\leq_W P_k$. We first prove that $\phi_\vecr(I_F(P_k,r))\subseteq I_F(P_{k-r_0},r_{-1})\otimes S(m,n)_{r_0}$. This is seen as follows. Let $W_\theta=P_k$ and, for any $N_{W,W_\theta}(a)\in I_F(P_k,r)=N_{W,W_\theta}(\End_{\sH_\theta}(V_F(m|n)^{\otimes r}))$, write by Lemma \[Mackey\] $$N_{W,W_\theta}(a)=\sum_{d\in\mathcal{D}_{\theta\rho}}N_{W_\rho,W^d_\theta\cap W_\rho}(\mathcal{T}_{d^{-1}}a\mathcal{T}_d).$$ If we put $W_\alpha=W^d_\theta\cap W_\rho=W_{\alpha_{(-1)}}\times W_{\alpha_{(0)}}$, where $\alpha_{(i)}$ is a composition of $\rho_{i}$ for $i=-1,0$, then $$\End_{\sH_\alpha}(V_F(m|n)^{\otimes r})=\End_{\sH_{\alpha_{(-1)}}}(V_F(m|n)^{\otimes r_{-1}})\otimes \End_{\sH_{\alpha_{(0)}}}(V_F(m|n)^{\otimes r_0l}).$$ So there exist $a_i(d)\in \End_{\sH_{\alpha_{(i)}}}(V_F(m|n)^{\otimes \rho_i})$ such that $\mathcal{T}_{d^{-1}}a\mathcal{T}_d=a_{-1}(d)\otimes a_0(d)$. Hence, $$N_{W,W_\theta}(a)=\sum_{d\in\mathcal{D}_{\theta\rho}}N_{W_{\rho_{-1}},W_{\alpha_{(-1)}}}(a_{-1}(d))\otimes N_{W_{\rho_{0}},W_{\alpha_{(0)}}}(a_{0}(d)).$$ If $P_{k-r_0}<_WP_{\alpha_{(-1)}}$, then $P_{\alpha_{(0)}}\leq_WP_{r_0-1}$ and so $\mathcal{F}^*(N_{W_{\rho_{0}},W_{\alpha_{(0)}}}(a_{0}(d)))=0$. Hence, $$\phi_\vecr(N_{W,W_\theta}(a))=\sum_{{d\in\mathcal{D}_{\theta\rho}}\atop{ P_{k-r_0}=_WP_{\alpha_{(-1)}(d)}}}N_{W_{\rho_{-1}},W_{\alpha_{(-1)}}}(a_{-1}(d))\otimes \mathcal{F}^*(N_{W_{\rho_{0}},W_{\alpha_{(0)}}}(a_{0}(d))),$$ which is in $ I_F(P_{k-r_0},r_{-1})\otimes S(m,n)_{r_0}$. We now prove the surjectivity. By the proof of Lemma \[image\], every basis element $x_{\mu_{(-1)},\la_{(-1)}d_{-1}}^*\otimes t_{\mu_{(0)},\la_{(0)}d_0}^*$ in $I_F(P_{k-r_0},r_{-1})\otimes S(m,n)_{r_0}$ has a pre-image $x_{\mu,\la d}^*$. Let $P_{\la d\cap\mu}$ be the maximal $l$-parabolic subgroup of $W_\la^d\cap W_\mu$. Then the same proof shows that $P_{\la d\cap\mu}=_WP_{k-r_0}\times P_{r_0}$ where $P_{k-r_0}$ is conjugate to the maximal $l$-parabolic subgroup of $W_{\la_{(-1)}}^{d_{-1}}\cap W_{\mu_{(-1)}}\leq \fS_{r_{-1}}$. Hence, $x_{\mu,\la d}^*\in I_F(P_k,r),$ proving the surjectivity. \[RBH\] Maintain the nation above. For $\vecr\in \mathcal R_r,$ let $\rho=\rho_{\vecr},\theta=\theta_{\vecr}$ be defined in and let $k\geq r_0$ with $kl\leq r$. Then $$\phi_{\vecr}(I_F(P_k,r))= \psi_\rho(I_F(P_k,\rho)).$$ If $r=sl+t$ with $0\leq t< l$, $r_0\leq k\leq s$, then one sees easily that $$I_F(P_k,\rho)=I_F(P_{k-r_0}, r_{-1})\otimes\sS_F(m|n,r_0l)+ \sum_{i=1}^{s-k}I_F(P_{k-r_0+i}, r_{-1})\otimes I_F(P_{r_0-i},r_0l).$$ Hence, $\psi_\rho(I_F(P_k,\rho))= \psi_\rho(I_F(P_{k-r_0}, r_{-1})\otimes\sS_F(m|n,r_0l))$ and, by Lemma \[new\], $$\phi_\vecr(I_F(P_k,r))=I_F(P_{k-r_0})\otimes S(m,n)_{r_0}=\psi_\rho(I_F(P_{k-r_0}, r_{-1})\otimes\sS_F(m|n,r_0l)).$$ Hence, the assertion follows. Applying the Brauer homomorphism $\phi_\vecr$ to the filtration gives rise to a filtration of ideals of $\sS_F(m|n,\vecr)$: $$\label{filtration1} 0\subseteq I_F(P_0,\vecr)\subseteq I_F(P_1,\vecr)\subseteq\cdots\subseteq I_F(P_{s-r_0},\vecr)=\sS_F(m|n,\vecr).$$ Like Definition \[defect group\], we may use this sequence to define the [*defect group*]{} $D(\bar e)$ of a primitive idempotent $\bar e$ of $\sS_F(m|n,\vecr)$. In particular, a primitive idempotent $\bar{e}\in\sS_F(m|n,\vecr)$ has the trivial defect group if $\bar{e}\in I_F(P_0,\vecr)$. \[ML3\] Let $e,e'$ be idempotents $\sS_F(m|n,r)$. Suppose $e$ is primitive with defect group $D(e)=P_k$. Suppose $\vecr=(r_{-1},r_0)\in\mathcal{R}_r$ and $P_\vecr\leq_WD(e)$ (so $r_0\leq k$). Then $$\phi_{\vecr}(e\sS_F(m|n,r)e')=\phi_{\vecr}(e)\sS_F(m|n,\vecr)\phi_{\vecr}(e')$$ and $\phi_\vecr(e)$ is also primitive with defect group $P_{k-r_0}$. In particular, if $D(e)=P_\vecr$, then $\phi_\vecr(e)$ has the trivial defect group. The first assertion is clear since $\phi_\vecr$ is an algebra homomorphism and $\sS_F(m|n,\vecr)$ is the homomorphic image of $\sS_F(m|n,r)$ under $\phi_{\vecr}$. To see the last assertion, we first notice that $\phi_\vecr(e)\neq0$ by Corollary \[BHP\]. Since $e$ is primitive, $e\sS_F(m|n,r)e$ is a local ring. Applying the surjective map $\phi_\vecr$ to this local ring yields that $\phi_{\vecr}(e\sS_F(m|n,r)e)=\phi_{\vecr}(e)\sS_F(m|n,\vecr)\phi_{\vecr}(e)$ is also a local ring. Hence, $\phi_{\vecr}(e)$ is a primitive idempotent in $\sS_F(m|n,\vecr)$ and $\phi_\vecr(e)\in I_F(P_{k-r_0},\vecr)$. The fact $D(\phi_\vecr(e))=P_{k-r_0}$ is clear from Isomorphism Theorems for ring homomorphisms, since $\text{ker}(\phi_\vecr)=I_F(P_{r_0-1},r)\subseteq I_F(P_{k},r)$ for all $k\geq r_0$. Classification of irreducible $\sS_F(m|n,r)$-modules ===================================================== We first interpret the algebra $\sS_F(m|n,\vecr)$ as an endomorphism algebra of a certain tensor space. By and the commutative diagram , we obtain for $\vecr=(r_{-1},r_0)$ and $\rho=(r_{-1},r_0l)$ $$\sS_F(m|n,\vecr)=\psi_\rho(\sS_F(m|n,\rho))\cong\sS_F(m|n,r_{-1})\otimes \bar\sS_F(m|n,r_0l),$$ where $\bar\sS_F(m|n,r_0l):=\sS_F(m|n,(0,r_0))=\sS_F(m|n,r_0l)/I_F(P_{r_0-1},r_0l).$ On the other hand, taking the dual of the Frobenius map $$\mathcal{F}|_{r_0}:A(m,n)_{r_0}\to\sA_q(m|n,r_0l)$$ induces an algebra isomorphism $$\label{homog} \bar\sS_F(m|n,r_0l)\cong\bigoplus_{i=0}^{r_0}(S(m,i)\otimes S(n,r_0-i))=S(m,n)_{r_0}.$$ Recall the even part $V_F(m|n)_0$ and the odd part $V_F(m|n)_1$ of the superspace $V_F(m|n)$. Fix $r_0\geq 0$ and define the subspace: $$(V_F(m|n)^{\otimes r_0})_\rmL=\bigoplus_{i=0}^{r_0} (V_F(m|n)_0)^{\otimes i}\otimes (V_F(m|n)_1)^{\otimes r_0-i}.$$ This is called the “Levi part” of $V_F(m|n)^{\otimes r_0}$. Let the product of symmetric groups $\fS_{\{1,\ldots,i\}}\times\fS_{\{i+1,\ldots,r_0\}}$ act on the summand $(V_F(m|n)_0)^{\otimes i}\otimes (V_F(m|n)_1)^{\otimes r_0-i}$ by place permutation for $\fS_{\{1,\ldots,i\}}$ and signed place permutation[^4] for $\fS_{\{i+1,\ldots,r_0\}}$. In this way, putting $$(\fS_{r_0})_\rmL:=\prod_{i=0}^{r_0} (\fS_{\{1,\ldots,i\}}\times\fS_{\{i+1,\ldots,r_0\}}),$$ $(V_F(m|n)^{\otimes r_0})_\rmL$ becomes an $(\fS_{r_0})_\rmL$-module. Now, the RHS of can be interpreted as the endomorphism algebra $\End_{(\fS_{r_0})_\rmL}((V_F(m|n)^{\otimes r_0})_\rmL)$ so that we have an algebra isomorphism $$\label{Brauer iso} \bar\sS_F(m|n,r_0l)\cong\End_{(\fS_{r_0})_\rmL}((V_F(m|n)^{\otimes r_0})_\rmL).$$ In general, for $\vecr=(r_{-1},r_0)\in \mathcal{R}_r$, let $$\aligned V_F(m|n)^{\boxtimes \vecr}&:=V_F(m|n)^{\otimes r_{-1}}\otimes (V_F(m|n)^{\otimes r_0})_L\\ &= \bigoplus_{i=0}^{r_0}V_F(m|n)^{\otimes r_{-1}}\otimes (V_F(m|n)_0)^{\otimes i}\otimes (V_F(m|n)_1)^{\otimes r_0-i}.\endaligned$$ Let $F(\fS_{r_0})_\rmL$ be the group algebra of $(\fS_{r_0})_\rmL$ and let $$\aligned \mathcal{H}_\vecr&:=\mathcal{H}_F(\fS_{r_{-1}})\otimes F(\fS_{r_0})_\rmL\\ &\cong \bigoplus_{i=0}^{r_0}\mathcal{H}_F(\fS_{r_{-1}})\otimes F \fS_{\{1,\ldots,i\}}\otimes F\fS_{\{i+1,\ldots,r_0\}}.\endaligned$$ Then $$\aligned \sS_F(m|n,\vecr)&=\bigoplus_{i=0}^{r_0}\sS_F(m|n,r_{-1})\otimes S(m,i)\otimes S(n,r_0-i)\\ &\cong \End_{\mathcal{H}_{\vecr}}(V_F(m|n)^{\boxtimes\vecr}). \endaligned$$ If $\bar{e}\in \sS_F(m|n,\vecr) $ is a primitive idempotent, then there are primitive idempotents $e_{-1}\in \sS_F(m|n,r_{-1})$, $e_0\otimes e_1\in S(m,i)\otimes S(n,r_0-i)$ for some $i$ such that $\bar e=e_{-1}\otimes e_0\otimes e_1$. By Lemma \[new\], $D(\bar{e})=D(e_{-1})$. We need the following result. \[MT\] (a) Let $e,e'$ be the idempotents of $\sS_F(m|n,r)$ with $e$ primitive and $D(e)=P_{\vecr}$ for some $\vecr=(r_{-1},r_0)\in\mathcal{R}_r$. Then $$V_F(m|n)^{\otimes r}e\mid V_F(m|n)^{\otimes r}e'\mbox{ if and only if } V_F(m|n)^{\boxtimes\vecr}\phi_{\vecr}(e)\mid V_F(m|n)^{\boxtimes\vecr}\phi_{\vecr}(e').$$ \(b) If $\bar{e}\in \sS_F(m|n,\vecr)$ is a primitive idempotent with $D(\bar{e})=1$, then $\bar{e}$ is equivalent to $\phi_{\vecr}(e')$ for some primitive idempotent $e'\in \sS_F(m|n,r)$ with $D(e')=P_{\vecr}$. \(c) Let $e$ be a primitive idempotent of $\sS_F(m|n,r)$. If $D(e)=1$, then $V_F(m|n)^{\otimes r}e$ is a projective indecomposable $\mathcal{H}_F$-module. With Corollary \[BHP\], Theorem \[ML3\], and Lemma \[new\], the proof for (a) and (b) is standard; see [@S 4.1]. So we omit it. \(c) By Theorem \[VAP\], the vertex of $V_F(m|n)^{\otimes r}e$ is $D(e)$. If $\mathcal{H}_F$-module $M$ has the trivial vertex, then Higman Criterion in Lemma \[Higman\](a) tells that $M$ is a projective $\mathcal{H}_F$-module. Let $\Lambda^+(r)$ (resp. $\Lambda^+(N,r)$, $\La_{l\reg}^+(r)$) be the set of partitions (resp., partitions with at most $N$ parts, $l$-regular partitions) of $r$. (A partition is called $l$-regular if no part is repeated $l$ or more times.) For each $\vecr=(r_{-1},r_0)\in\sR_r$, let $$\sP_\vecr=\begin{cases}\La_{l\reg}^+(r), &\text{ if } r_0=0;\\ \bigcup_{i=0}^{r_0}\{(\lambda,\xi,\eta)\mid \lambda\in \La_{l\reg}^+(r_{-1}),\xi\in \Lambda^+(m,i),\eta\in \Lambda^+(n,r_0-i)\},& \text{ if }r_0>0\\ \end{cases}$$ and let $$\label{index set} \sP_r=\bigcup_{\vecr\in\sR_r}\sP_\vecr.$$ Let $E_r^\vecr$ (resp., $E_\vecr^1$) be the set of nonequivalent primitive idempotents with defect group $P_\vecr$ (resp., the trivial defect group) in $\sS_F(m|n,r)$ (resp., $\sS_F(m|n,\vecr)$). Set $$E_r=\bigcup_{\vecr\in\sR_r}E_r^\vecr$$ to be the set of nonequivalent primitive idempotents of $\sS_F(m|n,r)$. \[ML4\] Assume $m+n\geq r$. 1. For each $\vecr\in\mathcal R_r$, the Brauer homomorphim $\phi_\vecr$ induces a bijection between the sets $E_r^\vecr$ and $E_\vecr^1$. 2. There is a bijection $\pi:E_\vecr^1\to\sP_\vecr$. Hence, there is a bijective map $\pi$ from $E_r$ to $\mathcal{P}_r$. For statement (1), let $\bar E_r^\vecr=\phi(E_r^\vecr)$ where $\phi=\phi_\vecr$. By Theorem \[ML3\], every element in $\bar E_r^\vecr$ is primitive with the trivial defect group. Thus, we may regard $\bar E_r^\vecr$ as a subset of $E_\vecr^1$ and consider the map $\phi:E_r^\vecr\to E_\vecr^1$. By Lemma \[MT\](a) we see that $\phi$ is injective, and by Lemma \[MT\](b), we see that $\phi$ is surjective, proving (1). We now prove (2). Pick $\bar e\in E_\vecr^1$. Then $D(\bar e)=1$ and there exists $i\in[1,r_0]$ and primitive idempotents $e_{-1}\in\sS_F(m|n,r_{-1})$, $e_0\in S(m,i)$, and $e_1\in S(n,r_0-i)$such that $\bar e=e_{-1}\otimes e_0\otimes e_1$. Since $D(e_{-1})=D(\bar e)=1$, by Lemma \[MT\](c), $V_F(m|n)^{r_{-1}}e_{-1}$ is a projective indecomposable $\mathcal{H}_{r_{-1}}$-module, where $\mathcal{H}_{r_{-1}}=\sH_F(\fS_{r_{-1}})$. Since the PIMs of $\sH_{r_{-1}}$ are labelled by $\La_{l\reg}^+(r_{-1})$ (see [@DJ]), $e_{-1}$ determines a unique $\la\in\La_{l\reg}^+(r_{-1})$. Similarly, idempotents $e_0$ and $e_1$ determines irreducible $S(m,i)$-module $L(\xi)$ and $S(n,r_0-i)$-module $L(\eta)$, respectively, where $\xi\in\Lambda^+(m,i)$ and $\eta\in\Lambda^+(n,r_0-i)$. Hence, $\bar e$ determines a unique triple $(\la,\xi,\eta)\in\sP_\vecr$ and, putting $\pi(\bar e)=(\la,\xi,\eta)$ defines a map $\pi:E_\vecr^1\to\sP_\vecr$. It is clear that $\pi$ is injective. For the subjectivity, we assume that $m+n\geq r$. In particular, $m+n\geq r_{-1}$. Thus, $\sH_F(\fS_{r_{-1}})$ is a direct summand of the tensor space $V_F(m|n)^{\otimes r_{-1}}$. Thus, for every triple $(\la,\xi,\eta)\in\sP_\vecr$, the PIM $Q_\la$ of $\sH_{r_{-1}}$ corresponding to $\la$ is a direct summand of the tensor space. Hence, there is a primitive idempotent $e_{-1}$ such that $Q_\la\cong V_F(m|n)^{\otimes r_{-1}}e_{-1}$. Similar, there exists primitive idempotents $e_0,e_1$ such that $$(V_F(m|n)_0)^{\otimes i}e_0\cong L(\xi)\text{ and } (V_F(m|n)_1)^{\otimes r_0-i} e_1\cong L(\eta),$$ where $L(\xi)$ (reps., $L(\eta)$) is the irreducible $S(m,i)$-module (reps. $S(n,r_0-i)$-module) with highest weight $\xi$ (reps., $\eta$). Hence, $\pi(e_{-1}\otimes e_0\otimes e_1)=(\la,\xi,\eta)$ and $\pi$ is surjective. For each $e\in E_r$, if $\pi(e)=(\la,\xi,\eta)$, define $$L_q(\la,\xi,\eta)=\sS_F(m|n,r)e/\text{Rad}(\sS_F(m|n,r)e).$$ When $\pi(e)=\la\in\La_{l\reg}^+(r)$, we write $L_q(\la)$. \[tensor product theorem\] Assume $m+n\geq r$. 1. The set $\{L(\lambda,\xi,\eta)\mid (\lambda,\xi,\eta)\in \mathcal{P}_r\}$ forms a complete set of non-isomorphic irreducible $\sS_F(m|n,r)$-modules. 2. If $(\lambda,\xi,\eta)\in \mathcal{P}_\vecr$ with $P_\vecr>1$, then the $\sS_F(m|n,\vecr)$-module $L_q(\la)\otimes L(\xi)\otimes L(\eta)$ becomes an $\sS_F(m|n,r)$-module by inflation and we have $$L_q(\la,\xi,\eta)\cong L_q(\la)\otimes L(\xi)\otimes L(\eta).$$ \[Alperin\] (1) Each element of $\mathcal{P}_r$ can also be regarded as a pair $(P, Q)$ where $P$ is an $l$-parabolic subgroup and $Q$ is a PIM of the quotient algebra defined by $P$. Equivalence classes associated with an equivalence relation on all such pairs play the role of “weights” as described in Alperin’s weight conjecture. \(2) If we know the classification of all primitive idempotents with the trivial defect group in $\sS_F(m|n,r)$ and label them by $\La_r$, then the condition $m+n\geq r$ can be removed by replacing the set $\La_{l\reg}^+(r_{-1})$ by the labelling set $\La_{r_{-1}}$. Moreover, when $m+n<r$, the classification can possibly be obtained by the Schur functor: $$\mathscr{S}_\vep:\sS_F(m'|n',r)\mathsf{-Mod}\longrightarrow \sS_F(m|n,r)\mathsf{-Mod}, M\longmapsto \vep M,$$ where $m'\geq m$, $n'\geq n$, $m'+n'\geq r$, and $\vep$ is given in Remark \[m’|n’ case\]. See [@BK 6.5] for the Schur superalgebra case and Appendix II for a comparison. \(3) Though our approach didn’t offer a construction for irreducible modules, the tensor product structure shown in Corollary \[tensor product theorem\](2) reduces the problem to the construction for $l$-restricted partitions. \(4) If $r<l$, then every idempotent $e\in\sS_F(m|n,r)$ has the trivial defect group and defines a PIM of the Hecke algebra $\sH_F$. Thus, there is a bijection between the PIMs of $\sH_F$ and the non-equivalent primitive idempotents in $\sS_F(m|n,r)$ where $m+n\geq r$. This is the semismple case similar to the situation described in Proposition \[DR8.1\]. In the classical (i.e., non-quantum) case, Donkin gave a classification of irreducible $S_k(m|n,r)$-modules under the assumption $m,n\geq r$, where $k$ is a field of a positive characteristic $p$. We end this section with a comparison. Assume now $m,n\geq r$. Let $S_k(m|n,r)=\sS(m|n,r)\otimes_\sZ k$ be the Schur superalgebra obtained by specialising $\up$ to 1. By [@D (3)], non-isomorphic irreducible $S_k(m|n,r)$-modules are in one-to-one correspondence with the set $$\{(\la,\mu)\in\La^+\times\La^+\mid r=|\la|+p|\mu|\},$$ where $\La^+$ denotes the set of all partitions. Clearly, this set is the same as the set $$\mathcal{P}'_r=\{(\tau,\nu)\mid \tau\in\Lambda^+(s),\nu\in\Lambda^+(t), s,t\in\mathbb N, s+lt=r\}.$$ With the assumption $m,n\geq r$ and taking $l=p$, the set $\sP_r$ defined in has the form $$\aligned \sP_r&=\{(\la,\xi,\eta)\mid \la\in \La_{l\reg}^+(r_{-1}),\xi\in \La^+(i),\eta\in \Lambda^+(r_0-i),i\in[0,r_0],(r_{-1},r_0)\in\sR_r\}\\ &=\{(\la,\xi,\eta)\mid \la\in \La_{l\reg}^+(i),\xi\in \La^+(j),\eta\in \Lambda^+(k),i, j,k\in\mathbb{N},i+l(j+k)=r\} \endaligned$$ We claim that there is a bijective map $g:\mathcal{P}_r\rightarrow \mathcal{P}'_r$ defined by setting $g((\lambda,\xi,\eta))=(\lambda^\iota+l\xi^\iota,\eta)$ for $(\lambda,\xi,\eta)\in\mathcal{P}_r$, where $\mu^\iota$ is the partition dual to $\mu$. Indeed, for $(\lambda,\xi,\eta)\in\mathcal{P}_r$, since $\lambda\in\Lambda^+_l{(i)},\xi\in\Lambda^+(j),\eta\in\Lambda^+(k)$, then $\lambda^\iota+l\xi^\iota\in\Lambda^+(i+lj)$. Thus, $(\lambda^\iota+l\xi^\iota,\eta)\in\sP'_r$. So $g$ is a map from $\mathcal{P}_r$ to $\mathcal{P}'_r$. For $(\mu,\alpha,\beta)\in\mathcal{P}_r$, if $(\mu,\alpha,\beta)\neq (\lambda,\xi,\eta)$, it is clear, by the fact that $\la$ and $\mu$ are $l$-regular partitions, that $(\mu^\iota+l\alpha^\iota,\beta)\neq (\lambda^\iota+l\xi^\iota,\eta)$. Hence $g$ is injective. For $(\tau,\nu)\in\mathcal{P}'_r$, suppose $\tau=(\tau_1,\tau_2,\cdots,\tau_m)$. Assume $\tau_j=\lambda_j+l\mu_j$ where $0\leq\lambda_j<l$, $j=1,\cdots,m$. Set $\lambda^\iota=(\lambda_1,\lambda_2,\cdots,\lambda_m),\mu^\iota=(\mu_1,\mu_2,\cdots,\mu_m)$. Then $(\lambda,\mu,\nu)\in\mathcal{P}_r$ and $g(\lambda,\mu,\nu)=(\tau,\nu)$. We have proved the claim. By the claim, we see that under the assumption $m,n\geq r$, our labelling set is the same as Donkin’s labelling set. Thus, this shows that the quantum classification in the $m,n\geq r$ case is a $q$-analogue of the classical classification. Appendix I: Brauer homomorphisms without Frobenius ================================================== In §11, the isomorphism was established by taking the dual of the Frobenius morphism. However, the Brauer homomorphism, originated from the group representation theory, has its own definition, see [@DU3 §3]. In this section, we provide a proof for without using the Frobenius morphism. Note that this proof is much simpler than the proof for $q$-Schur algebras given in [@DU3]. We first look at the structure of $\bar\sS_F(m|n,r_0l)$. The algebra $\bar\sS_F(m|n,r_0l)$ has a direct sum decomposition into centraliser subalgebras $$\bar\sS_F(m|n,r_0l)=\bigoplus_{i=0}^r\bar\sS_F(m|n,r_0l)_i.$$ Let $W=\fS_{r_0l}$ (i.e., we assume $r=r_0l$). The maximal $l$-parabolic subgroup of $W$ is unique and equal to $P_{r_0}$. By Theorem \[RNB\], $\sS_F(m|n,r_0l)$ has basis $$\mathcal{B}=\{N_{W,W^d_\lambda\cap W_\mu}(e_{\mu,\lambda d})\mid \mu,\lambda\in\Lambda(m|n,r_0l),d\in\mathcal{D}^\circ_{\lambda\mu}\}.$$ If the maximal $l$-parabolic subgroup $P_{\lambda d\cap \mu}$ of $W^d_\lambda\cap W_\mu$ is not equal to $P_{r_0}$, then it must be conjugate to a parabolic subgroup of $P_{r_0-1}$ and so $N_{W,W^d_\lambda\cap W_\mu}(e_{\mu,\lambda d})\in I_F(P_{r_0-1},r_0l)$. Thus, it suffices to look at the basis elements satisfying $P_{\lambda d\cap \mu}=P_{r_0}$. Then $P_{r_0}\leq W_\la ^d\cap W_\mu$. Since $P_{r_0}$ is the [*unique*]{} maximal $l$-parabolic subgroup of $W=\fS_{r_0l}$, $\la$ and $\mu$ must be of the form $$\aligned \lambda&=l\la'=(\lambda'_1l,\lambda'_2l,\cdots,\lambda'_ml\mid\lambda'_{m+1}l,\cdots,\lambda'_{m+n}l)\\\mu&=l\mu'=(\mu'_1l,\mu'_2l,\cdots,\mu'_ml\mid\mu'_{m+1}l,\cdots,\mu'_{m+n}l),\endaligned$$ where $\la',\mu'\in\Lambda(m|n,r_0)$ and $d\in N_{\fS_{r_0l}}(P_{r_0})$, the normaliser of $P_{r_0}$. Moreover, $N_{\fS_{r_0l}}(P_{r_0})=P_{r_0}\rtimes\widetilde{\fS}_{r_0}$, where $\widetilde{\fS}_{r_0}\cong {\fS}_{r_0}$ is the subgroup of $\fS_{r_0l}$ generated by $$\aligned \tilde s_1&=(1,l+1)(2,l+2)\cdots(l,2l), \\ \tilde s_2&=(l+1,2l+1)(l+2,2l+2)\cdots(2l,3l),\\ &\ldots\ldots,\\ \tilde s_{r_0-1}&=((r_0-2)l+1,(r_0-1)l+1)((r_0-2)l+2,(r_0-1)l+2)\cdots((r_0-1)l,r_0l).\endaligned$$ Thus, $d\in\mathcal{D}^\circ_{\lambda\mu}\cap \widetilde{\fS}_{r_0}$. This together with the even-odd trivial intersection property forces by the uniqueness of $P_{r_0}$ $$P_{r_0}=P^{00}_{\lambda d\cap \mu}\times P^{11}_{\lambda d\cap \mu}=P_{\mu^{(0)}}\times P_{\mu^{(1)}}=P_{\la^{(0)}}\times P_{\la^{(1)}}.$$ (Here $P^{ii}_{\lambda d\cap \mu}$ denotes the maximal $l$-parabolic subgroup of $W_{\la d\cap\mu}^{ii}$.) Hence, we must have $P_{\lambda^{(0)}}=P_{\mu^{(0)}}$ and $P_{\lambda^{(1)}}=P_{\mu^{(1)}}$, and consequently, $$|\lambda^{(0)}|=|\mu^{(0)}|=il\text{ and } |\lambda^{(1)}|=|\mu^{(1)}|=(r_0-i)l$$ for some $0\leq i\leq r_0$. This also forces that $d$ must have a decomposition $d=d_0d_1$ for some $d_0\in \mathcal{D}_{\lambda^{(0)}\mu^{(0)}}\cap W'$ and $ d_1\in \mathcal{D}_{\lambda^{(1)}\mu^{(1)}}\cap W''$, where $$W'=\widetilde{\fS}_{\{1,\ldots,i\}}\text{ and }W''=\widetilde{\fS}_{\{i+1,\ldots,r_0\}},$$ such that $W_{\la d\cap\mu}^{ii}=W_{\la^{(i)}}^{d_i}\cap W_{\mu^{(i)}}$ for $i=0,1$. Let $\zeta_{\mu,\la}^{d}$ be the image of $N_{W,W_\la^d\cap W_\mu}(e_{\mu,\la d})$ in $\bar\sS_F(m|n,r_0l)$. Then $\bar\sS_F(m|n,r_0l)$ has basis $\bar\sB=\cup_{i=0}^{r_0}\bar\sB_i$, where $$\bar\sB_i=\{\zeta_{\mu\la}^{d}\mid\lambda,\mu\in l\Lambda(m,i)\times l\Lambda(n,r_0-i),d\in \mathcal{D}^\circ_{\lambda\mu}\cap\widetilde\fS_{r_0}\}.$$ Let $\bar\sS_F(m|n,r_0l)_i$ be the subspace spanned by $\bar\sB_i$ and let $e_i=\sum_{\la:|\la^{(0)}|=il}\zeta_{\la\la}^1$. Then $1=\sum_{i=0}^{r_0}e_i$, $e_ie_j=0$ for $i\neq j$, and $\bar\sS_F(m|n,r_0l)_i=e_i \bar\sS_F(m|n,r_0l)e_i$. Our assertion follows. Let $\sT_\rmL$ be the subspace of $\sT=V_F(m|n)^{\otimes r}$ spanned by tensors of the form $$v_{j_1}^l\cdots v_{j_i}^lv_{j_{i+1}}^l\cdots v_{j_{r_0}}^l\quad(0\leq i\leq r_0)$$ such that $1\leq j_1,\ldots,j_{i}\leq m$ and $m+1\leq j_{i+1},\ldots,j_{r_0}\leq m+n$ and let $\sT_\rmL'$ be the subspace spanned by the rest of the tensors so that $\sT=\sT_\rmL\oplus\sT_\rmL'$. By looking at the action of elements in $\bar\sB$ on the quotient space $\sT/\sT_\rmL'$, we directly prove the following. We have algebra isomorphisms: $$\bar\sS_F(m|n,r_0l)\cong \End_{(\widetilde{\fS}_{r_0})_\rmL}(\sT_\rmL)\cong \End_{({\fS}_{r_0})_\rmL}(V_F(m|n)^{\otimes r_0})_\rmL).$$ The second isomorphism is clear. Suppose $N_{W, W_\nu}(e_{\mu,\lambda d})\in{\mathcal{B}}$ has image $\zeta_{\mu\la}^{d}\in\bar\sB$ ($\nu={\lambda d\cap \mu}$). We first inspect the action of $N_{W, W_\nu}(e_{\mu,\lambda d})$ on $v_\mu$: $$\begin{aligned} (v_\mu)N_{W, W_{\nu}}(e_{\mu,\lambda d}) &=(v_\mu)\sum_{x\in\mathcal{D}_{\nu}\cap W_\mu}\mathcal{T}_{x^{-1}} (e_{\mu,\lambda d})\mathcal{T}_{x}\\ &=\sum_{x_0x_1\in\mathcal{D}_{\nu}\cap (W_{\mu^{(0)}}\times W_{\mu^{(1)}})}q^{\ell(x_0)}(-q^{-1})^{\ell(x_1)}(-1)^{\widehat{dx_0x_1}-\widehat{d}}v_{\bsi_\lambda dx_0x_1} \end{aligned}$$ Since $d=d_0d_1$ with $d_0\in \mathcal{D}_{\lambda^{(0)}\mu^{(0)}}\cap W'$ and $ d_1\in \mathcal{D}_{\lambda^{(1)}\mu^{(1)}}\cap W''$, it follows that $(-1)^{\widehat{d}}=(-1)^{\ell(d_1)}$, $(-1)^{\widehat{dx_0x_1}}=(-1)^{\ell(d_1x_1)}=(-1)^{\ell(d_1)}(-1)^{\ell(x_1)}$, and $v_{\bsi_\lambda d}=v_{\bsi_{\lambda^{(0)}}d_0} v_ {\bsi_{\lambda^{(1)}}d_1}$. Also, since, for any $x_0x_1\in \mathcal{D}_{\nu}\cap W_{\mu}$ but $x_0x_1\not\in W'\times W''$, $v_{\bsi_{\lambda dx_0x_1}}\in\sT_\rmL'$ and since $l^2|\ell(x_0)$ and $l^2|\ell(x_1)$ for all $x_0\in W'$ and $x_1\in W''$, we have $$\aligned (v_\mu)N_{W, W_{\nu}}(e_{\mu,\lambda d})&\equiv\sum_{x_0x_1\in\mathcal{D}_{\nu}\cap W_{\mu}\cap (W'\times W'')}q^{\ell(x_0)}(q^{-1})^{\ell(x_1)}v_{\bsi_{\lambda ^{(0)}}d_0x_0} v_{\bsi_{\lambda^{(1)}}d_1x_1}(\text{mod }\sT_{\rmL'})\\ &=\sum_{x_0x_1\in\mathcal{D}_{\nu}\cap W_{\mu}\cap (W'\times W'')}v_{\bsi_{\lambda ^{(0)}}d_0x_0} v_{\bsi_{\lambda^{(1)}}d_1x_1}(\text{mod }\sT_{\rmL'}).\endaligned$$ This formula allows us to define an action of $\zeta_{\mu\la}^{d}$ on $\sT/\sT_\rmL'$ by setting, for any $\nu=l\nu'$ and $y\in\sD_\nu\cap\widetilde{\fS}_{r_0}$, $$\bar v_{\bsi_\nu y}\cdot \zeta_{\mu\la}^d=\delta_{\nu,\mu} \sum_{x_0x_1\in\mathcal{D}_{\nu}\cap W_{\mu}\cap (W'\times W'')}\bar v_{\bsi_{\lambda ^{(0)}}d_0x_0} \bar v_{\bsi_{\lambda^{(1)}}d_1x_1}y=\delta_{\nu,\mu}(\bar v_\mu)\xi_{\mu\la}^d y,$$ where $\bar v_\bsi=v_\bsi+\sT_\rmL'$ and $\xi_{\mu\la}^d=\sum_{x\in\mathcal{D}_{\nu}\cap W_{\mu}\cap (W'\times W'')} x^{-1}e_{\mu,\la d}x\in\End_{(\widetilde{\fS}_{r_0})_\rmL}(\sT_\rmL)$. Finally, it is easy to check that the linear isomorphism $$h:\bar\sS_F(m|n,r_0l)\longrightarrow \End_{(\widetilde{\fS}_{r_0})_\rmL}(\sT_\rmL),\; \zeta_{\mu\la}^d\longmapsto \xi_{\mu\la}^d$$ is an algebra homomorphism. Appendix II: a comparison with a result of Brundan–Kujawa ========================================================= In [@BK], Brundan–Kujawa obtained a complete classification of irreducible modules of Schur superalgebras for all $m,n,r$. Their approach is quite different. First, they used a construction of irreducible modules via Verma modules over the super hyperalgebra $U_k$ of $\mathfrak{gl}(m|n)$. Second, they used an isomorphism of the Hopf superalgebra $U_k$ and the distribution superalgebra Dist$(G)$ of the supergroup $G=GL(m|n)$ [@BK 3.2] and a category equivalence between $G$-supermodules and integrable $U_k$-supermodules [@BK 3.5] to obtain a complete set of irreducible $G$-supermodules. Finally, by determining the polynomial representations of $G$ and applying the Schur functor to Donkin’s result, they reached an explicit description of the index set $\La^{++}(m|n,r)$. We now show that, if $m+n\geq r$ and $l=p$, then there is a bijection from $\mathcal P_r$ to $\La^{++}(m|n,r)$. There are a quite few ingredients in order to describe the set $\La^{++}(m|n,r)$. Following [@BK], we identify $\lambda\in\La^+(r)$ with its Young diagram $$\lambda=\{(i,j)\in\mathbb{Z}_{>0}\times\mathbb{Z}_{>0} \mid j\leq \lambda_i\}$$ and refer to $(i,j)\in\lambda$ as the node in the $i$th row and $j$th column. Then the rim of $\lambda$ is defined to be the set of all nodes $(i,j)\in\lambda$ such that $(i+1,j+1)\notin\lambda$. The $p$-rim is a certain subset of the rim, defined as the union of the $p$-segments. The first $p$-segment is simply the first $p$ nodes of the rim, reading along the rim from left to right. The next $p$-segment is then obtained by reading off the next $p$ nodes of the rim, but starting from the column immediately to the right of the rightmost node of the first $p$-segment. The remaining $p$-segments are obtained by repeating this process. Of course, all but the last $p$-segment contain exactly $p$ nodes, while the last may contain less. Let $J(\lambda)$ be the partition obtained from $\lambda$ by deleting every node in the $p$-rim that is at the rightmost end of a row of $\lambda$ but that is not the $p$th node of a $p$-segment. Let $ j(\lambda) = |\lambda|-|J(\lambda)|$ be the total number of nodes deleted. Clearly, for partition $\mu,\nu$, $j(\mu + p\nu) = j(\mu).$ Let $$\Lambda^+(m|n,r)=\{\lambda\in \Lambda(m|n, r):\lambda_1\geq\cdots\geq \lambda_m,\lambda_{m+1}\geq\cdots\lambda_{m+n}\}.$$ For $\la\in\La^+(m|n,r)$, let $t(\la)=(\la_{m+1},\ldots,\la_{m+n})$. Note that, if we use the notation $\la= (\lambda^{(0)}\mid\lambda^{(1)})$, then $t(\la) =\la^{(1)}$. We can now describe the index set $\La^{++}(m|n,r)$ for irreducible modules of the Schur superalgebra: $$\label{BK,6.5} \Lambda^{++}(m|n,r)=\{\lambda\in\Lambda^+(m|n,r)\mid j(t(\lambda))\leq \lambda_m\}.$$ The functions $J$ and $j$ are also used to characterise the Mullineux conjugation. A partition $\lambda=(\lambda_1,\lambda_2,\cdots)$ is called restricted if $\lambda_i-\lambda_{i+1}<p$ for $i=1,2,\cdots$. Let $\mathcal{RP}(r)$ denote the set of all restricted partitions of $r$. The Mullineux conjugation is the bijective function ([@BK 6.1],[@Mu 4.1]): $$\label{BK,6.1} \ttM : \mathcal{RP}(r)\rightarrow \mathcal{RP}(r),\quad \lambda\longmapsto\ttM(\lambda)=(j(\la),j(J(\la)),j(J^2(\la)),\ldots).$$ We need another function. Let $G=GL(m|n)$ and let $T$ be the usual maximal torus of $G_{\text{ev}}:=GL(m)\times GL(n)$. The character group $X(T)=\Hom(T,G_m)$ is the free abelian group on generators $\varepsilon_1,\cdots,\varepsilon_m,\varepsilon_{m+1},\cdots,\varepsilon_{m+n}$. For $\lambda\in\Lambda(m+n):=\mathbb N^{m+n}$, we will identify $\lambda$ with $\sum_{i=1}^{m+n}\lambda_i\varepsilon_i\in X(T)$. Denote by $\mathcal{D}_{m,n}$ the set of representatives of the right coset of $\mathfrak{S}_m\times\mathfrak{S}_n$ in $\mathfrak{S}_{m+n}$. Let $B_1$ denote the standard Borel subgroup which is the stabiliser of the full flag associated with a fixed ordered basis $(v_1,\ldots,v_{m+n})$. For $w\in\sD_{m,n}$, let $B_w$ be the stabiliser of the full flag associated with $(v_{w(1)},\ldots,v_{w(m+n)})$. With $B_w$, define irreducible module $L_w(\la)$ for $\la\in X(T)$; see [@BK §2]. Then the relation $L_1(\lambda)\cong L_w(\mathbf{r}_w(\lambda))$ defines a bijective map $$\mathbf{r}_w:X(T)\longrightarrow X(T).$$ In particular, if we take $w$ to be the longest element $w_1$ in $\mathcal{D}_{m,n}$, the first part of the following lemma gives an explicit formula for $\mathbf{r}_{w_1}$. \[6.3,5.5\] (1) If $r\leq m,n$ and $\lambda\in \mathcal{RP}(r)$, then $\mathbf{r}_{w_1}(\lambda)=y(\ttM(\lambda))$, where $y:\mathcal{RP}(r)\rightarrow \Lambda^+(m|n,r)$ is defined by $y(\lambda)=\sum^{n}_{i=1}\lambda_i\varepsilon_{m+i}.$ \(2) Assume $r\leq M$. Then $$\Lambda^{++}(M|N,r)=\{\lambda\in\Lambda^+(M|N,r)\mid \lambda_{M+1}\equiv\cdots\equiv\lambda_{M+N}\equiv0( \mbox{mod } p)\}.$$ In particular, for $\la\in \Lambda^{++}(M|N,r)$, if we write $\lambda=\lambda^{0}+p\lambda^1$ such that $\lambda^0$ is $p$-restricted, then $\lambda^0_{M+1}=\cdots=\lambda^0_{M+N}=0$. Let $ w=\left( \begin{smallmatrix} 1&\cdots&m&m+1&\cdots&m+n&m+n+1&\cdots&m+2n\\ 1&\cdots&m&m+n+1&\cdots&m+2n&m+1&\cdots&m+n \end{smallmatrix} \right)\in \mathcal{D}_{m+n,n}. $ \[permutation\] Assume $r\leq m+n$. If $\lambda= \lambda^0+p\lambda^1\in\Lambda^{++}(m+n|n,r)$ with $\la^0$ $p$-restricted and $t(\lambda^0)=(\lambda^0_{m+1},\cdots,\lambda^0_{m+n})$, then $$\mathbf{r}_w(\lambda)=\sum^m_{i=1}\lambda^0_i\varepsilon_i+\sum^{n}_{i=1}(\ttM(t(\lambda^0)))_i\varepsilon_{m+n+i}+p\lambda^1.$$ Set $M = m + n$ and $ N = n$. Then $w$ is the $w_1$ with respect to the subgroup $\fS_{\{m+1,\ldots,m+n\}}\times\fS_{\{m+n+1,\ldots,m+2n\}}$ in $\fS_{\{m+1,\ldots,m+2n\}}$. Note that, first, $t(\lambda^0)$ is also $p$-restricted and $r\leq m+n$ implies $|t(\lambda^0)|\leq n$; second, by Lemma \[6.3,5.5\](2), $\lambda^0_{m+n+i}=0$ for $i=1,\cdots,n$. Thus, applying [@BK Lem. 4.2] and Lemma \[6.3,5.5\] yields[^5] (noting )$$\aligned \mathbf{r}_w(\lambda)&=\mathbf{r}_w(\la^0)+p\la^1=\sum^m_{i=1}(\lambda^0)_i\varepsilon_i+\sum^{n}_{i=1}j(J^{i-1}(t(\lambda^0)))\varepsilon_{m+n+i}+p\lambda^1\\ &=\sum^m_{i=1}(\lambda^0)_i\varepsilon_i+\sum^{n}_{i=1}(\ttM(t(\lambda^0)))_i\varepsilon_{m+n+i}+p\lambda^1,\endaligned$$ as required. Define $\tau:\Lambda(m|n,r)\rightarrow \Lambda(m+n|n,r)$ by setting $\tau(\lambda)=(\lambda^{(0)},0,\cdots,0|\lambda^{(1)})$. Let ${}^{\tau\!}\Lambda(m|n,r):=\tau(\Lambda(m|n,r))$. Assume $r\leq m+n$ and $l=p$. Let $w\in \mathcal{D}_{m+n,n}$ be the same one in Lemma \[permutation\]. Then there is a bijection from $\sP_r$ to ${}^{\tau\!}\Lambda^{++}(m|n,r)$. We first define a function $$h:\mathcal{P}_r\longrightarrow \La^{++}(m+n|n,r),\quad (\lambda,\xi,\eta)\longmapsto(\lambda^\iota+p\xi|p\eta).$$ By Lemma \[6.3,5.5\](2), $h$ is well defined and is injective. Since $\lambda$ is $p$-regular, so is $t(\lambda)=(\lambda_{m+1}, \cdots,\lambda_{m+n})$. Thus, $\ttM(t(\lambda)^\iota)$ is also $p$-restricted and $|t(\lambda)|\leq n$ since $r\leq m+n$. By Lemma \[permutation\], $$\mathbf{r}_w(h(\lambda,\mu,\xi))=\sum^m_{i=1}\lambda_i\varepsilon_i+ \sum^{n}_{i=1}\ttM(t(\lambda)^\iota)_i\varepsilon_{m+n+i}+p(\sum_{i=1}^m\xi_i\varepsilon_i+\sum_{i=1}^{n}\eta_i\varepsilon_{m+n+i}) .$$ So $\mathbf{r}_w(h(\lambda,\mu,\xi))\in {{}^{\tau\!}\Lambda^{+}(m|n,r)}$. Now, by the proof of [@BK 6.5], $$\label{polynomial weights relationship} \begin{aligned} {}^{\tau\!}\Lambda^{++}(m|n,r)&=\mathbf{r}_w(\Lambda^{++}(m+n|n,r))\cap{{}^{\tau\!}\Lambda^{+}(m|n,r)}\\ &=\{\mu\in {{}^{\tau\!}\Lambda^{+}(m|n,r)}\mid \mathbf{r}^{-1}_w(\mu)\in \Lambda^{++}(m+n|n,r)\}. \end{aligned}$$ Hence, $\mathbf{r}_w(h(\lambda,\mu,\xi))\in {{}^{\tau\!}\Lambda^{++}(m|n,r)}$ and $\mathbf{r}_w\circ h$ defines an injective map $$\mathbf{r}_w\circ h:\sP_r\longrightarrow {}^{\tau\!}\Lambda^{++}(m|n,r).$$ We now show that $\mathbf{r}_w\circ h$ is surjective. For $\mu\in {{}^{\tau\!}\Lambda^{++}(m|n,r)}$, write $\tau^{-1}(\mu) =(\mu^{(0)}|\mu^{(1)})\in \Lambda^{++}(m|n,r) $ and $\mu^{(1)}=\mu^{(1),0}+p\mu^{(1),1}$ with $\mu^{(1),0}$ $p$-restricted. By , $j(\mu^{(1)})\leq \mu_m$ (and so $|\mu^{(1)}|\leq n$). By (\[polynomial weights relationship\]), $\mathbf{r}^{-1}_w(\mu)\in \Lambda^{++}(m+n|n,r)$. This element is computed in the proof of [@BK 6.5] (noting ): $$\begin{aligned} \mathbf{r}^{-1}_w(\mu)&=\sum_{i=1}^m\mu_i\varepsilon_i+\sum_{i=1}^nj(J^{i-1}((\mu^{(1),0})))\varepsilon_{m+i}+ p\sum_{i=1}^{n}(\mu^{(1),1})_i\varepsilon_{m+n+i}\\ &=\sum_{i=1}^m\mu_i\varepsilon_i+\sum_{i=1}^n\ttM(\mu^{(1),0})_i\varepsilon_{m+i}+ p\sum_{i=1}^{n}(\mu^{(1),1})_i\varepsilon_{m+n+i}. \end{aligned}$$ Then $\la:=(\mathbf{r}^{-1}_w(\mu))^{(0)}=\sum_{i=1}^m\mu_i\varepsilon_i+\sum_{i=1}^nM(\mu^{(1),0})_i\varepsilon_{m+i}, (\mathbf{r}^{-1}_w(\mu))^{(1)}=p\mu^{(1),1} $. Since $\mu^{(1),0}$ is $p$-restricted, $\ttM(\mu^{(1),0})$ is $p$-restricted. Hence, if we write $$\la=\la^0+p\la^1,$$ where $\la^0$ is $p$-restricted, then $\la^1_{m+k}=0, k=1,\cdots,n.$ Set $d=|\la^1|$. Removing the $n$ zeros at the end of $\la^1$ produces a partition in $\Lambda^+(m,d)$. Thus, we obtain a triple $((\la^0)^\iota,\la^1,\mu^{(1),1})\in\sP_r$ such that $\mathbf{r}_w\circ h((\la^0)^\iota,\la^1,\mu^{(1),1})=\mu$, proving the desired surjectivity. Therefore, $\mathbf{r}_w\circ h$ is a bijection from $\sP_r$ to ${}^{\tau\!}\Lambda^{++}(m|n,r)$. [99]{} J.L. Alperin, *Weights for finite groups*, Proc. Symp. Pure Math., [**47**]{}, Amer. Math. Soc. (1987), 369–379. J. Brundan, A. Kleshchev, *Projective representations of symmetric groups via Sergeev duality*, Math. Z. [**239**]{} (2002), 27–68. J. Brundan, J. Kujawa, *A new proof of the Mullineux conjecture*, Journal of Algebraic Combinatorics [**18**]{} (2003), 13–39. B. M. Deng, J. Du, B. Parshall, J. P. Wang, *Finite Dimensional Alegebras and Quantum Groups*, Mathematical Surveys and Monographs, Vol. 150, Amer. Math. Soc, Providence R. I. (2008). R. Dipper, J. Du, *Trival and alternating source modules of Hecke algebras of type $A$*, Proc. London Math. Soc. [**66**]{} (1993) 479–506. R. Dipper, G. James, *Representations of Hecke algebras of general linear groups*, Proc. London Math. Soc. [**54**]{} (1987), 57–82. S. Donkin, *Symmetric and exterior powers, linear source modules and representations of Schur superalgebras*, Proc. London Math. Soc. [**83**]{} (2001), 647–680. J. Du,*The modular representation theory of q-Schur algebras*, Trans. Am. Math. Soc. [**329**]{} (1992), 253–271. J. Du,*The modular representation theory of q-Schur algebras II*, Math. Z. [**208**]{} (1991), 503–536. J. Du, *The green correspondence for the representations of Hecke algebras of type $A_{r-1}$*, Trans. Am. Math. Soc. [**329**]{} (1992), 273–287. J. Du, H. Rui,*Quantum schur superalgebras and Kazhdan-Lusztig combinatrics*, Journal of Pure and Applied algebra, [**215**]{} (2011), 2715–2737. W. Feit, *The Representation Theory of Finite Groups*, North-Holland, Amsterdam, 1982. G. James, *The representation theory of the symmetric groups*, Lecture Notes in Math. 682, Springer, Berlin, 1978. L. Jones, *Centers of generic algebras*, Trans. Am. Math. Soc. [**317**]{} (1990), 361–392. P. Landrock, *Finite Group Algebras and their Modules*, London Math. Soc. Lect. Note Series, Vol. 84, Cambridge University Press, Cambridge, 1983. Y. Manin, *Multiparameteric quantum deformation of the general linear supergroup*, Comm. Math. Physics [**123**]{} (1989),163–175. G. Mullineux, *Bijections of p-regular partitions and p-modular irreducibles of symmetric groups*, J. London Math. Soc. 20 (1979), 60¨C66. B. Parshall, J.-p. Wang, *Quantum linear groups*, Mem. Amer. Math. Soc.Vol. 89, No. 439 (1991). L. Scott, *Modular permutation representations*, Trans. Am. Math. Soc. [**175**]{} (1973), 101–121. H. El Turkey and J. Kujawa, *Presenting Schur superalgebras*, Pacific J. Math., to appear. [^1]: The first author gratefully acknowledge support from ARC under grant DP120101436. The work was completed while the second author was visiting UNSW. She would like to thank the China Scholarship Council and ARC for financial support and UNSW for its hospitality during her visit. [^2]: In this paper, for the Hom spaces $\Hom_{\sH_\la}(M,N)$ of [*right*]{} $\sH_\la$-modules $M,N$, the right hand function notation will be used to easily read the induced bimodule structure on the space: $(m)(xfy)=((m.x)f).y$ with $x,y\in\sH_\la,m\in M$, and $f\in\Hom_{\sH_\la}(M,N)$. [^3]: The $l$-parabolic subgroups play a role similar to $p$-subgroups in group representation theory. So the notation $P$ for an $l$-parabolic subgroup indicates this similarity. [^4]: In the Appendix, we will see that the signed permutation is naturally induced from the original Hecke algebra action. [^5]: In fact, we apply [@BK 4.2] repeatedly to the sequence of odd roots $$\begin{aligned} &\varepsilon_{m+n}-\varepsilon_{m+n+1},\varepsilon_{m+n-1}-\varepsilon_{m+n+1},\cdots,\varepsilon_{m+1}-\varepsilon_{m+n+1};\\ &\varepsilon_{m+n}-\varepsilon_{m+n+2},\varepsilon_{m+n-1}-\varepsilon_{m+n+2},\cdots,\varepsilon_{m+1}-\varepsilon_{m+n+2};\\ &\cdots\cdots\\ &\varepsilon_{m+n}-\varepsilon_{m+2n},\varepsilon_{m+n-1}-\varepsilon_{m+2n},\cdots,\varepsilon_{m+1}-\varepsilon_{m+2n}. \end{aligned}$$ and follow the proof of [@BK 6.3].
{ "pile_set_name": "ArXiv" }
--- abstract: 'We report on the production of a novel cold mixture of fermionic $^{53}$Cr and $^{6}$Li atoms delivered by two Zeeman-slowed atomic beams and collected within a magneto-optical trap (MOT). For lithium, we obtain clouds of up to $4 \,10^8$ atoms at temperatures of about $500\,\mu$K. A gray optical molasses stage allows us to decrease the gas temperature down to $45(5)\,\mu$K. For chromium, we obtain MOTs comprising up to $1.5\, 10^6$ atoms. The availability of magnetically trappable metastable $D$-states, from which $P$-state atoms can radiatively decay onto, enables to accumulate into the MOT quadrupole samples of up to $10^7$ $^{53}$Cr atoms. After repumping $D$-state atoms back into the cooling cycle, a final cooling stage decreases the chromium temperature down to $145(5)\,\mu$K. While the presence of a lithium MOT decreases the lifetime of magnetically trapped $^{53}$Cr atoms, we obtain, within a 5 seconds duty cycle, samples of about $4\, 10^6$ chromium and $1.5\,10^8$ lithium atoms. Our work provides a crucial step towards the production of degenerate Cr-Li Fermi mixtures.' author: - 'E. Neri' - 'A. Ciamei' - 'C. Simonelli' - 'I. Goti' - 'M. Inguscio' - 'A. Trenkwalder' - 'M. Zaccanti' title: Realization of a Cold Mixture of Fermionic Chromium and Lithium Atoms --- INTRODUCTION {#intro} ============ Ultracold Fermi gases represent nowadays a prominent platform for the implementation of model Hamiltonians and the exploration of a variety of many-body regimes, primarily in the context of superfluid pairing and magnetic ordering. Notable examples are the study of the BCS-BEC crossover in fermionic mixtures with balanced or imbalanced spin populations [@Chevy2010; @Radzihovsky2010; @Zwerger2012], and the disclosure of ferromagnetic [@Jo2009; @Sanner2012; @Valtolina2017; @Amico2018; @Scazza2019_ii] and anti-ferromagnetic [@Hart2015; @Parsons2016; @Boll2016; @Cheuk2016] correlations within itinerant and localized repulsive fermion systems, respectively. The introduction of a “heavy-light” mass asymmetry among the constituents of a fermionic mixture is considered as an appealing extension of this research field, that opens qualitatively new scenarios. At the few-particle level, two-fermion mixtures are predicted to exhibit a rich variety of N-body bound states and scattering resonances, both for free-space and trapped systems. Among others, a proper mass imbalance leads to the existence of the three-, four- and five-body Efimov effect [@Efimov1971; @Efimov1973; @Castin2010; @Bazak2017], non-Efimovian trimer states, both in free space [@Kartavtsev2007; @Endo2011; @Bazak2017] and in reduced or mixed dimensions [@Nishida2008; @Levinsen2009], as well as few-body clusters in trapped configurations, see e.g. Ref.[@Blume2012] and references therein. At the many-body level, the natural mismatch of the Fermi surfaces of a two-component mixture is predicted to favor superfluidity beyond the standard Cooper pairing mechanism [@Fulde1964; @Larkin1964; @Sarma1963; @Gubankova2003; @Liu2003; @Forbes2005; @Iskin2006; @Parish2007; @Baranov2008; @Baarsma2010; @Gezerlis2009; @Astrakharchik2012] at experimentally attainable temperatures. Moreover, a number of exotic normal many-body phases, such as ordered crystalline states [@Petrov2007] and trimer Fermi gases [@Nishida2008b; @Levinsen2009], have been theoretically investigated. Parallel to this, mass-imbalanced mixtures have recently emerged also as a promising framework for the quantum simulation of magnetic phenomena, arising in both itinerant [@Keyserlingk2011; @Cui2013; @Massignan2014] and localized [@Sotnikov2012; @Sotnikov2013] fermion systems with short-range repulsive interactions. In spite of the great interest for ultracold Fermi gases of chemically different species, their experimental investigation is still at a relatively early stage, only a very few degenerate Fermi systems with sizable mass imbalance being currently available: the $^{173}$Yb-$^6$Li [@Hara2011; @Gupta2019], the $^{40}$K-$^6$Li [@Taglieber2008; @Wille2008] and the $^{161}$Dy-$^{40}$K [@Ravensbergen2018; @Ravensbergen2018] mixtures. The first combination is primarily appealing for the creation of ground state polar molecules with both magnetic and electric dipole moment [@Micheli2006]. However, the extremely narrow nature of the predicted Feshbach resonances [@Brue2012], combined with a small background scattering length [@Hara2011] makes the reach of the strongly interacting regime in such a system extremely challenging. Furthermore, the very large mass ratio $M_{Yb}/m_{Li} \sim 29$ would support, at strong coupling, the existence of Efimov cluster states [@Efimov1971; @Efimov1973; @Kartavtsev2007; @Endo2011], whose intrinsically lossy nature could strongly limit the stability of a resonant Yb-Li mixture. In this respect, the $^{40}$K-$^6$Li system appears as a much more promising platform. Thanks to the availability of interspecies Feshbach resonances [@Wille2008; @Naik2011], interesting results have indeed been obtained both in the many-body [@Trenkwalder2011; @Kohstall2012; @Cetina2015; @Cetina2016] and in the few-body [@Jag2014; @Jag2016] context. In particular, the large but not extreme mass ratio of such a mixture, $M_{K}/m_{Li} \sim 6.6$, has enabled to experimentally unveil a resonant enhancement of $p$-wave K-KLi atom-dimer interactions on the BEC side of a K-Li Feshbach resonance [@Jag2014]. This exotic few-body feature starkly differs from the phenomenology typical of the Efimov scenario, as it exhibits a purely elastic, universal character: as such, it may represent an unforeseen opportunity to investigate in the future many-body regimes of ultracold fermionic matter in presence of non-perturbative few-body correlations. Furthermore, within a confined geometry [@Levinsen2011], this could enable to realize collisionally stable trimer Fermi gases with exotic topological properties [@Tajima2019]. However, the narrow character of the K-Li resonances, and especially the presence of non-zero two-body inelastic decay [@Naik2011], render the investigation of ground-state properties with such a mixture extremely challenging. These complications will be hopefully overcome in the near future by the recently realized $^{161}$Dy-$^{40}$K [@Ravensbergen2018] mixture, on which wide resonances in the system ground state, immune to two-body losses, have been discovered, and strongly interacting conditions have been attained [@Ravensbergen2019]. Yet, the very dense spectrum of both intra- and inter-species resonances in Dy-Dy and K-Dy collisions, connected with the anisotropic nature of the interatomic interactions [@Petrov2012], could complicate the production of highly degenerate samples and the controlled tuning of Dy-K elastic interactions. Further, the mass ratio $M_{Dy}/m_{K} \sim 4.0$ of such a mixture, significantly smaller than the K-Li one, may prevent, at least in three dimensions, the emergence of non-trivial few-body physics. Here, we report on the first important step towards the production of a novel Fermi mixture of $^{53}$Cr and $^6$Li atoms, namely the realization of magneto-optically trapped Cr-Li clouds in the cold regime. Our specific choice is motivated by the exceptional scattering properties of the Cr-Li system, so far overlooked, that cannot be obtained with any other atom-atom combination, and which open the way to a wealth of possibilities, going well beyond the scope of presently available systems. First and very importantly, the mass ratio of $^{53}$Cr and $^6$Li ($M_{Cr}/m_{Li}\sim 8.80$) is predicted to support one weakly bound (quasi bound) Cr$_2$Li trimer (Cr$_3$Li tetramer) state in the regime of strong Cr-Li repulsion [@Kartavtsev2007; @Endo2011; @Bazak2017], which may enable the resonant tuning of genuine few-body elastic interactions, on top of the standard two-body ones. Second, thanks to a purely quantum interference phenomenon, three-body recombination processes to weakly bound molecular states are drastically suppressed for the Cr-Li system, when compared to any other mixture [@Petrov2003], allowing for the realization of long-lived Fermi gases with strong interspecies repulsion, “immune” to the pairing instability inherent of such systems [@Pekker2011; @Sanner2012; @Amico2018]. This represents a key advantage in the experimental study of ferromagnetic phases with ultracold Fermi gases [@Jo2009; @Sanner2012; @Valtolina2017; @Amico2018; @Scazza2019_ii], possibly enabling polarized domains to develop over macroscopic lengthscales, in contrast with what has been observed in homonuclear systems [@Amico2018; @Scazza2019_ii]. Besides these two main points, it is also worth noticing that fermionic $^{53}$Cr is a yet poorly explored [@Chicireanu2006; @Naylor2015] but very interesting system on its own. Indeed, it exhibits a large magnetic dipole moment ($6\,\mu_B$, where $\mu_B$ is the Bohr’s magneton), and it is expected to feature few wide Feshbach resonances [@Simonipriv] which could enable precise tuning of homonuclear interactions. Finally, although no predictions are currently available for $^{53}$Cr-$^6$Li Feshbach resonances due to the lack of any experimental input, the rich hyperfine and Zeeman structure of these two atomic species, combined with the highly magnetic character of Cr atoms, is expected to yield rich resonance spectra, with a density about three-to-four times larger than the typical one featured by alkali mixtures [@Simonipriv]. In this regard, it is also important to stress that even relatively narrow Feshbach resonances could guarantee a good collisional stability for the resonantly interacting mixture [@Jag2016], and they would not limit the observability of the rich few-body properties of this system, thanks to the larger Cr-Li mass ratio, compared to the K-Li one [@Jag2014]. The paper is organized as follows. In section \[exp\] we present our new experimental setup and discuss its main features. In Section \[singlespecies\] we report on the performance of our apparatus when operating with one single atomic species. Section \[doublespecies\] discusses the first characterization of the cold Cr-Li mixture when the two atomic clouds are simultaneously loaded within a double-species magneto-optical trap. Finally, in Section \[outlook\] we summarize the outcome of our experimental studies, and discuss the prospects for the successive steps towards the ultracold regime, namely optical trapping and the implementation of evaporative and sympathetic cooling protocols. EXPERIMENTAL SETUP {#exp} ================== In the following, we summarize the main features of our new apparatus, whereas for a detailed description we refer the reader to Ref.[@Neri2019]. We first discuss the design strategy followed to build the vacuum and the magnetic coils setups. After recalling the optical transitions required for the laser cooling of the two atomic species, we then provide an overview of the specific optical setup for $^6$Li and $^{53}$Cr atoms implemented in our new machine. Vacuum and magnetic coils setup ------------------------------- Our vacuum apparatus employs two independent Zeeman slower (ZS) lines that connect the Li and Cr effusion cells to a common science chamber, see Fig.\[Fig1\]. This choice is firstly motivated by the significantly different sublimation temperature of Li (about $410^{\circ}C$) and Cr (of about $1500^{\circ}C$): While the former can be obtained with a custom-designed oven, for the latter we employ a commercial high-temperature effusion cell by CreaTec [@CrLi_sources_note]. Secondly, this choice is motivated by the fact that the Cr-Li chemical and collisional properties were initially unknown and, with this configuration, the two species do not affect each other up to the MOT region. ![image](VacuumSetup2.png){width="140mm"} Finally, our design conveniently enables to optimize the ZS characteristics for each species independently. The obvious drawback of this geometry is that four, rather than two, viewports of the main chamber are required for the ZS vacuum lines and optical beams. In consideration of this, we employ a large radius octagon chamber, with eight perimetral $CF40$ windows spaced out by $CF16$ ones. Two large $CF200$ re-entrant windows close the chamber along the vertical direction. Our vacuum apparatus comprises a $500\,$l/s NexTorr pump installed near the science chamber, and two differential pumping stages between the ZSs and the Cr and Li oven output, respectively. A $75\,$l/s VacIon Plus ion pump is installed at each oven. With this configuration, we ensure an ultra-high vacuum pressure below $10^{-10}\,$mbar in the MOT region. ![image](CrLi_lines_new.pdf){width="140mm"} The overall coil assembly is designed in such a way that the ZS axial fields merge with the MOT radial quadrupole field at the edge of the science chamber, forming two monotonically decreasing profiles. Such a non spin-flip configuration, already successfully implemented for single-species Li experiments \[see e.g. Ref.[@Serwane2007]\] but yet unexplored for Cr, is extremely advantageous for the fermionic isotope $^{53}$Cr, as it overcomes the difficulty connected with the existence of a “bad crossing” region along the ZS field profile near $25\,$G [@Chicireanu2006]: at this field, a non perfectly polarized ZS light efficiently depumps the atoms towards unaddressed hyperfine states, thus detrimentally reducing the flux of Cr atoms collected into the MOT. In our apparatus, the Zeeman slowed Cr atoms experience this field only $2\,$cm away from the quadrupole center, where they are already slow enough to be repumped and trapped by the MOT lights. Furthermore, this design minimizes transverse-velocity induced losses at the ZS output, as it zeroes the distance that atoms, decelerated down to very low velocities, must travel before reaching the MOT region. This is a key advantage in our setup, given the large radius of $13\,$cm of the science chamber. On the other hand, this ZS field configuration implies the use of relatively small ZS light detunings, which may sizably perturb the Cr MOT and limit its performances owing to enhanced light-assisted collisions [@Chicireanu2006]. A smooth merging of the ZS and MOT field profiles at the edge of the science chamber is enabled by two large MOT coils, characterized by an inner (outer) radius of 72 mm ($160\,$mm). These are made of a square-section hollow copper wire, which allows for very efficient water cooling, and they are embedded within a custom-designed resin structure that provides a sustaining frame. In addition to the MOT coils, our setup is equipped with two smaller coils (inner radius of $64\,$mm; outer radius of $85\,$mm), concentric with the MOT ones, and hosted within a water-flooded toroidal PEEK case, inserted within the re-entrance of the $CF200$ viewports, see section view on the top left in Fig.\[Fig1\]. This second pair of coils is primarily designed to create homogeneous Feshbach fields up to $1000\,$G with less than $200\,$A current. Moreover, when switched to anti-Helmoltz configuration, the Feshbach coils can provide a quadrupole field with fast switch-off timescales ($<1\,$ms), thanks to their small inductance. Three additional pairs of coils placed along three orthogonal directions enable to create fields of a few Gauss, to finely adjust the quadrupole position and to compensate for spurious offset fields. ![Simplified view of the optical setup employed to obtain the blue lights required for the laser cooling of $^{53}$Cr. The $425.5\,$nm light is produced from a $851\,$nm diode laser (1) through a frequency-doubling bow-tie cavity (3). A tapered amplifier chip hosted on a home-made mount (2) amplifies the infrared light, while a system of spherical and cylindrical lenses collimates the beam, see Ref.[@Neri2019] for details. A pickup beam at the cavity output is sent to the locking setup. An AOM along its path sets the light to be blue-detuned from the $^{52}$Cr $^7S_{3} \rightarrow {^7P_{4}}$ transition by $69\,$MHz. The remaining light is split into the cooling and repumper beam paths. Along the first one, one single-pass and two double-pass AOMs produce the ZS, MOT and TC beams, respectively. From the TC beam, the imaging beam is also derived. Along the repumper path, a series of three AOMs produces the $R_1$, $R_2$ and $R_3$ lights. From them, we produce the repumper beam for the MOT, containing all three lights, and the HP beam, containing only $R_1$ and $R_2$.[]{data-label="FigCROptSetup"}](CR_Opt_Setup.pdf){width="85mm"} Lithium Laser Setup ------------------- The atomic transitions employed for the laser cooling of Li atoms are schematically shown in Fig.\[Fig2\](a). The Li MOT is based on the $D_2$ ($^2S_{1/2}\rightarrow{^2P_{3/2}}$) atomic line at $671\,$nm, featuring a natural linewidth $\Gamma_{Li}/2\pi=5.87\,$MHz, and a saturation intensity $I_{S,Li}=2.54\,$mW/cm$^2$. The $D_2$ cooling and repumper lights, detuned by $228\,$MHz from each other, address the $F=3/2\rightarrow {F' =5/2}$ and the $F =1/2 \rightarrow {F' = 3/2}$ transition, respectively. Two additional laser lights addressing the hyperfine transitions associated with the $D_1$ ($^2S_{1/2}\rightarrow{^2P_{1/2}}$) atomic line, red-detuned by about $10\,$GHz from the $D_2$ line, are employed for a sub-Doppler cooling stage based on gray optical molasses [@Burchianti2014]. Our optical setup is essentially analogous to the one previously developed in the Lithium lab at LENS [@ValtolinaPhD]. The $D_2$ and the $D_1$ lights are provided by two separate master laser sources, and independently controlled by two acousto-optic modulators (AOMs) which act as fast switches. The two AOM output beams are immediately recombined on a common optical path. They are split into cooling and repumper lights, whose frequency and power are independently adjusted by means of two AOMs, and injected into two commercial tapered amplifiers, each one delivering about $300\,$mW. The cooling and repumper lights are then recombined on a $50:50$ non-polarizing beamsplitter, and sent to the MOT and ZS/imaging beam paths. Three fibers bring the MOT light to the science chamber: Three mutually orthogonal retroreflected beams with a $1/e^2$ radius of about $5.0\,$mm and total peak intensity of about $29\,I_{S,Li}$ realize the MOT configuration. The ZS/imaging path includes an additional single-pass AOM, and its frequency is toggled to inject the light into either the ZS or the imaging fiber. At the entrance of the vacuum chamber, the ZS beam, slightly focused, features a $1/e^2$ radius of about $2.2\,$mm, with total peak intensity of about $200\,I_{S,Li}$. Chromium Laser Setup -------------------- The optical setup required for the laser cooling of fermionic $^{53}$Cr atoms is rather complex, owing to the presence of metastable $D$-states into which atoms can leak from the cooling cycle, and to a rich hyperfine structure \[see Fig.\[Fig2\](b)\]. Moreover, a transverse cooling (TC) stage at the oven chamber is known to be crucial [@Chicireanu2006] to increase the atomic beam collimation, and thus the flux of Zeeman slowed atoms at the MOT region. The laser cooling scheme is based on the $^7S_{3} \rightarrow{^7P_{4}}$ atomic line at $425.5\,$nm, with natural linewidth $\Gamma_{Cr}/2\pi=5.06\,$MHz and saturation intensity $I_{S,Cr} =8.52\,$mW/cm$^2$. While for bosonic Cr isotopes one single frequency is needed to operate the MOT, the four hyperfine sub-levels of the $^{53}$Cr fine states require additional repumper beams, besides the $F = 9/2 \rightarrow {F' = 11/2}$ cooling light. Up to three blue repumpers, denoted by $R1$, $R2$ and $R3$ in Fig.\[Fig2\](b), are in principle needed to bring excited $P$-state atoms back into the cooling cycle. Moreover, even with all blue repumpers on, the MOT transition remains slightly leaky, since optically excited atoms can decay from the $^7P_4$ state onto the underlying $^5D_{3,4}$ metastable states. Therefore, additional “red” repumpers are needed to fully close the cooling cycle [@Chicireanu2006]. These lights, at a wavelength of $654\,$nm and $663\,$nm, pump atoms from the metastable $^5D_3$ and $^5D_4$ states, respectively, back to the ground state via the $^7P_3$ level, see red solid arrows in Fig.\[Fig2\](b). In the following, we provide a detailed description of our Cr optical setup and its distinctive features, as compared to the one employed by the only other group that has successfully laser cooled $^{53}$Cr thus far [@Chicireanu2006]. The setup is schematically shown in Fig.\[FigCROptSetup\]. The blue light is generated by frequency-doubling a high-power laser source at $851\,$nm using a non-linear lithium triborate (LBO) crystal within a custom-designed bow-tie optical cavity [@Neri2019]. Up to $3\,$W of $851\,$nm light are obtained by seeding a tapered amplifier with a commercial diode laser. With an overall conversion efficiency of $30$%, the doubling cavity, injected with $2\,$W of infrared light, delivers about $600\,$mW at $425.5\,$nm. $10\,$mW of this light are used to lock the master laser, via a standard saturation spectroscopy scheme on a hollow cathode lamp, such that the blue light is resonant with the $^{52}$Cr $^7S_{3} \rightarrow {^7P_{4}}$ transition. The remaining $425.5\,$nm light is split into two main optical paths for the cooling and the repumpers, respectively. The cooling beam is further split into ZS, MOT and TC paths, see Fig.\[FigCROptSetup\]. Similarly to the case of Li, our MOT light configuration consists of three mutually orthogonal retroreflected beams. These are characterized by a $1/e^2$ radius of $3.0\,$mm and an average peak intensity of $2.3\,I_{S,Cr}$. The TC light is sent near the output of the Cr oven, where it is split into two retroreflected orthogonal beams, perpendicular to the ZS direction. In order to maximize the interaction time of the atoms with the cooling light, the TC beams have an elliptical shape of $1 \div 3$ aspect ratio, with the large waist of $4.1\,$mm oriented along the ZS direction. A total peak intensity of $20\,I_{S,Cr}$ is employed for the two TC beams, and their frequency is set close to resonance with the $F=9/2 \rightarrow {F'=11/2}$ transition. A pickup of the TC light is coupled into a fiber, delivering $200\,\mu$W, employed for absorption imaging. The ZS beam, characterized by a waist of $3.1\,$mm at the MOT region, is slightly focusing and it is directed towards the oven output by means of an in-vacuum mirror. The light is circularly polarized as to address mostly $\sigma_+$ transitions, with an intensity of $32\,I_{S,Cr}$ at the MOT region. From the repumper path we obtain the $R_1$, $R_2$ and $R_3$ lights, indicated in Fig.\[FigCROptSetup\] by green, red and yellow arrows, respectively. Their frequencies are adjusted by means of the three AOMs, and they are kept constant throughout the experimental cycle. By combining all repumper lights at a polarizing beam splitter we obtain two beams. The first one, containing all three lights, is sent to the science chamber to create a single retroreflected beam, crossing the MOT region with an angle of $22.5$ degrees with respect both to one of the horizontal MOT directions and to the ZS axis. It is characterized by a $1/e^2$ radius of $3.4\,$mm and total peak intensities of $12$ ($R_1$), $6$ ($R_2$) and $6\,I_{S,Cr}$ ($R_3$), respectively. The second beam, containing only $R_1$ and $R_2$, is sent to the Cr oven output. It is exploited to increase the atomic population of the lowest hyperfine $F=9/2$ manifold, and thus the flux of Zeeman slowed atoms. This “hyperfine pumping” (HP) beam has a $1/e^2$ radius of about $2.2\,$mm ($0.7\,$mm) along the longitudinal (transverse) direction of the atomic beam, and a total peak intensity of $90\,I_{S,Cr}$, equally distributed between $R_1$ and $R_2$. Finally, besides the blue lights, our setup also comprises two “red” repumpers, realized by two commercial diode lasers at $654\,$nm and $663\,$nm. Their proximity to the Li atomic reference at $671\,$nm enables a simple and cost-effective locking scheme of their frequency, based on a commercial Fabry-Pérot resonator as transfer cavity \[see Ref.[@Neri2019] for details\]. EXPERIMENTAL RESULTS: SINGLE SPECIES {#singlespecies} ==================================== In the following, we discuss the experimental procedures exploited to produce individual clouds of cold Li and Cr atoms, and summarize the performance of our machine under optimized conditions for single species operation. Lithium ------- Our strategy to laser cool and manipulate $^6$Li atoms is essentially analogous to the one already developed in the Lithium lab at LENS [@Burchianti2014], the only difference being the non spin-flip configuration of the ZS employed in our setup. Optimization of the experimental cycle is obtained by monitoring the MOT fluorescence signal on a calibrated photo-detector, and by acquiring time-of-flight (TOF) absorption images of the Li clouds on a CCD camera, from which we deduce the gas temperature. Fig.\[FigLi\](a) schematically illustrates the sequence employed in the experiment, which consists of the initial MOT loading, followed by two compressed-MOT (CMOT) phases, all operating on the $D_2$ transition. These are followed by a gray optical molasses stage based on the $D_1$ line. Table \[TabLi\] summarizes the corresponding atom number, peak density and temperature of the cloud, measured at each step of the routine after $7\,$s loading and with the following experimental parameters: oven temperature set to $410^{\circ}\,$C, ZS total power (detuning) of $40\,$mW ($-21.3\,\Gamma$), total MOT peak intensity of $29\,I_{S,Li}$. During the loading stage, similarly to Ref.[@Burchianti2014], we keep the detuning of the MOT cooling (repumper) light at about $-6.8\,\Gamma$ ($-7.5\,\Gamma$), to maximize the number of trapped atoms while limiting the gas temperature to about $2.5\,$ mK. -------------- ---------- -------------- --------- ----------- Atom number 2.0(2) 2.0(2) 1.7(2) 1.27(12) Peak density 0.64(12) 4(1) 21(4) 24(6) Temperature [2500]{} [900(100)]{} 500(50) [45(5)]{} -------------- ---------- -------------- --------- ----------- : Typical atom number, peak density and temperature of our $^6$Li cloud recorded after each stage of the experimental cycle, for a $7\,$s loading time and the experimental settings detailed in the text.[]{data-label="TabLi"} We then turn off both the ZS field and light, and we reduce the temperature of the Li cloud by performing two successive stages of CMOT, during which we decrease both the detuning and intensity of the MOT beams, see Fig.\[FigLi\](a). A first CMOT stage, lasting for $100\,$ms, enables to lower the temperature to about $0.9(1)\,$mK. Simultaneously, we adiabatically transfer the cloud from the MOT to the Feshbach coils quadrupole. Owing to their much smaller inductance, indeed, the Feshbach coils allow for a fast switch-off time of about $500\,\mu$s at typical operating currents. A second CMOT stage, lasting about $4\,$ms, further cools the Li MOT below $500\,\mu$K, with atom losses below $20$%. The average $1/e$ MOT radius drops from the initial $3.8(3)\,$mm value down to $1.20(14)\,$mm and, correspondingly, the peak density increases from $6.4(1.2)\,10^{8}\,$atoms/cm$^3$ to $2.1(4)\,10^{10}\,$ atoms/cm$^3$. We then turn off the $D_2$ MOT light and quadrupole field while turning on the $D_1$ cooling and repumper lights, both blue-detuned with respect to their corresponding resonances. We have tested the performance of $D_1$ molasses upon varying the light parameters, by monitoring the temperature and the atom number through TOF absorption imaging, resonant with the $D_2$ $F=1/2 \rightarrow {F’=3/2}$ transition. Fig.\[FigLi\](b) shows the evolution of temperature and atom number, after 2 ms of gray molasses, as a function of the relative detuning, $\delta = \delta_1 - \delta_2$. Here, $\delta_1$ and $\delta_2$ are the detuning of the $D_1$ repumper and cooling lights from the $F=1/2 \rightarrow {F'=3/2}$ and $F=3/2 \rightarrow {F'=3/2}$ transitions, respectively \[see Fig.\[Fig2\](a)\]. The cloud temperature as a function of $\delta$ exhibits an asymmetric Fano profile with sub-natural width centered around the Raman resonance ($\delta=0$). This trend represents the distinctive feature of such a sub-Doppler cooling technique, that arises from the Sisyphus effect on the blue $F \rightarrow {F' = F}$ transition combined with the emergence of a coherent dark state, see e.g. Ref.[@Grier2013] for details. In particular, at the Raman condition, the cloud temperature is found to reach a minimum value of $T=45(5)\, \mu$K, with a cooled fraction $N/N_0$ of about $75$%, in excellent agreement with what was previously observed in the Lithium lab at LENS [@Burchianti2014]. Chromium -------- In contrast with the well-established Li system, fermionic Cr is still a relatively poorly explored species in the cold regime. Thus far, indeed, only one group worldwide successfully laser-cooled $^{53}$Cr atoms [@Chicireanu2006] and, more recently, brought it to quantum degeneracy by means of sympathetic cooling with the most abundant bosonic isotope $^{52}$Cr [@Naylor2015]. For this reason, in the following we provide a detailed description of our experimental protocol, which enabled to realize samples of almost $10^7$ $^{53}$Cr atoms and temperatures below 150 $\mu$K, discussing the similarities and differences between our working conditions and the Paris ones [@Chicireanu2006]. A few peculiar features make laser cooling of Cr isotopes rather challenging. First, the anomalously large light-assisted collision rates of this species limit the $^{52}$Cr ($^{53}$Cr) steady-state MOT population to a few $10^7$ ($10^5$) \[see [@Chicireanu2006] and references therein\]. Secondly, as discussed in Section \[exp\], the radiative decay of $P$-state atoms towards metastable $D$-states requires the implementation of a laser setup significantly more complex than for alkali or lanthanide atoms. Thirdly, when the fermionic $^{53}$Cr isotope is considered, its rich hyperfine structure and the relatively small natural abundance strongly decreases the flux of Zeeman-slowed atoms that can be experimentally achieved. In fact, while relatively high loading rates of $\Gamma_L \sim 3\, 10^8$ atoms/s can be obtained for $^{52}$Cr, $^{53}$Cr MOTs usually exhibit $\Gamma_L$ of the order of $10^6$ atoms/s [@Chicireanu2006]. Such a reduction of more than two orders of magnitude arises from the combination of a $9$-fold lower natural abundance, together with the fact that only one among the 28 Zeeman sublevels of the electronic ground state, specifically the $|F,m_F\rangle {=} |9/2,+9/2\rangle$ state, can be decelerated within the ZS. In our experiment, we have first maximized the number of collected $^{53}$Cr atoms by thoroughly scanning all MOT, TC, HP and ZS parameters. Examples of such a characterization are presented in the contour plots of Fig.\[FigCr\]. There we show the dependence of the steady-state MOT atom number on both the peak intensity and detuning of the cooling light MOT \[Fig.\[FigCr\](a)\] and ZS \[Fig.\[FigCr\](b)\] beams. The observed trend nicely matches the results of previous studies [@Chicireanu2006]. In particular, an almost constant maximum value of the MOT population is obtained for detunings $3.4\lesssim|\delta|/\Gamma_{Cr}\lesssim 4.2$ at single-beam intensities $3 \lesssim I/I_{S,Cr} \lesssim 5$. The decrease of the MOT performances for $|\delta|< 3.4\,\Gamma_{Cr}$ can be understood by considering that smaller detunings cause an increase of both the atomic density and the light-assisted collision rate [@ChicireanuPhD]. Moving to the opposite limit, we observe a decrease in the atom number for $|\delta| >5 \, \Gamma_{Cr}$ which cannot be compensated any more upon increasing the beams intensity. From this we deduce that the MOT capture volume is limited in this regime by our MOT beams size. Under the working conditions yielding the largest atom number, we obtain samples of up to $3.5(2)\,10^5$ atoms. By fitting the initial slope of the measured loading curves \[see example in Fig.\[CrLoading\](a)\], characterized by a typical $1/e$ time of about 65(5) ms, we infer $^{53}$Cr loading rates of about $\Gamma_L \sim 4-6 \, 10^6\,$atoms/s for an oven temperature of $1500^{\circ}$C. These values are comparable with, and even slightly larger than, the ones previously reported by the Paris group [@Chicireanu2006; @ChicireanuPhD] under similar MOT gradient and cooling light parameters. --- --- --------- --------- --------- --------- 0 0 0.27(3) 0.54(4) 0.21(6) 0.42(5) --- --- --------- --------- --------- --------- : **Top panel.** Effect of the different beams on the relative $^{53}$Cr atom number in the steady-state MOT without red repumpers. The same factors apply to the population of magnetically-trapped D-state atoms. **Bottom panel.** Light parameters for an optimum MOT loading at a ZS current of $14.5\,$A and a MOT current of $25\,$A, corresponding to a radial gradient of $12.6\,$G/cm. The detuning of each light is measured from the corresponding relevant transition.[]{data-label="Table2"} ------------------------ ------ ------ ZS cooling 32 6.94 TC cooling 20 0.86 MOT cooling (per beam) 2.3 4.02 HP R$_1$ 45 0.94 HP R$_2$ 45 0.43 MOT R$_1$ 12.4 0.94 MOT R$_2$ 5.9 0.43 MOT R$_3$ 6.7 0.63 ------------------------ ------ ------ : **Top panel.** Effect of the different beams on the relative $^{53}$Cr atom number in the steady-state MOT without red repumpers. The same factors apply to the population of magnetically-trapped D-state atoms. **Bottom panel.** Light parameters for an optimum MOT loading at a ZS current of $14.5\,$A and a MOT current of $25\,$A, corresponding to a radial gradient of $12.6\,$G/cm. The detuning of each light is measured from the corresponding relevant transition.[]{data-label="Table2"} Despite these similarities with the Paris experiment, we emphasize that in our case this MOT performance is obtained under very different conditions of ZS and repumping light parameters. First of all, in our setup no repumping lights are required in the ZS beam to reach the aforementioned loading rate. This is due to our non spin-flip configuration, where the “bad crossing” near $25$ G along the ZS field profile is produced very close to the MOT region, as detailed in Section \[exp\]. Secondly, our ZS optimally works at a much smaller light detuning, of about $-7\,\Gamma_{Cr}$ \[see Fig.\[FigCr\](b)\], corresponding to a ZS exit velocity at zero magnetic field of $15\,$m/s. Such a low value is not constrained by the MOT parameters of our setup, and it appears to be inherently related to the hyperfine structure of $^{53}$Cr. In fact, although not discussed in the present paper, while investigating the bosonic $^{52}$Cr isotope [@Neri2019], our MOT could capture atoms traveling at a threefold higher velocity. In the case of fermionic $^{53}$Cr, atoms exit our ZS in the $|9/2,+9/2\rangle$ Zeeman level of the lowest hyperfine manifold and, after crossing the quadrupole center, they must scatter many $\sigma_-$ photons to be captured in the MOT. During this final deceleration stage, transitions to hyperfine levels other than $F'=11/2$ are not forbidden, and they are even enhanced by the residual Doppler shift, making the Zeeman slowed atoms likely to fall into $F \neq 9/2$ states, out of the cooling cycle. This presumably does not happen with the spin-flip ZS of the Paris setup, which mostly delivers atoms in the $|9/2, -9/2\rangle$ state, for which undesired transitions are suppressed by the MOT light polarization. As a result, all three repumpers are essential to optimally capture the $^{53}$Cr Zeeman slowed atoms within our MOT, as summarized in the top row of Table \[Table2\]. Besides $R_1$, we need at least a second repumper to form the $^{53}$Cr MOT, and the removal of the $R_2$ ($R_3$) light causes a three- (two-) fold reduction of the collected atoms. This starkly contrasts with what observed in the Paris experiment [@Chicireanu2006], where removing $R_2$ caused only a $30$% reduction of the MOT atom number, and the effect of $R_3$ was inferred to be negligible. We ascribe this discrepancy to the opposite polarization of the atoms exiting the ZS in the two setups. In particular, even a relatively small residual Doppler shift makes the $R_1$ repumper light resonant with the undesired $F=7/2\rightarrow {F'=7/2}$ transition, further increasing the leak towards higher-lying hyperfine manifolds, and thus making the additional repumpers essential for operating our MOT. This interpretation is supported by noticing that $R_2$ and $R_3$ are irrelevant for the blue MOT, when this is seeded by cold $D$-state atoms trapped in the magnetic quadrupole \[see discussion below\], and characterized by thermal velocities more than $50$ times lower than our ZS exit velocity. Differently from the Paris experiment, we shine the repumper lights onto the MOT via a single retroreflected beam. We have verified that this setup yields a repumping efficiency equivalent to that of a three-beams configuration. Furthermore, with such a scheme one has the advantage that $R_1$, essentially coincident with the $^{52}$Cr cooling light, does not create an undesired bosonic MOT on top of the fermionic one. Another relevant difference in our experimental strategy with respect to the one exploited by the Paris group is the implementation of the HP beam, combined with the TC stage, at the Cr oven output and in absence of external magnetic fields. The increment in the MOT atom number produced by the HP and TC stages is relevant, yielding about a two-fold and five-fold increase of the atomic flux, respectively \[see top row of Table \[Table2\]\]. We have found that both the HP efficiency and the MOT performances are essentially constant within the range of detunings $-1.5\lesssim \delta/\Gamma_{Cr}\lesssim0$ for the three blue repumpers. This allows us to derive the repumper lights for the HP and the MOT beams from the same AOM setup \[see Fig.\[FigCROptSetup\]\]. By contrast, the final temperature of the $^{53}$Cr cloud can be strongly reduced by employing slightly red-detuned repumpers during a final CMOT stage \[see discussion at the end of this section\]. As such, the detunings of $R_1$, $R_2$ and $R_3$ are fixed, throughout the experimental cycle, to the small values listed in the bottom part of Table \[Table2\], which summarizes our optimum light parameters. We now move to discuss how the collected atom number can be substantially increased by employing two additional red repumper lights. Indeed, as already discussed in the previous sections, the MOT transition is not perfectly closed even when all blue repumpers are exploited, due to the leak from the excited $^7P_4$ state towards metastable $^5D$ states, with an overall rate of a few hundred Hz [@Chicireanu2006; @ChicireanuPhD]. The red repumping lights at $663\,$nm and $654\,$nm, addressing the $^5D_{4}\rightarrow {^7P_3}$ and the $^5D_{3}\rightarrow {^7P_3}$ transitions, respectively, bring $D$-state atoms back into the ground state via the $^7P_3$ state, thus closing the cooling cycle. In principle, fermionic Cr, contrarily to the bosonic isotopes lacking hyperfine structure, would require three, rather than two, distinct red repumpers, as atoms excited to the $F=11/2$ hyperfine manifold of the $^7P_4$ level, can decay onto the $F=9/2$ manifold of the $^5D_{3}$ state, and onto the $F=11/2$ and $F=9/2$ manifolds of the $^5D_{4}$ one \[see Fig.\[Fig2\](b)\]. We have experimentally located the three red repumper resonances by scanning the frequency of each of the two lasers, continuously shined onto the atomic cloud. Hitting one resonance is signaled by a sudden increase of the MOT fluorescence, as the red repumper light prevents the accumulation of atoms into the dark $D$-states. The results of this spectroscopic study are in excellent agreement with the findings of the Paris group [@ChicireanuPhD]. Although the shift of about $250\,$MHz between the two $^5D_{4}\rightarrow {^7P_3}$ transitions could be easily bridged by one additional AOM, for simplicity in this work we have not employed the $663\,$nm repumper addressing the $F=9/2$ manifold, as this gives the smallest atom number increase. The impact of the red repumpers on the MOT properties can be noticed by comparing the loading curves recorded with only the blue light on, see Fig.\[CrLoading\](a), and with the addition of the red lights, constantly illuminating the cold sample, see Fig.\[CrLoading\](b). One can see how in the latter case the steady-state MOT population features a three-fold increase, relative to the former one. Parallel to this, while the loading rate $\Gamma_L$ remains unchanged within our accuracy, the $1/e$ characteristic time of the MOT loading curve increases from about $70\,$ms to $200\,$ms. These results, fully consistent with previous observations [@Chicireanu2006], can be understood by considering that the time evolution of the MOT population, under fixed loading conditions, is determined by the concurrent action of radiative decay and light-assisted collisions: $$\begin{aligned} N_{M}^{'}(t) =\Gamma_{L}-\gamma_D N_{M}(t)-\beta^{'}_{MM}N_{M}^2(t). \label{MOTeq}\end{aligned}$$ Here $\gamma_D=\Pi_P\gamma_{P\rightarrow {D}}$ is the decay rate towards the $D$-states, weighted by the relative population of $P$-state atoms $\Pi_P$, which in turn depends on the photon scattering rate within the MOT. The two-body loss term is $\beta^{'}_{MM}=\frac{\beta_{MM}}{V_{MM}}$, where $\beta_{MM}$ is the rate coefficient per unit volume for light-assisted collisions, and $V_{MM}= N_M^2/(\int d\textbf{r} \,n_M^2(\textbf{r}))$ is the MOT collisional volume. This is assumed to be constant in time, an approximation fully validated by early experimental studies [@Bradley2000]. We remark that Eq. disregards two-body losses induced by collisions of MOT atoms with $D$-states ones, which can persist in the MOT region due to the trapping effect of the MOT quadrupole gradient. This approximation is justified by the large mismatch, up to two orders of magnitude, between the large volume occupied by the magnetically trapped $D$-state atoms and the one characteristic of the MOT, as we derive in the following. From Eq. it is straightforward to verify that the steady-state population of the MOT monotonically increases for decreasing $\gamma_{P\rightarrow {D}}$ values, up to the saturation value $\sqrt{\Gamma_{D}/\beta'_{MM}}$ for zero one-body losses, ideally allowed by the presence of red repumper lights. Correspondingly, the typical loading time also increases, owing to the reduced number of loss channels. An accurate analysis of the MOT dynamics goes beyond the scope of the present work. Nonetheless, the results of Fig.\[CrLoading\](a) and (b) allow us to estimate the parameters $\gamma_{P\rightarrow {D}}$ and $\beta_{MM}$, under our working conditions. To this end, we proceed as follows. We first fit the solution of the rate equation to the datapoints of the MOT loading without red repumper light \[Fig.\[CrLoading\](a)\]. The best fit is obtained for vanishing $\beta'_{MM}$, meaning that we are not sensitive to 2-body collisions under these conditions. By fixing $\beta_{MM}$, hence $\beta'_{MM}$, to the literature value $1.5\,10^{-9}$cm$^3\,$/s [@ChicireanuPhD], we obtain the best-fit result $\Pi_P\gamma_{P\rightarrow {D}}=11.7(1)\,$Hz. The collisional volume is calculated from the *in situ* size of the MOT derived from absorption images. From the knowledge of our MOT light absolute intensity within a $30$% uncertainty, we obtain $\Pi_P=0.04(1)$, and hence $\gamma_{P\rightarrow {D}}=290(60)\,$Hz, which is in good agreement with previous findings [@ChicireanuPhD]. We then move to the data taken in the presence of red repumpers \[Fig.\[CrLoading\](b)\], to which we fit the same model. In this case, the one-body loss rate is determined by the decay to the $F=9/2$ manifold of the $^5D_4$ state, which is not addressed by our repumper lights. Based on the relative strength of the repumper transitions, consistently confirmed by our spectroscopy scans, we estimate this decay to be $1/4$ of the total one. By fixing the one-body loss rate to such a value, we obtain $\beta_{MM}=5\,10^{-10}\,$ cm$^3$/s. This is in reasonable agreement with, although slightly smaller than, the value reported in literature [@ChicireanuPhD] for analogous experimental settings. The experimental routine followed to produce the data of Fig.\[CrLoading\](b) allows us to attain steady-state MOT containing up to $1.6 \,10^6$ atoms, a value upper limited by two-body light-assisted losses. A further sizable increase of the population of our cold $^{53}$Cr sample can be obtained by accumulating low-field seeking metastable $D$-state atoms within the magnetic MOT quadrupole. Indeed, also for the fermionic Cr isotope, previous studies [@Chicireanu2006] suggest that $D$-state atoms feature two-body loss rates more than one order of magnitude smaller than the one induced by light-assisted collisions in the MOT. Similarly to what was already done in previous experiments [@Stuhler2001; @Chicireanu2006], we thus load the $^{53}$Cr MOT with no red repumpers \[see Fig.\[CrLoading\](a)\] for a variable time. We then apply a $30\,$ms-long pulse of the red repumpers onto the blue MOT, after which we monitor the atomic cloud through absorption imaging after a short time of flight. This allows us to reveal, besides the MOT cloud, also the population of magnetically trapped $D$-state atoms as a function of time. A typical loading curve, recorded under optimum MOT loading parameters \[see Table \[Table2\]\], is presented in Fig.\[CrLoading\](c). Compared to the MOT one, the latter dynamics features completely different timescales and saturation values, see Fig.\[CrLoading\](a,b). In this case, by fitting the loading curve to a phenomenological exponential rise $N_{\infty} (1- e^{-t/\tau})$, we obtain a $1/e$ time $\tau=11.7(8)\,$s, and a steady-state number of $N_{\infty}=8.6(4)\,10^6$, corresponding to a loading rate of about $\Gamma=7.3(5) \, 10^5\,$atoms/s. Namely, in spite of a six-fold reduced loading rate compared to the MOT one, almost $10^7$ $D$-state $^{53}$Cr atoms can be collected within the magnetic trap, a number about $30$ times larger than the one of our steady-state MOT. We now apply a more quantitative analysis on the results of Fig.\[CrLoading\](c), by exploiting the following model rate equation, coupled to Eq.: $$\begin{aligned} N_{D}^{'}(t) = \eta \gamma_D N_{M}(t) -\tau_D^{-1} N_D(t) -\beta^{'}_{DD} N_{D}^2(t). \label{Deq}\end{aligned}$$ The first term accounts for the seeding of the quadrupole trap from the MOT, and the factor $\eta <1$ represents the fraction of $D$-state Zeeman sublevels that can be trapped by the quadrupole gradient. Magnetically trapped $D$-state atoms undergo one-body losses \[second term of Eq.\] with a rate $\tau_D^{-1}$, and two-body losses \[third term of Eq.\] with a rate $\beta'_{DD}=\beta_{DD}/V_{DD}$. $\beta_{DD}$ is the rate coefficient per unit volume, and $V_{DD}$ the collisional volume of the $D$-state cloud. The absence of the third red repumper during these measurements, hence the not complete repumping of metastable atoms, is taken into account by a scaling factor for the solution $N_{D}(t)$. Also in this case we neglect collisions between MOT and metastable Cr atoms. In order to quantify the inelastic parameters $\beta_{DD}$ and $\tau_D$ we proceed as follows. We first derive the one-body loss rate $\tau_D^{-1}$ by monitoring the decay dynamics of magnetically trapped clouds for two different density conditions, realized under either an optimum or non-optimum loading of the magnetic trap. The latter is obtained by suppressing HF, TC and $R_3$ lights. In both cases, we continuously load the magnetic trap for 4 s, after which we wait for a variable hold time with the MOT lights off. Then, by flashing the red repumpers for 30 ms, we recapture the atoms in the blue MOT, and we acquire an absorption image of the sample in time of flight. An example of such a characterization is shown in Fig.\[Crdecay\] (blue squares). A fit of these data to an exponentially decaying function yields $\tau_D=48(2)\,$s, irrespective of the initial $D$-state cloud density. We remark that such a lifetime is not affected by collisions with hot background Cr atoms, as $\tau_D$ does not vary with the presence or absence of the thermal atomic beam, which can be blocked by an in-vacuum shutter. The extracted $\tau_D$ is significantly larger than the $8\,$s lifetime measured in the Paris experiment [@ChicireanuPhD], and it might be the key factor that allows us to obtain significantly larger $^{53}$Cr samples, containing almost ten times more $D$-state atoms. By fixing $\tau_D$ and $\gamma_D$ to the previously determined values, a fit of the data in Fig.\[CrLoading\](c) to Eq., with only $\eta$ and $\beta'_{DD}$ as free parameters, yields $\eta=0.25(1)$ and $\beta'_{DD}=5.6(2.2)\,10^{-9}\,$s$^{-1}$, respectively. The fitted value of $\eta$ corresponds to an average magnetic moment of $6.2(2)\,$MHz/G for the trapped $D$-state atoms. From this we can estimate the collisional volume $V_{DD}$, and thus obtain the two-body loss rate parameter per unit volume $\beta_{DD}$. Assuming the magnetically trapped atoms to be in thermal equilibrium at a temperature of 1/3 of the MOT one, as shown by previous work on $^{52}$Cr [@Stuhler2001], we derive $\beta_{DD}=5.7(2.3)\,10^{-11}\,$cm$^3$/s, a value close to the one we measured for the bosonic $^{52}$Cr isotope, of $7\,10^{-11}\,$cm$^3$/s [@Neri2019]. We emphasize that the extracted value of $\beta'_{DD}$ is almost three orders of magnitude lower than the light-assisted loss term $\beta'_{MM}$ obtained for the MOT dynamics. This allows us to greatly increase the number of cold $^{53}$Cr atoms, by letting them fall into the $D$ states and accumulate for a few seconds in the magnetic quadrupole, and by recapturing them in the MOT only at the end upon shining the red repumpers, while setting the cooling light at an optimum detuning of $-5.0\,\Gamma_{Cr}$. This procedure enables to maximize the atom number, but it results in relatively high MOT cloud temperatures of $T=530(70)\,\mu$K, about four times higher than the Doppler one, $T_D=124\,\mu$K. In order to decrease the temperature of our sample, we apply a final $10\,$ms CMOT stage, reducing the cooling light detuning to $\delta=-1.4 \,\Gamma_{Cr}$. By varying the cooling intensity applied at this stage, while keeping constant all the other parameters, we have pinpointed the best conditions that minimize the cloud temperature while not causing substantial atom losses, see Fig.\[CMOTCr\]. In particular, for a single-beam MOT intensity of about $1.8\,I_{S, Cr}$, we can reach temperatures of $145(5)\,\mu$K, less than 20$\%$ above the Doppler limit, with a surviving fraction exceeding 75$\%$. The CMOT stage sizeably shrinks the atomic cloud, decreasing its *in situ* $1/e$ radius from $270(30)\,\mu$m down to $170(15)\,\mu$m. Correspondingly, it greatly increases the cloud peak density, reaching values of about $10^{11}\,$atoms/cm$^3$. Our final $^{53}$Cr CMOT temperature is comparable to previously reported values [@Chicireanu2006], whereas the peak densities featured by our clouds are considerably larger than those achieved by the Paris group. EXPERIMENTAL RESULTS: DOUBLE SPECIES {#doublespecies} ==================================== Having detailed our strategies to produce individual Li and Cr cold clouds, we now discuss how the two species, collected within the same MOT region, affect one another. First of all, we have checked that neither the Li $671\,$nm light nor the Li atomic beam affects the Cr sample. Similarly, we have found that application of the Cr $425\,$nm lights, as well as of the red repumpers, do not modify the features of our Li MOT. Moreover, while the simultaneous presence of the two ZS fields modifies the quadrupole center and the position of each MOT, the optimum loading performances can be easily retrieved by slightly realigning the optical beams. We then move to inspect the mutual effect of Li and Cr atoms in the cold regime, when simultaneously collected within a double-species MOT. In particular, since the number of Li atoms is typically two orders of magnitude higher than the Cr one, we focus on the loading dynamics of the chromium sample. Although, in practice, the two clouds can be easily displaced one from the other by tweaking the polarization and alignment of the beams, we intentionally investigate the Cr cloud, loaded into the MOT and within the magnetic trap, under the worst conditions possible, namely once its overlap with a Li MOT, of roughly $10^8$ atoms at mK temperatures, is maximized. We have first monitored the Cr MOT loading, both without and with red repumpers, in the presence of an underlying Li cloud. No appreciable reduction of our Cr MOT loading efficiencies is observed in either case. This fact represents a major advantage of our novel biatomic system: in contrast with the case of other available ones \[see, e.g., Ref.[@Vaidya2015] for the Rb-Yb case\], we can conveniently perform a simultaneous loading of the two species within the same spatial region. This automatically provides an optimum starting point for the subsequent optical trapping and evaporation stages, and it considerably shortens the experimental cycle. Finally, we discuss how the presence of the Li cloud affects the stability of magnetically trapped samples of $D$-state Cr atoms. Also for this measurement, we intentionally put ourselves in the least favourable conditions, by first loading the Li MOT centered on the magnetic quadrupole, and then by collecting a relatively small cloud, of less than two million metastable $^{53}$Cr atoms, within the subsequent three seconds. Once the Cr cloud has been produced in the magnetic trap, we turn off the Cr lights, wait for a variable time, after which we turn on both red repumpers and blue MOT beams for $30\,$ms, and finally apply a 3-ms-long CMOT stage on chromium. We then let both the Li and Cr clouds ballistically expand for $0.6\,$ms, after which we acquire two consecutive absorption images. The resulting decay curve for Cr is shown in Fig.\[Crdecay\], red circles, and it is contrasted with the dynamics of Cr atoms in the absence of the Li MOT, blue squares, obtained through the same routine while keeping the Li oven shutter closed. By comparing the two data sets, one can see how the Li MOT causes a moderate but detectable decrease, of about 25$\%$, of the quadrupole population loaded after the $3\,$s loading. Parallel to this, the $1/e$ lifetime of the magnetically trapped cloud drops, although not dramatically, from $48(2)$ in the single species case, down to $13.0(7)\,$s when the Li MOT is present. We remark that two-body losses connected with intra-species collisions between $D$-state atoms are negligible for these Cr densities: identical decay times, in absence of Li atoms, are indeed observed for a Cr cloud with a ten-fold lower atom number \[not shown in the figure\]. The Li sample features a decay with a typical lifetime of about $25\,$s, unaffected by the presence of the Cr cloud. Although we cannot exclude other loss mechanisms, we ascribe the observed reduction of the Cr lifetime to multiple spin-exchange collisions between $D$-state Cr with Li atoms. These processes, which release energies on the order of a few hundreds of $\mu$K, do not cause Li losses from the MOT, but promote the Cr atoms towards high-field seeking Zeeman sublevels of the $^5D_4$ and $^5D_3$ states, causing their loss from the magnetic trap. By applying the same model as the one exploited for Cr alone, see again Eq., to the Cr decay data in Fig.\[Crdecay\], we extract a value of $\beta'_{Cr,Li}=9.8(4)\,10^{-10}\,$Hz for the inter-species two-body loss rate. With a collisional volume estimated from the in situ images of the lithium cloud and from the volume of $D$-state Cr atoms derived in Section \[singlespecies\] , this result yields a value of $\beta_{Cr,Li}=5.0(3)\,10^{-11}\,$cm$^3$/s for the rate coefficient per unit volume. Our study indicates that the protocols devised for optimum Cr loading, relying on the accumulation of $D$-state atoms into the MOT quadrupole, can be suitably employed also to produce cold Cr-Li mixtures. In particular, we have verified that a sequential loading of the two species \[both with the Li MOT produced first, and vice versa\] leads to samples analogous to those obtained with the simplest simultaneous loading scheme, which enables to create clouds of about $1.5\,10^8$ Li and $4 \, 10^6$ Cr atoms on a $5\,$s-long experimental cycle. CONCLUSIONS AND OUTLOOK {#outlook} ======================= In conclusion, we have produced a novel mixture of $^6$Li and $^{53}$Cr fermionic atoms in the cold regime. For lithium, our setup enables to realize clouds of more than 10$^8$ atoms at sub-Doppler temperatures within a few seconds cycle time. Most importantly, we have demonstrated the possibility to realize comparatively large clouds of several millions of $^{53}$Cr near the Doppler temperature, by relying on the accumulation of dilute samples of metastable $D$-state atoms within the MOT magnetic quadrupole, followed by a final CMOT stage. The presence of a large cloud of about 10$^8$ Li atoms, overlapped with the Cr one, does not dramatically compromise the loading performance observed for the single Cr gas, allowing for a simultaneous loading of the two species with a fast cycle time. This appears as a very promising starting point to implement evaporative cooling of $^6$Li and sympathetic cooling of $^{53}$Cr atoms within an optical dipole trap, following a strategy conceptually analogous to the one already successfully implemented for the $^6$Li-$^{40}$K mixture [@Spiegelhalder2010]. In particular, we expect the small size of the Cr MOT to enable a very large collection efficiency [@Naylor2015] within our recently realized optical trapping potential [@Simonelli2019]. Although the scattering properties of the Cr-Li system are totally unknown, yet, as mentioned already in Section \[intro\], the complex hyperfine and Zeeman structure of chromium, combined with its highly magnetic character, is expected to guarantee a relatively rich spectrum of heteronuclear Feshbach resonances [@Simonipriv]. These could be employed to enhance the interspecies scattering cross section, and thus to optimize the sympathetic cooling efficiency even when small background Cr-Li scattering lengths were found. Finally, we also emphasize that the realization of large and cold $^{53}$Cr clouds, demonstrated in this work, may also enable, in the future, to evaporatively cool down to quantum degeneracy binary spin mixtures of fermionic chromium, with no need to rely on inter-species collisions with a second atomic component. We thank T. Pfau, B. Laburthe-Tolra, L. Vernac and E. Maréchal for insightful discussions and for sharing their know-how on chromium. We acknowledge A. Simoni and D. Petrov for stimulating discussions and for sharing their data on the two- and three-body properties Cr-Li mixtures. Special thanks to F. DiNoia, A. Cosco, M. Seminara and M. Jag for their contribution in the early stage of the experiment, and to the whole Quantum Gases Group at LENS, in particular G. Roati, for the continuous support. This work was supported by the ERC through grant no.637738 PoLiChroM and by the Italian MIUR through the FARE grant no.R168HMHFYM. [75]{}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 [****, ()](\doibase 10.1088/0034-4885/73/11/112401) [****,  ()](http://stacks.iop.org/0034-4885/73/i=7/a=076501) , ed., @noop [**]{}, , Vol.  (, ) [****,  ()](\doibase 10.1126/science.1177112) [****,  ()](\doibase 10.1103/PhysRevLett.108.240404) [****,  ()](\doibase 10.1038/nphys4108) [****,  ()](\doibase 10.1103/PhysRevLett.121.253602) [****,  ()](\doibase 10.1103/PhysRevA.101.013603) [****,  ()](\doibase 10.1038/nature14223) [****,  ()](\doibase 10.1126/science.aag1430) [****,  ()](\doibase 10.1126/science.aag1635) [****,  ()](\doibase 10.1126/science.aag3349) @noop [****,  ()]{} [****,  ()](\doibase https://doi.org/10.1016/0375-9474(73)90510-1) [****,  ()](\doibase 10.1103/PhysRevLett.105.223201) [****,  ()](\doibase 10.1103/PhysRevLett.118.083002) [****,  ()](\doibase 10.1088/0953-4075/40/7/011) [****,  ()](\doibase 10.1007/s00601-011-0229-6) [****, ()](\doibase 10.1103/PhysRevLett.101.170401) [****,  ()](\doibase 10.1103/PhysRevLett.103.153202) [****,  ()](\doibase 10.1088/0034-4885/75/4/046401) @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} [****,  ()](\doibase 10.1103/PhysRevLett.94.017001) [****,  ()](\doibase 10.1103/PhysRevLett.97.100404) [****, ()](http://dx.doi.org/10.1038/nphys520) [****,  ()](\doibase 10.1103/PhysRevA.78.033620) [****, ()](\doibase 10.1103/PhysRevA.82.013624) [****,  ()](\doibase 10.1103/PhysRevLett.103.060403) [****,  ()](\doibase 10.1103/PhysRevB.86.174518) [****, ()](\doibase 10.1103/PhysRevLett.99.130407) [****,  ()](\doibase 10.1103/PhysRevLett.100.090405) [****,  ()](\doibase 10.1103/PhysRevA.83.053625) @noop [****,  ()]{} [****,  ()](http://stacks.iop.org/0034-4885/77/i=3/a=034401) [****, ()](\doibase 10.1103/PhysRevLett.109.065301) [****,  ()](\doibase 10.1103/PhysRevA.87.053602) [****,  ()](\doibase 10.1103/PhysRevLett.106.205304) [ ()](http://arxiv.org/abs/1912.04874) [****,  ()](\doibase 10.1103/PhysRevLett.100.010401) [****,  ()](\doibase 10.1103/PhysRevLett.100.053201) [****,  ()](\doibase 10.1103/PhysRevA.98.063624) [****, ()](\doibase 10.1038/nphys287) [****,  ()](\doibase 10.1103/PhysRevLett.108.043201) [****,  ()](\doibase 10.1140/epjd/e2010-10591-2) [****,  ()](\doibase 10.1103/PhysRevLett.106.115304) [****,  ()](http://dx.doi.org/10.1038/nature11065) [****,  ()](\doibase 10.1103/PhysRevLett.115.135302) [****,  ()](\doibase 10.1126/science.aaf5134) [****,  ()](\doibase 10.1103/PhysRevLett.112.075302) [****,  ()](\doibase 10.1103/PhysRevA.94.062706) [****,  ()](\doibase 10.1140/epjd/e2011-20071-x) [****,  ()](\doibase 10.1007/s10909-018-2104-z) @noop [ ()]{},  [****, ()](\doibase 10.1103/PhysRevLett.109.103002) [****,  ()](\doibase 10.1103/PhysRevA.67.010703) [****,  ()](\doibase 10.1103/PhysRevLett.106.050402) [****,  ()](\doibase 10.1103/PhysRevA.73.053406) [****,  ()](\doibase 10.1103/PhysRevA.91.011603) @noop () @noop ,  (),  @noop @noop ,  (),  [****,  ()](\doibase 10.1103/PhysRevA.90.043408) @noop ,  (),  [****,  ()](\doibase 10.1103/PhysRevA.87.063411) @noop ,  (), [****,  ()](\doibase 10.1103/PhysRevA.61.053407) [****,  ()](\doibase 10.1103/PhysRevA.64.031405) [****,  ()](\doibase 10.1103/PhysRevA.92.043604) [****,  ()](\doibase 10.1103/PhysRevA.81.043637) [****,  ()](\doibase 10.1364/OE.27.027215)
{ "pile_set_name": "ArXiv" }
--- abstract: '$\rm LaCoO_3$ (LCO) nanoparticles were synthesized and their magnetic and structural properties were examined using SQUID magnetometery and neutron diffraction. The nanoparticles exhibit ferromagnetic long-range order beginning at $T_C \approx 87$ K that persists to low temperatures. This behavior is contrasted with the ferromagnetism of bulk LCO, which also starts at $T_C \approx 87$ K but is suppressed below a second transition at $T_o \approx 37$ K. The ferromagnetism in both systems is attributed to the tensile stress from particle surfaces and impurity phase interfaces. This stress locally increases the Co-O-Co bond angle $\gamma$. It has recently been shown that LCO loses long-range ferromagnetic order when $\gamma$ decreases below the critical value $\gamma_c=162.8$$^\circ$. Consistent with this model, we show that $\gamma$ in nanoparticles remains larger than $\gamma_c$ at low temperatures, likely a consequence of all spins being in close proximity to surfaces or interfaces.' author: - 'A. M. Durand' - 'D. P. Belanger' - 'F. Ye' - 'S. Chi' - 'J. A. Fernandez-Baca' - 'C. H. Booth' - 'M. Bhat' bibliography: - 'magnetism.bib' title: 'Magnetism in nanoparticle $\rm LaCoO_3$' --- The magnetism of $\rm LaCoO_3$ (LCO) is well-known to be unusual. Both ferromagnetic and antiferromagnetic interactions have been proposed in the bulk material. [@g58; @akg01] The dominant interaction above $T=100$ K is antiferromagnetic; however, in our previous study, we did not find evidence of antiferromagnetic ordering upon decreasing $T$. [@dbbycfb13] We thus concluded that the antiferromagnetic interactions are strongly frustrated. We have also recently shown [@dbbycfb13] that weak ferromagnetic order occurs below $T_C=87$ K in small applied fields, $H \le 100$ Oe, yet is lost below $T_o \approx 37$ K. Studies of LCO nanoparticles [@fmmpwtvhvg08; @zhzgzs09; @zszhyz07; @wzwf12] and thin films [@fpssnssmrl07; @hrsd09; @fapssl08; @srkkmsw12; @fdaeaeksgl09; @fms08; @prkkmsfw09; @pfmwalsns08; @pbsskzwld11; @rhsd10] also show the presence of a ferromagnetic phase transition near $87$ K, though these materials do not exhibit the same loss of ferromagnetic long-range order seen in bulk LCO below $37$ K. The unusual low temperature magnetism in the bulk has been shown to correlate well with the behavior of the Co-O-Co bond angle, $\gamma$. Experimental measurements [@dbbycfb13] indicate that ferromagnetism only exists when $\gamma$ is greater than a critical value $\gamma_c= 162.8^\circ$, which happens only for $T>T_o$. Above $T_o$, all of the lattice parameters, including $\gamma$, show power-law behavior in $T-T_o$. For $T<T_o$, the lattice parameters show only a small, linear $T$ dependence. The existence of ferromagnetism only for $\gamma > 162.8^\circ$ is consistent with LCO thin film studies. [@fapssl08] Furthermore, recent band-structure calculations [@lh13] indicate that LCO is magnetic only for $\delta y < 0.52$, where $\delta y$ is a measure of the rhombohedral distortion of the lattice. For the measured LCO lattice parameters, $\gamma_c$ corresponds to $\delta y = 0.53$, which indicates excellent agreement between the calculations and experiments for bulk LCO. Thin film studies  [@fdaeaeksgl09; @rhsd10; @pbsskzwld11; @prkkmsfw09; @fms08; @srkkmsw12; @fapssl08; @hrsd09; @fpssnssmrl07; @spd12] have also found supporting evidence for a lattice distortion causing ferromagnetism. LCO material deposited on substrates that result in tensile strain show ferromagnetic order below $T \approx 87$ K, and the strength of the net ferromagnetic moment increases with the value of $\gamma$. Based on the results of experiments on bulk LCO and LCO films grown on various substrates, a model has been proposed in which surface-induced lattice stress increases $\gamma$ near the surfaces. [@dbbycfb13] This induces a transition to long-range ferromagnetic order at $T_C$ throughout the LCO lattice. For moments far from the surfaces, $\gamma$ becomes lower than $\gamma_c$ below $37$ K in bulk LCO. At these low temperatures the lattice loses ferromagnetic order, except near the surfaces where $\gamma > \gamma_c$. We extend our previous study on bulk LCO to nanoparticles using the same neutron scattering and magnetometry techniques. The lattice parameters for both materials were determined over a range of temperatures $10 \le T \le 300$ K, and the field-cooled and zero-field-cooled magnetization was tracked in fields of $20$ Oe and $60$ kOe. We show that nanoparticles of LCO exhibit a phase transition very close to the transition temperature found in bulk LCO. However, the magnetic order persists to low $T$, sharply contrasting the bulk behavior where the ferromagnetism collapses below $37$ K. The ferromagnetic moment is much larger than in the bulk particles and continues to increase as the temperature decreases. We will show that these behaviors can be understood with the same Co-O-Co bond angle and surface-induced distortion model developed for the bulk behavior. LCO nanoparticles were synthesized using the amorphous heteronuclear complex DTPA as a precursor, [@ftyjyc00] using a method similar to that described in Ref. [@jbsbamz09]. A NaOH solution at 1.0 M concentration was added by drops to a mixture solution of $\rm La(NO_3)_3$ and $\rm Co(NO_3)_3$ to prepare fresh hydroxides. A stoichiometric amount of NaOH was used to ensure complete reaction of the metal cations. The excess Na ions were then removed via dialysis over approximately 24 hours. Equimolar amounts of DTPA were then added to the metal hydroxides to synthesize the complex precursor. The mixture was stirred as it was heated to 80$^o$C. The resulting transparent solution was vaporized slowly at 80$^o$C until a dark purple resin-like gel formed. This precursor was decomposed in air at 350$^o$C for 1.5 hours to burn off the organic components. The resulting ash-like material was then heated at a calcination temperature of 620$^o$C for 4 hours. Zero-field neutron diffraction measurements were carried out with the US/Japan wide-angle neutron diffractometer (WAND) at the Oak Ridge National Laboratory High Flux Isotope Reactor with $\lambda$ = 1.48 Å  using vanadium sample cans. Rietveld refinements were performed using FullProf. [@r90] Small-angle x-ray scattering (SAXS) measurements to determine nanoparticle size and x-ray diffraction measurements were performed using a Rigaku SmartLab x-ray diffractometer. The magnetization, $M(T)$, was measured for $H \le 60$ kOe using a Quantum Design SQUID magnetometer. ![Neutron diffraction intensity vs $2\theta$ with FullProf refinements using the R$\overline{3}$c perovskite structure and differences for nanoparticles (upper) and LCO bulk (lower) at $T=10$ K. Nanoparticles show impurity peaks including those corresponding to $\rm Co_3O_4$ (box), which were fitted. A small amount of $\rm Co_3O_4$ was observed for the bulk and a small peak was observed at the antiferromagnetic position of $\rm CoO$ (circle). \[fig:LCOnanobulk\_refine\] ](LCOnanobulk_Fullprof_fig_noCoO.png){height="3.4in"} The neutron diffraction intensity is shown versus $2\theta$ for the bulk and nanoparticle powders in Fig. \[fig:LCOnanobulk\_refine\] at $T = 10$ K. Refinements were more difficult for the nanoparticles than for the bulk, likely due to a distribution of small, non-spherical particles which results in wider peaks that are difficult to model precisely. Noticeable Bragg peaks correspond to impurity phases that result from the low firing temperature used to create small LCO particles. The impurity phases will be addressed further below. SAXS measurements yield an average agglomerate particle size of 75 nm with standard deviation of 21 nm. A Debye-Scherrer analysis of the x-ray diffraction data yields LCO nanoparticle crystal sizes between 25-40 nm. Hence, we conclude that the LCO nanoparticle crystals agglomerate into larger particles during the heating process. No evidence of Bragg magnetic peaks in bulk or nanoparticles has been observed in experiments at WAND, which may be a consequence of the weak net FM moment. The diffraction patterns for both the bulk and nanoparticle LCO show small peaks corresponding to a $\rm Co_3O_4$ phase. The CoO phase observed in the bulk was not evident in the nanoparticles. Refinements indicate a weight fraction of $< 3.5\%$ $\rm Co_3O_4$ in the nanoparticles. The effect of these phases on the bulk has been previously discussed [@dbbycfb13] and it applies similarly to the nanoparticles. In addition to the $\rm Co_3O_4$ phase, there are also some Bragg peaks that were not identified as expected oxides of La or Co. When comparing the bulk to the nanoparticle patterns, the lattice structures of these phases appear distinct from that of the LCO bulk. Unless there were a significant amount of impurity peaks directly overlapping the LCO peaks, we would not expect them to significantly affect the quality of the LCO parameter refinements. The magnetization data do not show any unusual behavior which can be readily attributed to these impurity phases, although calculations of the average effective Co moment may reflect the presence of non-LCO phases. Fitted lattice parameters $a$ and $c$ for the hexagonal unit cell are shown for $0<T<300$ K in Fig. \[fig:LCObulknano\_params\] and are compared to previous results obtained for bulk LCO. The Co-O-Co bond angle ($\gamma$), and the parameter $\delta y = \frac{d}{a}\cos(\gamma/2)$, where $d$ is the Co-O bond length, are also shown. The parameter $\delta y$ describes the deviation of the oxygen position from the straight line connecting neighboring Co ions and characterizes the amount of rhombohedral distortion of the lattice. [@rkfk99; @dbbycfb13; @mkfoiyk99] Although $\delta y$ is a derived quantity from $a$ and $\gamma$, we include it in the fits so as to be consistent with other relevant papers which interpret their results in the context of $\delta y$. [@lh13; @dbbycfb13] ![Lattice parameters $a$, $c$, $\delta y$, and the Co-O-Co angle ($\gamma$) for LCO nanoparticles and bulk powder. Solid curves for the bulk represent power-law behavior in ($T-T_o$) and dashed lines linear behavior. The solid curves for the nanoparticles are fits using the Gruneisen expression with the Einstein model power-law behavior is not apparent and $\gamma > \gamma_c$ for all $T$ for the nanoparticle case. \[fig:LCObulknano\_params\] ](LCOBulknano_params_4fig_Grunpower.png){height="3.4in"} Contrasting the sharp changes observed near $T_o \approx 37$ K in bulk LCO, $\delta y$, $\gamma$, $a$ and $c$ show no abrupt change in slope in the nanoparticles. The value of $\gamma$ observed for the nanoparticles remains well above $\gamma_c$ throughout the entire range of temperatures studied. [l\*[5]{}[c]{}r]{} Parameter & $y(0) $ && $\alpha$ && $T_E (\rm K) $\ a & 5.405(4) && 2.7${\times 10^{-4}}$ && 50.0\ c & 12.966(1) && 7.3${\times 10^{-4}}$ && 58.6\ $\delta y$ & 0.0512(1) && -7.5${\times 10^{-3}}$ && 98.7\ Co-O-Co ($\gamma$) & 163.384(4) && 8.0${\times 10^{-4}}$ && 96.1\ The temperature dependence of the nanoparticle lattice parameters were fit using the Gruneisen expression with the Einstein model for thermal lattice expansion, $$y(T) = y(0)[1 + \alpha(\coth(\frac{T_E}{2T}) - 1)] \quad , \label{eq:Grun}$$ where the lattice parameter being fit is $y(T)$, $y(0)$ is its value at $T=0$, $T_E$ is the Einstein temperature, and $\alpha$ is the thermal expansion coefficient for $T >> T_E$. [@crbmrsrfg99] The lattice parameters are well fit by the thermal expansion model, but $T_E$ is rather small and inconsistent among the various parameters. It is not surprising that data for lattice parameters that exhibit no sharp features can be fit by the simple temperature dependence of Eq. \[eq:Grun\]. No power-law behavior or phase transition is observed for nanoparticles, unlike the bulk LCO. $M/H$ vs $T$ for the $\rm LaCoO_3$ bulk and nanoparticles at $H=20 \pm 1$ Oe is shown in Fig. \[fig:LCO\_bulknanomag\] a) and c). The nanoparticle magnetic phase transition occurs at nearly the same temperature as the bulk, indicating that the magnetism in both materials is of the same origin. The field-cooled (FC) and zero-field-cooled (ZFC) behaviors for both materials diverge just below $T_C$. The nanoparticles show a much larger FC moment and a much larger difference between the FC and ZFC behaviors. The magnetization below a FM phase transition is expected to exhibit the power law behavior $t^\alpha$ as the individual moments align with each other, where $t=(T_C-T)/T_C$ ($t << 1$) and $\alpha < 1/2$. Hence, we normally expect significant curvature below $T_C$. However, the net moment in bulk LCO shows little curvature ($\alpha \approx 1$). This likely reflects the very small net FM moment as well as either a decrease in the local moment size or the strength of the interaction between moments as $\gamma$ decreases with temperature towards $T_o$. Although the nanoparticle magnetization exhibits greater curvature with a value $\alpha \approx 0.8$ for fits over the range $32 < T < 84$ K, the curvature is still not as large as normally expected for a FM phase transition. The nanoparticle moments and interactions vary less than in the bulk, but the net FM moment remains small (though much larger than in the bulk). Overall, the rounding of the transition near $T_C$ and significant non-critical contributions well below $T_C$ preclude meaningful fits to the critical behavior. The shapes of the ZFC curves in the bulk and nanoparticles show a similar paramagnetic-like tail below $T=20$ K, except that the nanoparticle moment appears to saturate at $0.054$ $\rm \frac{emu}{Oe \cdot mol}$. Although the individual nanoparticles are fairly well ordered, there are likely moments at grain boundaries that contribute to the paramagnetic tail. The nanoparticle ZFC magnetization between $T=20$ K and $T_C$ is much larger than that of the bulk and probably represents randomly oriented net ferromagnetic moments of the nanoparticle grains which are not readily aligned by the applied field. ![$M/H$ vs. $T$ for nanoparticle and bulk LCO. The magnetic behavior is shown in a) and c) for $H=20 \pm 1$ Oe and b) and d) for 60 kOe. The ZFC data are shown by open symbols and the FC data by closed symbols. ZFC and FC data overlap closely for the 60 kOe measurements. The inset expands the nanoparticle ZFC data. Note the different vertical scales for the bulk and nanoparticles, and in c) the 0 is offset for clarity. The behaviors for $T>100$ K are similar in magnitude and field independent for all cases. \[fig:LCO\_bulknanomag\] ](LCO_bulknano_mag_new.png){height="3.4in"} The $M/H$ nanoparticle behavior for $H=60$ kOe sharply contrasts that of the bulk particles. Whereas the bulk magnetization decreases with $T$ below $90$ K, with a minimum just below $40$ K (Fig. \[fig:LCO\_bulknanomag\] b), the nanoparticle moment increases monotonically as $T$ dereases over the entire temperature range (Fig. \[fig:LCO\_bulknanomag\] d). The transition is rounded for $H=60$ kOe because the critical point is at $H=0$. However, there is a slight inflection point near $T_C$, which is more clear from $H/M$ versus $T$ data shown in Fig. \[fig:LCObulk\_invMH\]. This indicates a strong influence of the ferromagnetic interactions in high field. The bulk data indicate no such signifcant FM interaction at high fields, consistent with the small net ferromagnetic moment observed in the bulk and the disappearance of the ferromagnetism below the transition at $T_o$. ![$H/M$ vs. $T$ for nanoparticle LCO at $H=20$ Oe and $60$ kOe. The linear region in both fields is fit to Curie-Weiss behavior (dashed line). \[fig:LCObulk\_invMH\] ](LCOnano_invMH_fig.png){width="2.9in"} Figure \[fig:LCObulk\_invMH\] shows $H/M$ versus $T$ for the nanoparticles at $H=60$ kOe and $20$ Oe. The steep drop in $H/M$ with decreasing temperature beginning at $T = 100$ K is indicative of ferromagnetism for $20$ Oe. The slight dip at the same temperature for $60$ kOe demonstrates that the ferromagnetic interactions are still significant. Data at both fields show nearly the same paramagnetic behavior above $170$ K. The straight line fit for $170 \leq T \leq 300$ K is interpreted as Curie-Weiss (CW) behavior using $H/M=1/\chi = (T-\theta_{CW})/C$. We obtain a CW temperature $\theta_{CW}=-121(3)$ K and Curie constant $C=1.14(1)$ emu$\cdot$K$\cdot$mol$^{-1}$ for the nanoparticles. This value of $\theta_{CW}$ is consistent with $T_C=87$ K (CW calculations yield upper limits for transition temperatures) and suggests dominant antiferromagnetic interactions above $170$ K. Both values are similar to those obtained for bulk LCO, where $\theta_{CW} = -138(4)$ and $C = 1.31(2)$. Above the magnetic transition at $87$ K, the effective moment is $\mu_{eff} = 3.03(2)$ $\mu_{B}$ per Co ion. This is smaller than the bulk value of $3.23(2)$ $\mu_B$. The less negative $\theta_{CW}$ and smaller $\mu_{eff}$ perhaps indicate that fewer moments are participating due to the less perfect crystallinity of the nanoparticle powder compared to the bulk. The impurities present in the nanoparticles may also contribute to an overestimation of the number of participating Co ions. A larger FM interaction in the nanoparticles could also result in a less negative $\theta_{CW}$. Notably, Fita *et al.* found an even smaller value of $\mu_{eff} = 2.44 \mu_B$ for their LCO nanoparticles, as well as a less negative $\theta_{CW}$ of -48 K. [@fmmpwtvhvg08] In several previous studies, the ferromagnetism in LCO nanoparticles has been attributed to ferromagnetic ordering of the surface, [@yzg04_a; @httski07] surface-induced lattice strain, [@fmmpwtvhvg08] and unit-cell expansion. [@wzwf12; @zhzgzs09] Yan *et al.* [@yzg04_a] found that the magnetic susceptibility of their samples increased as the surface-to-volume ratio increased and attributed this to localized spins on the surface of the material. Although they ruled out double exchange between Co(III) and Co(IV) atoms as the mechanism for the surface ferromagnetism, their study was inconclusive as to the origin of the ferromagnetic interaction. Harada *et al.* [@httski07] found similar results; the magnetization increased with decreasing particle size and they suggested the source to be chemisorption of oxygen atoms at the surface. Again, the mechanism leading to ferromagnetism was not made explicit. Fita *et al.* [@fmmpwtvhvg08] examined the lattice parameters of LCO nanoparticles and found that they increase with decreasing particle size but did not identify surfaces as the source of ferromagnetism. Instead, they pointed to the surface-induced lattice expansion which persists throughout the material as the cause. In the above studies, LCO crystals and powders were synthesized using several different methods: floating-zone single-crystal synthesis, solid-state reaction, crushing the single-crystal into a powder, and a citric acid solution method. It is interesting to note that in all but one case, the magnetic susceptibility increased monotonically with surface area and the transition temperature remained fairly constant at $T_C \approx 85$ K. Wei *et al.* [@wzwf12] found that their nanoparticles showed a decrease in magnetization and $T_C$ with particle size, despite also noting a similar increase in lattice parameters with the smaller sizes. It is possible that the citric acid sol-gel method employed in this case differed slightly from the one used by Harada *et al.*, [@httski07] but it is unclear how that would produce results that differ qualitatively from the others. Tensile stress from substrates allows LCO thin films to order ferromagnetically. The resulting FM order is found throughout the material for films of order 100 nm [@fdaeaeksgl09] and the net moment increases with film thickness. [@fpssnssmrl07] In nanoparticles, all of the LCO material is well within 100 nm of a surface. Hence, tensile stress from these surfaces should result in a larger average $\gamma$ for the particles, although the exact process by which the strain changes $\gamma$ requires further investigation. Our experiments show that $\gamma$ does remain well above $\gamma_c$ for all $T$ (Fig. \[fig:LCObulknano\_params\]), which is consistent with the net FM moment being much larger than in bulk LCO and not collapsing below $T=37$ K. Note that the lowest value of $\gamma$ in the nanoparticles is comparable to the bulk value for $T \approx 200$ K, so we would expect the system to remain ferromagnetic for $T<87$ K in small $H$. This model, which correlates the stability of ferromagnetism with the stress-induced increase in $\gamma$, is consistent with the observations made by most of studies noted above. The ferromagnetic transition at $T=87$ K is a result of the stress at surfaces, but long-range ferromagnetic order takes place throughout the LCO lattice. Only for bulk LCO is the long-range order observed to disappear below $T=40$ K because only in this system is $\gamma$ observed to decrease below $\gamma_c$. Although $\gamma$ is a useful parameter by which to gauge the degree of magnetism in a sample, the mechanism by which it controls the ferromagnetism remains unclear. Calculations by Lee and Harmon indicate $\delta y$ (closely related to $\gamma$) to be the controlling parameter in LCO and note that the amount of rhombohedral lattice distortion determines whether the ground state is magnetic.[@lh13] According to this model, the degree of orbital overlap between the Co and O ions affects the balance between the repulsive Coulomb interaction and the exchange interaction: more distortion (and less overlap) leads to a non-magnetic ground state. Goodenough proposes that the local configurations of low-spin and high-spin Co$^{3+}$ and Co$^{4+}$ ions result in areas of ferromagnetism, antiferromagnetism, and paramagnetism.[@g58] It is also conceivable in this interpretation that the Co-O-Co bond angle affects the spin state of the Co ions or groups of ions by allowing for increased or decreased overlap of the cobalt and oxyen orbitals. In summary, we have shown that, for $T>87$ K, the interaction strengths and paramagnetic behavior are comparable in bulk and nanoparticle LCO. On the other hand, the observed magnetic behaviors observed for $T<87$ K in LCO bulk and nanoparticles are very different. Nevertheless, the magnetic behaviors of bulk and nanoparticles can be modeled within the same framework; the interactions between spins in LCO are strongly dependent on the Co-O-Co bond angle $\gamma$. Magnetic order can only be sustained when $\gamma$ is larger than a critical value $\gamma_c =162.8$ $^\circ$. In bulk LCO, this occurs only for $T>T_o$, where $T_o \approx 37$ K. For $T<T_o$, magnetism in bulk LCO is associated only with regions of tensile stress near surfaces and interfaces with impurity phases. In LCO thin films and nanoparticles, all moments are near to surfaces so $\gamma$ is always larger than $\gamma_c$ and long-range ferromagnetic order is present for all $T<87$ K. We thank Y. Abdollahian, F. Bridges, C. de la Cruz, A. Elvin, B. Harmon, J. Howe, S. Shastry, and N. Sundaram for helpful discussions and$/$or assistance with measurements. The work at ORNL is supported by the DOE BES Office of Scientific User Facilities. Work at Lawrence Berkeley National Laboratory was supported by the Director, Office of Science (OS), Office of Basic Energy Sciences (OBES), of the U.S. Department of Energy (DOE) under Contract No. DE-AC02-05CH11231. Some X-ray data in this work were recorded on an instrument supported by the NSF Major Research Instrumentation (MRI) Program under Grant DMR-1126845.
{ "pile_set_name": "ArXiv" }
Introduction {#sec:Introduction} ============ Quantum Monte Carlo (QMC) methods are powerful and general tools for calculating the ground-state electronic properties of atoms, molecules, and solids. Since the computational cost increases only as the cube of the number of particles, it is possible to study systems containing hundreds of electrons subject to periodic boundary conditions. This is enough to model real condensed matter with surprising precision, as shown by the accuracy of 0.1 eV per atom or better achieved in QMC calculations of the cohesive energies of solids. By comparison, the errors in local density functional calculations of cohesive energies are often of the order of 1 eV per atom. The two most widely used QMC methods are variational Monte Carlo (VMC) and diffusion Monte Carlo (DMC). [@dmc; @hammond] In VMC a trial many-electron wave function is chosen and expectation values are evaluated using Monte Carlo integration, which is more efficient than grid-based quadrature methods for high-dimensional integrals. Most VMC simulations of solids use trial wave functions containing a number of adjustable parameters, the values of which are determined by minimizing the energy or its variance. DMC is a stochastic method for evolving a solution of the imaginary-time Schrödinger equation. The imaginary-time evolution gradually enhances the ground-state component of the solution relative to the excited-state components, but the algorithm does not maintain the fermionic symmetry of the starting state. The solution therefore converges towards the overall ground state, which is bosonic. This difficulty is known as the sign problem. Although several exact solutions to the sign problem have been proposed, none has the statistical efficiency required to study the large systems of interest to condensed matter physicists. Most DMC simulations therefore use the approximate fixed-node method, [@anderson] which is numerically stable and often very accurate. The details will be described in Sec. \[sec:DMC\], but the basic idea is quite simple. A real trial many-electron wave function is chosen and used to define a trial nodal surface, which is the surface on which the trial function is zero and across which it changes sign. In a three-dimensional system containing $N$ electrons, the trial wave function is a function of $3N$ variables, and the trial nodal surface is $3N-1$ dimensional in general. The fixed-node DMC algorithm maintains the nodal surface of the trial wave function, so enforcing the fermionic symmetry and producing the lowest energy many-electron wave function consistent with that nodal surface. Although VMC and DMC are principally ground-state methods, they can also provide some information about excited states. In particular, they can be used to study the lowest energy state of each distinct symmetry. In VMC this is done by choosing a trial wave function which possesses the required symmetry for all values of the variational parameters. The energy obtained after optimizing the trial function is therefore greater than or equal to the eigenvalue of the lowest energy eigenstate of that symmetry. A similar technique is also used in DMC, although this is much harder to justify. The problem is that the DMC trial function is only used to define the trial nodal surface, which may not be sufficient to fix the symmetry of the state produced by the stochastic DMC algorithm. In any case, practical tests have shown that this approach often gives excellent results. Examples are the study of excitations of the hydrogen molecule by Grimes [*et al.*]{}, [@grimes] and calculations of excitation energies in diamond [@mitas2; @mitas3] and silicon. [@sibands] If the trial function used in an excited-state DMC simulation has no definite symmetry, the only certainty is that the DMC energy must be greater than or equal to the many-electron ground-state energy. [@hammond] In cases when the DMC trial function does have a definite symmetry, however, it is normally assumed that the fixed-node DMC solution has the same symmetry as the trial function, and hence that the DMC energy is greater than or equal to the eigenvalue of the lowest energy eigenstate of that symmetry. This symmetry-constrained variational principle is widely accepted, but we show by constructing a specific example that it is not always correct: the fixed-node DMC solution need not have the same symmetry as the trial function; and the fixed-node DMC energy may be lower than the energy of the lowest exact eigenstate of that symmetry. The symmetry-constrained DMC variational principle is guaranteed to hold only when the trial function transforms according to a one-dimensional irreducible representation of the symmetry group of the Hamiltonian. The corresponding eigenstate is then non-degenerate, or has only accidental degeneracies. If the trial function transforms according to a multi-dimensional irreducible representation, corresponding to a degenerate energy level, the DMC energy may lie below the energy of the lowest eigenstate of that symmetry. In such cases a weaker variational principle may be obtained by choosing a trial function that transforms according to a one-dimensional irreducible representation of a subgroup of the full symmetry group. The DMC energy is then greater than or equal to the eigenvalue of the lowest exact eigenstate with that subgroup symmetry. This provides a strict variational lower bound for the DMC energy, but one that usually lies below the energy of the degenerate eigenstate of interest. As an example, consider the case of a crystalline solid. Any trial function with a definite crystal momentum ${\bf k}$ satisfies the many-electron version of Bloch’s theorem and so transforms according to a one-dimensional irreducible representation of the translation group, which is a subgroup of the full symmetry group. The weaker variational principle therefore guarantees that the DMC energy must be greater than or equal to the energy of the lowest exact eigenstate with crystal momentum ${\bf k}$. Unfortunately, most Bloch states are complex and so cannot be used as fixed-node DMC trial functions. Real linear combinations of Bloch functions and their complex conjugates can be used instead, but in most cases these do not transform according to one-dimensional irreducible representations and do not lead to useful variational principles. This is illustrated in Sec. \[sec:sepexample\], where we show that the DMC energy obtained using such a trial function may lie below the energy of the lowest eigenstate with crystal momentum ${\bf k}$. The weaker variational principle is useful, but relies on a very careful choice of trial functions and cannot explain all the past successes of the fixed-node DMC method for excited states. The real explanation of these successes, we believe, is that although the DMC algorithm does not always preserve the symmetry of the trial function, the imposed nodal surface acts as such a strong restriction that the DMC solution cannot stray “too far” from that symmetry. The calculated energy is therefore close to the variational value that would have been obtained if the symmetry had been preserved. In cases when the excited state of interest satisfies the strong variational principle, the errors in the ground and excited state energies are guaranteed to have the same sign and tend to cancel, so improving the accuracy of the calculated energy difference. The rest of this paper is organized as follows: Sec. \[sec:DMC\] contains brief explanations of the DMC method and the fixed-node approximation for ground states; Sec. \[sec:GenVar\] shows that no general symmetry-constrained variational principle exists; Sec. \[sec:Lowest\] shows that a variational theorem holds for the lowest energy state of each symmetry provided that the trial function transforms according to a one-dimensional irreducible representation of the group of spatial transformations of the Hamiltonian; Sec. \[sec:d&gt;1\] introduces a weaker variational principle which may give energy bounds even when the trial function transforms according to an irreducible representation of dimension greater than one; Sec. \[sec:tiling\] shows that a generalization of the tiling theorem [@tiling] holds in every case when we can prove that the DMC energy obeys a variational bound; Secs. \[sec:tbexample\] and \[sec:sepexample\] give examples which demonstrate the absence of a general symmetry-constrained variational principle and illustrate the application of the weaker variational principle; and Sec. \[sec:Conclusions\] summarizes and concludes. Fixed-Node Diffusion Monte Carlo for Ground States {#sec:DMC} ================================================== In this section we summarize the known results concerning the application of the fixed-node DMC method to ground states. [@dmc; @hammond] The aim is to evaluate expectation values with an antisymmetric wave function $\Phi(X)$, where $X \equiv (x_1, x_2, \ldots, x_N)$ lists the coordinates of all $N$ electrons, and $x_i = ({\bf r}_i, s_{zi})$ specifies the position and spin projection of electron $i$. We choose wave functions with a fixed total $S_z = \sum_{i=1}^{N} s_{zi}$. The expectation value of a spin-independent symmetric operator $\hat{A}({\bf R})$ is given by $$\label{eq:Aspin} \langle A \rangle = \frac {\sum_S \int \Phi^*(X) \hat{A}({\bf R}) \Phi(X) \, d{\bf R}} {\sum_S \int \Phi^*(X) \Phi(X) \, d{\bf R}} \;,$$ where ${\bf R} \equiv ({\bf r}_1,{\bf r}_2,\ldots,{\bf r}_N)$. For each spin configuration, $S$, the antisymmetric wave function $\Phi(X) = \Phi(x_1,x_2,\ldots,x_N)$ may be replaced by a version with permuted arguments, $\Phi(x_{i_1}, x_{i_2}, \ldots, x_{i_N})$, where the permutation is chosen such that the first $N_{\uparrow}$ arguments are spin-up and the last $N_{\downarrow}=N-N_{\uparrow}$ are spin-down. Since ${\bf R}$ is a dummy variable, we can relabel $({\bf r}_{i_1},{\bf r}_{i_2},\ldots, {\bf r}_{i_N})$ as $({\bf r}_1,{\bf r}_2,\ldots,{\bf r}_N)$, after which the sums over spin configurations can be removed. The expectation value may then be written as $$\label{eq:Anospin} \langle A \rangle = \frac {\int \tilde{\Phi}^*({\bf R}) \hat{A}({\bf R}) \tilde{\Phi}({\bf R}) \, d{\bf R}} {\int \tilde{\Phi}^*({\bf R}) \tilde{\Phi}({\bf R}) \, d{\bf R}} \;,$$ where $$\begin{aligned} \lefteqn { \tilde{\Phi}({\bf r}_1,{\bf r}_2,\ldots,{\bf r}_N) \; \equiv } \nonumber & & \\ & \;\;\; & \Phi({\bf r}_1,\uparrow; {\bf r}_2,\uparrow;\ldots; {\bf r}_{N_{\uparrow}},\uparrow; {\bf r}_{N_{\uparrow}+1},\downarrow; \ldots;{\bf r}_{N},\downarrow) \;.\end{aligned}$$ The function $\tilde{\Phi}({\bf r}_1,\ldots,{\bf r}_{N})$ is antisymmetric under interchange of any two of the up-spin arguments ${\bf r}_1,\ldots,{\bf r}_{N_{\uparrow}}$ or any two of the down-spin arguments ${\bf r}_{N_{\uparrow}+1},\ldots,{\bf r}_{N_{\uparrow}+N_{\downarrow}}$, but has no definite symmetry under interchange of up- and down-spin arguments. It still obeys the Schrödinger equation, but the up- and down-spin electrons are now treated as distinguishable, allowing us to avoid explicit reference to the spin variables. For simplicity, all wave functions in the following discussion will be chosen to be of this type. DMC is a method for solving the imaginary-time Schrödinger equation, $$\label{eq:its} \left ( - \frac{1}{2}\nabla_{\bf R}^2 + V({\bf R}) - E_{\rm S} \right ) \Psi({\bf R},\tau) = - \frac{\partial}{\partial \tau} \Psi({\bf R},\tau) \;,$$ where $\nabla_{\bf R}^2$ is shorthand for $\sum_{i=1}^N \nabla_i^2$. The potential $V({\bf R})$ includes the electron-electron interactions as well as the external potential terms, and is assumed to be a local function of ${\bf R}$. It has Coulomb singularities whenever two charged particles approach each other, but is finite everywhere else. The constant energy shift $E_{\rm S}$ has been introduced to set a convenient zero of energy as explained below. The variable $\tau$ (which is real) is usually called the imaginary time, but we will often abbreviate “imaginary time” to “time” in what follows. If the starting state $\Psi({\bf R},\tau=0)$ is written as a linear combination of energy eigenfunctions, $$\label{eq:lincomb} \Psi({\bf R},\tau=0) = \sum_i c_i \Psi_i({\bf R}) \; ,$$ the large $\tau$ limit of the solution of Eq. (\[eq:its\]) takes the form, $$\Psi({\bf R},\tau \rightarrow \infty) = c_l \Psi_l({\bf R}) e^{-(E_l - E_{\rm S}) \tau} \; ,$$ where $\Psi_l({\bf R})$ is the lowest energy eigenfunction appearing in Eq. (\[eq:lincomb\]) and $E_l$ is the corresponding eigenvalue. In applications to many-electron systems, $\Psi({\bf R},\tau=0)$ is antisymmetric and $\Psi_l({\bf R})$ is usually the many-electron ground state $\Psi_0({\bf R})$. We will assume that the Hamiltonian $\hat{H}$ is real (possesses time-reversal symmetry), in which case $\Psi_l({\bf R})$ may also be chosen real and may be obtained by following the imaginary-time evolution of a real function $\Psi({\bf R},\tau)$. If the Hamiltonian is complex (does not possess time-reversal symmetry), as when there is an applied magnetic field, the fixed-node DMC method discussed in this paper does not apply and it is necessary to use the fixed-phase method of Ortiz [*et al.*]{} [@ortiz] DMC solves Eq. (\[eq:its\]) using a stochastic algorithm, the efficiency of which is much improved by an importance sampling transformation. [@reynolds] A real trial wave function $\Phi_{\rm T}({\bf R})$ is chosen and Eq. (\[eq:its\]) is recast in terms of the product $$\label{eq:f} f({\bf R},\tau) = \Psi({\bf R},\tau) \Phi_{\rm T}({\bf R})$$ which is also real. After some straightforward algebra, $f({\bf R},\tau)$ is found to satisfy the equation, $$\begin{aligned} - \frac{\partial}{\partial \tau}f({\bf R},\tau) & = & - \frac{1}{2}\nabla_{\bf R}^{2}f({\bf R},\tau) \; + \; \nabla_{\bf R} \cdot \left[{\bf F}({\bf R})f({\bf R},\tau)\right] \nonumber \\ && \; + \; \left(E_{\rm L}({\bf R}) - E_{\rm S}\right)f({\bf R},\tau) \; , \label{eq:fits}\end{aligned}$$ where $E_{\rm L}({\bf R}) \equiv \Phi_{\rm T}^{-1}({\bf R}) \hat{H} \Phi_{\rm T}({\bf R})$ is known as the local energy and ${\bf F}({\bf R}) \equiv \Phi_{\rm T}^{-1}({\bf R}) \nabla_{\bf R}\Phi_{\rm T}({\bf R})$ as the quantum force. If the chosen trial wave function is close to the exact ground-state wave function, the local energy is nearly constant and the statistical efficiency of the algorithm is optimized. If $f({\bf R},\tau)$ is constrained to be positive, Eq. (\[eq:fits\]) may be interpreted as describing the time evolution of the density of a population of “random walkers” multiplying or dying out as they diffuse and drift through a $3N$-dimensional “configuration space”. The constraint that $f$ is positive is known as the fixed-node approximation, [@anderson] because it forces the nodal surface of $\Psi({\bf R}, \tau)$ to be the same as that of $\Phi_{\rm T}({\bf R})$. In practice, the distribution $f$ is represented by a few hundred walkers propagating stochastically according to rules derived from Eq. (\[eq:fits\]). The growth or decay rate of the total number of walkers in the simulation depends on the average value of $E_{\rm L}({\bf R}) - E_{\rm S}$, and the constant energy shift $E_{\rm S}$ is chosen to ensure that the population remains stable on average. The initial walker positions are normally picked from the probability distribution $f({\bf R},\tau=0) = (\Phi_{\rm T}({\bf R}))^2$, resulting in an initial population scattered throughout the configuration space. The nodal surface of $\Phi_{\rm T}({\bf R})$ divides this space into different nodal pockets, among which the walkers are distributed. During the simulation, the walkers never cross the fixed nodal surface separating one pocket from another, and so the fixed-node DMC algorithm proceeds independently in each pocket. In the long-time limit, the probability density $f$ of the walkers within nodal pocket $v_{\alpha}$ becomes proportional to $\phi_{\alpha}({\bf R})\Phi_{\rm T}({\bf R})$, where the “pocket ground state” $\phi_{\alpha}({\bf R})$ is the lowest energy real normalized wave function which is zero outside $v_{\alpha}$ and satisfies the fixed-node boundary conditions on the surface of $v_{\alpha}$. This function generally has gradient discontinuities across the surface of $v_{\alpha}$, and the action of the kinetic energy operator on these discontinuities produces delta function terms which will be denoted $\delta_{\alpha}$. The pocket ground state therefore satisfies the equations: $$\begin{aligned} \label{eq:eigensoln} \left. \begin{array}{c} \hat{H}\phi_{\alpha}({\bf R}) = \epsilon_{\alpha}\phi_{\alpha} ({\bf R}) + \delta_{\alpha} \\ \phi_{\alpha}({\bf R})\Phi_{\rm T}({\bf R}) \geq 0 \end{array} \right\} & \;\;\; \rm{when} \; {\bf R} \in \; {\it v}_{\alpha} \nonumber \\ \phi_{\alpha}({\bf R}) = 0 \;\;\;\;\;\;\; & \;\;\;\;\; \rm{when} \; {\bf R} \not\in \; {\it v}_{\alpha} \;.\end{aligned}$$ Since the walker density is positive, the sign of $\phi_{\alpha}({\bf R})$ within $v_{\alpha}$ is the same as that of $\Phi_{\rm T}({\bf R})$, implying that $\phi_{\alpha}({\bf R})$ and $\Phi_{\rm T}({\bf R})$ have a non-zero overlap. The form of $\phi_{\alpha}({\bf R})$ does not depend on the details of $\Phi_{\rm T}({\bf R})$, but only on the shape of its nodal surface. The energy given by the DMC simulation in nodal pocket $v_{\alpha}$ is equal to the pocket eigenvalue $\epsilon_{\alpha}$. This need not equal the exact ground-state eigenvalue $E_0$ unless the trial nodal surface is the same as that of the exact ground state $\Psi_0({\bf R})$, in which case $\phi_{\alpha}({\bf R})$ is proportional to $\Psi_0({\bf R})$ within $v_{\alpha}$. These results were used by Reynolds [*et al.*]{} [@reynolds] and Moskowitz [*et al.*]{} [@moskowitz] to prove that the fixed-node DMC energy is greater than or equal to the exact ground-state energy. Reynolds’ proof starts from the solution in a single nodal pocket, and uses the permutations $P$ that do not flip spins to construct a real wave function antisymmetric with respect to interchanges of electron coordinates of the same spin, $$\label{eq:antiwave} \bar{\Psi}_{\alpha}({\bf R}) = \frac{1}{N_{\uparrow}!N_{\downarrow}!} \sum_P (-1)^P \phi_{\alpha} (P{\bf R}) \equiv \hat{\mathcal{A}}\phi_{\alpha}({\bf R})\;.$$ The function $\bar{\Psi}_{\alpha}({\bf R})$ cannot equal zero everywhere since, at any point ${\bf R}$, all terms contributing to the sum in Eq. (\[eq:antiwave\]) have the same sign. This follows from the antisymmetry of the trial function $\Phi_{\rm T}({\bf R})$, which guarantees that all permutations $P$ for which the point $P{\bf R}$ lies in nodal pocket $v_{\alpha}$ have the same parity (positive if the sign of the trial function at ${\bf R}$ is the same as the sign of the trial function in $v_{\alpha}$, negative otherwise). The function $\bar{\Psi}_{\alpha}({\bf R})$ is however equal to zero everywhere on the trial nodal surface. This can be deduced from the knowledge that $\phi_{\alpha}({\bf R})$ is equal to zero there (as well as in many other places), and that the nodal surfaces of antisymmetric functions such as $\Phi_{\rm T}({\bf R})$ are not altered by permutations. The real antisymmetric function $\bar{\Psi}_{\alpha}({\bf R})$ is now substituted into the standard quantum mechanical variational principle to give, $$\begin{aligned} \label{eq:var} E_0 & \leq & \frac{\int \bar{\Psi}_{\alpha} {\hat{H}} \bar{\Psi}_{\alpha}\, d{\bf R}}{\int \bar{\Psi}_{\alpha} \bar{\Psi}_{\alpha}\, d{\bf R}} \, = \, \frac{\int \bar{\Psi}_{\alpha} {\hat{H}} {\hat{\mathcal A}} {\phi}_{\alpha}\, d{\bf R}}{\int \bar{\Psi}_{\alpha} {\hat{\mathcal A}} {\phi}_{\alpha}\, d{\bf R}} \nonumber \\ & = & \frac{\int \bar{\Psi}_{\alpha} {\hat{H}} {\phi}_{\alpha}\, d{\bf R}}{\int \bar{\Psi}_{\alpha} {\phi}_{\alpha}\, d{\bf R}} \, = \, \epsilon_{\alpha} \; ,\end{aligned}$$ where we have used the fact that $\hat{\mathcal A}$ commutes with $\hat{H}$, that it is self-adjoint, and that it is idempotent (so that $\hat{\mathcal A}\bar{\Psi}_{\alpha}({\bf R}) = \bar{\Psi}_{\alpha}({\bf R})$). The delta function terms appearing in $\hat{H} \phi_{\alpha}$ do not contribute to the energy expectation value because they occur on the fixed nodal surface where $\bar{\Psi}_{\alpha}({\bf R})=0$. If the nodal surface of $\Phi_{\rm T}({\bf R})$ is the same as the nodal surface of the exact ground state, the equality holds and the pocket eigenvalue $\epsilon_{\alpha}$ is equal to $E_0$; but if $\Phi_{\rm T}({\bf R})$ does not have the correct nodal surface then $\epsilon_{\alpha} > E_0$. The energy $\epsilon_{\alpha}$ produced by the DMC simulation in nodal pocket $v_{\alpha}$ is therefore minimized and equal to the exact ground-state energy when the trial nodal surface is exact. Since $\epsilon_{\alpha}$ is also expected to depend smoothly on the shape of the nodal surface, it follows that the error in $\epsilon_{\alpha}$ is in general second order in the error in the nodal surface. For systems with reasonably well behaved local potentials and real ground-state wave functions, the tiling theorem [@tiling] states that all the ground-state nodal pockets are related by permutation symmetry. A derivation of this theorem appears in Sec. \[sec:tiling\]. The tiling theorem holds even when the ground state is degenerate, in which case every possible real linear combination of the degenerate ground states possesses the tiling property. Many DMC simulations use trial wave functions with the same nodal surface as the density-functional ground state. Since the density-functional Hamiltonian has a local potential, such trial states always satisfy the tiling theorem. This guarantees that the value of $\epsilon_{\alpha}$ is the same in every nodal pocket, and hence that the energy obtained in the fixed-node DMC simulation cannot depend on how the walkers are distributed among the pockets. In cases when the trial function does not possess the tiling property, the walker population grows most rapidly in nodal pockets with low values of $\epsilon_{\alpha}$ (or, equivalently, low average values of $E_{\rm L}({\bf R})-E_{\rm S}$). Since $E_{\rm S}$ is chosen such that the total walker population stays roughly constant, and since fixed-node DMC walkers never cross from one nodal pocket into another, the walkers in less favorable pockets are gradually annihilated. Although many different nodal pockets may have contained walkers initially, the population becomes more and more concentrated in the pocket or pockets with the lowest value of $\epsilon_{\alpha}$, and it becomes more and more likely that only these pockets will be occupied. The DMC energy is therefore almost certain to converge to the lowest of the pocket eigenvalues of the nodal volumes initially occupied with walkers, $E_{\rm DMC} = \rm{min}_{\alpha}(\epsilon_{\alpha})$. This means that the energy obtained in a DMC simulation which does not possess the tiling property may depend on the initial distribution of walkers. Absence of a General Variational Principle for Fixed-Node DMC Calculations of Excited States {#sec:GenVar} ============================================================================================ In a few cases, the exact nodal surface of an excited-state wave function can be determined using symmetry arguments alone. An example is the first excited state of a particle in a one-dimensional square well, which has a single nodal point located at the well center. The trial wave function can then be chosen to have exactly the same nodal surface as the excited state, and the fixed-node DMC simulation gives the exact excited-state eigenvalue. This result holds whether or not the trial wave function is orthogonal to all the lower energy eigenstates. In practice, however, the exact nodal surface is rarely determined by symmetry alone, and it is rarely possible to choose a trial function with exactly the same nodal surface as the excited state of interest. Furthermore, the wave function of an arbitrary excited state need not possess the tiling property, and so the DMC energy may depend on the initial distribution of the walkers. As an example, consider a hydrogen atom in its $2{\rm s}$ state, $\Psi_{2{\rm s}}(r)$, with eigenvalue $E_{2{\rm s}}$. The exact nodal surface is a sphere of radius $r_0$, the value of which cannot be determined using symmetry arguments. If the imposed nodal surface is exact, the pocket eigenvalues in the inner and outer pockets will both be exactly equal to $E_{2{\rm s}}$; but if the nodal surface is a sphere of radius $a \neq r_0$, the pocket eigenvalue $\epsilon_>$ of the wave function $\Psi_>$ in the outer pocket will not equal the pocket eigenvalue $\epsilon_<$ of the wave function $\Psi_<$ in the inner pocket. Consider the case when the fixed node is too close to the nucleus, $a < r_0$. According to the variational principle applied to the outer pocket, the value of $\epsilon_>$ must be bounded above by the energy expectation value of the trial function, $$\Phi_{\rm T}^{>}(r) = \left \{ \begin{array}{ll} 0 \;\;\;\;\;\; & a \leq r \leq r_0 \\ \Psi_{2{\rm s}}(r) & r > r_0 \; . \end{array} \right .$$ This expectation value is the exact $2{\rm s}$ eigenvalue $E_{2{\rm s}}$. Clearly, we could construct a lower energy trial function by removing the kink in $\Phi_{\rm T}^{>}(r)$ at $r = r_0$, and hence $\epsilon_> < E_{2{\rm s}}$. Similarly, we can take the pocket ground state from the inner pocket, $\Psi_<$, and use it as a variational trial function within the exact $2{\rm s}$ nodal surface $r = r_0$, $$\Phi_{\rm T}^{<}(r) = \left \{ \begin{array}{ll} \Psi_<(r) \;\;\; & r < a \\ 0 & a \leq r \leq r_0 \; . \end{array} \right .$$ In this case the energy expectation value of the trial function is $\epsilon_<$, while the minimum possible expectation value for a wave function with a node at $r_0$ is $E_{2{\rm s}}$. Again, we could construct a lower energy trial function by removing the kink at $r = a$, and hence $\epsilon_< > E_{2{\rm s}}$. Note that variational arguments like these can be used whenever the exact nodal pocket completely encloses the trial pocket or vice-versa, irrespective of dimension or symmetry. This will prove useful in Sec. \[sec:sepexample\]. The last paragraph showed that if $a \! < \! r_0$ then $\epsilon_> < E_{2{\rm s}} < \epsilon_<$; if $a > r_0$, a similar derivation gives $\epsilon_< < E_{2{\rm s}} < \epsilon_>$. The two pockets have different energies unless $a = r_0$. How does this affect the fixed-node DMC algorithm? As long as the lower energy pocket contains plenty of walkers initially, the walker population in the higher energy pocket will almost certainly die out. The fixed-node DMC energy will then tend to the pocket eigenvalue of the lower energy pocket, which is always less than or equal to $E_{2{\rm s}}$. The DMC energy is therefore *maximized* when $a = r_0$. If $a$ is increased through $r_0$, the $\tau \rightarrow \infty$ DMC walker population switches from the outer nodal pocket to the inner one, and the slope of the graph of the DMC energy versus $a$ changes discontinuously. It follows that the error in the fixed-node DMC energy is first order in the error in the nodal surface, not second order as it is for the ground state. This example shows that there is no variational principle when fixed-node DMC is used to study general excited states: the error in the DMC energy of the excited state may increase linearly with the error in the nodal surface; and the DMC energy need not be minimized when the nodal surface is exact. The Lowest Energy Eigenstate of Each Symmetry {#sec:Lowest} ============================================= We now address the question of whether there is a variational principle for fixed-node DMC simulations of the lowest energy state of each symmetry. We denote by ${\mathcal G}$ the group of spatial transformations ${\mathcal T}$ (combinations of rotations, reflections, translations, and inversions of all electrons simultaneously) which leave the many-electron Hamiltonian invariant. For simplicity, we consider this group to be finite [@lie] and of order $g$. The full symmetry group of the Hamiltonian is the direct product of ${\mathcal G}$ with the permutation and time-reversal groups. All the symmetry arguments in this paper can easily be recast in terms of the full symmetry group instead of ${\mathcal G}$, but real many-electron wave functions have such simple time-reversal and permutation symmetries (they transform according to the one-dimensional identity representation of the time-reversal group and the one-dimensional antisymmetric representation of the permutation group) that this is unnecessary. Since the arguments based on the spatial group ${\mathcal G}$ are somewhat easier to grasp, we choose to work with this group in what follows. We begin by showing that a symmetry-constrained variational principle holds whenever the chosen trial function is real and has an invariant nodal surface. This theorem was stated without proof in a paper by Caffarel and Claverie. [@Caffarel] Here we sketch a proof. A real trial function $\Phi_{\rm T}({\bf R})$ is said to have an invariant nodal surface if the transformed function, $$\hat{Q}({\mathcal T})\Phi_{\rm T}({\bf R}) \equiv \Phi_{\rm T}({\mathcal T}^{-1}{\bf R}) \; ,$$ has the same nodal surface as $\Phi_{\rm T}({\bf R})$ for all coordinate transformations ${\mathcal T} \in {\mathcal G}$. Note that the nodal surface is defined as the surface on which $\Phi_{\rm T}({\bf R})$ is zero and across which it changes sign; $\Phi_{\rm T}({\bf R})$ may also have additional zeros where the sign does not change, but these are not on the nodal surface and need not be invariant. The proof relies on the properties of the function $\chi({\bf R})$ defined by, $$\chi({\bf R}) = \left \{ \begin{array}{rl} 0 & {\rm on\;the\;trial\;nodal\;surface} \\ +1 & {\rm in\;nodal\;pockets\;where\;} \Phi_{\rm T} \geq 0 \\ -1 & {\rm in\;nodal\;pockets\;where\;} \Phi_{\rm T} \leq 0 \; . \end{array} \right .$$ Given any point ${\bf R}$ not on the nodal surface, and any spatial transformation ${\mathcal T} \in {\mathcal G}$, it is clear that $\chi({\bf R}) = \eta \chi({\mathcal T}^{-1}{\bf R})$, where $\eta=\pm 1$. Furthermore, since the functions $\chi({\bf R})$ and $\chi({\mathcal T}^{-1}{\bf R})$ have the same nodal surface and so change sign together as ${\bf R}$ changes, the sign of $\eta$ must be independent of ${\bf R}$. This shows that all symmetries ${\mathcal T} \in {\mathcal G}$ either leave $\chi({\bf R})$ unchanged or multiply it by $-1$, from which it follows that $\chi({\bf R})$ transforms according to a one-dimensional irreducible representation $\Gamma^{r}_{\mathcal G}$ of ${\mathcal G}$. We can now adapt the proof of the ground-state variational principle given in Sec. \[sec:DMC\]. Take the pocket ground state $\phi_{\alpha}$ and antisymmetrize it to obtain a function $\bar{\Psi}_{\alpha}$ as in Eq. (\[eq:antiwave\]). Now apply the group theoretical projection operator, [@Cornwell] $$\label{eq:proj} \hat{\mathcal P}^{r} = \frac{1}{g} \sum_{{\mathcal T} \in \mathcal{G}} \Gamma^{r}_{\mathcal G}({\mathcal T}) \hat{Q}({\mathcal T}) \; ,$$ where $\Gamma^r_{\mathcal G}({\mathcal T})$ is the one-by-one matrix representing ${\mathcal T}$. The application of $\hat{\mathcal P}^{r}$ produces a new antisymmetric state $\bar{\Psi}_{\alpha}^r$ which transforms according to the one-dimensional irreducible representation $\Gamma^r_{\mathcal G}$. Since the original pocket ground state $\phi_{\alpha}({\bf R})$ had a non-zero overlap with $\chi({\bf R})$, which is itself an antisymmetric function of symmetry $\Gamma^r_{\mathcal G}$, the antisymmetric $\Gamma^r_{\mathcal G}$ component projected out by applying $\hat{\mathcal A}$ and then $\hat{\mathcal P}^{r}$ to ${\phi}_{\alpha}$ cannot be zero. The energy expectation value of $\bar{\Psi}_{\alpha}^r$ is greater than or equal to the energy $E_0^r$ of the lowest antisymmetric state of symmetry $\Gamma^r_{\mathcal G}$, $$\begin{aligned} \label{eq:Ealphap} E^{r}_0 & \leq & \frac{\int \bar{\Psi}_{\alpha}^{r*}{\hat{H}} \bar{\Psi}_{\alpha}^{r}\, d{\bf R}}{\int \bar{\Psi}_{\alpha}^{r*} \bar{\Psi}_{\alpha}^{r}\, d{\bf R}} \, = \, \frac{\int \bar{\Psi}_{\alpha}^{r*}{\hat{H}} \hat{\mathcal P}^{r} {\hat{\mathcal A}} {\phi}_{\alpha}\, d{\bf R}}{\int \bar{\Psi}_{\alpha}^{r*} \hat{\mathcal P}^{r} {\hat{\mathcal A}} {\phi}_{\alpha}\, d{\bf R}} \nonumber \\ & = & \frac{\int \bar{\Psi}_{\alpha}^{r*}{\hat{H}} {\phi}_{\alpha}\, d{\bf R}}{\int \bar{\Psi}_{\alpha}^{r*} {\phi}_{\alpha}\, d{\bf R}} \, = \, \epsilon_{\alpha} \; .\end{aligned}$$ In analogy with the ground-state proof of Sec. \[sec:DMC\], these manipulations rely on the self-adjointness and idempotency of the operators $\hat{\mathcal A}$ and $\hat{\mathcal P}^r$, both of which commute with $\hat{H}$. It is also important that $\bar{\Psi}_{\alpha}^{r*}({\bf R})$ is equal to zero everywhere on the surface of $v_{\alpha}$, so that the delta function terms in the expression for $\hat{H} {\phi}_{\alpha}({\bf R})$ from Eq. (\[eq:eigensoln\]) do not contribute to the expectation value. This is guaranteed because the fixed nodal surface is invariant. Note that the symmetry of the trial function played no part in this derivation; the only thing that mattered was the invariance of the trial nodal surface. In most DMC simulations, however, the trial function does have a definite symmetry, and so transforms according to a specific irreducible representation of ${\mathcal G}$. Given that functions transforming according to different irreducible representations are orthogonal, and that the definition of $\chi({\bf R})$ ensures that $\langle \chi | \Phi_{\rm T} \rangle > 0$, it follows that any trial function with a definite symmetry and an invariant nodal surface must transform according to the same one-dimensional irreducible representation as the corresponding $\chi({\bf R})$. This implies that a real trial function transforming according to an irreducible representation of dimension greater than one cannot have an invariant nodal surface. The converse is also true: any real trial function $\Phi_{\rm T}({\bf R})$ transforming according to a one-dimensional irreducible representation must have an invariant nodal surface. This follows from the transformation law, $$\hat{Q}({\mathcal T})\Phi_{\rm T}({\bf R}) = \Gamma^r_{\mathcal G}({\mathcal T}) \Phi_{\rm T}({\bf R}) \; ,$$ and the observation that the real normalized function $\Phi_{\rm T}({\bf R})$ remains real and normalized under all transformations in ${\mathcal G}$. The one-by-one matrix $\Gamma^r_{\mathcal G}({\mathcal T})$ is therefore equal to $\pm 1$, and the nodal surface of $\hat{Q}({\mathcal T})\Phi_{\rm T}({\bf R})$ is the same as that of $\Phi_{\rm T}({\bf R})$. Putting everything together, we can now conclude that whenever the real trial function transforms according to a one-dimensional irreducible representation of ${\mathcal G}$, the nodal surface is invariant, and the DMC energy is greater than or equal to the energy of the lowest exact eigenfunction with that symmetry. This is the symmetry-constrained variational principle mentioned in the introduction. If the trial function transforms according to an irreducible representation of dimension greater than one, the nodal surface cannot be invariant, and the delta functions produced when $\hat{H}$ is applied to $\phi_{\alpha}$ need not all occur where $\bar{\Psi}_{\alpha}^{r*}=0$. The above proof of the symmetry-constrained variational principle therefore breaks down. In Sec. \[sec:sepexample\] we give an example of a system with a trial function transforming according to an irreducible representation $\Gamma^r_{\mathcal G}$ of dimension $d_r = 2$ for which $E_{\rm DMC} < E^{r}_0$. The restriction of the variational principle to trial functions transforming according to one-dimensional irreducible representations can be understood in a very simple way. The problem being solved in a fixed-node DMC simulation is not the imaginary-time Schrödinger equation of Eq. (\[eq:its\]), but the imaginary-time Schrödinger equation subject to the additional boundary conditions specified by the trial nodal surface. As a result, the relevant symmetry group is not the group ${\mathcal G}$ of symmetries of $\hat{H}$, but the subgroup ${\mathcal G}_{\rm FN}$ that leaves both the Hamiltonian and the fixed nodal surface (boundary conditions) invariant. The fixed-node eigenfunctions need only conform to the symmetries in ${\mathcal G}_{\rm FN}$, and their transformation properties should be analyzed in terms of the irreducible representations of ${\mathcal G}_{\rm FN}$, not ${\mathcal G}$. If the real trial function transforms according to a one-dimensional irreducible representation of ${\mathcal G}$, the fixed nodal surface is invariant and the groups ${\mathcal G}$ and ${\mathcal G}_{\rm FN}$ are the same. The symmetry-constrained variational principle discussed in this section does not apply unless this is the case. Irreducible Representations of Dimension Greater than One {#sec:d>1} ========================================================= Can anything be said when the trial function transforms according to an irreducible representation of dimension greater than one? Suppose that the real trial function $\Phi_{{\rm T},l}^{r}({\bf R})$ transforms as the $l$th row of an irreducible representation $\Gamma^r_{\mathcal G}$ with dimension $d_r > 1$. We already know that the nodal surface of $\Phi_{{\rm T},l}^{r}({\bf R})$ cannot be invariant with respect to all the operations in ${\mathcal G}$, but it may be invariant under a subset of those operations. Any such subset forms a proper subgroup, ${\mathcal G}_{\rm FN}$, which in general depends on the row index $l$. As in Sec. \[sec:Lowest\], we define the function $\chi({\bf R})$ which is equal to $+1$ in all nodal pockets where $\Phi_{{\rm T},l}^{r}({\bf R}) \geq 0$, equal to $-1$ in all nodal pockets where $\Phi_{{\rm T},l}^{r}({\bf R}) \leq 0$, and equal to zero everywhere on the nodal surface of $\Phi_{{\rm T},l}^{r}({\bf R})$. By construction, $\chi({\bf R})$ transforms according to a one-dimensional irreducible representation of ${\mathcal G}_{\rm FN}$ (although not of ${\mathcal G}$). Since all the pocket ground states $\phi_{\alpha}$ have non-zero overlaps with $\chi({\bf R})$, they all have non-zero components of the same subgroup symmetry as $\chi({\bf R})$. We can therefore re-use the symmetry-projection argument leading to Eq. (\[eq:Ealphap\]) to show that the fixed-node DMC energy is greater than or equal to the eigenvalue of the lowest energy exact eigenfunction with the same subgroup symmetry as $\chi({\bf R})$. This provides a rigorous variational principle, but the subgroup ${\mathcal G}_{\rm FN}$ that leaves the trial nodal surface invariant is usually small and its irreducible representations provide a correspondingly limited set of symmetry labels. The bounds obtained are therefore weak and this variational principle is of little use unless optimized as explained below. If the nodes of $\Phi_{{\rm T},l}^{r}({\bf R})$ were exact one could form a trial function from any real linear combination of different rows and always obtain the same DMC energy. When the nodes are not exact, however, the DMC method breaks this $d_r$-fold degeneracy, and the variational lower bound on the DMC energy depends on the precise linear combination of rows chosen in constructing the trial wave function. This freedom can be exploited to improve the weak variational principle. The $d_r$ functions $\Phi_{{\rm T},l}^{r}({\bf R})$ $(l=1,2,\ldots,d_r)$ are a basis for the irreducible representation $\Gamma^r_{\mathcal G}$ of ${\mathcal G}$, and hence transform into linear combinations of each other under all coordinate transformations ${\mathcal T} \in {\mathcal G}$. This implies that they also transform into linear combinations of each other under all coordinate transformations ${\mathcal T} \in {\mathcal G}_s$, where ${\mathcal G}_s$ is any proper subgroup of ${\mathcal G}$. The subset of the matrices $\Gamma^r_{\mathcal G}({\mathcal T})$ for which ${\mathcal T} \in {\mathcal G}_s$ is therefore a representation of ${\mathcal G}_s$. This representation is not in general irreducible. To enumerate the different possibilities, begin by determining all the proper subgroups ${\mathcal G}_s$ of ${\mathcal G}$. For each such subgroup, consider the representation of ${\mathcal G}_s$ consisting of the matrices $\Gamma^r_{\mathcal G}({\mathcal T})$ with $T \in {\mathcal G}_s$. This representation may be decomposed into its irreducible components with respect to ${\mathcal G}_s$, $$\label{eq:Compat} \Gamma^{r}_{\mathcal G} = n_{q_1} \Gamma^{q_1}_{{\mathcal G}_s} \oplus n_{q_2} \Gamma^{q_2}_{{\mathcal G}_s} \oplus \cdots \oplus n_{q_m} \Gamma^{q_m}_{{\mathcal G}_s} \; ,$$ where the positive integer $n_{q_i}$ is the number of times the irreducible representation $\Gamma^{q_i}_{{\mathcal G}_s}$ of ${\mathcal G}_s$ appears, so $d_r = \sum_i n_{q_i} d_{q_i}$. The irreducible representations of ${\mathcal G}_s$ appearing in Eq. (\[eq:Compat\]) are said to be compatible [@Cornwell] with $\Gamma^r_{\mathcal G}$. Note that since the group containing only the identity element is always a subgroup of ${\mathcal G}$, it is always possible to find at least one subgroup for which the reduction of $\Gamma^r_{\mathcal G}$ contains a one-dimensional irreducible representation $\Gamma^{q_i}_{{\mathcal G}_s}$. A trial function transforming as the $l$th row of $\Gamma^r_{\mathcal G}$ may contain components along all the compatible representations of ${\mathcal G}_s$, but it is always possible to construct linear combinations, $$\label{eq:lincom} \Phi_{\rm T}^{q_i}({\bf R}) = \sum_{l=1}^{d_r} c_l^{q_i} \Phi_{{\rm T},l}^{r}({\bf R}) \; ,$$ transforming according to each particular compatible representation $\Gamma^{q_i}_{{\mathcal G}_s}$. Every real function $\Phi_{\rm T}^{q_i}({\bf R})$ corresponding to a one-dimensional representation of ${\mathcal G}_s$ has an invariant nodal surface with respect to ${\mathcal G}_s$. A DMC energy calculated using the nodes of such a function therefore satisfies the variational principle $E_{\rm DMC} \geq E^{q_i}_0$. Clearly, the strength of the variational principle obtained depends on the choices of the subgroup ${\mathcal G}_s$ and the one-dimensional representation $\Gamma^{q_i}_{{\mathcal G}_s}$. In some cases, one can find a subgroup ${\mathcal G}_s$ with a one-dimensional representation $\Gamma^{q_i}_{{\mathcal G}_s}$ which is compatible with $\Gamma^r_{\mathcal G}$ but incompatible with all those irreducible representations $\Gamma^{r'}_{\mathcal G}$ for which $E_0^{r'}<E_0^r$. A DMC energy calculated using the trial function $\Phi_{\rm T}^{q_i}({\bf R})$ is then guaranteed to be greater than or equal to $E_0^r$. In general, if one knows the ordering of the energy levels beforehand, one can use the compatibility analysis to find the one-dimensional irreducible representation $\Gamma^{q_i}_{{\mathcal G}_s}$ which gives the most stringent energy bound for the eigenvalue of interest. As an application of this symmetry-constrained variational principle, consider a crystal with space group ${\mathcal G}$. We wish to establish whether it is possible to use fixed-node DMC to obtain a variational estimate of the eigenvalue $E_{0}({\bf k})$ of the lowest energy eigenstate with crystal momentum ${\bf k}$. The relevant subgroup ${\mathcal G}_s$ is the translation group, which is Abelian and has only one-dimensional irreducible representations. These are labeled by the crystal momentum ${\bf k}$, and so the statement of the symmetry-constrained variational principle is very straightforward: if the DMC trial function has crystal momentum ${\bf k}$, the DMC energy is greater than or equal to $E_0({\bf k})$. It is worth noting that this statement holds whether or not the lowest energy state of crystal momentum ${\bf k}$ is degenerate. In the degenerate case the trial function transforms according to a multi-dimensional irreducible representation of ${\mathcal G}$, but it still has crystal momentum ${\bf k}$ and so still transforms according to a one-dimensional irreducible representation of the translation subgroup ${\mathcal G}_s$. This is sufficient to guarantee that the DMC energy is greater than or equal to $E_0({\bf k})$. The symmetry-constrained DMC variational principle for the lowest energy state of crystal momentum ${\bf k}$ is unfortunately much less useful than it appears, because most Bloch states are complex and cannot be used as fixed-node DMC trial functions. Instead, the standard approach is to use a real linear combination of a Bloch function and its complex conjugate. (This is justified by the assumption of time-reversal invariance, which guarantees that if $\Psi_{\bf k}$ is an eigenfunction with eigenvalue $E_0({\bf k})$ then so is $\Psi^{*}_{\bf k}$.) Since such trial functions contain components with two different wave vectors, ${\bf k}$ and $-{\bf k}$, they do not transform according to a single irreducible representation of the translation group. The symmetry-constrained variational theorem is therefore inapplicable, and it is possible that the DMC energy may lie below $E_0({\bf k})$. In Sec. \[sec:sepexample\] we show by means of a specific example that such calculations may indeed produce a DMC energy which is lower than $E_0({\bf k})$. An important exception arises when the wave vector ${\bf k}$ is equal to half a reciprocal lattice vector, in which case ${\bf k}$ and $-{\bf k}$ are alternative labels for the same irreducible representation of the translation group. The linear combination of $\Psi_{\bf k}$ and $\Psi_{\bf k}^{*}$ is then a pure Bloch function, and the normal proof of the symmetry-constrained variational principle applies. As long as ${\bf k}$ equals half a reciprocal lattice vector, the DMC energy is greater than or equal to the energy of the lowest exact eigenstate with crystal momentum ${\bf k}$. Generalizations of the Tiling Theorem {#sec:tiling} ===================================== In his paper on fermion nodes, [@tiling] Ceperley stated that the tiling theorem could be generalized to the case where there were other discrete symmetries present. A more precise statement is that, given a Hamiltonian with a reasonable local potential and a symmetry group ${\mathcal G}$, the tiling theorem applies to any real state which is the lowest energy eigenfunction of a symmetry $\Gamma^r_{\mathcal G}$ with dimension $d_r = 1$. This can be demonstrated via a simple generalization of Ceperley’s proof [@tiling] of the ground-state tiling theorem. Consider the real antisymmetric state $\Psi_0^r({\bf R})$ which is the lowest energy eigenfunction transforming according to the one-dimensional irreducible representation $\Gamma_{\mathcal G}^r$. This function may have many nodal pockets, but pick one at random and color it blue. Now apply a symmetry operator from the group containing the spatial symmetries (rotations, translations, reflections and inversions) and the permutations. Since the nodal surface is invariant, this symmetry operator maps the region of space within the blue nodal pocket into itself or into one of the other nodal pockets. If the blue pocket is mapped into some other nodal pocket, this pocket is equivalent to the blue one by symmetry and is also colored blue. Repeat this process for every operator in the group, until all the pockets equivalent to the original one have been found and colored blue. There are now two possibilities: the blue regions may fill the entire configuration space, in which case the nodal pockets are all equivalent by symmetry and the tiling theorem holds; or there may be other inequivalent nodal pockets which have not yet been found. We can rule out the second possibility using the following argument. Assume that the blue nodal pockets do not fill the configuration space, and that the local potential $V({\bf R})$ and all its derivatives are finite except at the Coulomb singularities occurring when two charged particles approach each other. As long as the system is not one-dimensional (in which case different arguments are required), this ensures that almost every point on the nodal surface lies a finite distance away from the nearest singularity in the potential. The eigenfunction $\Psi_0^r({\bf R})$ may therefore be expanded as a power series with a finite radius of convergence about almost any point ${\bf R}_s$ on the nodal surface. If the gradient of $\Psi_0^r({\bf R})$ is assumed to be zero over any finite area of the nodal surface surrounding ${\bf R}_s$, it can be shown that every term in this series has to be zero, and hence that $\Psi_0^r({\bf R}) = 0$ everywhere within the radius of convergence. Since any solution of the Schrödinger equation can be analytically continued around the isolated singularities in the potential, this further implies that $\Psi_0^r({\bf R})$ is zero everywhere. We therefore conclude that the gradient of $\Psi_0^r({\bf R})$ must be non-zero almost everywhere on the nodal surface. Now consider the trial function $\tilde{\Psi}^r({\bf R})$, which is defined to equal $\Psi_0^r({\bf R})$ within the blue pockets and zero elsewhere. This trial function is antisymmetric and transforms according to the irreducible representation $\Gamma_{\mathcal G}^r$, but has gradient discontinuities on the nodal surfaces separating the blue pockets from the rest of configuration space. It satisfies the Schrödinger-like equation, $$\label{eq:newdelta} \hat{H}\tilde{\Psi}^r({\bf R}) = E_0^r \tilde{\Psi}^r({\bf R}) + \delta^r \;,$$ where the symbol $\delta^r$ denotes the delta functions produced by the action of the kinetic energy operator on the gradient discontinuities. The delta functions occur where $\tilde{\Psi}^r = 0$ and so do not affect the energy expectation value, $$E_0^r = \frac{ \langle \Psi^r_0 | \hat{H} | \Psi^r_0 \rangle } { \langle \Psi^r_0 | \Psi^r_0 \rangle } = \frac{ \langle \tilde{\Psi}^r | \hat{H} | \tilde{\Psi}^r \rangle } { \langle \tilde{\Psi}^r | \tilde{\Psi}^r \rangle } \; . \label{eq:tilvar}$$ We know, however, that a state which has gradient discontinuities almost everywhere on a finite area of the nodal surface cannot be an eigenfunction unless the potential is infinite almost everywhere on that area. Since we are assuming that this is not the case, the function $\tilde{\Psi}^r$ must contain excited-state components of symmetry $\Gamma^r_{\mathcal G}$ and cannot have the same energy expectation value as the lowest energy state of that symmetry. This conclusion contradicts Eq. (\[eq:tilvar\]), and so the assumption that the blue nodal pockets do not fill the configuration space must have been incorrect. All the nodal pockets of $\Psi_0^r$ are therefore equivalent by symmetry. This proves the tiling theorem for any real state which is the lowest energy eigenfunction of a symmetry $\Gamma_{\mathcal G}^r$ with dimension $d_r = 1$. An obvious corollary is that there is also a tiling theorem for the lowest energy state transforming according to any one-dimensional irreducible representation of any subgroup ${\mathcal G}_s$ of ${\mathcal G}$. This statement is analogous to the weaker variational principle discussed in Sec. \[sec:d&gt;1\]. In every case when we have demonstrated the existence of a DMC variational principle, we have therefore also been able to demonstrate the existence of a tiling theorem. Our analysis has been restricted to the physically interesting case of a local potential which is finite everywhere except at Coulomb singularities, but our conclusions may be somewhat more general than this suggests. The familiar many-fermion ground-state tiling theorem may be viewed as a special case of the subgroup tiling theorem mentioned above. The permutation group is always a subgroup of the full symmetry group (which contains both spatial and permutation symmetries), and the many-fermion ground state is the lowest energy state which transforms according to the one-dimensional antisymmetric irreducible representation of that subgroup. The subgroup tiling theorem therefore guarantees that the many-fermion ground state possesses the tiling property. Note that the tiling property holds with respect to the permutation subgroup, not the full symmetry group. This means that it is only the elements of the permutation subgroup that need be applied to the initial blue pocket to find all equivalent pockets and turn the whole configuration space blue. The above derivation of the subgroup tiling theorem only applies to states that transform according to one-dimensional irreducible representations of the chosen subgroup ${\mathcal G}_s$. Such states may also transform according to multi-dimensional irreducible representations of the full symmetry group ${\mathcal G}$, so the tiling theorem is not restricted to non-degenerate energy levels. In systems with degenerate many-fermion ground states, for example, any real linear combination of the degenerate ground states is antisymmetric and so possesses the tiling property with respect to the permutation subgroup. This result holds even though the nodal surface is not invariant under all the operators from the full symmetry group. The most important consequence of the generalized tiling theorem is exactly as in the ground-state case. It is common for a fixed-node DMC trial wave function to have the same nodal surface as an energy eigenfunction calculated using an approximate method such as local-density-functional theory. The approximate Hamiltonian is chosen to have the same symmetries as the exact Hamiltonian, but may also have extra symmetries which are not relevant to the argument and may be ignored. If the approximate Hamiltonian has a reasonably well behaved local potential, its eigenstates have the same tiling properties with respect to ${\mathcal G}$ as the corresponding exact eigenstates. The lowest energy eigenstate transforming according to any one-dimensional irreducible representation of ${\mathcal G}$ or any subgroup of ${\mathcal G}$ therefore satisfies the tiling theorem. This ensures that the energy produced by a DMC simulation using the nodal surface of such a state is independent of the initial walker distribution. Tight-Binding Example {#sec:tbexample} ===================== This section and the next describe simple examples that show what happens when a trial wave function transforming according to an irreducible representation with $d_r\!>\!1$ is used to define the trial nodal surface for a DMC simulation. We find that the DMC energy may indeed be lower than the eigenvalue of the lowest exact eigenstate with the same symmetry as the trial function. This demonstrates that the variational principle of Eq. (\[eq:Ealphap\]) does not apply when $d_r\!>\!1$. In searching for a suitable example system, we found it convenient to impose the following restrictions: (i) the group of the Hamiltonian is finite; (ii) the trial wave function and its nodal surface can be easily visualized; and (iii) the exact eigenstates of the Hamiltonian and the pocket ground states of Eq. (\[eq:eigensoln\]) can be calculated without numerical error. These restrictions are unnecessary, but make the analysis much simpler. To satisfy criterion (ii), we must choose a system containing at most three electrons in one dimension or a single electron in two or three dimensions. The nodal surfaces of one-dimensional systems are not easily altered because they are almost entirely determined by the antisymmetry of the wave function, so we decided to concentrate on one-electron systems in two and three spatial dimensions. The first system we studied was an electron confined to a rectangular box in two or three dimensions. Although this system does have multi-dimensional irreducible representations, we failed to find an example in which the DMC energy calculated using a trial state of a given symmetry was less than the eigenvalue of the lowest energy eigenstate of that symmetry. The second system we tried was more successful. We start by discussing a simple tight-binding realization of this system, and then in Sec. \[sec:sepexample\] present an alternative realization based on a separable solution of the Schrödinger equation. Consider a molecule containing one electron moving in the potential of three protons fixed at the corners of an equilateral triangle. The symmetry group of the Hamiltonian of this system is called $D_{3{\rm h}}$. [@Cornwell] A convenient way to generate DMC trial wave functions with specific symmetries is to solve the Schrödinger equation within a tight-binding approximation using a single spherically-symmetric atomic-like orbital $\xi(|{\bf r}|)$ centered on each proton. Once a tight-binding eigenfunction has been found and used to define a trial nodal surface, the fixed-node DMC algorithm can be used to solve the original Hamiltonian exactly subject to the fixed-node constraint. The ground state of the molecule is a non-degenerate nodeless function transforming according to the identity representation. Some of the excited-state eigenfunctions must be doubly degenerate, however, since the symmetry group $D_{3{\rm h}}$ has two two-dimensional irreducible representations, one of which is called $\Gamma^3_{D_{3{\rm h}}}$. The tight-binding Hamiltonian has only three eigenstates, the nodeless ground state and a degenerate pair of excited states which can be written as the Bloch functions, $$\begin{aligned} \Psi_{+}({\bf r}) & = & \sum_{j=0}^{2} \xi(|{\bf r} - {\bf d}_j |) e^{i2\pi j/3} \;, \nonumber \\ \Psi_{-}({\bf r}) & = & \sum_{j=0}^{2} \xi(|{\bf r} - {\bf d}_j |) e^{-i2\pi j/3}\;,\end{aligned}$$ where ${\bf d}_0$, ${\bf d}_1$, and ${\bf d}_2$ are the position vectors of the three protons. These two Bloch functions transform into linear combinations of each other under the operations of the point group (there is no need to include time-reversal symmetry) and form a basis for $\Gamma^3_{ D_{3{\rm h}}}$. Taking linear combinations of $\Psi_{+}$ and $\Psi_{-}$, one can form the real functions, $$\begin{aligned} \label{eq:psi1_psi2} \Psi_{1}({\bf r}) & = & \xi(|{\bf r} - {\bf d}_1|) - \xi(|{\bf r} - {\bf d}_2|) \; , \nonumber \\ \Psi_{2}({\bf r}) & = & 2\xi(|{\bf r} - {\bf d}_0|) - \xi(|{\bf r} - {\bf d}_1|) - \xi(|{\bf r} - {\bf d}_2|) \; ,\end{aligned}$$ which form an alternative basis for the same irreducible representation. Fig. (\[fig1\]) shows the nodal surfaces of $\Psi_1$ and $\Psi_2$ along with the expansion coefficients from Eq. (\[eq:psi1\_psi2\]). The nodal surface of $\Psi_1$ does not depend on the particular spherically-symmetric tight-binding basis functions chosen and turns out to be an exact excited-state nodal surface. The nodal surface of $\Psi_2$ is not exact, however, and its precise shape depends on the details of the atomic-like orbitals used in the tight-binding model. Different choices of $\xi(|{\bf r}|)$ give trial functions with the same symmetries but different nodal surfaces. By changing $\xi(|{\bf r}|)$, it is possible to change the relative sizes of the two nodal pockets of $\Psi_2$, causing one to grow at the expense of the other. As in the example of the $2{\rm s}$ state of the H atom discussed earlier, this suggests that in many cases the pocket eigenvalue of one of the two nodal pockets will be too high while that of the other is too low. The DMC energy will then lie below the energy of the exact eigenstate of interest. In Sec. \[sec:sepexample\] we numerically solve a specific example with the same symmetry properties as this system and observe exactly this behavior. Here we apply the weaker variational principle described in the previous section to the degenerate excited state with symmetry $\Gamma^3_{ D_{3{\rm h}}}$ of the group $D_{3{\rm h}}$. The largest subgroups $C_{3{\rm h}}$ and $D_3$ each have six elements. If the subgroup $D_3$ is used, the compatibility relation of Eq. (\[eq:Compat\]) becomes, $$\Gamma^3_{D_{3{\rm h}}} = \Gamma^3_{D_3} \;,$$ where $\Gamma^3_{D_3}$ is two dimensional. This result is not useful since we are seeking a reduction that contains one-dimensional representations. Using the subgroup $C_{3{\rm h}}$ results in two one-dimensional representations $$\Gamma^3_{D_{3{\rm h}}} = \Gamma^5_{C_{3{\rm h}}} \oplus \Gamma^6_{C_{3{\rm h}}} \;.$$ However, any trial wave function that transforms as either $\Gamma^5_{C_{3{\rm h}}}$ or $\Gamma^6_{C_{3{\rm h}}}$ is complex and therefore unsuitable for fixed-node DMC. If we use the smaller subgroup, $C_{2{\rm v}}$, which has four elements, we obtain the compatibility relation $$\Gamma^3_{D_{3{\rm h}}} = \Gamma^1_{C_{2{\rm v}}} \oplus \Gamma^4_{C_{2{\rm v}}} \;,$$ from which we can construct real trial wave functions and apply the weaker variational principle. In fact, the trial functions $\Psi_1$ and $\Psi_2$ already have the correct transformation properties: $\Psi_1$ transforms as the one-dimensional representation $\Gamma^4_{C_{2{\rm v}}}$, while $\Psi_2$ transforms as the one-dimensional representation $\Gamma^1_{C_{2{\rm v}}}$. The group $C_{2{\rm v}}$ therefore preserves the nodal surfaces of $\Psi_1$ and $\Psi_2$. Because the symmetry corresponding to the representation $\Gamma^1_{C_{2{\rm v}}}$ is compatible with the ground-state symmetry as well as with $\Gamma^3_{D_{3{\rm h}}}$, a DMC simulation using the trial function $\Psi_2$ satisfies only the weaker variational principle $E_{\rm DMC} \geq E_0$, where $E_0$ is the overall ground-state energy. The representation $\Gamma^4_{C_{2{\rm v}}}$ is not compatible with the ground-state symmetry, however, and so a simulation using the trial function $\Psi_1$ gives a stronger variational principle. For the example studied in the next section, it turns out that the $\Gamma^3_{D_{3{\rm h}}}$ state of interest is the lowest energy exact eigenstate with which the representation $\Gamma^4_{C_{2{\rm v}}}$ is compatible. The strong variational principle therefore applies and $E_{\rm DMC}\geq E_0^3$, where $E_0^3$ is the exact $\Gamma^3_{D_{3{\rm h}}}$ eigenvalue. It is important to appreciate that this is not a general result; it is not always possible to pick a trial function that maintains the strong variational principle for a given symmetry. Note that the strong variational principle applies because the trial wave function $\Psi_1$ transforms according to the irreducible representation $\Gamma^4_{C_{2{\rm v}}}$ of the subgroup $C_{2{\rm v}}$. If we had chosen a different pair of trial functions, constructed by taking linear combinations of $\Psi_1$ and $\Psi_2$, both would have contained components along $\Gamma^4_{C_{2{\rm v}}}$ and $\Gamma^1_{C_{2{\rm v}}}$. The only subgroup preserving the nodes would then have been the group of the identity, and the only variational principle would have been with respect to the overall ground-state energy, $E_{\rm DMC} \geq E_0$. This illustrates the general rule that the strongest variational principles are obtained by choosing trial functions which transform according to specific one-dimensional irreducible representations of specific subgroups of ${\mathcal G}$. Separable Example {#sec:sepexample} ================= We now present an explicit solution of a different example with the same $D_{3{\rm h}}$ symmetry group as the triangular molecule discussed above. Consider a particle of unit mass moving in a triangular potential in three dimensions. The wave function $\Psi({\bf r})$ obeys the Schrödinger equation, $$\label{eq:Hamil} \left ( - \frac{1}{2} \nabla^2 + V({\bf r}) \right ) \Psi({\bf r}) = E \Psi({\bf r}) \; , \label{eq:sepham}$$ where $$\label{eq:Poten} V({\bf r}) = \frac{\cos(3\theta)}{\rho^2} \;,$$ with $\rho$, $\theta$ and $z$ the usual cylindrical coordinates. The boundary conditions are $\Psi(\rho,\theta,z) = 0$ for $\rho \geq 1$ or $|z| \geq \pi/2$, confining the particle within a cylinder. Writing $\Psi({\bf r}) = R ( \rho ) \Theta ( \theta ) Z(z)$, the Schrödinger equation separates into: $$\begin{aligned} - \frac{1}{2} \frac{d^2 Z}{dz^2} & = & E_z Z \; , \label{eq:zeq} \\ - \frac{1}{2} \frac{d^2 \Theta}{d \theta^2} + \cos ( 3\theta ) \Theta & = & E_{\theta} \Theta \; , \label{eq:thetaeq} \\ - \frac{1}{2\rho} \frac{d}{d \rho} \left ( \rho \frac{dR}{d\rho} \right ) + \left ( E_z + \frac{E_{\theta}}{\rho^2} \right ) R & = & E R \;\; . \label{eq:rhoeq}\end{aligned}$$ The lowest energy eigenfunction of Eq. (\[eq:zeq\]) is $$Z(z) = \sqrt{\frac{2}{\pi}} \cos(z) \;,$$ with eigenvalue $E_z = 1/2$. Eq. (\[eq:rhoeq\]) may be simplified by the substitutions $r = \sqrt{2(E - E_z)} \rho$ and $\nu = \sqrt{2 E_{\theta}}$, which yield Bessel’s equation, $$r^2 \frac{dR}{dr^2} + r \frac{dR}{dr} + ( r^2 - \nu^2 ) R = 0 \; .$$ The general solutions are the Bessel and Neumann functions, $J_{\nu}(r)$ and $N_{\nu}(r)$, but only $J_{\nu}(r)$ is well behaved at the origin. Hence $$R(\rho) = J_{\sqrt{2E_{\theta}}} \left ( \sqrt{2(E - E_z)} \rho \right ) \; ,$$ with the energy $E$ determined by the boundary condition at $\rho = 1$, $$J_{\sqrt{2E_{\theta}}} ( \sqrt{2(E - E_z )} ) = 0 \; .$$ Eq. (\[eq:thetaeq\]) can be transformed into Mathieu’s equation [@Abram] by a simple change of variables, but here we solve it numerically by expanding the eigenfunctions in (normalized) sines and cosines, $$\Theta(\theta) = a_0 \frac{1}{\sqrt{2\pi}} + \sum_{n=1}^{M} \left ( a_n \frac{\cos(n \theta)}{\sqrt{\pi}} + b_n \frac{\sin(n \theta)}{\sqrt{\pi}} \right ) \; ,$$ and diagonalizing the corresponding Hamiltonian matrix. The results converge rapidly with $M$, and choosing $M\!=\!50$ gives very accurate eigenvalues for the lowest few eigenstates. The three lowest energy angular eigenfunctions are shown in Fig. (\[fig2\]). As stated in Sec. \[sec:tbexample\], the lowest energy eigenstate is a nodeless function invariant under all elements of the symmetry group of the Hamiltonian. The next two states are a degenerate pair forming a basis for the irreducible representation $\Gamma^3_{{\rm D}_{3{\rm h}}}$. One of the two has nodes at $\theta = 0$ and $\theta = \pi$, and will be called $\Psi_1$ in analogy with the corresponding state from Sec. \[sec:tbexample\]; the other has nodes at $\theta = \pm 1.7934$ radians, and will be called $\Psi_2$. Note that $\Psi_1$ is the lowest energy eigenfunction transforming according to the one-dimensional irreducible representation $\Gamma^4_{C_{2{\rm v}}}$ of the subgroup $C_{2{\rm v}}$; it therefore possesses the tiling property with respect to $C_{2{\rm v}}$. The $\Gamma^1_{C_{2{\rm v}}}$ symmetry of $\Psi_2$ is shared by the overall ground state, however, so $\Psi_2$ is not the lowest eigenstate of that subgroup symmetry and does not satisfy a tiling theorem. Consider how DMC might be used to find the eigenvalue of the lowest energy $\Gamma^3_{{\rm D}_{3{\rm h}}}$ doublet. We do not want to impose the exact nodal surface since DMC would then give the exact answer and we would learn little about variational principles, so we seek a trial function with a different nodal surface but the same symmetry. We choose to generate such a trial function by solving the Schrödinger equation for a different triangular potential, $$\hat{H} = - \frac{1}{2} \nabla^2 + \frac{\cos(3\theta)}{\rho^2} + \mu \frac{\cos(6\theta)}{\rho^2} \; ,$$ where $\mu$ is an adjustable parameter. The first three angular eigenfunctions of this Hamiltonian when $\mu=5$ are shown in Fig. (\[fig3\]). Since the $\cos(6\theta)$ term does not change the symmetry group, these eigenfunctions still belong to the same irreducible representations. However, the nodal angle of $\Psi_2$ has moved slightly in response to the perturbation. One of the two nodal volumes of this trial wave function is therefore slightly too small, while the other is slightly too large. The trial nodal pocket that is too large (small) encloses (is enclosed by) the corresponding exact nodal pocket. We can therefore apply the variational argument used in the discussion of the $2{\rm s}$ state of the hydrogen atom to show that the pocket eigenvalue from the pocket that is too small must be greater than the exact eigenvalue, while the pocket eigenvalue from the pocket that is too large must be less than the exact eigenvalue. As in the H $2{\rm s}$ example, therefore, the DMC energy is always $\leq$ the exact eigenvalue. The maximum of the DMC energy, equal to the exact $\Gamma^3_{{\rm D}_{3{\rm h}}}$ ground-state eigenvalue, is attained only when the nodal angle is exact; and the slope of the graph of DMC energy against nodal angle changes discontinuously at this point. The error in the DMC energy is first order, not second order, in the error in the nodal angle. This analysis is confirmed by the results of a full calculation given in Fig. (\[fig4\]), which shows how the angular pocket eigenvalues $E_{\theta}$ of the two pockets depend on the angular half width of the nodal pocket centered on $\theta = 0$. As expected, $E_{\theta}$ is too large when the nodal pocket is too small and vice-versa. Fig. (\[fig5\]) shows the dependence of the total pocket eigenvalues $E$ on the angular half width of the nodal pocket centered on $\theta = 0$, confirming that an increase in angular half width gives rise to a decrease in total eigenvalue and vice-versa. The numerical results therefore support the conclusions of the variational argument. Note, finally, that Eq. (\[eq:thetaeq\]) can be interpreted as the Schrödinger equation for a one-dimensional crystal with periodic boundary conditions, in which case $\Psi_1$ and $\Psi_2$ are real linear combinations of Bloch waves with equal and opposite crystal momenta. Seen from this viewpoint, the degeneracy of $\Psi_1$ and $\Psi_2$ arises from the time-reversal (complex-conjugation) symmetry of the real Hamiltonian rather than from its spatial symmetry, but the failure of the symmetry-constrained variational principle is still apparent. This confirms the assertion made in Sec. \[sec:d&gt;1\]: a DMC simulation using a real trial state constructed from Bloch states with equal and opposite crystal momenta may yield an energy below that of the lowest exact eigenstate with that crystal momentum. Conclusions {#sec:Conclusions} =========== The main lesson to be learned from this work is that symmetry arguments cannot be applied to fixed-node DMC unless the symmetries of both the Hamiltonian and the nodal surface of the trial wave function are taken into account. The fixed-node DMC algorithm solves the Hamiltonian subject to the boundary conditions imposed by the trial nodal surface, and so the relevant symmetry group ${\mathcal G}_{{\rm FN}}$ contains only those symmetry operations that leave both the Hamiltonian and the trial nodal surface (boundary conditions) invariant. This “group of the fixed-node Hamiltonian” is a subgroup of the more familiar “group of the Hamiltonian”, ${\mathcal G}$. If the trial function transforms according to a real one-dimensional irreducible representation $\Gamma^r$ of ${\mathcal G}$, all symmetry operations in ${\mathcal G}$ simply multiply the trial function by a real number. This does not change the nodal surface and hence ${\mathcal G} = {\mathcal G}_{{\rm FN}}$. The symmetry-constrained variational principle then implies: (i) that the fixed-node DMC energy is greater than or equal to the eigenvalue of the lowest energy exact eigenstate that transforms according to the same one-dimensional irreducible representation as the trial state; and (ii) that the error in the DMC energy is in general second order in the difference between the nodal surfaces of the lowest energy exact eigenstate that transforms as $\Gamma^r$ and the trial function. If the irreducible representation to which the trial function belongs is of dimension greater than one, it is inevitable that some of the symmetry operations from ${\mathcal G}$ will change the trial nodal surface. The nodal surfaces of the states $\Psi_1$ and $\Psi_2$ shown in Fig. (\[fig1\]) are examples of this. The symmetry-constrained variational principle need not apply in such cases, because the symmetrized state $\bar{\Psi}_{\alpha}^{r}$ used in its derivation need not be zero everywhere on the trial nodal surface. The delta functions produced when the kinetic energy operator is applied to the fixed-node pocket ground state $\phi_{\alpha}$ from Eq. (\[eq:eigensoln\]) may therefore contribute to the expectation value in Eq. (\[eq:Ealphap\]). In such cases, a weaker version of the symmetry-constrained variational principle can be obtained by re-analyzing the problem using only the symmetries in the subgroup ${\mathcal G}_{\rm FN}$. The idea is to forget all the symmetry operations which change the nodes of the trial function, and consider only those which leave the trial nodal surface invariant. The symmetry-constrained variational principle then applies as long as the symmetries are labeled using the irreducible representations of ${\mathcal G}_{\rm FN}$. In summary, a useful DMC variational principle exists whenever the trial state transforms according to a one-dimensional irreducible representation $\Gamma^r$ of ${\mathcal G}$ or any subgroup of ${\mathcal G}$. In many DMC simulations, the nodal surface of the trial state is the same as that of an eigenstate of an approximate Hamiltonian such as the local-density-functional Hamiltonian. If the state used to define the nodal surface is the lowest energy eigenstate with symmetry $\Gamma^r$ of an approximate Hamiltonian with a reasonable local potential, the generalized tiling theorem discussed in Sec. \[sec:tiling\] shows that all the nodal pockets are equivalent by symmetry. The DMC energy is therefore independent of the initial distribution of walkers among the nodal pockets. The ordinary fixed-node approximation provides a good example of these ideas. The many-electron ground state is never the overall ground state (which is bosonic), and may be degenerate, in which case we cannot prove the existence of a variational principle by analyzing the problem using the full symmetry group. We can, however, use the permutation group, which is always a subgroup of the full symmetry group. The many-electron trial function transforms according to the one-dimensional antisymmetric irreducible representation of this subgroup. The weaker variational principle therefore guarantees that the fixed-node DMC energy is greater than or equal to the energy of the many-electron ground state; and the generalized tiling theorem guarantees that the exact many-electron ground state possesses the tiling property with respect to permutations. This shows that the ground-state versions of the fixed-node variational principle [@reynolds] and tiling theorem [@tiling] may be regarded as special cases of the more general versions discussed in this paper. The different members of a set of trial functions forming a basis for a multi-dimensional irreducible representation of ${\mathcal G}$ have different nodal surfaces and need not all produce the same fixed-node DMC energy. The strength of the weaker variational principle may therefore be optimized by using specific linear combinations of these basis functions. The best linear combinations transform according to one-dimensional irreducible representations of subgroups of ${\mathcal G}$, and may be found following the procedure explained in Sec. \[sec:d&gt;1\]. If this procedure is not carried out, trial functions belonging to multi-dimensional irreducible representations of ${\mathcal G}$ usually have nodal surfaces with very little spatial symmetry. In many cases, the only symmetry operations that leave the nodal surface invariant are the elements of the permutation group, and the only variational principle that survives is the one relating to the many-electron ground state. Acknowledgements {#sec:Acknowledgements} ================ We thank the Engineering and Physical Sciences Research Council, UK, for financial support under grants GR/L40113 and GR/M05348. Present address: Lawrence Livermore National Laboratory, Livermore, CA 94550, USA. D. Ceperley, G. Chester, and M. Kalos, Phys. Rev. B [**16**]{}, 3081, (1971). B. L. Hammond, W. A. Lester, Jr., and P. J. Reynolds, [*Monte Carlo Methods in Ab Initio Quantum Chemistry*]{}, (World Scientific, Singapore, 1994). J. B. Anderson, J. Chem. Phys. [**65**]{}, 4121, (1976). R. M. Grimes, B. L. Hammond, P. J. Reynolds, and W. A. Lester, Jr., J. Chem. Phys. [**85**]{}, 4749, (1986). L. Mitáš, Comput. Phys. Commun. [**96**]{}, 107, (1996). L. Mitáš, [*Electronic Properties of Solids Using Cluster Methods*]{}, eds. T. A. Kaplan and S. D. Mahanti (Plenum, New York, 1995), p151. A. J. Williamson, R. Q. Hood, R. J. Needs, and G. Rajagopal, Phys. Rev. B [**57**]{}, 12140, (1998). D. Ceperley, J. Stat. Phys. [**63**]{}, 1237 (1991). G. Ortiz, D. M. Ceperley, and R. M. Martin, Phys. Rev. Lett. [**71**]{}, 2777 (1993). P. J. Reynolds, D. M. Ceperley, B. J. Alder, and W. A. Lester, Jr., J. Chem. Phys. [**77**]{}, 5593 (1982). J. W. Moskowitz, K. E. Schmidt, M. A. Lee, and M. H. Kalos, J. Chem. Phys. [**77**]{}, 343 (1982). The results presented in this section also hold for infinite order compact Lie groups, and hence are sufficiently general to encompass most fixed-node DMC simulations of real electronic systems. M. Caffarel and P. Claverie, J. Chem. Phys. [**88**]{}, 1100 (1987). J. F. Cornwell, [*Group Theory in Physics*]{}, (Academic Press, London, 1984), Vol. 1. See, for example, [*Handbook of Mathematical Functions*]{}, eds. M. Abramowitz and I. A. Stegun (Dover, New York, 1972), p722.
{ "pile_set_name": "ArXiv" }
--- abstract: 'The Rastall gravity is a modification of Einstein’s general relativity, in which the energy-momentum conservation is not satisfied and depends on the gradient of the Ricci curvature. It is in dispute whether the Rastall gravity is equivalent to the general relativity (GR). In this work, we constrain the theory using the rotation curves of Low Surface Brightness (LSB) spiral galaxies. Through fitting the rotation curves of LSB galaxies, we obtain the parameter $\beta$ of the Rastall gravity. The $\beta$ values of LSB galaxies satisfy Weak Energy Condition (WEC) and Strong Energy Condition(SEC). Combining the $\beta$ values of type Ia supernovae and gravitational lensing of elliptical galaxies on the Rastall gravity, we conclude that the Rastall gravity is equivalent to the general relativity.' author: - Meirong Tang Zhaoyi Xu Jiancheng Wang title: Observational Constraints on the Rastall gravity from Rotation Curves of Low Surface Brightness Galaxies --- INTRODUCTION ============ One hundred years ago, Einstein proposed the general theory of relativity describing the gravity successfully. One of the important fundamental of the GR is to assume that the covariant derivative of the energy-momentum tensor is zero, and the GR naturally satisfies the equivalence principle. Rastall generalized the covariant conservation of energy-momentum tensor [@1972PhRvD...6.3357R; @1976CaJPh..54...66R], and obtained the conservation equation of energy-momentum tensor with the form $T^{\mu\nu}_{;\mu}=\lambda R^{,\nu}$, where $T^{\mu\nu}$ is energy momentum tensor, $R$ is Ricci curvature (or Ricci scalar) and $\lambda$ is the parameter of the Rastall gravity. This theory can reduce to the GR in asymptotically flat space-time. However, the Rastall gravity is still a controversial gravitational theory. One view is that the Rastall gravity is equivalent to the GR [@2018PhLB..782...83V], the parameter $\lambda$ represents the re-arrangement of perfect fluid matter. According to this view, we just need to redefine the energy-momentum tensor for satisfying the covariant conservation. The only change is the addition of matter fields with different distributions in space-time. On the contrary, [@2018EPJC...78...25D] considered that the Rastall gravity is not equivalent to the GR. The Rastall gravity strengthens the role of Mach principle in gravity theory [@2006gr.qc....10070M], in which the local structure depends on the distribution of matter in entire space-time. Although the nature of the Rastall gravity is not clear enough, we try to constrain its property by observational events. On the cosmological scale, [@2013EPJC...73.2425B] used the data of type Ia supernovae to analyse the gravity model and achieved some good results. Recently, *Li and Xu et al.* [^1] acquired the measured value of the Rastall gravity parameter on the scale of elliptical galaxies, and supposed that the Rastall gravity can explain the mass distribution of elliptical galaxies. The measured value is consistent with one constrained by energy condition, but the result of the Rastall gravity in accordance with the GR on a large scale is the requirement of the theory itself. To test the result, we need the data of galaxies on a smaller scale, such as galaxy nucleus and spiral galaxies. In this study, we will use the data of rotation curves from low surface brightness (LSB) galaxies to constrain the Rastall gravity. This paper is organized as follows. In Section \[metric\], we introduce the perfect fluid dark matter (PFDM) and Rastall gravity. In Section \[motion\], we constrain the Rastall gravity model from the rotation curves of LSB spiral galaxies. In Section \[discussion\], we discuss our results and compare them with previous results. Finally, in Section \[conclusion\], we present our conclusions. PERFECT FLUID DARK MATTER AND RASTALL GRAVITY {#metric} ============================================= Perfect fluid dark matter in GR ------------------------------- [@2003CQGra..20.1187K] obtained the black hole solution of perfect fluid matter in GR. They assumed that the equation of state defined by the ratio of pressure to density of perfect fluid matter $\omega=p/\rho$ is a constant, where $p$ and $\rho$ are the pressure and density of perfect fluid matter, respectively. The expression of black hole solution is $$ds^{2}=-f(r)dt^{2}+\dfrac{1}{g(r)}dr^{2}+r^{2}d\Omega^{2}, \label{BH_solution_GR}$$ where $f(r)$ and $g(r)$ are written as $$f(r)=g(r)=1-\dfrac{2M}{r}-\dfrac{\alpha}{r^{1+3\omega}}. \label{BH1_solution_GR}$$ $M$ represents the mass of black hole, and $\alpha$ is the intensity parameter of perfect fluid matter around black hole. If the equation of state is given by $\omega=-1/3$, the black hole solution represents a Schwarzschild black hole under the perfect fluid dark matter background [@2003gr.qc.....3031K; @2017PhRvD..95f4015X; @2010PhLB..694...10R; @2000astro.ph..3105G]. In general, the perfect fluid dark matter (PFDM) is quintessence matter, because only in this kind of matter, the equation of state is possibly equal to $-1/3 $. At the same time, the solution can also be understood as the black hole solution under the PFDM model. In this situation, the flatness of rotation curves of spiral galaxies at a long distance can be explained naturally. Here, we assume that this property continues to be valid in the Rastall gravity. Perfect fluid dark matter in Rastall gravity -------------------------------------------- [@2017PhLB..771..365H] generalized [@2003CQGra..20.1187K] solution from the GR to the Rastall gravity, and obtained the spherically symmetric black hole solution in perfect fluid matter. This solution has a form of $$f(r)=g(r)=1-\dfrac{2M}{r}-\alpha r^{-\dfrac{1+3\omega-6\beta(1+\omega)}{1-3\beta(1+\omega)}}, \label{BH1_solution_Rg}$$ where $\kappa\lambda$ is a parameter of the Rastall gravity, which determines the distribution of perfect fluid matter. For convenience, we write $\kappa\lambda$ as $\beta$ throughout this article, i.e. $\beta=\kappa\lambda$. For the PFDM ($\omega=-1/3$), the energy density $\rho_{DM}$ can be derived from the Einstein equation. Because the motion velocity of dark matter particle is much smaller than the speed of light, the energy density of the PFDM can approximate to the mass density. Here, from [@2017PhRvL.119k1102K], the baryon matter can be treated as an index disk, i.e. $\rho_{b}=\Sigma_{0}exp[-r/r_{d}]\delta(z)$, where $\Sigma_{0}$ and $r_{d}$ are the central surface density and scale radius of the disk, respectively. In this space-time metric, using the mass density of the PFDM halo and baryon disk, we can calculate the total mass function described as $M(r) =4\pi\int_{0}^{r}\rho_{DM}r^{2}dr+2\pi\int_{0}^{r}\rho_{b}rdr$. We then obtain the rotation velocity of stars on the equatorial plane written as [@2018EPJC...78..513X] $$\upsilon(r) =\sqrt{\dfrac{GM(r)}{r}} =\sqrt{\dfrac{G\alpha}{2}\dfrac{1-4\beta}{1-2\beta}r^{\dfrac{4\beta}{1-2\beta}}-2\pi G\Sigma_{0}r_{d}exp[\dfrac{-r}{r_{d}}](1+\dfrac{r_{d}}{r})} . \label{rotation_velocity}$$ where G is the gravitational constant. We will use this equation to fit the rotation curves of LSB galaxies, and get the $\beta$ values of the Rastall gravity. The parameter $r_{d}$ is $2 kpc$ [@2017PhRvL.119k1102K] in section 3. Energy condition in Rastall gravity ----------------------------------- In the theory of gravity, it is extremely difficult to solve the equation of gravitational field. Through the Einstein field equation, we can know that the distribution of energy-momentum tensor determines the structure of space-time. Due to the complexity of matter distribution, the energy- momentum tensor can not be expressed by a specific form. Therefore, the certain conditions that the energy density is greater than or equal to zero were used to study the gravitational field equation. In 1955, Raychaudhuri formally proposed the basic equation of energy condition, such as weak energy condition and strong energy condition. Under these energy conditions, the fundamental properties of gravity are satisfied. In the references [@2017PhLB..771..365H] and [@2018EPJC...78..513X], they made a specific study on the energy conditions of the Rastall gravity, and found that the constraint of weak energy condition and strong energy condition on the Rastall parameter $\beta$ is equal in the assumption of perfect fluid. They can be given by $$(3\beta(1+\omega)-3\omega)(1-4\beta)\geq0. \label{WEC_SEC}$$ If $\omega=-1/3$, the perfect fluid matter is described by the PFDM model, and the range of the Rastall parameter $\beta$ is $-1/2<\beta<1/4$. If the $\beta$ obtained by fitting the observation data is within this range and is a constant on the scale of spiral galaxy, elliptical galaxy and cosmology, the Rastall gravity will be supported. In contrast, the model will be excluded. CONSTRAINTS FROM ROTATION CURVES OF LSB GALAXIES {#motion} ================================================ In this section, according to the Eq.(\[rotation\_velocity\]), we adopt the Bayesian method [@2010arXiv1008.4686H] to fit the rotation curves of 16 LSB spiral galaxies, and obtain the good fits overall, with $\chi^{2}/dof<1$ for 15 galaxies ( F563-1, F568-3, F583-1, F571-8, F579-v1, F583-4, F730-v1, U5750, U11454, U11616, U11648, U11819, ESO0140040, ESO2060140, ESO3020120 ), and $\chi^{2}/dof<2$ for one galaxy ( ESO4250180 ). Here, the predicted velocity $\upsilon_{pre}$ is from Eq.(\[rotation\_velocity\]) as $\upsilon(r)$, and the observed velocity $\upsilon_{obs}$ is from an astronomical website. For each galaxy, we assume that it has $i$ data points. Therefore, the likelihood function can be expressed as $$ln\mathcal{L}=-\dfrac{1}{2}\sum\limits_{i= 0}^{16}[\dfrac{(\upsilon_{pre}^{i}-\upsilon_{obs}^{i})^{2}}{s_{i}^{2}}+ln(2\pi s_{i}^{2})], \label{likelihood_function}$$ where $$s_{i}^{2}=\delta^{2}+(\upsilon_{err}^{i})^{2}, \label{likelihood1_function}$$ and $\delta$ is the intrinsic scatter between $\upsilon_{pre}$ and $\upsilon_{obs}$, which is considered as a free parameter in our Bayesian analysis. $\upsilon_{err}$ is the measurement error of $\upsilon_{obs}$. Now, the posterior probability function is written as $$p(\alpha, \beta, \Sigma_{0}, \delta|{\upsilon_{obs}})=\mathcal{L}({\upsilon_{obs}|\alpha, \beta, \Sigma_{0}, \delta})p(\alpha, \beta, \Sigma_{0}, \delta). \label{posterior_probability}$$ Here, for each LSB galaxy, we choose a flat prior $p(\alpha, \beta, \Sigma_{0}, \delta)$, and use the Python implementation named Emcee [@2013PASP..125..306F] along with four free parameters $\alpha, \beta, \Sigma_{0}, \delta$ to fit the rotation curves. Our results are shown in Table 1. Table 1. Best fitting results from the rotation curves of 16 LSB spiral galaxies using the eq.(\[rotation\_velocity\]). Column (1), (2) and (3 ) are the name of galaxy, the fitting values of Rastall parameter $\beta$ and the $\chi^{2}$ values, respectively.\ \[2mm\] ------------ --------- ---------------- Galaxy $\beta$ $\chi^{2}/dof$ (1) (2) (3) F563-1 0.054 0.879 F568-3 0.154 0.866 F583-1 0.15 0.836 F571-8 0.143 0.882 F579-v1 0.048 0.13 F583-4 0.141 0.211 F730-v1 0.095 0.53 U5750 0.146 0.82 U11454 0.118 0.826 U11616 0.122 0.808 U11648 0.132 0.356 U11819 0.147 0.958 ESO0140040 0.083 0.757 ESO2060140 0.1 0.812 ESO3020120 0.135 0.397 ESO4250180 0.121 1.675 ------------ --------- ---------------- DISCUSSION ========== In this study, we constrain the parameter $\beta$ of the Rastall gravity by fitting the data of the rotation curves from 16 LSB spiral galaxies. Comparing the previous results of type Ia supernovae and elliptical galaxies, we support that the Rastall gravity is equivalent to the GR. On the scale of spiral galaxy, the values of parameter $\beta$ we obtained are of the order of $10^{-1}$ and within the limits of strong energy condition in the Rastall gravity. On the cosmological scale, [@2013EPJC...73.2425B] used the data of type Ia supernovae to constrain the parameter $\beta$, and found that the $\beta$ value is of the order of $10^{-4}$. This is inconsistent with our results by three orders of magnitude. If our analysis is correct, then the value of the Rastall parameter will cause some troubles. For example, the value of the parameter $\beta$ on the scale of spiral galaxy is not same as the value on the cosmological scale, which indicates that the parameter $\beta$ is not universally applicable. On the other hand, *Li and Xu et al.* recently used the gravitational Lensing data of elliptical galaxies to constrain the Rastall parameter. Their $\beta$ values obtained by fitting the observed data are consistent with ours, which implies that the parameter values of the Rastall gravity are consistent between elliptical galaxies and spiral galaxies. At the same time, according to the Rastall gravity model, the result of the Rastall gravity in accordance with the GR on a large scale is the requirement of the theory itself, and the difference only appears on a smaller scale. Our results will be more convincing. The nature of the Rastall gravity is in dispute. What is the explanation of our results for this? If the Rastall gravity is equivalent to the GR, the $\beta$ only represents the distribution of matter, and the different systems will have different values, therefore it is easy to understand the difference of the $\beta$ between the galactic scale and the cosmological scale. If the Rastall gravity is not equivalent to the GR, on the basis of the assumption of the Rastall gravity, the $\beta$ values should be same for the galactic scale and the cosmological scale, but this is contrary to our analysis results. So we will conclude that the Rastall parameter $\beta$ can only be understood as the parameter determining the distribution of matter in space-time. Thus our results support that the Rastall gravity is equivalent to the GR. CONCLUSIONS {#conclusion} =========== In this work, using the rotation curves of 16 LSB spiral galaxies, we obtained the values of parameter $\beta$ in the Rastall gravity model. These values are about 0.1 and satisfy the strong energy condition. And after comparing the results of type Ia supernovae and elliptical galaxies, we found that our results support that the Rastall gravity is equivalent to the GR, and then the values of parameter $\beta$ can only be understood as the re-arrangement of matter in space-time, the difference between the galactic scale and the cosmological scale can be easily explained. We thank Kuzio de Naray, R., McGaugh, S.S., and de Blok for providing us the rotation curve data and thank Li Rui for his help in programming. We acknowledge the anonymous referee for a constructive report that has significantly improved this paper. We acknowledge the financial support from the National Natural Science Foundation of China under grants No. 11503078, 11573060 and 11661161010. [99]{} Batista, C. E. M., Fabris, J. C., Piattella, O. F., & Velasquez-Toribio, A. M. 2013, European Physical Journal C, 73, 2425 Darabi, F., Moradpour, H., Licata, I., Heydarzade, Y., & Corda, C. 2018, European Physical Journal C, 78, 25 Foreman-Mackey, D., Hogg, D. W., Lang, D., & Goodman, J. 2013, , 125, 306 Guzman, F. S., Matos, T., Nunez, D., & Ramirez, E. 2000, arXiv:astro-ph/0003105 Heydarzade, Y., & Darabi, F. 2017, Physics Letters B, 771, 365 Kamada, A., Kaplinghat, M., Pace, A. B., & Yu, H.-B. 2017, Physical Review Letters, 119, 111102 Kiselev, V. V. 2003, Classical and Quantum Gravity, 20, 1187 Kiselev, V. V. 2003, arXiv:gr-qc/0303031 Majernik, V., & Richterek, L. 2006, arXiv:gr-qc/0610070 Rahaman, F., Nandi, K. K., Bhadra, A., Kalam, M., & Chakraborty, K. 2010, Physics Letters B, 694, 10 Rastall, P. 1976, Canadian Journal of Physics, 54, 66 Rastall, P. 1972, , 6, 3357 Visser, M. 2018, Physics Letters B, 782, 83 Xu, Z., Hou, X., Gong, X., & Wang, J. 2018, European Physical Journal C, 78, 513 Xu, Z., & Wang, J. 2017, , 95, 064015 Hogg, D. W., Bovy, J., & Lang, D. 2010, arXiv:1008.4686 [^1]: this article is being reviewed, but not on arXiv
{ "pile_set_name": "ArXiv" }
--- abstract: 'Kippenhahn’s Theorem asserts that the numerical range of a matrix is the convex hull of a certain algebraic curve. Here, we show that the joint numerical range of finitely many Hermitian matrices is similarly the convex hull of a semi-algebraic set. We discuss an analogous statement regarding the dual convex cone to a hyperbolicity cone and prove that the class of bases of these dual cones is closed under linear operations. The result offers a new geometric method to analyze quantum states.' author: - Daniel Plaumann - Rainer Sinn - Stephan Weis date: 'June 19th, 2020' title: 'Kippenhahn’s Theorem for Joint Numerical Ranges and Quantum States' --- Introduction {#sec:intro} ============ Let $H_d$ be the real vector space of Hermitian $d\times d$ matrices. We denote the set of [*density matrices*]{} by $$\label{eq:density-matrices} {\mathcal{B}}={\mathcal{B}}_d =\{\rho\in H_d \colon \rho\succeq 0, \operatorname{tr}(\rho)=1\},$$ where $A\succeq 0$ means that $A\in H_d$ is positive semi-definite. The letter ${\mathcal{B}}$ underlines that the set ${\mathcal{B}}$ is a base of the cone of positive semi-definite matrices (see ). We use the bilinear form $\langle A,B\rangle=\operatorname{tr}(AB)$, $A,B\in H_d$, to identify $H_d$ and its dual space. Let $A_1,A_2,\dots,A_n\in H_d$, $n\in{\mathbb{N}}$, and define $$\label{eq:jnr} W = W_{A_1,A_2,\ldots,A_n} = \{(\langle \rho, A_1\rangle,\dots,\langle \rho, A_n\rangle) \colon \rho\in{\mathcal{B}}_d\},$$ a convex compact subset of the dual space $({\mathbb{R}}^n)^\ast$ to ${\mathbb{R}}^n$. The set $W$ has been called [*joint numerical range*]{} in operator theory, see Section 2 of [@BonsallDuncan1971] (also joint algebraic numerical range [@Mueller2010]). Our motivation for this paper is matrix theory and quantum mechanics. Physicists call density matrices [*quantum states*]{}, as density matrices describe the physical states of a quantum system including all statistical properties [@Holevo2011]. Hence, our results contribute to the geometry of quantum states [@AubrunSzarek2017; @BengtssonZyczkowski2017]. presents numerical range methods in quantum mechanics. Perhaps more commonly, the term joint numerical range refers to $$\label{eq:jnr-pure} W^\sim = W^\sim_{A_1,A_2,\ldots,A_n} = \{(\langle\psi|A_1\psi\rangle,\dots,\langle\psi|A_n\psi\rangle) \colon \psi\in{\mathbb{C}}^d, \langle \psi|\psi\rangle=1\},$$ where $\langle\varphi|\psi\rangle$ is the standard inner product of $\varphi,\psi\in{\mathbb{C}}^d$, see [@ChienNakazato2010; @LiPoon2000]. A [*pure state*]{} is a projection $\rho$ onto the span of a unit vector $\psi\in{\mathbb{C}}^d$. Since $\langle\rho,A\rangle=\langle\psi|A\psi\rangle$ holds for all $A\in H_d$, the set $W^\sim$ is a linear image of the set of pure states. As the pure states are the extreme points of the set of density matrices ${\mathcal{B}}$, the joint numerical range $W$ is the convex hull of $W^\sim$. Thus, $W=W^\sim$ holds when $W^\sim$ is convex. This is the case if by the Toeplitz-Hausdorff theorem [@Hausdorff1919; @Toeplitz1918], whose 100th anniversary we celebrated at the time of writing. Note that $W_{A_1,A_2}^\sim$ is the standard [*numerical range*]{} $W(A) = \{\langle\psi|A\psi\rangle : \psi\in{\mathbb{C}}^d, \langle \psi|\psi\rangle=1\} \subset{\mathbb{C}}$ of $A=A_1+{\operatorname{i}}A_2$. Also, $W^\sim$ is convex if $n=3$ and $d\geq 3$ [@Au-YeungPoon1979]. The convexity of $W^\sim$ is an open problem for $n>3$, see [@Gutkin2004; @LiPoon2000]. Numerical ranges are generally nonconvex if the complex field is replaced with the skew field of the quaternions [@Rodman2014 p. 39]. Algebraic geometry has been employed to study numerical ranges since the 1930s. We consider the determinant $$p=\det( x_0 {\mathds{1}}+ x_1 A_1 + \cdots + x_n A_n ) \in{\mathbb{R}}[x_0,x_1,\dots,x_n],$$ where ${\mathds{1}}$ denotes the $d\times d$ identity matrix, and the complex projective hypersurface $${\mathcal{V}}(p)= \{x\in{\mathbb{P}}^n \mid p(x) = 0 \}.$$ If $n=2$, then ${\mathcal{V}}(p)\subset{\mathbb{P}}^2$ is an algebraic curve. Murnaghan [@Murnaghan1932] showed that the eigenvalues of the matrix $A_1+{\operatorname{i}}A_2$ are the foci of the curve $$T=\{ y_1+{\operatorname{i}}y_2 \mid y_1,y_2\in{\mathbb{R}},(1:y_1:y_2)\in {\mathcal{V}}(p)^\ast\}\subset{\mathbb{C}},$$ where $X^\ast\subset(\mathbb{P}^n)^\ast$ denotes the dual variety parametrizing hyperplanes tangent to a variety $X\subset\mathbb{P}^n$ (cf.  for more details). Kippenhahn recognized the meaning of the convex hull of the curve $T$. \[thm:k\] The numerical range of $A_1+{\operatorname{i}}A_2$ is the convex hull of the curve $T$, in other words, $W(A_1+{\operatorname{i}}A_2)=\operatorname{conv}(T)$. is a well-known tool in matrix analysis [@Bebiano-etal2005; @Fiedler1981; @GallaySerre2012; @JoswigStraub1998]. The curve $T\subset{\mathbb{C}}$ is called the [*Kippenhahn curve*]{} or the [*boundary generating curve*]{} of the numerical range $W(A_1+{\operatorname{i}}A_2)$. The curve $T$ has proven useful in classifications of numerical ranges of $3$-by-$3$ matrices [@Keeler-etal1997; @Kippenhahn1951] and $4$-by-$4$ matrices [@Camenga-etal2019; @ChienNakazato2012], and it has been studied for special matrices [@ChienNakazato2018; @GauWu2004; @Gau-etal2013; @Militzer-etal2017]. We sketch a proof of in a manner that may help to explain the geometry behind the proof of later on: Consider the convex set $S=\{(x_1,x_2)^T\in{\mathbb{R}}^2\mid {\mathds{1}}+ x_1 A_1+x_2 A_2\succeq 0\}$ (a [*spectrahedron*]{}). Assume for simplicity that $S$ is compact, $X={\mathcal{V}}(p)$ is smooth, and that the degree $d$ of $p$ is even. The curve $X$ is hyperbolic, i.e. its real points consist of $\frac d2$ nested ovals in the real projective plane. The innermost oval is the boundary of $S$. All but finitely many points of the dual curve $X^\ast$ correspond to simple tangent lines to $X$. The set of real points of the dual curve $X^\ast$ (of degree $d(d-1)$) again consists of $\frac d2$ nested connected components, together with at most finitely many isolated real (singular) points. The tangent lines to the boundary of $S$ now correspond to the outermost oval of $X^\ast$, since all other tangent lines to $X$ do not pass through $S$ (see ). The outermost oval therefore bounds the convex dual $S^\circ$ of $S$, which is exactly the numerical range $W(A_1+{\operatorname{i}}A_2)$. The claim of Kippenhahn’s theorem follows if we can show that none of the isolated real singularities of $X^\ast$ lie outside of $W$; see . Chien and Nakazato [@ChienNakazato2010] provided a more rigorous proof of compared to Kippenhahn’s. They also found a triple of Hermitian $3\times 3$-matrices for which the literal analogue of fails in dimension $n=3$. We will see that the last part of the above sketch in the plane case, the position of singular points of the dual curve, is exactly what causes the failure of the theorem in higher dimensions. This can also be seen in the counterexample by Chien and Nakazato, . By removing all singular points from the projective dual variety ${\mathcal{V}}(p)^\ast$ and taking Euclidean closure, we will prove that a modified version of is valid in all dimensions. Let $X_1,\dots,X_r$ denote the irreducible components of the hypersurface ${\mathcal{V}}(p)$. We consider the set $(X_i^\ast)_{\mathrm{reg}}$ of the regular points of the dual variety $X_i^\ast$, the set $$T_i = \left\{ (y_1,\dots,y_n) \in ({\mathbb{R}}^n)^\ast \mid (1:y_1:\dots:y_n)\in (X_i^\ast)_{\mathrm{reg}}\right\}, \qquad i=1,\dots,r,$$ and the Euclidean closure $T^\sim=\operatorname{clos}(T_1\cup\cdots\cup T_r)$ of the union $T_1\cup\cdots\cup T_r$. Our main result is as follows. \[eq:modified-K\] The joint numerical range $W$ is the convex hull of $T^\sim$. \[rem:main-theorem\] 1. implies that the set $T^\sim$ contains all extreme points of the compact, convex set $W$. We show in the proof of that $T^\sim$ contains the exposed points of $W$, and hence all of the extreme points by a limiting argument (Straszewicz’s Theorem). But just as in Kippenhahn’s original theorem, $T^\sim$ is not necessarily contained in the boundary of $W$, only in $W$. 2. We point out that , which implies the theorem stated here, holds more generally for hyperbolic hypersurfaces rather than just determinantal hypersurfaces. While this makes no difference in the plane, by the Helton-Vinnikov theorem, the statement is indeed more general in higher dimensions, and the proof relies purely on the real geometry of hyperbolic polynomials. 3. The joint numerical range $W$ is a semi-algebraic set as it is a linear image of the semi-algebraic set ${\mathcal{B}}$ by quantifier elimination (see e.g. [@BCR Thm. 2.2.1]). The set $T_1\cup\cdots\cup T_r$ and hence its Euclidean closure $T^\sim$ are semi-algebraic sets as well. 4. not only describes a semi-algebraic set that contains the extreme points of $W$ but more precisely the Zariski closure of the set of extreme points: The union of the dual varieties $X_i^\ast$ of the irreducible components $X_i$ of the algebraic boundary of the hyperbolicity cone of $p$ is the Zariski-closure of the set of extreme points of $W$, see for details. We organize the article as follows. collects preliminaries from convex geometry and real algebraic geometry. presents a detailed discussion of the remarkable fact, proved by the second author in [@Sinn2015], that the dual convex cone $C^\vee$ to a hyperbolicity cone $C$ is the closed convex cone generated by a particular semi-algebraic set. This implies that every base of $C^\vee$ is the closed [*convex hull*]{} of a section of that semi-algebraic set. The same is true for linear images of the bases as we show in , because (up to isomorphism) they are bases of dual convex cones to sections of $C$, which are hyperbolicity cones themselves. Returning to the cone of positive semi-definite matrices in , we obtain a proof of and discuss examples. We analyze the case $n=2$ separately in , which yields a proof of Kippenhahn’s original result as stated in . Connections to Quantum Mechanics {#sec:qm} ================================ Physicists refer to linear images of certain subsets of the set of quantum states ${\mathcal{B}}_d$ as *numerical ranges*. Often they consider images under a map ${\mathcal{B}}_d\to{\mathbb{R}}^n$, $\rho\mapsto\langle \rho, A_i\rangle_{i=1}^n$, where $A_1,A_2,\dots,A_n\in H_d$ are Hermitian matrices. We discuss examples where algebraic geometry could help solving problems of quantum mechanics in the context of numerical ranges. In this paper, we ignore numerical ranges outside the pattern of linear images of subsets of ${\mathcal{B}}_d$, for example higher-rank numerical ranges [@Choi-etal2006]. Linear images of the set of all quantum states {#sec:qm-all} ---------------------------------------------- The joint numerical range $W_{A_1,A_2,\dots,A_n}$, as defined in , appears in problems of experimental and theoretical physics. The geometry of the joint numerical range has been of direct interest to the experimentalists Xie et al. [@Xie-etal2019]. Their drawings of data from photonic experiments show ellipses on the boundary of the joint numerical range $W_{A_1,A_2,A_3}$ of three $3$-by-$3$ matrices $A_1,A_2,A_3\in H_3$, clearly in agreement with the classification: the exposed faces of positive dimensions are ellipses and segments assembling one of ten configurations [@Szymanski-etal2018]. The possibility to carry out experiments with $4$-level quantum systems calls for a similar classification of $W_{A_1,A_2,A_3}$ for $4$-by-$4$ matrices $A_1,A_2,A_3\in H_4$. Using and analyzing the Kippenhahn curve, Kippenhahn [@Kippenhahn1951] obtained a classification of the numerical range $W(A_1+{\operatorname{i}}A_2)=W_{A_1,A_2}$ in terms of flat portions on the boundary of $W_{A_1,A_2}$ for all Hermitian matrices $A_1,A_2\in H_3$, see also [@Keeler-etal1997]. A similar approach has been taken for $4$-by-$4$ matrices [@Camenga-etal2019; @ChienNakazato2012]. We are invited to describe the exposed faces of positive dimensions of the joint numerical range $W_{A_1,A_2,A_3}$, employing and studying the semi-algebraic set $T^\sim$, whose convex hull is $W_{A_1,A_2,A_3}$. This should reproduce the classification of the set $W_{A_1,A_2,A_3}$ for $3$-by-$3$ matrices [@Szymanski-etal2018] and lead to a classification for $4$-by-$4$ matrices. Quantum thermodynamics [@YungerHalpern-etal2016] describes equilibrium states with multiple conserved quantities $A_1,A_2,\dots,A_n$ in terms of *generalized Gibbs states* $$\rho_x=\tfrac{e^{x_1A_1+\dots+x_nA_n}}{\operatorname{tr}e^{x_1A_1+\cdots+x_nA_n}}\in{\mathcal{B}}_d, \qquad x=(x_1,\dots,x_n)^T\in{\mathbb{R}}^n.$$ The “boundary at infinity” ($|x|\to\infty$) to the manifold $\{\rho_x : x\in{\mathbb{R}}^n\}$ may not be closed in the Euclidean topology if the matrices $A_1,A_2,\dots,A_n$ fail to commute [@WeisKnauf2012]. The physical meaning of this topological problem remains mysterious [@Chen-etal2015]. Mathematically, the discontinuity depends on the geometry of the joint numerical range $W_{A_1,A_2,\dots,A_n}=\operatorname{conv}(T^\sim)$ [@Chen-etal2015; @Rodman-etal2016; @SpitkovskyWeis2018]. Hence, the semi-algebraic set $T^\sim$ explains infinitesimal properties of the manifold of generalized Gibbs states near the boundary at infinity. A connection between functional analysis and algebraic geometry awaits further investigation. The *Wigner distribution* of a quantum state $\rho\in{\mathcal{B}}_d$ with respect to $A_1,A_2,\dots,A_n$ is the tempered distribution ${\mathcal W}_\rho$ on ${\mathbb{R}}^n$ that satisfies $$\textstyle \int \operatorname{d}\!a \,{\mathcal W}_\rho(a_1,\dots,a_n) f(x_1 a_1 + \cdots + x_n a_n) = \langle\rho,f(x_1 A_1 + \cdots + x_n A_n)\rangle$$ for all $x\in{\mathbb{R}}^n$ and all infinitely differentiable functions $f:{\mathbb{R}}\to{\mathbb{C}}$. The Wigner distribution is a common tool in quantum optics and theoretical physics. Schwonnek and Werner [@SchwonnekWerner2018] showed that the distribution ${\mathcal W}_\rho$ is compactly supported on the joint numerical range $W_{A_1,A_2,\dots,A_n}$ and that the singularities of ${\mathcal W}_\rho$ lie in the semi-algebraic set $T^\sim$. Linear images of subsets of the set of quantum states {#sec:qm-subsets} ----------------------------------------------------- Linear images of semi-algebraic subsets of ${\mathcal{B}}_d$ are amenable to algebraic geometry as well. We mention examples relevant to quantum mechanics. Algebraic geometry [@Conca-etal2015] has proven helpful in the theory of pure state tomography [@Heinosaari-etal2013], the reconstruction of a pure state $\rho\in{\mathcal{B}}_d$ from its expected value tuple $\langle\rho,A_i\rangle_{i=1}^n\in W^\sim$ in the linear image $W^\sim$ of the set of pure states defined in . A different topic, for example, in density functional theory [@Chen-etal2012], is describing the extreme points of the joint numerical range $W$. Both $W^\sim$ and $T^\sim$ are semi-algebraic subsets of $W$ that contain the extreme points of $W$. The set $T^\sim$ is especially suitable to study the extreme points of $W$, see . Many-particle systems are fascinating due to interaction and correlation between the units. The simplest example in the quantum domain is the two-qubit system with Hilbert space ${\mathbb{C}}^4={\mathbb{C}}^2\otimes{\mathbb{C}}^2$. Physicists [@Chen-etal2016; @Gawron-etal2010] have studied the *joint product numerical range* of $A_1,A_2,\dots,A_n\in H_4$, $$\Pi =\{\langle\psi\otimes\varphi|A_i\psi\otimes\varphi\rangle_{i=1}^n \mid \psi,\varphi\in{\mathbb{C}}^2, \langle \psi|\psi\rangle=\langle \varphi|\varphi\rangle=1\},$$ a linear image of the set of product states $\sigma\otimes\tau$, where $\sigma,\tau\in{\mathcal{B}}_2$ are pure states. The convex hull of the product states is the set of [*separable*]{} states, the states that lack the genuine quantum correlation called [*entanglement*]{} [@AubrunSzarek2017; @BengtssonZyczkowski2017]. Hence, the set $\Pi$ and its convex hull allow us to study quantum correlations. Two-qubit density matrices offer insights into statistical mechanics. As per the [*quantum de Finetti theorem*]{} [@Lewin-etal2014; @Stormer1969], the two-particle marginals of an infinite bosonic qubit-system are convex combinations of symmetric product states $\sigma\otimes\sigma$, where $\sigma\in{\mathcal{B}}_2$ is a pure state. The ground state energy of an energy operator with two-party interactions on an infinite bosonic qubit-system is the distance of the origin from a supporting hyperplane to the set $$\Pi^\mathrm{sym}_{A_1,A_2,\dots,A_n} = \{\langle\psi\otimes\psi|A_i\psi\otimes\psi\rangle_{i=1}^n \mid \psi\in{\mathbb{C}}^2, \langle \psi|\psi\rangle=1\}$$ for suitable matrices $A_1,A_2,\dots,A_n\in H_4$. Notably, ruled surfaces on the boundary of the convex hull of $\Pi^\mathrm{sym}_{A_1,A_2,A_3}\subset{\mathbb{R}}^3$ are expressions of phase transitions [@Chen-etal2016; @Zauner-etal2016]. An analogue to regarding the set $\Pi^\mathrm{sym}_{A_1,A_2,A_3}$ would be helpful for the analysis of bosonic qubit-systems. Preliminaries {#sec:prelim} ============= We collect terms, basic results, and references to the literature regarding convex geometry and (real) algebraic geometry. Convex Geometry {#sec:conv-geo} --------------- We discuss various notions of cones in a finite-dimensional real vector space $V$: cones (which may be nonconvex), convex cones, and normal cones. Additionally, we define affine cones over complex projective varieties in . As a general reference for convex geometry, we recommend [@Barvinok2002; @Rockafellar1970]. A subset $C$ of $V$ is a *cone* if $\lambda x\in C$ whenever $\lambda>0$ and $x\in C$. A subset $K$ of $V$ is a *convex cone* if $K$ is a nonempty convex set and if $\lambda x\in K$ whenever $\lambda\geq 0$ and $x\in K$. The *affine hull* $\operatorname{aff}(S)$, *convex hull* $\operatorname{conv}(S)$, *cone*, *convex cone* $\operatorname{cc}(S)$, and *closed convex cone* generated by a subset $S\subset V$ is the smallest affine space, convex set, cone, convex cone, and closed convex cone, respectively, containing $S$. A subset $B\subset V$ is a *base* of a cone $C\subset V$ if $B$ is the intersection of $C$ with an affine hyperplane, $0\not\in\operatorname{aff}(B)$, and for all nonzero points $x\in C$ there exists $y\in B$ and $\lambda>0$ such that $x=\lambda y$. A convex cone $K\subset V$ is *pointed* if $K\cap (-K) = \{0\}$, i.e. if $K$ contains no lines. We denote the dual vector space of $V$ by $V^\ast$. The *annihilator* of a subset $S\subset V$ is $$S^\perp=\{\ell\in V^\ast \colon \ell(x)=0\,\text{ for all } x\in S\},$$ the *dual convex cone* to $S$ is $$S^\vee = \{\ell\in V^\ast \colon \ell(x)\geq 0\,\text{ for all }x\in S \},$$ and the *dual convex set* to $S$ is $$S^\circ = \{\ell\in V^\ast \colon 1+\ell(x)\geq 0\,\text{ for all } x\in S \}.$$ We denote intersections of $S$ with affine hyperplanes avoiding the origin by $$\label{eq:base} \operatorname{h}_\ell(S)=\{x\in S\colon \ell(x)=1\}, \qquad \ell\in V^\ast, \ell\neq 0.$$ \[lem:base-affine\] Let $C\subset V$ be a cone and let $B\neq\emptyset$ be a base of $C$. Then there exists a nonzero functional $\ell\in V^\ast$ such that $B=\operatorname{h}_\ell(C)$. If $C$ admits a compact base, then $C\cup\{0\}$ is closed. Let $B$ be a nonempty base of the cone $C$ and let $X$ be the linear span of $C$. It follows from the definition of a base that there is a nonzero linear functional $\hat{\ell}\in X^\ast$ such that $B\subset\operatorname{h}_{\hat{\ell}}(X)$, hence $B=\operatorname{h}_{\hat{\ell}}(C)$. Identifying the dual space $X^\ast$ with any subspace complementary to the annihilator $X^\perp$ in $V^\ast$ and extending $\hat{\ell}$ to a functional $\ell\in V^\ast$ by the Hahn-Banach theorem, we obtain $B=\operatorname{h}_\ell(C)$. Let $(x_i)\subset C$ be a sequence converging to a nonzero point $x\in V$. If the base $\operatorname{h}_\ell(C)$ of $C$ is compact, then the sequence $(\tfrac{x_i}{\ell(x_i)})\subset\operatorname{h}_\ell(C)$ has a converging subsequence with limit $y\in\operatorname{h}_\ell(C)$. As $\ell(x)=\lim_{i\to\infty}\ell(x_i)\geq 0$ and as $x=\lim_{i\to\infty}\ell(x_i)\tfrac{x_i}{\ell(x_i)}=\ell(x)y$, the point $x$ lies in $C$. The biduality theorem for closed convex cones follows from the separation theorem in convex geometry, see for example [@Rockafellar1970 Theorem 14.1]. \[thm:biduality-cone\] Let $K\subset V$ be a closed convex cone. Then $(K^\vee)^\vee = K$. We describe the family of bases of a closed convex cone. \[lem:char-compact-base\] Let $K\subset V$ be a closed convex cone. The following assertions are equivalent for all nonzero points $x\in V$. 1. The point $x$ is an interior point of $K$. 2. The set $\operatorname{h}_x(K^\vee)$ is a base of $K^\vee$. If one of these equivalent assertions is true, then the set $\operatorname{h}_x(K^\vee)$ is compact. Let $x\in V$ be nonzero. By Theorem 13.1 in [@Rockafellar1970] the point $x$ is an interior point of $K$ if and only if $\ell(x)<\delta^\ast(\ell|K)$ holds for all $\ell\in V^\ast\setminus\{0\}$, where $$\delta^\ast(\ell|K) =\sup_{x\in K}\ell(x) =\left\{\begin{array}{ll} 0 & \text{if $\ell\in-K^\vee$,}\\ \infty & \text{if $\ell\not\in-K^\vee$,} \end{array}\right. \qquad\ell\in V^\ast,$$ is the support function of $K$. Hence the assertion (1) is equivalent to $\ell(x)>0$ for all $\ell\in K^\vee\setminus\{0\}$, which is equivalent to $\operatorname{h}_x(K^\vee)$ being a base of $K^\vee$. The compactness follows from properties of the *recession cone* $0^+(\operatorname{h}_x(K^\vee))$, the set of vectors $\ell'\in V^\ast$ such that $\ell+\lambda\ell'$ lies in $\operatorname{h}_x(K^\vee)$ for all $\ell\in\operatorname{h}_x(K^\vee)$ and $\lambda\geq 0$. We can assume that the set $\operatorname{h}_x(K^\vee)$ is nonempty. In this case we have $0^+(\operatorname{h}_x(K^\vee))=x^\perp\cap K^\vee$ by Coro. 8.3.2 and 8.3.3 of [@Rockafellar1970] as the recession cone of $\operatorname{h}_x(V^\ast)$ is $x^\perp$. Since $\operatorname{h}_x(K^\vee)$ is a base of $K^\vee$, we have $\ell(x)>0$ for all $\ell\in K^\vee\setminus\{0\}$, hence $x^\perp\cap K^\vee=\{0\}$. Now [@Rockafellar1970 Thm. 8.4] shows that $\operatorname{h}_x(K^\vee)$ is bounded, hence compact. Cones that are contained in a pointed closed convex cone behave nicely. \[lem:cn-ccc\] Let $K\subset V$ be a closed convex cone with nonempty interior. Let $C\subset K^\vee$ be a cone and let $x$ be a nonzero interior point of $K$. Then $$\label{eq:compact-base} \operatorname{h}_x(\operatorname{cc}(\operatorname{clos}(C))) =\operatorname{conv}(\operatorname{h}_x(\operatorname{clos}(C))) =\operatorname{conv}(\operatorname{clos}(\operatorname{h}_x(C)))$$ is a compact base of the pointed, closed convex cone $\operatorname{cc}(\operatorname{clos}(C))=\operatorname{clos}(\operatorname{cc}(C))$. Let $x\neq 0$ be an interior point of $K$. Then $\operatorname{h}_x(K^\vee)$ is a compact base of $K^\vee$ by . *A fortiori*, $\operatorname{h}_x(C)$ is a base of the cone $C\subset K^\vee$; thus $\operatorname{h}_x(\operatorname{cc}(C))\subset\operatorname{conv}(\operatorname{h}_x(C))$. The converse inclusion is clear and proves the first equality sign in after replacing $C$ with $\operatorname{clos}(C)$. The inclusion $\operatorname{h}_x(\operatorname{clos}(C))\subset\operatorname{clos}(\operatorname{h}_x(C))$ holds, again as $\operatorname{h}_x(C)$ is a base of $C$. The converse inclusion is clear and proves the second equality sign in . As $\operatorname{h}_x(\operatorname{clos}(C))$ is compact, its convex hull is compact by Theorem 17.2 in [@Rockafellar1970]. Hence $\operatorname{h}_x(\operatorname{cc}(\operatorname{clos}(C)))$ is a compact base of the convex cone $\operatorname{cc}(\operatorname{clos}(C))$, which is closed by . This proves $\operatorname{clos}(\operatorname{cc}(C))\subset\operatorname{cc}(\operatorname{clos}(C))$; the converse inclusion is clear. allows us to focus on irreducible varieties in . \[pro:dual-of-intersection\] Let $K_1,\ldots,K_r\subset V$ be convex cones and let $e\neq 0$ be an interior point of $K=K_1\cap K_2\cap\dots\cap K_r$. Let $C_i\subset V^\ast$ be a cone such that $K_i^\vee=\operatorname{clos}(\operatorname{cc}(C_i))$, that is to say, the dual convex cone to $K_i$ is the closed convex cone generated by $C_i$ for $i=1,\ldots,r$. Then $$\begin{aligned} \label{eq:inter-dual} K^\vee = \operatorname{cc}\big(\operatorname{clos}(C_1\cup C_2\cup\dots\cup C_r)\big)\end{aligned}$$ and $$\begin{aligned} \label{eq:inter-dual-cl} \operatorname{h}_e(K^\vee) = \operatorname{conv}\Big(\operatorname{clos}\big(\operatorname{h}_e(C_1)\cup\operatorname{h}_e(C_2)\cup\dots\cup\operatorname{h}_e(C_r)\big)\Big).\end{aligned}$$ Corollary 16.4.2 of [@Rockafellar1970] shows $K^\vee=K_1^\vee+K_2^\vee+\cdots+K_r^\vee$ as $e$ is an interior point of $K_i$ for all $i=1,\ldots,r$, hence $K^\vee=\operatorname{cc}(K_1^\vee\cup K_2^\vee\cup\cdots\cup K_r^\vee)$. This proves , as we have, by assumption and by , $$K_i=\operatorname{clos}(\operatorname{cc}(C_i))=\operatorname{cc}(\operatorname{clos}(C_i)), \qquad i=1,\ldots,r.$$ Equation follows from the equation $\operatorname{h}_e(\operatorname{cc}(\operatorname{clos}(C)))=\operatorname{conv}(\operatorname{clos}(\operatorname{h}_e(C)))$ provided in using $C=\bigcup_{i=1}^rC_i$ and from . We discuss faces and normal cones. Let $C\subset V$ be a convex subset. A subset $F\subset C$ is a [*face*]{} of $C$ if $F$ is convex and whenever $(1-\lambda)x+\lambda y\in F$ for some $\lambda\in(0,1)$ and $x,y\in C$, then $x$ and $y$ are also in $F$. A subset $F\subset C$ is an [*exposed face*]{} of $C$ if there is a linear functional $\ell\in V^\ast$ such that $x\in F$ if and only if $\ell(x)=\inf_{y\in C}\ell(y)$ for all $x\in C$. A point $x\in C$ is an [*extreme point*]{} (resp. [*exposed point*]{}) of $C$ if $\{x\}$ is a face (resp. exposed face) of $C$. Let ${\mathbb{R}}_+=\{\lambda\in{\mathbb{R}}:\lambda\geq 0\}$. If $x\in C$ is a nonzero point, and ${\mathbb{R}}_+x$ is a face (resp. exposed face) of $C$, then the set ${\mathbb{R}}_+x$ is called an [*extreme ray*]{} (resp. [*exposed ray*]{}) of $C$. We denote the set of all faces of $C$ by ${\mathcal{F}}(C)$. The family ${\mathcal{F}}(C)$ is a complete lattice of finite length under the partial ordering of set inclusion [@Tam1985]. The [*duality operator*]{} of the closed convex cone $C$ is the map $$N_C:{\mathcal{F}}(C)\to{\mathcal{F}}(C^\vee), \qquad F\mapsto F^\perp\cap C^\vee.$$ The map $N_C$ is antitone, that is to say, $F\subset G$ implies $N_C(F)\supset N_C(G)$ for all faces $F,G\in{\mathcal{F}}(C)$. The image of $N_C$ is the set of nonempty exposed faces of $C^\vee$ and [@Tam1985 Prop. 2.4] shows $F\subset N_{C^\vee}\circ N_C(F)$ for all faces $F\in{\mathcal{F}}(C)$. We have just confirmed that the pair of duality operators $N_C,N_{C^\vee}$ defines a *Galois connection* between ${\mathcal{F}}(C)$ and ${\mathcal{F}}(C^\vee)$. Theorem 20 in Section V.8 of [@Birkhoff1973] then proves the following assertion, which we use in . \[lem:lattice-iso\] Let $C\subset V$ be a closed convex cone. The duality operator $N_C$ restricts to an antitone lattice isomorphism from the set of nonempty exposed faces of $C$ to the set of nonempty exposed faces of $C^\vee$. The inverse isomorphism is the restricted duality operator $N_{C^\vee}$. It is easy to see that the exposed face $N_C(F)$ of $C^\vee$ is the (inner) [*normal cone*]{} to the closed convex cone $C$ at $F$, $$N_C(F) =\{\ell\in V^\ast \colon \ell(y-x)\geq 0\; \forall y\in C \; \forall x\in F\},$$ for all nonempty faces $F\subset C$. In addition, for every relative interior point $x$ of $F$, we have $$\label{eq:nc-x} N_C(F) =x^\perp\cap C^\vee =\{\ell\in V^\ast \colon \ell(y-x)\geq 0\; \forall y\in C\},$$ hence we also refer to $N_C(F)$ as the (inner) normal cone to $C$ at $x$. Real Algebraic Geometry {#sec:rag} ----------------------- We are working in the setup of real algebraic geometry. A *(real) affine variety* for us is a subset of ${\mathbb{C}}^n$ (for some $n\in{\mathbb{N}}$) that is defined by a finite number of polynomial equations $p_1 = p_2 = \cdots = p_r = 0$, $r\in{\mathbb{N}}$, with real coefficients $p_i\in{\mathbb{R}}[x_1,x_2,\dots,x_n]$, $i=1,\dots,r$. The affine varieties in ${\mathbb{C}}^n$ are the closed sets of the *Zariski topology* on ${\mathbb{C}}^n$. So the Zariski closure of a set $S\subset {\mathbb{C}}^n$ is the smallest real affine variety containing $S$. A *(real) projective variety* for us is a subset of projective space ${\mathbb{P}}^n$ that is defined by a finite number of homogeneous polynomial equations $p_1 = p_2 = \cdots = p_r = 0$, $r\in{\mathbb{N}}$, with real coefficients $p_i\in{\mathbb{R}}[x_0,x_1,\dots,x_n]$, $i=1,\dots,r$. The projective varieties in ${\mathbb{P}}^n$ are the closed sets of the *Zariski topology* on ${\mathbb{P}}^n$. Identifying points in ${\mathbb{P}}^n$ with lines in ${\mathbb{C}}^{n+1}$, a projective variety can be seen as an affine variety in ${\mathbb{C}}^{n+1}$ which is an algebraic cone. The *affine cone* ${\widehat}{S}$ over a subset $S\subset{\mathbb{P}}^n$ is the union of all lines in ${\mathbb{C}}^{n+1}$ spanned by a vector $(x_0,x_1,\dots,x_n)^T$ such that $(x_0:x_1:\dots:x_n)\in S$. Conversely, the projective variety ${\mathbb{P}}(X)\subset{\mathbb{P}}^n$ associated with an algebraic cone $X\subset{\mathbb{C}}^{n+1}$ consists of the points $(x_0:x_1:\dots:x_n)\in{\mathbb{P}}^n$ for which the vector $(x_0,x_1,\dots,x_n)^T$ is included in $X$. These notions are explained in introductory textbooks on algebraic geometry like [@Harris1992] with the caveat that affine and projective varieties are usually complex varieties, i.e. defined by finitely many polynomial equations with complex coefficients. A point $x\in{\mathbb{P}}^n$ is [*real*]{} if the line ${\widehat}{\{x\}}\subset{\mathbb{C}}^{n+1}$ contains a nonzero real point. We denote the set of real points of a subset $S\subset{\mathbb{C}}^n$ or $S\subset{\mathbb{P}}^n$ by $S({\mathbb{R}})$. The dual projective space $({\mathbb{P}}^n)^\ast = {\mathbb{P}}(({\mathbb{C}}^{n+1})^\ast)$ is the projective space over the dual vector space so that the hyperplanes in ${\mathbb{P}}^n$ are in one-to-one correspondence with points in $({\mathbb{P}}^n)^\ast$. We specify a functional $\ell\in({\mathbb{P}}^n)^\ast$ in terms of its hyperplane $$H\ =\ {\mathcal{V}}(\ell)\ \subset\ {\mathbb{P}}^n$$ by writing $\ell=[H]$. Identifying $(({\mathbb{P}}^n)^\ast)^\ast={\mathbb{P}}^n$, a point $x\in{\mathbb{P}}^n$ defines a hyperplane in $({\mathbb{P}}^n)^\ast$ which we denote by $${\mathcal{V}}(x)\ =\ \{y\in({\mathbb{P}}^n)^\ast \colon x_0y_0 + x_1y_1 + \cdots + x_ny_n = 0 \}.$$ Let $X\subset{\mathbb{P}}^n$ be an irreducible projective variety. We define the *projective dual variety* $X^\ast$ of $X$ as the Zariski closure of the set of hyperplanes that are tangent to $X$ at some regular point, i.e. the closure of $$\{[H] \in ({\mathbb{P}}^n)^\ast\colon T_x X\subset H \text{ for some } x\in X_{\mathrm{reg}}\}.$$ An instructive and especially nice case of duality occurs for hypersurfaces defined by quadratic forms of full rank. \[exm:quadric\] Let $q \in{\mathbb{R}}[x_0,x_1,\dots,x_n]$ be a quadratic form and let $M_q$ be the real symmetric $(n+1)\times(n+1)$ matrix representing $q$, i.e. with $$q = (x_0,x_1,\dots,x_n)M_q (x_0,x_1,\dots,x_n)^T.$$ The projective variety $X = {\mathcal{V}}(q)\subset {\mathbb{P}}^n$ is smooth if and only if the rank of $M_q$ is $n+1$. We compute the dual variety of $X$ under the assumption that $X$ is smooth. Let $x = (x_0:x_1:\dots:x_n)\in X$ be a point. The differential $\ell_x = 2x^TM_q\in({\mathbb{P}}^n)^\ast$ of $q$ at $x$ defines the tangent hyperplane $T_xX = \{y\in{\mathbb{P}}^n\colon \ell_x(y) = 0\}$ to $X$ at $x$. In other words, the dual variety to $X$ is the Zariski closure of the set $\{\ell_x\colon x\in X\}\subset ({\mathbb{P}}^n)^\ast$. The condition $x\in X$ is $0 = x^TM_q x = (\ell_x^T M_q^{-1}) M_q (M_q^{-1}\ell_x)$. We conclude that $X^\ast$ is the quadratic hypersurface defined by $M_q^{-1}$. For irreducible algebraic varieties, the famous biduality theorem holds. \[thm:biduality\] If $X\subset {\mathbb{P}}^n$ is an irreducible projective variety, then $(X^\ast)^\ast = X$ under the canonical identification of the bidual of ${\mathbb{P}}^n$ with ${\mathbb{P}}^n$ itself. This theorem has several useful consequences like the following. \[rem:dense-C\] Let $X\subset {\mathbb{P}}^n$ be an irreducible projective variety. For all points $x$ of $X$ in a dense subset in the Euclidean topology of $X$, the point $x$ is regular, the hyperplane ${\mathcal{V}}(x)\subset({\mathbb{P}}^n)^\ast$ is tangent to $X^\ast$ at a regular point $\ell$, and the hyperplane ${\mathcal{V}}(\ell)\subset{\mathbb{P}}^n$ is tangent to $X$ at $x$. This is an application of the conormal variety $CN(X)$, defined as the Zariski closure of $$CN_0(X)\ =\ \{(x,[H])\in{\mathbb{P}}^n\times ({\mathbb{P}}^n)^\ast\colon x\in X_{\mathrm{reg}}, T_xX\subset H\}.$$ The projection $\pi_1 \colon CN_0(X) \to X_{\rm reg}$ is the conormal bundle of $X$, which shows that $CN_0(X)$ is an irreducible and smooth variety. The biduality theorem is often proven as a consequence of the fact that $CN(X) = CN(X^\ast)$, see [@GKZ1994 Chapter 1]. The biduality theorem implies that the subset $$U\ =\ \{(x,[H])\in CN_0(X) \colon [H]\in (X^\ast)_{\mathrm{reg}}\}$$ is a non-empty Zariski open subset of $CN(X)$. If $x\in X_{\rm reg}$ and $[H]\in (X^\ast)_{\rm reg}$ are regular points, then $H$ is tangent to $X$ at $x$ if and only if ${\mathcal{V}}(x)$ is tangent to $X^\ast$ at $[H]$, see [@Tevelev2005 Thm. 1.7(b)]. This shows $$U\ =\ CN_0(X) \cap CN_0(X^\ast).$$ By definition, the right-hand side consists of pairs $(x,[H])$ of regular points $x\in X_{\rm reg}$ and $[H]\in (X^\ast)_{\rm reg}$ such that ${\mathcal{V}}(x)$ is tangent to $X^\ast$ at $[H]$ and $H$ is tangent to $X$ at $x$. Since $U$ is dense in $CN(X)$ in the Euclidean topology [@Mumford1976 Thm. 2.33], the claim follows as the projection from $CN(X)$ to the first factor $X$ is continuous and surjective. When passing to real points, the direct analogue of Remark \[rem:dense-C\] fails: The set of regular real points $X_{\mathrm{reg}}({\mathbb{R}})$ of an irreducible projective variety $X\subset {\mathbb{P}}^n$ may not be dense in $X({\mathbb{R}})$ with respect to the Euclidean topology, even if it is non-empty, see for example [@BCR Section 3.1] or below. This is addressed in the following remark. \[rem:bidualitytangency\] Let $X\subset {\mathbb{P}}^n$ be an irreducible projective variety. For all real regular points $x$ of $X$ in a dense subset of $X_{\mathrm{reg}}({\mathbb{R}})$ in the Euclidean topology, the hyperplane ${\mathcal{V}}(x)\subset({\mathbb{P}}^n)^\ast$ is tangent to $X^\ast$ at a real regular point $\ell$ and the hyperplane ${\mathcal{V}}(\ell)\subset{\mathbb{P}}^n$ is tangent to $X$ at $x$. This claim is trivial if $X$ has no regular real points. We resume the discussion from assuming that $X$ does contain a regular real point. The variety $CN_0(X)$ is smooth and contains real points, since $X$ contains smooth real points. Since $CN_0(X)\setminus CN_0(X^\ast)=CN_0(X)\setminus U$ is a Zariski closed proper subset relative to $CN_0(X)$, it is of lower dimension. As $CN_0(X)({\mathbb{R}})$ is a real analytic manifold of dimension $\dim(CN_0(X))$, see [@BCR Prop. 3.3.11], the set $U({\mathbb{R}})$ is dense in $CN_0(X)({\mathbb{R}})$ in the Euclidean topology. This proves the claim, because the projection of $CN_0(X)({\mathbb{R}})$ onto the first factor $X$ is onto $X_{\mathrm{reg}}({\mathbb{R}})$. \[def:central-point\] We call a real point $x$ of an algebraic variety $X\subset {\mathbb{C}}^n$ *central* if it is in the Euclidean closure of the set of regular and real points of $X$, i.e. if $x$ is in the Euclidean closure of $X_{\mathrm{reg}}({\mathbb{R}})$. \[rem:central-pt-local-dim\] For an irreducible algebraic variety $X\subset{\mathbb{C}}^n$, a point $x\in X({\mathbb{R}})$ is central if and only if the local dimension of $x$ in $X({\mathbb{R}})$ is equal to $\dim(X)$, see [@BCR Prop. 7.6.2]. Dual Hyperbolicity Cones {#sec:duality-cones} ======================== We discuss a result by the second author [@Sinn2015] more explicitly. The result is that the dual convex cone to a hyperbolicity cone is the convex cone generated by a particular semi-algebraic cone. This semi-algebraic cone is the Euclidean closure of the cone of regular real points on the dual variety to the hyperbolic hypersurface that lie in the right half-space. The algebraic boundary of the hyperbolicity cone allows us to simplify this semi-algebraic cone. We prove a stronger result for three-dimensional hyperbolicity cones. A homogeneous polynomial $p\in{\mathbb{R}}[x_0,x_1,\dots,x_n]$ of degree $d$ is called *hyperbolic* with respect to a fixed point $e\in{\mathbb{R}}^{n+1}$ if $p(e)\neq 0$ and the polynomial $p(te-a)$ in one variable $t$ has only real roots for every point $a\in{\mathbb{R}}^{n+1}$. Without loss of generality, we fix the sign at $e$ and always assume $p(e)>0$. The roots of $p(te-a)$ are sometimes called the eigenvalues of $a$ with respect to $p$ and $e$, in analogy with characteristic polynomials of Hermitian matrices. Given any such polynomial $p$, the set $$C_e(p)\ =\ \bigl\{ a\in{\mathbb{R}}^{n+1} \colon \mbox{all roots of $p(te-a)$ are non-negative} \bigr\}$$ is a closed convex cone called the *(closed) hyperbolicity cone of $p$ with respect to $e$*, and $e$ is an interior point of $C_e(p)$, see [@Renegar2006]. Our goal is to describe the dual convex cone $$C_e(p)^\vee\ =\ \bigl\{ \ell\in ({\mathbb{R}}^{n+1})^\ast \colon \mbox{$\ell(x)\geq 0$ for all $x\in C_e(p)$} \bigr\}.$$ An essential technique is projective duality. A general approach is described in the paper [@Sinn2015] by the second author. The goal of this section is to explain this method more explicitly for the special case of hyperbolicity cones. An important example of a hyperbolic polynomial is the determinant of a matrix pencil, i.e. $p=\det(x_0A_0+\cdots+x_nA_n)$ for Hermitian $d\times d$ matrices $A_0,\dots,A_n\in H_d$, which is hyperbolic with respect to $e=(e_0,\dots,e_n)^T$ provided the matrix $e_0A_0+\cdots+e_nA_n$ is positive definite. In this case, $C_e(p)$ is the *spectrahedral cone* defined by $A_0,\dots,A_n$, see . However, the discussion in this section does not require such a determinantal representation and we consider general hyperbolic polynomials. The proof of our main result, , makes use of the following on hyperbolic polynomials. For the sake of completeness, we include a short proof based on the Helton-Vinnikov theorem on determinantal representations of hyperbolic curves; see [@PlaumannVinzant2013 Lemma 2.4] for a direct proof of a special case. Let $f\in{\mathbb{R}}[x_0,x_1,\dots,x_n]$ be a polynomial and $x\in{\mathbb{C}}^{n+1}$ be a point. The *multiplicity* of $x$ on ${\mathcal{V}}(f)\subset{\mathbb{C}}^{n+1}$ is the smallest degree of a non-zero homogeneous term in the Taylor expansion of $f$ around $x$. \[lem:multiplicity\] Let $p\in{\mathbb{R}}[x_0,\dots,x_n]$ be hyperbolic with respect to $e$ and let $x\in{\mathbb{R}}^{n+1}$. If $x$ has multiplicity $m$ on ${\mathcal{V}}(p)$, the hyperbolic hypersurface defined by $p$, then $t=0$ is a root of multiplicity $m$ of $p(x+te)\in{\mathbb{R}}[t]$. Moreover, $t=0$ is also a root of multiplicity $m$ of $p(x+t(e-x))\in{\mathbb{R}}[t]$. If the multiplicity of $x$ on ${\mathcal{V}}(p)$ is $m$, then $\frac{\partial^m}{\partial s^m}p(x+sy)|_{s=0}\neq 0$ for generic $y\in{\mathbb{R}}^{n+1}$. Fix such $y$ in the interior of $C_e(p)$ and consider the hyperbolic polynomial $p(rx+sy+te)$ in three variables $r,s,t$. By the Helton-Vinnikov theorem ([@HeltonVinnikov2007 Thm. 2.2]), this polynomial has a determinantal representation $$p(rx+sy+te) = \det(rA + sB + tC)$$ with real symmetric matrices $A,B,C$, where $B$ and $C$ are positive definite, hence factor as $B=UU^T$ and $C=VV^T$, with $U$ and $V$ invertible. Now $s=0$ is a root of $p(x+sy)=\det(A+sB)$ of multiplicity $m$, which means that $U^{-1}A(U^T)^{-1}$ has $m$-dimensional kernel. But then so does $V^{-1}A(V^T)^{-1}$, hence the root $t=0$ of $p(x+te)=\det(A+tC)$ has multiplicity $m$ as well. The second part of the claim follows from the part we have just proved. Indeed, write $f(s,t) = p(sx + te)\in {\mathbb{R}}[s,t]$ which is the homogenization of $p(x+te)$ because $p(e)\neq 0$. Since $t = 0$ is a root of multiplicity $m$ of $p(x+te)$, we can write $p(x + te) = t^mq(t)$ with $q(0)\neq 0$. Since $p$ is hyperbolic with respect to $e$, the polynomial $q$ factors into linear terms over ${\mathbb{R}}$, say $q = c (1- \lambda_1 t)\cdot\ldots\cdot (1-\lambda_{d-m}t)$ for some nonzero $c\in{\mathbb{R}}$. Here we used $q(0)\neq 0$ because the $\lambda_i$ are the reciprocals of the roots of $q$. So we get that $f(s,t) = c \cdot t^m \prod (s-\lambda_i t)$. The polynomial $p(sx+t(e-x))$ is the same as $f(s-t,t) = c \cdot t^m \prod (s- (\lambda_i + 1)t)$. Dehomogenizing this again shows that $t=0$ is a root of multiplicity $m$ of $p(x+t(e-x))$. \[Cor:multiplicity\] If $x$ is a regular real point of a hyperbolic hypersurface ${\mathcal{V}}(p)$, then the line incident with $x$ and the hyperbolicity direction $e$ is not tangent to ${\mathcal{V}}(p)$ at $x$, i.e. is not contained in $T_x\bigl({\mathcal{V}}(p)\bigr)$. If a line $L$ is tangent to ${\mathcal{V}}(p)$ at $x$, then the multiplicity of $x$ in $L\cap {\mathcal{V}}(p)$ is greater than the multiplicity of $x$ in ${\mathcal{V}}(p)$. This is impossible if $e\in L$, by the previous . We prove basics from differential and convex geometry regarding hyperbolicity cones. \[lem:some-hyper\] Let $p\in{\mathbb{R}}[x_0,\dots,x_n]$ be an irreducible hyperbolic polynomial with respect to $e\in{\mathbb{R}}^{n+1}$. Then $M=\partial C_e(p)\cap{\mathcal{V}}(p)_{\mathrm{reg}}$ is an $n$-dimensional real analytic manifold, which is open and dense in the Euclidean boundary $\partial C_e(p)$ of $C_e(p)$ in the Euclidean topology. The (inner) normal cone to $C_e(p)$ at any point $x\in M$ is the ray ${\mathbb{R}}_+\ell$, where $\ell=\nabla p(x)^T\in C_e(p)^\vee\subset({\mathbb{R}}^{n+1})^\ast$. Since $\partial C_e(p)\subset {\mathcal{V}}(p)({\mathbb{R}})$ and since the set of singular points of ${\mathcal{V}}(p)({\mathbb{R}})$ is a variety of dimension at most $n-1$, see [@BCR Prop. 3.3.14], the complement $M=\partial C_e(p)\cap{\mathcal{V}}(p)_{\mathrm{reg}}$ is open and dense in $\partial C_e(p)$ in the Euclidean topology. As ${\mathcal{V}}(p)_{\mathrm{reg}}({\mathbb{R}})$ is an analytic manifold of dimension $n$, see [@BCR Prop. 3.3.11], and as the eigenvalues depend continuously on $x\in{\mathbb{R}}^{n+1}$, see [@Renegar2006], the set $M$ is an analytic manifold of dimension $n$. Let $x\in M$. As $x$ is a regular point of ${\mathcal{V}}(p)$, the functional $\ell=\nabla p(x)^T$ is non-zero. Hence, , case $m=1$, shows $\ell(e)\neq 0$. Since $M$ is an $n$-dimensional analytic manifold included in $C_e(p)$, the normal cone $N(x)$ of $C_e(p)$ at $x$ is a subset of the line ${\mathbb{R}}\ell$, hence $N(x)={\mathbb{R}}_+\ell$ or $N(x)=-{\mathbb{R}}_+\ell$ as $x\in\partial C_e(p)$. The derivative polynomial $p'\in{\mathbb{R}}[x_0,\dots,x_n]$, defined by $$p'(y)=\tfrac{\partial}{\partial t}p(y+te)|_{t=0}=\nabla p(y)^Te, \qquad y\in{\mathbb{R}}^{n+1},$$ is hyperbolic with respect to $e$ and $C_e(p)\subset C_e(p')$ holds, see [@Renegar2006]. This proves $\ell(e)=p'(x)>0$ and rules out $N(x)=-{\mathbb{R}}_+\ell$, as $N(x)\subset C_e(p)^\vee$ by . This proves the claim. The second author obtained and below in [@Sinn2015 Example 3.15]. We slightly abuse notation in the following statements. If $X\subset{\mathbb{C}}^{n+1}$ is an algebraic cone, then we write $X^\ast\subset ({\mathbb{C}}^{n+1})^\ast$ for the affine cone over the projective dual variety ${\mathbb{P}}(X)^\ast$ of ${\mathbb{P}}(X)$. Let $H_{e,+}\subset({\mathbb{R}}^{n+1})^\ast$ denote the half-space $H_{e,+} = \{\ell\in ({\mathbb{R}}^{n+1})^\ast\colon \ell(e)\geq 0\}$ for nonzero $e\in{\mathbb{R}}^{n+1}$. \[thm:dual-hyp-cone\] Let $p\in{\mathbb{R}}[x_0,x_1,\dots,x_n]$ be an irreducible hyperbolic polynomial with respect to $e\in{\mathbb{R}}^{n+1}$. Then we have $$C_e(p)^\vee\ =\ \operatorname{clos}\Big( \operatorname{cc}\big( ({\mathcal{V}}(p)^\ast)_{\mathrm{reg}}({\mathbb{R}}) \cap H_{e,+} \big) \Big).$$ We prove the inclusion “$\supset$”. Since $C_e(p)^\vee$ is a closed convex cone, it is enough to show that $S=({\mathcal{V}}(p)^\ast)_{\mathrm{reg}}({\mathbb{R}}) \cap H_{e,+}$ is contained in $C_e(p)^\vee$. By , it is enough to prove $\ell\in C_e(p)^\vee$ for all $\ell\in S$ such that the hyperplane ${\mathcal{V}}(\ell)$ is tangent to ${\mathcal{V}}(p)$ at a regular real point $x$, i.e. $\ell = \nabla p(x)^T\in ({\mathbb{R}}^{n+1})^\ast$. By above, case $m=1$, we have $0\neq\frac{\partial}{\partial t}p(x+te)|_{t=0} = \ell(e)$ for such an $\ell$ and hence $e\not\in{\mathcal{V}}(\ell)$. As $p$ is hyperbolic with respect to every interior point of the hyperbolicity cone [@Renegar2006], it follows that the interior of $C_e(p)$ is disjoint from ${\mathcal{V}}(\ell)$. This means that $\ell$ has constant sign on $C_e(p)$. Since $\ell\in H_{e,+}$, it follows that $\ell \in C_e(p)^\vee$. The inclusion “$\subset$” follows from [@Sinn2015 Coro. 3.14]. We repeat the argument, adapted to our setup, for completeness. Since $C_e(p)^\vee$ is the convex cone generated by its extreme rays and the right-hand side is a convex cone, it suffices to prove that every extreme ray of $C_e(p)^\vee$ is contained in the right-hand side. By Straszewicz’s Theorem [@Rockafellar1970 Thm. 18.6], which says that every extreme ray is a limit of exposed rays, it suffices to prove the claim for every exposed ray of $C_e(p)^\vee$, because the right-hand side is closed. Let ${\mathbb{R}}_+ \ell$ be an exposed ray of the convex cone $C_e(p)^\vee$. It is enough to show that $\ell$ is a central point of ${\mathcal{V}}(p)^\ast({\mathbb{R}})$, i.e. $\ell$ lies in the (Euclidean) closure of $({\mathcal{V}}(p)^\ast)_{\rm reg}({\mathbb{R}})$. Let $F_\ell = \{x\in C_e(p)\colon \ell(x) = 0\}$ be the exposed face of $C_e(p)$ corresponding to $\ell$ and let $x$ be a point in the relative interior of $F_\ell$. As proven in , the analytic manifold $M=\partial C_e(p)\cap{\mathcal{V}}(p)_{\mathrm{reg}}$ is dense in the Euclidean boundary $\partial C_e(p)$ of the hyperbolicity cone $C_e(p)$. Hence, there is a sequence $(y_j)\subset M$ converging to $x$. shows that after slightly moving the members of the sequence $(y_j)$ within $M$ without changing the limit $x$, the hyperplane ${\mathcal{V}}(y_j)$ is tangent to ${\mathcal{V}}(p)^\ast$ at a regular real point $\ell_j\in({\mathcal{V}}(p)^\ast)_{\mathrm{reg}}({\mathbb{R}})$ and ${\mathcal{V}}(\ell_j)$ is tangent to ${\mathcal{V}}(p)$ at $y_j$ for all $j\in{\mathbb{N}}$. After scaling $\ell_j$ with a nonzero real number, we have $\ell_j=\nabla p(y_j)^T$. Hence, the ray ${\mathbb{R}}_+\ell_j$ lies in the dual convex cone $C_e(p)^\vee$ by for all $j\in{\mathbb{N}}$. After normalizing and passing to a subsequence, the sequence $(\ell_j)$ converges to a point $\ell'$ in the compact unit sphere of $({\mathbb{R}}^{n+1})^\ast$. We have $\ell'\in C_e(p)^\vee$ and $\ell'(x) = 0$, the latter as $$\begin{aligned} \ell'(x) = \ell_j (x-y_j) + (\ell'-\ell_j) y_j + (\ell'-\ell_j) (x-y_j)\end{aligned}$$ holds for all $j\in{\mathbb{N}}$ and since $\ell_j\to\ell'$ and $y_j\to x$ as $j\to\infty$. Since ${\mathbb{R}}_+ \ell$ is an exposed ray of $C_e(p)^\vee$, the lattice isomorphism of and show ${\mathbb{R}}_+ \ell=F_\ell^\perp\cap C_e(p)^\vee=x^\perp\cap C_e(p)^\vee$. This proves ${\mathbb{R}}_+ \ell={\mathbb{R}}_+ \ell'$. Hence $\ell$ is a central point of ${\mathcal{V}}(p)^\ast({\mathbb{R}})$. In other words, says that the dual convex cone to the hyperbolicity cone $C_e(p)$ is the closed convex cone generated by the regular real points of the dual variety ${\mathcal{V}}(p)^\ast$ lying in the appropriate half-space $H_{e,+}$. a)![a) Cayley cubic. b) Steiner surface with three singular lines.[]{data-label="fig:steiner"}](Cayley.pdf "fig:"){height="6cm"}b)![a) Cayley cubic. b) Steiner surface with three singular lines.[]{data-label="fig:steiner"}](Roman.png "fig:"){height="6cm"} The well-known Steiner surface explains why singular points of the dual variety ${\mathcal{V}}(p)^\ast$ have to be excluded from the statement of . \[ex:Cayley\] The Cayley cubic is the cubic hypersurface in ${\mathbb{P}}^3$ defined by the polynomial $$p=\det \begin{pmatrix} x_0 & x_1 & x_3 \\ x_1 & x_0 & x_2 \\ x_3 & x_2 & x_0 \end{pmatrix}.$$ This polynomial is irreducible and hyperbolic with respect to the point $(1,0,0,0)$ in ${\mathbb{R}}^4 = \{(x_0,x_1,x_2,x_3)\}$. Its hyperbolicity cone is the homogenization of the elliptope $\mathcal{E}_3$, which is the feasible set of the Goemans-Williamson semidefinite relaxation of the MAX-CUT problem (for graphs with three vertices). The dual convex cone is the closed convex cone generated by the regular real points with $y_0 > 0 $ on the Steiner surface given by the equation $$q=y_1^2y_2^2 + y_1^2y_3^2 + y_2^2y_3^2 - 2 y_0y_1y_2y_3.$$ The singular locus of this quartic surface is the union of three real lines in $({\mathbb{P}}^3)^\ast$, which are not contained in the dual convex cone. See , where we draw the real affine parts of the varieties ${\mathcal{V}}(p)$ and ${\mathcal{V}}(q)$, that is to say, the set of points $(x_1,x_2,x_3)^T\in{\mathbb{R}}^3$ for which $(1:x_1:x_2:x_3)$ lies in ${\mathcal{V}}(p)$ in drawing a) and the set of points $(y_1,y_2,y_3)\in({\mathbb{R}}^3)^\ast$ for which $(1:y_1:y_2:y_3)$ lies in ${\mathcal{V}}(q)$ in drawing b). Convex geometry suffices to generalize from irreducible polynomials to arbitrary hyperbolic polynomials. \[cor:dual-hyp-cone\] Let $p\in{\mathbb{R}}[x_0,x_1,\dots,x_n]$ be a hyperbolic polynomial with respect to $e\in{\mathbb{R}}^{n+1}$. Let $X_1,X_2,\dots,X_r$ be the irreducible components of the hyperbolic hypersurface ${\mathcal{V}}(p)$. Then we have $$\label{eq:dual-hyp-cone} C_e(p)^\vee\ = \operatorname{cc}\big(\operatorname{clos}(S_1\cup S_2\cup\dots\cup S_r)\big),$$ where $S_i = (X_i^\ast)_{\mathrm{reg}}({\mathbb{R}}) \cap H_{e,+}$ for $i=1,\dots,r$. If $p=p_1^{m_1}p_2^{m_2}\dots p_r^{m_r}$ is a factorization of $p$ into irreducible factors, then $C_e(p)=C_e(p_1)\cap C_e(p_2)\cap\dots\cap C_e(p_r)$. The claim follows from and from in . 1. The cones $S_i$ in , their union $S_1\cup S_2\cup\dots\cup S_r$, and the Euclidean closure $\operatorname{clos}(S_1 \cup S_2 \cup\dots\cup S_r)$, see [@BCR Prop. 2.2.2], are semi-algebraic sets. Hence, the closed convex cone $C_e(p)^\vee$ is the convex cone generated by the semi-algebraic cone $\;\operatorname{clos}(S_1 \cup S_2 \cup\dots\cup S_r)$. 2. The Euclidean closure of the set $S_i$ in is the set of central points of $X_i^\ast({\mathbb{R}})$ lying in $H_{e,+}$ for all $i = 1,2,\dots,r$. Hence, writing $\mathrm{cent}(X)$ for the set of central real points of $X$, we can rephrase as $$C_e(p)^\vee\ =\ \operatorname{cc}\left( \bigcup_{i=1}^r \mathrm{cent}(X_i^\ast({\mathbb{R}}))\cap H_{e,+} \right).$$ We can strengthen to the homogeneous version of Kippenhahn’s Theorem, the statement of , if the hyperbolicity cone has dimension three. Chien and Nakazato [@ChienNakazato2010] observed that this stronger version is false in higher dimensions, see . \[thm:dim3\] Let $p\in {\mathbb{R}}[x_0,x_1,x_2]$ be an irreducible hyperbolic polynomial with respect to $e\in{\mathbb{R}}^3$. Then $C_e(p)^\vee = \operatorname{cc}\big({\mathcal{V}}(p)^\ast({\mathbb{R}})\cap H_{e,+}\big)$. The inclusion $C_e(p)^\vee \subset \operatorname{cc}\big({\mathcal{V}}(p)^\ast({\mathbb{R}})\cap H_{e,+}\big)$ follows from and . We prove the opposite inclusion by contradiction based on two observations. Let $\ell$ be a nonzero functional that lies in $\operatorname{cc}\big({\mathcal{V}}(p)^\ast({\mathbb{R}})\cap H_{e,+}\big)$ but not in $C_e(p)^\vee$. First, the hyperplane ${\mathcal{V}}(\ell)\subset{\mathbb{C}}^3$ intersects the interior of the hyperbolicity cone $C_e(p)$, as it holds for all functionals $\hat{\ell}\in H_{e,+}\setminus C_e(p)^\vee$. Since $\hat{\ell}\in H_{e,+}$, we have $\hat{\ell}(e)\geq 0$, and since $\hat{\ell}\not\in C_e(p)^\vee$ there is a point $x\in C_e(p)$ such that $\hat{\ell}(x)<0$. Hence, there is $\lambda\in[0,1)$ such that the point $y=(1-\lambda)e+\lambda x$ lies on ${\mathcal{V}}(\hat{\ell})$. As $e$ is an interior point of $C_e(p)$ so is $y$ [@Rockafellar1970 Thm. 6.1]. Secondly, the line ${\mathcal{V}}(\ell)\subset{\mathbb{P}}^2$ is tangent to ${\mathcal{V}}(p)$ at a point $x\in{\mathcal{V}}(p)$. Applying to the projective variety $X={\mathcal{V}}(p)^\ast$, we can choose a sequence of regular points $(\ell_j)$ in ${\mathcal{V}}(p)^\ast$ converging to $\ell$ and a sequence of regular points $(x_j)$ of ${\mathcal{V}}(p)$ such that the line ${\mathcal{V}}(\ell_j)$ is tangent to ${\mathcal{V}}(p)$ at $x_j$ for all $j\in{\mathbb{N}}$. Because of the compactness of the projective space ${\mathbb{P}}^2$, we can assume $(x_j)$ converges to a point $x\in{\mathcal{V}}(p)$. Since ${\mathcal{V}}(\ell_j)$ is tangent to ${\mathcal{V}}(p)$ at $x_j$, the line ${\mathcal{V}}(\ell)$ is tangent to ${\mathcal{V}}(p)$ at $x$, see [@Fischer2001 Sec. 8.2]. The real line ${\mathcal{V}}(\ell)\subset{\mathbb{P}}^2$ intersects the hyperbolicity cone $C_e(p)$ in an interior point by the first observation. Since the polynomial $p$ is hyperbolic with respect to this interior point and since the point $x\in{\mathcal{V}}(p)$ constructed above lies on the line ${\mathcal{V}}(\ell)$, it follows that $x$ is a real point. then shows that ${\mathcal{V}}(\ell)$ is not tangent to ${\mathcal{V}}(p)$ at $x$, which contradicts the second observation. Again, convex geometry suffices to generalize from irreducible polynomials to arbitrary hyperbolic polynomials. \[cor:dim3\] Let $p\in {\mathbb{R}}[x_0,x_1,x_2]$ be a hyperbolic polynomial with respect to $e\in{\mathbb{R}}^3$. Then $C_e(p)^\vee = \operatorname{cc}({\mathcal{V}}(p)^\ast({\mathbb{R}})\cap H_{e,+})$. If $p=p_1^{m_1}p_2^{m_2}\dots p_r^{m_r}$ is a factorization of $p$ into irreducible factors, then $$C_e(p)=C_e(p_1)\cap C_e(p_2)\cap\dots\cap C_e(p_r)$$ and $${\mathcal{V}}(p)^\ast={\mathcal{V}}(p_1)^\ast\cup{\mathcal{V}}(p_2)^\ast \cup\dots\cup{\mathcal{V}}(p_r)^\ast.$$ The claim follows from and in . Not all components in the union ${\mathcal{V}}(p)^\ast={\mathcal{V}}(p_1)^\ast\cup{\mathcal{V}}(p_2)^\ast \cup\dots\cup{\mathcal{V}}(p_r)^\ast$ are needed in the statements of and . The selection can be described as follows. \[def:alg-boundary\] Let $S\subset {\mathbb{R}}^n$ be a semi-algebraic set. The *algebraic boundary* of $S$, denoted $\partial_a S$, is the Zariski closure in ${\mathbb{C}}^n$ of the Euclidean boundary $\partial S$ of $S$. Determining the algebraic boundary of the hyperbolicity cone of a hyperbolic polynomial $p\in{\mathbb{R}}[x_0,x_1,\dots,x_n]$ amounts to computing a factorization of $p$ into irreducible factors and picking the correct subset of the factors. \[rem:boundary-hyperbolicity\] 1. Let $p\in{\mathbb{R}}[x_0,x_1,\dots,x_n]$ be irreducible and hyperbolic with respect to $e$. The algebraic boundary of the hyperbolicity cone $C_e(p)$ is the algebraic hypersurface ${\mathcal{V}}(p) = \{x\in{\mathbb{C}}^{n+1}\colon p(x) = 0\}$. 2. If $p$ is hyperbolic with respect to $e$, but factors as $p = p_1 p_2\dots p_r$ into irreducible factors, then ${\mathcal{V}}(p) = {\mathcal{V}}(p_1)\cup {\mathcal{V}}(p_2)\cup \dots\cup {\mathcal{V}}(p_r)$ is the decomposition of the hypersurface ${\mathcal{V}}(p)$ into its irreducible components. The algebraic boundary of $C_e(p)$ is a union of some, not necessarily all, of the irreducible hypersurfaces ${\mathcal{V}}(p_i)$. The hypersurfaces in this union are the irreducible components of $\partial_a C_e(p)$. 3. If $p$ is squarefree, i.e. $p_j\neq p_i$ for $i\neq j$, then the factors $p_i$ that contribute to the algebraic boundary of $C_e(p)$ are exactly those with the property that the hyperbolicity cone of $\prod_{j\neq i}p_j$ is strictly larger than $C_e(p)$, because $C_e(p) = \bigcap_i C_e(p_i)$. a)![a) Intersection of four filled ellipses (gray area). b) Convex hull of the dual ellipses (gray area). []{data-label="fig:four-ellipses"}](FourEllipses.pdf "fig:"){height="5cm"}b)![a) Intersection of four filled ellipses (gray area). b) Convex hull of the dual ellipses (gray area). []{data-label="fig:four-ellipses"}](FourEllipsesDual.pdf "fig:"){height="5cm"} \[ex:four\_ellipses\] Consider the four ellipses depicted in a). The intersection of the filled ellipses is an area with nonempty interior, which is isometric to the base of a hyperbolicity cone *via* the embedding ${\mathbb{R}}^2\to{\mathbb{R}}^3$, $(x_1,x_2)^T\mapsto(1,x_1,x_2)^T$. Note that the red ellipse does not contribute to the algebraic boundary of the hyperbolicity cone, but the other ellipses do. Homogenizing the polynomials, we obtain four conics in ${\mathbb{P}}^2$ from the ellipses. The real affine parts of the dual conics, *via* the embedding $({\mathbb{R}}^2)^\ast\to({\mathbb{P}}^2)^\ast$, $(y_1,y_2)\mapsto(1:y_1:y_2)$, are depicted in b). Their convex hull is a base of the dual convex cone to the hyperbolicity cone. We return to these bases in . Note that the red ellipse in b) is redundant, as the other three ellipses generate the same convex hull as all four together. \[cor:dual-hyp-cone-bound\] Let $p\in{\mathbb{R}}[x_0,x_1,\dots,x_n]$ be a hyperbolic polynomial with respect to $e\in{\mathbb{R}}^{n+1}$. Let $X_1,X_2,\dots,X_r$ be the irreducible components of the algebraic boundary $\partial_a C_e(p)$. Then we have $$C_e(p)^\vee\ = \operatorname{cc}\big(\operatorname{clos}(S_1\cup S_2\cup\dots\cup S_r)\big),$$ where $S_i = (X_i^\ast)_{\mathrm{reg}}({\mathbb{R}}) \cap H_{e,+}$ for $i=1,\dots,r$. This follows from Part 3) of and . \[rem:zariski-extreme\] The algebraic boundary of the hyperbolicity cone $C_e(p)$ is algebraically the optimal description of the set of extreme rays of $C_e(p)^\vee$. Using the same techniques and ideas that are presented in this paper, one can argue that the Zariski closure of the set of extreme rays of $C_e(p)^\vee$ is the union of the dual varieties $X_i^\ast$ for which the varieties $X_i$ belong to the algebraic boundary $\partial_a C_e(p)$, see [@Sinn2015 Corollary 3.5]. The statements above in this section require employing the duality theory of real algebraic geometry, which is more subtle than the duality theory of algebraic geometry. The reason is that the set of Hermitian matrices is just a real vector space, not a complex vector space. \[rem:complex-duality\] The image of the set $H_d^{(1)}$ of Hermitian matrices of rank $1$ under the projection map $$\pi(M) = ({\langle M,A_1\rangle},{\langle M,A_2\rangle},\ldots,{\langle M,A_n\rangle})$$ from $H_d$ to ${\mathbb{R}}^n$ is of interest to us (see ), as its convex hull is the joint numerical range. The complexification of the real vector space $H_d$ is the complex vector space of all complex $d\times d$ matrices. The Zariski closure of $H_d^{(1)}$ is the variety $R_d^{(1)}$ of complex $d\times d$ matrices of rank at most $1$. Therefore, the images of $H_d^{(1)}$ and $R_d^{(1)}$ under the map $\pi$ have the same Zariski closure. To evaluate ${\langle M,A_i\rangle}$ at a complex matrix $M$, we write $M = {\rm Re}(M) + {\operatorname{i}}{\rm Im}(M)$, where ${\rm Re}(M)$ and ${\rm Im}(M)$ are Hermitian, and define ${\langle M,A_i\rangle} = \operatorname{tr}({\rm Re}(M) A_i) + {\operatorname{i}}\operatorname{tr}({\rm Im}(M)A_i)$ (which is ${\mathbb{C}}$-linear). If the map $\pi$ restricted to $R_d^{(1)}$ is an isomorphism, then Prop. 4.1 in [@GKZ1994] implies roughly speaking that $\pi(R_d^{(1)})$ is projectively dual to the intersection of the orthogonal complement of the kernel of $\pi$ and the dual variety of $R_d^{(1)}$. This dual variety is the determinantal hypersurface in the space of complex $d\times d$ matrices. The assumption that $\pi$ restricted to $R_d^{(1)}$ is an isomorphism is generically satisfied provided that $n$ is sufficiently large relative to the size $d$ of the matrices; specifically the kernel of $\pi$ must not intersect the secant variety of $R_d^{(1)}$, which is the variety of matrices of rank at most $2$. Bases of Dual Hyperbolicity Cones {#sec:duality-selfdual} ================================= The results of carry over from the dual convex cone of a hyperbolicity cone to all its bases, by replacing the convex cone generated by a semi-algebraic cone with the convex hull of a base of this semi-algebraic cone. Remarkably, the results also apply to all linear images of these bases, because we can interpret these linear images as the bases of the dual convex cones to linear sections of the original hyperbolicity cone. \(m) \[matrix of math nodes,row sep=0.1em,column sep=0em,minimum width=0em, column 1/.style=[anchor=base east]{}, column 2/.style=[anchor=base west]{}\] [ e&\[0em\] V &\[4em\] \^n &\[4em\] \^n\ &&\ && C = \_0\^[-1]{}()\ && C\^= \_0(\^)\ &&\ =\_e(\^) \^& V\^& (\^n)\^& (\^n)\^& ()\ ]{}; (m-1-3) edge node \[below\][$\phi_0=\pi_0^\ast$]{} (m-1-2); (m-1-4) edge node \[below\][$\phi'=\pi_2^\ast$]{} (m-1-3); (m-6-2) edge node \[above\][$\pi_0$]{} (m-6-3); (m-6-3) edge node \[above\][$\pi_2$]{} (m-6-4); (m-1-4) edge node \[below\][$\phi=\pi^\ast$]{} (m-1-2); (m-6-2) edge node \[above\][$\pi$]{} (m-6-4); Before returning to hyperbolicity cones at the end of the section, we discuss the necessary convex geometry. See for a summary of our setup. \[def:construction\] Let $V$ be a finite-dimensional real vector space. Let ${\mathcal{K}}\subset V$ be a closed convex cone and let $e\neq 0$ be an interior point of ${\mathcal{K}}$. The set $${\mathcal{B}}= \operatorname{h}_e({\mathcal{K}}^\vee) = \{\ell\in {\mathcal{K}}^\vee \colon \ell(e) = 1\}$$ introduced in is a compact convex base of the dual convex cone ${\mathcal{K}}^\vee$ by . We study the image of ${\mathcal{B}}$ under an arbitrary linear map $$\pi \colon V^\ast \to ({\mathbb{R}}^n)^\ast.$$ Let $\pi_0 \colon V^\ast \to ({\mathbb{R}}\oplus{\mathbb{R}}^n)^\ast$, $\ell\mapsto(\ell(e),\pi(\ell))$ and let $\pi_2 \colon ({\mathbb{R}}\oplus{\mathbb{R}}^n)^\ast \to ({\mathbb{R}}^n)^\ast$, $(y_0,y)\mapsto y$ denote the projection onto the second summand. We denote the dual map to $\pi$, $\pi_0$, and $\pi_2$ by $\phi:{\mathbb{R}}^n\to V$, $\phi_0:{\mathbb{R}}\oplus{\mathbb{R}}^n\to V$, and $\phi':{\mathbb{R}}^n\to{\mathbb{R}}\oplus{\mathbb{R}}^n$, respectively. We will describe $\pi({\mathcal{B}})$ in terms of the set $$C = \phi_0^{-1}({\mathcal{K}})\subset{\mathbb{R}}\oplus {\mathbb{R}}^n.$$ Let $b_0,b_1,\dots,b_n$ denote the standard basis of ${\mathbb{R}}^{n+1}={\mathbb{R}}\oplus{\mathbb{R}}^n$. It is easy to see that the set $C$ is a closed convex cone containing $b_0$ as an interior point. In addition, $C$ can be seen as a linear section of the convex cone ${\mathcal{K}}$ through the interior point $e$. To be more precise, let $v_1,v_2,\ldots,v_n\in V$ and let $\pi_{v_1,\ldots,v_n}:V^\ast\to({\mathbb{R}}^n)^\ast$ be defined by $$\label{eq:pi-explicit} \pi_{v_1,\ldots,v_n}(\ell)=(\ell(v_1),\ldots,\ell(v_n)), \qquad \ell\in V^\ast.$$ If $\pi=\pi_{v_1,\ldots,v_n}$, then $$\phi_0((x_0,x)^T)=x_0e+x_1v_1+\cdots+x_nv_n$$ holds for all $(x_0,x)^T\in{\mathbb{R}}\oplus{\mathbb{R}}^n$ where $x=(x_1,\ldots,x_n)^T$. The image of $\phi_0$ is the subspace $\operatorname{im}(\phi_0)=\operatorname{span}(e,v_1,\ldots,v_n)$, which intersects the convex cone ${\mathcal{K}}$ in the interior point $e$. If $e,v_1,v_2,\ldots,v_n$ are linearly independent, then the monomorphism $\phi_0$ identifies $C$ with a linear section of ${\mathcal{K}}$ through $e$. shows that the set $\operatorname{h}_{b_0}(C^\vee)=\{(y_0,y)\in C^\vee\colon y_0=1\}$ is a (compact, convex) base of the dual convex cone $C^\vee$, as $b_0$ is an interior point of $C$. Obviously, the set $\operatorname{h}_{b_0}(C^\vee)$ is isometric under $\pi_2$ to the set $$\pi_2\circ\operatorname{h}_{b_0}(C^\vee) =\{y\in({\mathbb{R}}^n)^\ast\colon (1,y)\in C^\vee\}.$$ We show this set is a linear image of the base ${\mathcal{B}}$ of ${\mathcal{K}}^\vee$. \[prop:pro-cone\] The image of the convex cone ${\mathcal{K}}^\vee$ under the linear map $\pi_0$ is the closed convex cone $\pi_0({\mathcal{K}}^\vee)=C^\vee$. The image of the set ${\mathcal{B}}$ is the compact base $\pi_0({\mathcal{B}})=\operatorname{h}_{b_0}(C^\vee)$ of $C^\vee$. In addition we have $\pi({\mathcal{B}})=\pi_2\circ\operatorname{h}_{b_0}(C^\vee)$. To prove the equation $C^\vee = \pi_0({\mathcal{K}}^\vee)$, first let $y = \pi_0(\ell)\in\pi_0({\mathcal{K}}^\vee)$ for some $\ell\in {\mathcal{K}}^\vee$. For all $x\in C$ we have $\phi_0(x)\in {\mathcal{K}}$ and hence $$y(x) =\pi_0(\ell)(x) =\ell(\phi_0(x)) \geq 0.$$ This shows that $y\in C^\vee$ and therefore $\pi_0({\mathcal{K}}^\vee)\subset C^\vee$. We prove a partial converse by duality, i.e. we show $\left(\pi_0({\mathcal{K}}^\vee)\right)^\vee \subset C$. Let $x\in \left(\pi_0({\mathcal{K}}^\vee)\right)^\vee$. Then $$\ell(\phi_0(x))=\pi_0(\ell)(x)\geq 0 \qquad \text{for all $\ell\in {\mathcal{K}}^\vee$,}$$ which implies $\phi_0(x)\in {\mathcal{K}}$ or in other words $x\in \phi_0^{-1}({\mathcal{K}}) = C$. We finish proving $C^\vee = \pi_0({\mathcal{K}}^\vee)$ by showing that the convex cone $\pi_0({\mathcal{K}}^\vee)$ is closed. By it suffices to prove that $\pi_0({\mathcal{B}})$ is a compact base of $\pi_0({\mathcal{K}}^\vee)$. The set $\pi_0({\mathcal{B}})$ is compact as ${\mathcal{B}}$ is compact. The set ${\mathcal{K}}^\vee\setminus\{0\}$ is the cone generated by ${\mathcal{B}}$, hence the set $\pi_0({\mathcal{K}}^\vee)\setminus\{0\}$ is the cone generated by $\pi_0({\mathcal{B}})$. It remains to show that the origin does not lie in the affine hull of $\pi_0({\mathcal{B}})$. Since $\ell(e)=1$ holds for all $\ell\in{\mathcal{B}}$, the affine hull of ${\mathcal{B}}$ does not intersect the annihilator $e^\perp=\{\ell\in V^\ast:\ell(e)=0\}$. As $e\in\operatorname{im}(\phi_0)$ and $\ker(\pi_0)=\operatorname{im}(\phi_0)^\perp$, we have $\ker\pi_0\subset e^\perp$. This shows $\operatorname{aff}({\mathcal{B}})\cap\ker\pi_0=\emptyset$, which proves that the origin does not lie in $\pi_0(\operatorname{aff}({\mathcal{B}}))=\operatorname{aff}(\pi_0({\mathcal{B}}))$. Using the equation $\pi_0({\mathcal{K}}^\vee)=C^\vee$ and the projection $\pi_1:({\mathbb{R}}\oplus{\mathbb{R}}^n)^\ast\to{\mathbb{R}}^\ast$, $(y_0,y)\mapsto y_0$ onto the first summand, we get $$\begin{aligned} \pi_0({\mathcal{B}}) &= \{\pi_0(\ell)\in({\mathbb{R}}\oplus{\mathbb{R}}^n)^\ast \colon \ell\in{\mathcal{K}}^\vee, \ell(e)=1\}\\ &= \{\pi_0(\ell)\in({\mathbb{R}}\oplus{\mathbb{R}}^n)^\ast \colon \ell\in{\mathcal{K}}^\vee, \pi_1(\pi_0(\ell))=1 \}\\ &= \{(y_0,y)\in C^\vee \colon y_0=1 \} = \operatorname{h}_{b_0}(C^\vee).\end{aligned}$$ Finally, $\pi({\mathcal{B}})=\pi_2\circ\pi_0({\mathcal{B}})=\pi_2\circ\operatorname{h}_{b_0}(C^\vee)$ completes the proof. induces a duality of convex sets. Let $b_0^\ast,b_1^\ast,\dots,b_n^\ast$ denote the standard basis of $({\mathbb{R}}^{n+1})^\ast=({\mathbb{R}}\oplus{\mathbb{R}}^n)^\ast$. Let $(x_0,x)^T\mapsto x$ denote the projection onto the second summand. Clearly, the affine slice $\operatorname{h}_{b_0^\ast}(C)=\{(x_0,x)^T\in C\colon x_0=1\}$ of the convex cone $C$ is isometric under $\widetilde{\pi_2}$ to the set $$\widetilde{\pi_2}\circ\operatorname{h}_{b_0^\ast}(C)=\{x\in{\mathbb{R}}^n\colon (1,x)^T\in C\}.$$ We show this set is the dual convex set to a linear image of the set ${\mathcal{B}}$. \[cor:convex-duality\] The dual convex set to $\pi({\mathcal{B}})$ is $\widetilde{\pi_2}\circ\operatorname{h}_{b_0^\ast}(C)=\big(\pi({\mathcal{B}})\big)^\circ$. If the set $\pi({\mathcal{B}})$ contains the origin, then $\pi({\mathcal{B}})=\big(\widetilde{\pi_2}\circ\operatorname{h}_{b_0^\ast}(C)\big)^\circ$. Let $x\in{\mathbb{R}}^n$. Then $$\begin{aligned} x\in (\pi({\mathcal{B}}))^\circ &\iff x\in (\pi_2\circ\operatorname{h}_{b_0}(C^\vee))^\circ\\ &\iff 1+y(x)\geq 0 && \forall y\in\pi_2\circ\operatorname{h}_{b_0}(C^\vee)\\ &\iff (y_0,y)(1,x)^T\geq 0 && \forall (y_0,y)\in\operatorname{h}_{b_0}(C^\vee)\\ &\iff (1,x)^T\in C\\ &\iff x\in\widetilde{\pi_2}\circ\operatorname{h}_{b_0^\ast}(C).\end{aligned}$$ The first equivalence follows from . Regarding the fourth equivalence, note that $\operatorname{h}_{b_0}(C^\vee)$ is a base of the convex cone $C^\vee$. The remaining equivalences follow immediately from the definitions. As $\pi({\mathcal{B}})$ is a closed convex set, the second statement follows from the equation $S=(S^\circ)^\circ$, which holds for all closed convex subsets $S\subset({\mathbb{R}}^n)^\ast$ containing the origin [@Rockafellar1970]. The convex duality of has been used earlier [@ChienNakazato2010; @Henrion2010; @HeltonSpitkovsky2012] in the context of the joint numerical range. The conic duality of has advantages in our situation. If the convex set $\pi({\mathcal{B}})$ contains the origin, we could use the convex duality $$\label{eq:conv-duality} \pi({\mathcal{B}}) =(\widetilde{\pi_2}\circ\operatorname{h}_{b_0^\ast}(C))^\circ$$ of to describe the set $\pi({\mathcal{B}})$. and present an example where fails as $0\in\pi({\mathcal{B}})$ fails. A remedy would be to translate $\pi({\mathcal{B}})$ along a vector $(\lambda_1,\ldots,\lambda_n)\in({\mathbb{R}}^n)^\ast$ and apply to the translated set. Indeed, if $v_1,\ldots,v_n\in V$, then the map in yields $$\pi_{v_1+\lambda_1 e,\ldots,v_n+\lambda_n e}({\mathcal{B}}) =\pi_{v_1,\ldots,v_n}({\mathcal{B}})+(\lambda_1,\ldots,\lambda_n).$$ The translation is unnecessary if we use the conic duality $$\label{eq:cone-duality} \pi({\mathcal{B}}) =\pi_2\circ\operatorname{h}_{b_0}(C^\vee)$$ of , which is valid even if $0\in\pi({\mathcal{B}})$ fails. The main reason why we employ the conic duality of is that it conveys the algebraic geometry of the dual hyperbolicity cone $C^\vee$ described in to the convex set $\pi({\mathcal{B}})$ directly. The algebraic geometry of the hyperbolicity cone $C$ is hidden behind a convex duality if we use the to describe the set $\pi({\mathcal{B}})$. a) (-1.2,2.4) – (0,0) – (2.4,-0.8) – (2.4,2.4); (-1.2,2.4) – (0,0) – (2.4,-0.8); (-0.5,1) – (2.1,1); (2.1,1) – (2.8,1); (-0.5,0) – (2.1,0); (2.1,0) – (2.4,0); (-1.3,0) – (2.6,0) node\[right\] [$x_1$]{}; /in [-1/-1, -.5/[-]{}, 1/1, 2/2]{} (0pt,2pt) – (0pt,-2pt) node\[below\] ; (0,-.9) – (0,2.6) node\[above\] [$x_0$]{}; /in [1/1, 2/2]{} (2pt,0pt) – (-2pt,0pt) node\[left\] ; (0,0) node\[above right\] ; (2.4,2.4) node\[below left\] [$C$]{}; (1,1) node\[above\] [$\operatorname{h}_{b_0^\ast}(C)$]{}; b) (.8,2.4) – (0,0) – (2.4,1.2) – (2.4,2.4); (.8,2.4) – (0,0) – (2.4,1.2); (.333,1) – (2,1); (.333,0) – (2,0); (-.5,0) – (2.6,0) node\[right\] [$y_1$]{}; /in [.333/, 1/1, 2/2]{} (0pt,2pt) – (0pt,-2pt) node\[below\] ; (0,-.9) – (0,2.6) node\[above\] [$y_0$]{}; /in [1/1, 2/2]{} (2pt,0pt) – (-2pt,0pt) node\[left\] ; (0,0) node\[below left\] ; (2.4,2.4) node\[below left\] [$C^\vee$]{}; (1.3,1) node\[above\] [$\operatorname{h}_{b_0}(C^\vee)$]{}; \[ex:AstarnotW\] Let ${\mathcal{K}}=\{(\xi_1,\xi_2)^T\in{\mathbb{R}}^2:\xi_1,\xi_2\geq 0\}$ be the nonnegative quadrant of the plane $V={\mathbb{R}}^2$ and let $e=(1,1)^T\in V$. The segment $${\mathcal{B}}=\operatorname{h}_e({\mathcal{K}}^\vee) =\big[(0,1),(1,0)\big]$$ is a base of the dual convex cone ${\mathcal{K}}^\vee=\{z^T:z\in{\mathcal{K}}\}$. Let $v=(2,\tfrac{1}{3})^T\in V$. The image of ${\mathcal{B}}$ under the map $\pi:V^\ast\to{\mathbb{R}}^\ast$, $z\mapsto z(v)$ is the interval $[\tfrac{1}{3},2]$, which does not contain the origin, $0\not\in\pi({\mathcal{B}})$. We visualize , , and the failure of in . We formulate the results from in terms of bases of cones, taking the convex cone ${\mathcal{K}}\subset V$ in equal to a hyperbolicity cone. \[thm:piB\] Let $p$ be a hyperbolic polynomial on $V$ with hyperbolicity direction $e\in V$ and hyperbolicity cone $C_e(p)\subset V$. The image of the compact convex base $\operatorname{h}_e(C_e(p)^\vee)$ of the dual convex cone $C_e(p)^\vee$ under the linear map $\pi$ is $$\pi\Big(\operatorname{h}_e\big(C_e(p)^\vee\big)\Big) =\operatorname{conv}\big(\operatorname{clos}(T_1\cup T_2\cup\dots\cup T_r)\big).$$ Here, $p_1^{m_1}p_2^{m_2}\dots p_r^{m_r}$ is a factorization of the pullback $p\circ\phi_0\in{\mathbb{R}}[x_0,x_1,\dots,x_n]$ into irreducible polynomials, and $$T_i = \{ (y_1,\dots,y_n) \in ({\mathbb{R}}^n)^\ast \mid (1:y_1:\dots:y_n)\in ({\mathcal{V}}(p_i)^\ast)_{\mathrm{reg}}\}, \quad i=1,\dots,r.$$ The pullback $p\circ\phi_0$ is a hyperbolic polynomial on ${\mathbb{R}}\oplus{\mathbb{R}}^n$ with hyperbolicity direction $b_0$. As $\phi_0((x_0,x)^T)=x_0e+\phi(x)$ holds for all points $(x_0,x)^T\in{\mathbb{R}}\oplus{\mathbb{R}}^n$, we have $p\circ\phi_0(b_0)=p(e)>0$. The equation $$\label{eq:pull-back} p\circ\phi_0(tb_0-a) =p(te-\phi_0(a))$$ shows that the polynomial $p\circ\phi_0(tb_0-a)$ in one variable $t$ has only real roots for every point $a\in{\mathbb{R}}\oplus{\mathbb{R}}^n$ so that $p\circ\phi_0$ is indeed hyperbolic with respect to the point $b_0$. also shows $$C_{b_0}(p\circ\phi_0) =\phi_0^{-1}\left(C_e(p)\right).$$ Using the hyperbolicity cone ${\mathcal{K}}=C_e(p)$ in , we get $$\pi\Big(\operatorname{h}_e\big(C_e(p)^\vee\big)\Big) =\pi_2\circ\operatorname{h}_{b_0}\big(C_{b_0}(p\circ\phi_0)^\vee\big).$$ If $p_1^{m_1}p_2^{m_2}\dots p_r^{m_r}=p\circ\phi_0$ is a factorization into irreducible polynomials, then $$C_{b_0}(p\circ\phi_0) =C_{b_0}(p_1)\cap C_{b_0}(p_2)\cap\dots\cap C_{b_0}(p_r)$$ holds. proves $C_{b_0}(p_i)^\vee=\operatorname{clos}(\operatorname{cc}(C_i))$, $i=1,\dots,r$, where $$C_i =\big\{ (y_0,y_1,\dots,y_n) \in ({\mathbb{R}}\oplus{\mathbb{R}}^n)^\ast \mid (y_0:y_1:\dots:y_n)\in ({\mathcal{V}}(p_i)^\ast)_{\mathrm{reg}}, y_0\geq 0 \big\}.$$ Now in yields $$\operatorname{h}_{b_0}\Big(C_{b_0}(p\circ\phi_0)^\vee\Big) = \operatorname{conv}\Big(\operatorname{clos}\big(\operatorname{h}_{b_0}(C_1)\cup\operatorname{h}_{b_0}(C_2) \cup\dots\cup\operatorname{h}_{b_0}(C_r)\big)\Big).$$ The claim follows since the map $\pi_2:({\mathbb{R}}\oplus{\mathbb{R}}^n)^\ast\to({\mathbb{R}}^n)^\ast$ restricts to an affine isomorphism from $\operatorname{h}_{b_0}(({\mathbb{R}}\oplus{\mathbb{R}}^n)^\ast)$ onto $({\mathbb{R}}^n)^\ast$, and since $\pi_2\circ\operatorname{h}_{b_0}(C_i)=T_i$ holds for $i=1,\dots,r$. Like in , some of the factors of the pullback $p\circ\phi_0=p_1^{m_1}p_2^{m_2}\dots p_r^{m_r}$ are redundant in . Let $I\subset\{1,2,\dots,r\}$ be the subset such that $i\in I$ if and only if the hypersurface ${\mathcal{V}}(p_i)$ belongs to the algebraic boundary $\partial_a C_{b_0}(p\circ\phi_0)$ of the hyperbolicity cone $C_{b_0}(p\circ\phi_0)$ for $i=1,\dots,r$. Then $\pi({\mathcal{B}})=\operatorname{conv}\Big(\operatorname{clos}\big(\bigcup_{i\in I}T_i\big)\Big)$. The Cone of Positive-Semidefinite Matrices {#sec:PSD-cone} ========================================== Here we prove the theorems stated in the introduction by applying our results to the hyperbolicity cone of positive semi-definite matrices. We discuss block diagonalization *versus* factorization of the determinant. We present examples, among them the promised example by Chien and Nakazato. Let $V=H_d$ be the space of Hermitian $d\times d$ matrices. The determinant is a hyperbolic polynomial on $V$ with respect to any positive definite matrix. The hyperbolicity cone is the set of positive semi-definite matrices $${\mathcal{K}}=C_{\mathds{1}}(\det) =\{A\in H_d:A\succeq 0\}.$$ This convex cone is a self-dual convex cone, i.e. ${\mathcal{K}}^\vee = {\mathcal{K}}$, as we identify the dual space $V^\ast$ with $V$ using the scalar product $\langle A,B\rangle=\operatorname{tr}(AB)$, $A,B\in H_d$. In the notation of , the base $\operatorname{h}_{\mathds{1}}({\mathcal{K}}^\vee)$ of ${\mathcal{K}}^\vee$ is the set of density matrices introduced in , $${\mathcal{B}}=\operatorname{h}_{\mathds{1}}({\mathcal{K}}^\vee) =\{\rho\in H_d \colon \rho\succeq 0, \operatorname{tr}(\rho)=1\}.$$ Let $A_1,\dots,A_n\in H_d$ be Hermitian matrices and let $\pi:H_d\to({\mathbb{R}}^n)^\ast$ be the linear map defined by $\pi(A)=\langle A,A_i\rangle_{i=1}^n$ for all $A\in H_d$. The image of the set ${\mathcal{B}}$ under $\pi$ is the joint numerical range defined in , $$W=\pi({\mathcal{B}})=\{\langle \rho,A_i\rangle_{i=1}^n:\rho\in{\mathcal{B}}\}.$$ The dual map $\phi=\pi^\ast$ has the values $\phi(x)=x_1 A_1+\cdots+x_n A_n$ and the map $\phi_0:{\mathbb{R}}\oplus{\mathbb{R}}^n\to V$, see , has the values $\phi_0[(x_0,x)^T]=x_0{\mathds{1}}+\phi(x)$ for all $x_0\in{\mathbb{R}}$ and $x=(x_1,\dots,x_n)^T\in{\mathbb{R}}^n$. Hence, the pullback $\det\circ\phi_0$ takes the form $$\label{eq:p-detphi} p =\det\circ\phi_0 =\det(x_0{\mathds{1}}+x_1 A_1+\cdots+x_n A_n) \in{\mathbb{R}}[x_0,x_1,\dots,x_n].$$ With this preparation, follows from immediately. To obtain Kippenhahn’s original result, as stated in , we use instead of in the proof of . In the matrix setting of this section, the convex cone $\phi_0^{-1}({\mathcal{K}})$ is the *spectrahedral cone* [@Netzer2012] $$\{x\in{\mathbb{R}}^{n+1} : x_0{\mathds{1}}+x_1 A_1+\cdots+x_n A_n \succeq 0 \},$$ which is the hyperbolicity cone $C_{b_0}(p)$ of the polynomial in , as we observed more generally in the proof of above. Researchers have studied the relationships between the possibility to block diagonalize the Hermitian matrices $A_1,\ldots,A_n\in H_d$ simultaneously and to factor the determinant $p$ in . The matrices $A_1,\ldots,A_n$ are called [*unitarily reducible*]{} if there is a $d$-by-$d$ unitary $U$ and an integer $0<j<d$ such that $UA_iU^\ast=A_{i,1}\oplus A_{i,2}$ is a block diagonal matrix, where $A_{i,1}\in H_j$ and $A_{i,2}\in H_{d-j}$ for all $i=1,\dots,n$. Otherwise the matrices $A_1,\ldots,A_n$ are [*unitarily irreducible*]{}. Clearly, the polynomial $p\in{\mathbb{R}}[x_0,x_1,x_2]$ is the product of the polynomials corresponding to the two diagonal blocks for every pair of unitarily reducible matrices $A_1,A_2\in H_d$. Interestingly, $p$ may even factor when the matrices $A_1,A_2$ are unitarily irreducible. \[ex:Laffey\] Kippenhahn [@Kippenhahn1951 Prop. 28a] conjectured that if $p$ has a repeated factor, then the matrices $A_1,A_2$ are unitarily reducible. Laffey [@Laffey1983] found a first counterexample to the conjecture by presenting a pair of unitarily irreducible hermitian $8$-by-$8$ matrices $A_1,A_2\in H_8$, where $p$ is the square of a polynomial of degree four. The paper [@Buckley2016] summarizes more recent work on the topic and shows that every pair of hermitian $6$-by-$6$ matrices $A_1,A_2\in H_6$ is unitarily reducible if $p$ is the square of a polynomial of degree three that defines a smooth cubic. Similarly, one may ask whether the polynomial $p\in{\mathbb{R}}[x_0,x_1,x_2]$ can be a product of several irreducible factors of multiplicity one if the hermitian matrices $A_1,A_2\in H_d$ are unitarily irreducible. Such questions have been studied in detail by Kerner and Vinnikov ([@Kerner-Vinnikov2011; @Kerner-Vinnikov2012]), resulting in a number of necessary and sufficient conditions. For example, the union of three lines in the projective plane admits a unitarily irreducible determinantal representation (see [@Kerner-Vinnikov2011 Remark 3.6]). Here is a simple example demonstrating that the dual varieties to the irreducible components of the hypersurface ${\mathcal{V}}(p)$ may have different dimensions. \[ex:drop\] Let $$A_1=\left( \begin{array}{ccc} 0 & 1 & 0 \\ 1 & 0 & 0 \\ 0 & 0 & 2 \end{array}\right), \qquad A_2=\left( \begin{array}{ccc} 0 & -{\operatorname{i}}& 0 \\ {\operatorname{i}}& 0 & 0 \\ 0 & 0 & 0 \end{array}\right), \qquad A_3=\left( \begin{array}{ccc} 1 & 0 & 0 \\ 0 & -1 & 0 \\ 0 & 0 & 0 \end{array}\right).$$ The irreducible components of ${\mathcal{V}}(p)$ are $X_1=\{x\in{\mathbb{P}}^3:x_1^2+x_2^2+x_3^2=x_0^2\}$ and $X_2=\{x\in{\mathbb{P}}^3:x_0+2 x_1=0\}$. The dual varieties are $X_1^\ast=\{y\in({\mathbb{P}}^3)^\ast:y_1^2+y_2^2+y_3^2=y_0^2\}$, by , and $X_2^\ast=\{(1:2:0:0)\}$. Both $X_1^\ast$ and $X_2^\ast$ are smooth, so $T_1=\{y\in({\mathbb{R}}^3)^\ast:y_1^2+y_2^2+y_3^2=1\}$ and $T_2=\{(2,0,0)\}$. The joint numerical range $W=\operatorname{conv}(T_1\cup T_2)$ is the convex hull of a sphere and a point outside the sphere. a)![Pictures for : a) Surface $T$ (blue), singular locus $y_1$-axis (orange), and ellipse on the boundary of the joint numerical range (black). b) The hyperplane $y_2=0$ intersects $T$ in the union of an ellipse (yellow) and the $y_1$-axis (orange).[]{data-label="fig:ChienNakazato"}](ChienNakazato1.png "fig:"){height="3.6cm"}\ b)![Pictures for : a) Surface $T$ (blue), singular locus $y_1$-axis (orange), and ellipse on the boundary of the joint numerical range (black). b) The hyperplane $y_2=0$ intersects $T$ in the union of an ellipse (yellow) and the $y_1$-axis (orange).[]{data-label="fig:ChienNakazato"}](ChienNakazato2.png "fig:"){height="3.6cm"} In our last example, we discuss the original example given by Chien and Nakazato from the point of view of real algebraic geometry in detail. The example explains why singular points of the dual varieties ${\mathcal{V}}(p_i)^\ast$ have to be excluded from the statement of . \[ex:ChienNakazato\] Let $$A_1=\left( \begin{array}{ccc} 1 & 0 & 0 \\ 0 & -1 & 1 \\ 0 & 1 & 0 \end{array}\right), \qquad A_2=\left( \begin{array}{ccc} 0 & 0 & -{\operatorname{i}}\\ 0 & 0 & 0 \\ {\operatorname{i}}& 0 & 0 \end{array}\right), \qquad A_3=\left( \begin{array}{ccc} 0 & 0 & 0 \\ 0 & 0 & 0 \\ 0 & 0 & 1 \end{array}\right).$$ The hyperbolic cubic form $$\begin{aligned} p &= \det( x_0 {\mathds{1}}+ x_1 A_1 + x_2 A_2 + x_3 A_3 )\\ &= x_0^3 + x_0^2x_3 - 2x_0x_1^2 - x_0x_2^2 - x_1^3 - x_1^2x_3 + x_1x_2^2\end{aligned}$$ is irreducible. The dual variety to $X=\{x\in{\mathbb{P}}^3: p(x)=0\}$ is the hypersurface $X^\ast=\{y\in({\mathbb{P}}^3)^\ast: q(y)=0\}$ defined by the homogeneous quartic form $$\begin{aligned} q = & \, 4 y_0^2 y_3^2 + 8 y_0 y_1 y_3^2 - 4 y_0 y_2^2 y_3 - 24 y_0 y_3^3 + 4 y_1^2 y_3^2 - 4 y_1 y_2^2 y_3 - 8 y_1 y_3^3\\ & \, + y_2^4 + 8 y_2^2 y_3^2 + 20 y_3^4.\end{aligned}$$ It is easy to show that the singular locus $X^\ast\setminus(X^\ast)_{\mathrm{reg}}$ of $X^*$ is the line $\{(y_0,y_1,y_2,y_3)\in({\mathbb{P}}^3)^\ast: y_2=y_3=0\}$. The surface $T=\{(y_1,y_2,y_3)\in({\mathbb{R}}^3)^\ast \mid (1:y_1:y_2:y_3)\in X^\ast\}$, depicted in , is the zero-set of the polynomial $Q(y_1,y_2,y_3)=q(1,y_1,y_2,y_3)$. All points $(y_1,0,0)\in T$ on the line of singular points $y_2=y_3=0$ in the interval $|y_1|\leq 1$ are central points of $X^\ast$. This can be seen from a parametrization of $X$ described in [@SchwonnekWerner2018]. In addition, the points $(y_1,0,0)\in T$ with $|y_1|>1$ are not central. This can be seen from the roots of the quadratic polynomial $R_{y_1,y_3}\in{\mathbb{R}}[y_2]$ satisfying $R_{y_1,y_3}(y_2^2)=Q(y_1,y_2,y_3)$. In fact, if $(y_1,y_3)\in({\mathbb{R}}^2)^\ast$, then the line $\{(y_1,\lambda,y_3) \colon \lambda\in{\mathbb{R}}\}$ intersects $T$ if and only if $R_{y_1,y_3}$ has a non-negative root $y_2$. It is not hard to see (for example using Sturm’s theorem [@BCR]) that, under the assumption $y_3\neq 0$, the polynomial $R_{y_1,y_3}$ has a non-negative root if and only if $(y_1,y_3)$ lies in the convex hull of the union of the singleton $\{(1,0,0)\}$ with the ellipse given by $y_1^2 + 5 y_3^2 - 2 y_1 y_3 + 2 y_1 - 6 y_3 + 1 = y_2 = 0$, the yellow curve in  b). Therefore, the points $(y_1,0,0)$ with $|y_1|>1$ are not central. These are the points which are removed from $X^\ast({\mathbb{R}})$ in the statement of . The convex hull of the remainder of $T$ is the joint numerical range $W$. For a general, regular point $(y_0:y_1:y_2:y_3)$ on $X^*$, the dual hyperplane in ${\mathbb{P}}^3$ defined by $x_0y_0 + x_1y_1 + x_2y_2 + x_3y_3 = 0$ intersects $X$ in an irreducible but singular cubic. The hyperplane sections of the hypersurface $X$ corresponding to points of the line of singular points of $X^\ast$ are special in the following way: The sections are reducible curves in ${\mathbb{P}}^2$ that factor into a conic and a line, which is tangent to the conic at a real point. Not all of the points on this line are central points of $X^*$. The central points of $X^*$ on this line are exactly those that we can perturb such that the dual hyperplane section of $X$ deforms to a real cubic with a real singularity. The other points on this line can only be perturbed on $X^*$ to complex points, which means that the dual hyperplane section of $X$ only deforms to complex singular hyperplane sections that are complex cubics with complex singularities. To see this explicitly, we compute the restrictions of $p$ to hyperplanes in ${\mathbb{P}}^3$ defined by $x_0 y_0 + x_1 y_1 = 0$. Let us assume for simplicity for now that $y_0 \neq 0$ so that $x_0 = a x_1$ for some $a\in {\mathbb{R}}$. Then $p$ factors $$p(a x_1,x_1,x_2,x_3) = x_1 \left( (a^3 - 2a - 1) x_1^2 + (1 - a) x_2^2 + (a^2 - 1) x_1 x_3 \right).$$ So in the plane defined by $x_0y_0 + x_1y_1=0$, we see the conic $$(a^3 - 2a - 1) x_1^2 + (1 - a) x_2^2 + (a^2 - 1) x_1 x_3 = 0$$ and the line defined by $x_0 = x_1 = 0$, which is tangent to the conic at the point $(0:0:0:1)$, which is a singular point of $X$. (The computation for the case $y_0 = 0$ is similar.) For every real value of the parameter $a$, the conic is real and indefinite of full rank except for $a=-1$ and $a=1$. These two values of $a$ bound the interval of central points on the line of singular points on $X^\ast$; see . Section 6 of [@Szymanski-etal2018] presents further examples of surfaces analogous to the surface $T$ in that contain straight lines. These lines are identified in the paper `arXiv:1603.06569v3 [math.FA]` on the preprint server arXiv. So why do naive generalizations of Kippenhahn’s Theorem fail in higher dimensions? The reason is that hyperplanes and lines are only the same in ${\mathbb{P}}^2$. A central point of the argument in the proof of Kippenhahn’s Theorem is , which argues that there are no lines through the interior of the hyperbolicity cone that are tangent to the hyperbolic hypersurface (). The hyperplane corresponding to a real point of the dual variety may well meet an interior point of the hyperbolicity cone as long as the hyperplane contains no tangent line to the hyperbolic hypersurface incident with this interior point. Our generalization holds essentially because this can only happen for hyperplanes that are not central points of the dual variety. Acknowledgements {#acknowledgements .unnumbered} ---------------- We thank H. Nakazato for helpful comments. SW thanks K. [Ż]{}yczkowski for discussions of numerical ranges in quantum mechanics. We would like to thank the anonymous referees for their suggestions to improve the paper. [10]{} Y.-H. Au-Yeung and Y.-T. Poon, [*A remark on the convexity and positive definiteness concerning Hermitian matrices*]{}, Southeast Asian Bull. Math. [**3**]{} (1979), 85–92. G. Aubrun and S.J. Szarek, [*Alice and Bob Meet Banach: The Interface of Asymptotic Geometric Analysis and Quantum Information Theory*]{}, AMS, Providence, 2017. A. Barvinok, [*A Course in Convexity*]{}, AMS, Providence, 2002. N. Bebiano, R. Lemos, J. da Providência, G. Soares, [*On the geometry of numerical ranges in spaces with an indefinite inner product*]{}, Linear Algebra and Its Applications [**399**]{} (2005), 17–34. I. Bengtsson and K. [Ż]{}yczkowski, [*Geometry of Quantum States*]{}, 2nd ed., Cambridge University Press, Cambridge, 2017. G. Birkhoff, [*Lattice Theory*]{}, Third edition, AMS, Providence, 1973. J. Bochnak, M. Coste, M.-F. Roy, *Real Algebraic Geometry*, vol. 36 of *Ergebnisse der Mathematik und ihrer Grenzgebiete (3)*, Springer-Verlag, Berlin, 1998. F.F. Bonsall and J. Duncan, [*Numerical Ranges of Operators on Normed Spaces and of Elements of Normed Algebras*]{}, Cambridge University Press, London, 1971. A. Buckley, [*Indecomposable matrices defining plane cubics*]{}, Operators and Matrices [**10**]{} (2016), 1059–1072. K. A. Camenga, L. Deaett, P. X. Rault, T. Sendova, I. M. Spitkovsky, R. B. Johnson Yates, [*Singularities of base polynomials and Gau-Wu numbers*]{}, Linear Algebra and its Applications [**581**]{} (2019), 112–127. J. Chen, Z. Ji, M.B. Ruskai, B. Zeng, and D.-L. Zhou, [*Comment on some results of Erdahl and the convex structure of reduced density matrices*]{}, Journal of Mathematical Physics [**53**]{} (2012), 072203. J. Chen, Z. Ji, C.-K. Li, Y.-T. Poon, Y.-Shen, N. Yu, B. Zeng, and D. Zhou, [*Discontinuity of maximum entropy inference and quantum phase transitions*]{}, New Journal of Physics [**17**]{} (2015), 083019. J. Chen, C. Guo, Z. Ji, Y. -T. Poon, N. Yu, B. Zeng, J. Zhou, [*Joint product numerical range and geometry of reduced density matrices*]{}, Science China Physics, Mechanics & Astronomy [**60**]{} (2017), 020312. M.-T. Chien and H. Nakazato, [*Joint numerical range and its generating hypersurface*]{}, Linear Algebra and its Applications [**432**]{} (2010), 173–179. M.-T. Chien and H. Nakazato, [*Singular points of the ternary polynomials associated with 4-by-4 matrices*]{}, Electronic Journal of Linear Algebra [**23**]{} (2012) 755–769. M.-T. Chien and H. Nakazato, [*Unitary similarity of the determinantal representation of unitary bordering matrices*]{}, Linear Algebra and its Applications [**541**]{} (2018), 13–35. M.-D. Choi, D. W. Kribs, K. [Ż]{}yczkowski, [*Quantum error correcting codes from the compression formalism*]{}, Reports on Mathematical Physics [**58**]{} (2006), 77–91. A. Conca, D. Edidin, M. Hering, C. Vinzant, [*An algebraic characterization of injectivity in phase retrieval*]{}, Applied and Computational Harmonic Analysis [**38**]{} (2015), 346–356. M. Fiedler, [*Geometry of the numerical range of matrices*]{}, Linear Algebra and Its Applications [**37**]{} (1981), 81–96. G. Fischer, [*Plane Algebraic Curves*]{}, AMS, Providence, 2001. H.-L. Gau and P.Y. Wu, [*Companion matrices: Reducibility, numerical ranges and similarity to contractions*]{}, Linear Algebra and its Applications [**383**]{} (2004), 127–142. H.-L. Gau, M.-C. Tsai, H.-C. Wang, [*Weighted shift matrices: Unitary equivalence, reducibility and numerical ranges*]{}, Linear Algebra and its Applications [**438**]{} (2013), 498–513. T. Gallay and D. Serre, [*Numerical measure of a complex matrix*]{}, Communications on Pure and Applied Mathematics [**65**]{} (2012), 287–336. P. Gawron, Z. Pucha[ł]{}a, J. A. Miszczak, [Ł]{}. Skowronek, K. [Ż]{}yczkowski, [*Restricted numerical range:  A versatile tool in the theory of quantum information*]{}, Journal of Mathematical Physics [**51**]{} (2010), 102204. I.M. Gel’fand, M. M. Kapranov, A. V. Zelevinsky, [*Discriminants, Resultants and Multidimensional Determinants*]{}, Birkhäuser, Boston, 1994. E. Gutkin, E. A. Jonckheere, M. Karow, [*Convexity of the joint numerical range: topological and differential geometric viewpoints*]{}, Linear Algebra and its Applications [**376**]{} (2004), 143–171. J. Harris, [*Algebraic Geometry: A First Course*]{}, Springer-Verlag, New York, 1992. F. Hausdorff, [*Der Wertvorrat einer Bilinearform*]{}, Math. Z. [**3**]{} (1919), 314–316. T. Heinosaari, L. Mazzarella, M. M. Wolf, [*Quantum tomography under prior information*]{}, Communications in Mathematical Physics [**318**]{} (2013), 355–374. J.W. Helton and I.M. Spitkovsky, [*The possible shapes of numerical ranges*]{}, Operators and Matrices [**6**]{} (2012), 607–611. J.W. Helton and V. Vinnikov, [*Linear matrix inequality representation of sets*]{}, Communications on Pure and Applied Mathematics [**60**]{} (2007), 654–674. D. Henrion, [*Semidefinite geometry of the numerical range*]{}, Electronic Journal of Linear Algebra [**20**]{} (2010), 322–332. A. Holevo, [*Probabilistic and Statistical Aspects of Quantum Theory*]{}, 2nd edition, Edizioni della Normale, Pisa, 2011. M. Joswig and B. Straub, [*On the numerical range map*]{}, Journal of the Australian Mathematical Society [**65**]{} (1998), 267–283. D.S. Keeler, L. Rodman, I. M. Spitkovsky, [*The numerical range of $3\times 3$ matrices*]{}, Linear Algebra and its Applications [**252**]{} (1997), 115–139. D. Kerner and V. Vinnikov, [*Decomposability of local determinantal representations of hypersurfaces*]{}, arXiv:1009:2517 (preprint 2011, unpublished). D. Kerner and V. Vinnikov, [*Determinantel representations of singular hypersurfaces in ${\mathbb{P}}^n$*]{}, Advances in Mathematics [**231**]{} (2012), 1619–1654. R. Kippenhahn, [*Über den Wertevorrat einer Matrix*]{}, Mathematische Nachrichten [**6**]{} (1951), 193–228. T.J. Laffey, [*A counterexample to Kippenhahn’s conjecture on Hermitian pencils*]{}, Linear Algebra and its Applications [**51**]{} (1983), 179–182. M. Lewin, P. T. Nam, N. Rougerie, [*Derivation of Hartree’s theory for generic mean-field Bose systems*]{}, Advances in Mathematics [**254**]{} (2014), 570–621. C.-K. Li and Y.-T. Poon, [*Convexity of the joint numerical range*]{}, SIAM Journal on Matrix Analysis and Applications [**21**]{} (2000), 668–678. E. Militzer, L. J. Patton, I. M. Spitkovsky, M.-C. Tsai, [*Numerical ranges of 4-by-4 nilpotent matrices: Flat portions on the boundary*]{}, in: D. Bini et al. (eds.), Large Truncated Toeplitz Matrices, Toeplitz Operators, and Related Topics, Birkhäuser, Cham, 2017, 561–591. V. M[ü]{}ller, [*The joint essential numerical range, compact perturbations, and the Olsen problem*]{}, Studia Mathematica [**197**]{} (2010), 275–290. D. Mumford, [*Algebraic Geometry. 1: Complex Projective Varieties*]{}, Corr. 2. print, Springer, Berlin, 1976. F.D. Murnaghan, [*On the field of values of a square matrix*]{}, PNAS [**18**]{} (1932), 246–248. T. Netzer, [*Spectrahedra and Their Shadows*]{}, Habil.-Schr., Leipzig, Germany, 2012. D. Plaumann and C. Vinzant, [*Determinantal representations of hyperbolic plane curves: An elementary approach*]{}, Journal of Symbolic Computation [**57**]{} (2013), 48–60. J. Renegar, [*Hyperbolic programs, and their derivative relaxations*]{}, Foundations of Computational Mathematics [**6**]{} (2006), 59–79. R.T. Rockafellar, [*Convex Analysis*]{}, Princeton University Press, Princeton, 1970. L. Rodman, [*Topics in Quaternion Linear Algebra*]{}, Princeton University Press, Princeton, 2014. L. Rodman, I. M. Spitkovsky, A. Szko[ł]{}a, S. Weis, [*Continuity of the maximum-entropy inference:  Convex geometry and numerical ranges approach*]{}, Journal of Mathematical Physics [**57**]{} (2016), 015204. R. Schwonnek and R.F. Werner, [*Properties of the Wigner distribution for n arbitrary operators*]{}. `arXiv:1802.08343 [quant-ph]` R. Sinn, [*Algebraic boundaries of convex semi-algebraic sets*]{}, Mathematical Sciences (2015) 2:3. `https://doi.org/10.1186/s40687-015-0022-0` I.M. Spitkovsky and S. Weis, [*Signatures of quantum phase transitions from the boundary of the numerical range*]{}, Journal of Mathematical Physics [**59**]{} (2018), 121901. E. St[ø]{}rmer, [*Symmetric states of infinite tensor products of C\*-algebras*]{}, Journal of Functional Analysis [**3**]{} (1969), 48–68. K. Szyma[ń]{}ski, S. Weis, K. [Ż]{}yczkowski, [*Classification of joint numerical ranges of three Hermitian matrices of size three*]{}, Linear Algebra and its Applications [**545**]{} (2018), 148–173. B.-S. Tam, [*On the duality operator of a convex cone*]{}, Linear Algebra and its Applications [**64**]{} (1985), 33–56. E.A. Tevelev, [*Projective Duality and Homogeneous Spaces*]{}, Springer, Berlin, 2005. O. Toeplitz, [*Das algebraische Analogon zu einem Satze von Fejér*]{}, Mathematische Zeitschrift [**2**]{} (1918), 187–197. S. Weis and A. Knauf, [*Entropy distance: New quantum phenomena*]{}, Journal of Mathematical Physics [**53**]{} (2012), 102206. J. Xie, A. Zhang, N. Cao, H. Xu, K. Zheng, Y.-T. Poon, N.-S. Sze, P. Xu, B. Zeng, L. Zhang [*Observing geometry of quantum states in a three-level system*]{}.\ [`arXiv:1909.05463 [quant-ph]`](https://arxiv.org/abs/1909.05463) N. Yunger Halpern, P. Faist, J. Oppenheim, A. Winter, [*Microcanonical and resource-theoretic derivations of the thermal state of a quantum system with noncommuting charges*]{}, Nat. Commun. [**7**]{} (2016), 12051. V. Zauner, D. Draxler, L. Vanderstraeten, J. Haegeman, F. Verstraete, [*Symmetry breaking and the geometry of reduced density matrices*]{}, New Journal of Physics [**18**]{} (2016), 113033.
{ "pile_set_name": "ArXiv" }
IFT-UAM/CSIC-09-08\ 27 February 2009 [**Cosmological solutions from fake $N=2$ EYM supergravity**]{}\ [**Patrick Meessen and Alberto Palomo-Lozano**]{}\ [*Instituto de Física Teórica UAM/CSIC, Facultad de Ciencias C-XVI\ C.U. Cantoblanco, E-28049 Madrid, Spain*]{}\ [**abstract**]{}\ > [ We characterise the (fake) supersymmetric solutions of Wick-rotated $N=2$ $d=4$ gauged supergravity coupled to non-Abelian vector multiplets. In the time-like case we obtain generalisations of Kastor [&]{} Traschen’s cosmological black holes: they have a specific time-dependence and the base-space must be 3-dimensional hyperCR/Gauduchon-Tod space. In the null-case, we find that the metric has a holonomy contained in $\mathrm{Sim}(2)$, give a general characterisation of the solutions, and give some examples. Finally, we point out that in some cases the solutions we found are non-BPS solutions to $N=2$ $d=4$ supergravity coupled to vector multiplets. ]{} In 1992 Kastor [&]{} Traschen [@Kastor:1992nn] created a cosmological multi-black hole solution to Einstein-Maxwell-De Sitter gravity by observing that the extreme Reissner-Nordstrom-De Sitter black hole solution written in spherical coordinates could be transformed to the time-dependent conforma-static form $$\label{eq:KT1} ds^{2} \; =\; \Omega^{-2}d\tau^{2} \ -\ \Omega^{2}\ d\vec{x}^{2}_{(3)} \hspace{.6cm}\mbox{with}\hspace{.6cm} \Omega \ =\ H\tau \ +\ \frac{m}{r} \; ,$$ where $3H^{2}$ is the cosmological constant and we introduced the further coordinate transformation $H\tau =e^{Ht}$; as the $r$-dependent part of $\Omega$ is a spherically symmetric harmonic function, the multi-bh solutions can be created by changing it to a general harmonic function. Seeing the similarity of the above solution and the supersymmetric solutions to minimal $N=2$ $d=4$ supergravity [@Tod:1983pm], whose bosonic part is just EM-theory, Kastor [&]{} Traschen showed [@Kastor:1993mj] that their multi-bh solution solved the spinorial equations[^1] $$\label{eq:FM6a} \nabla_{a}\epsilon_{I} \, =\, -\textstyle{iH\over 2}\gamma_{a}\ \varepsilon_{IJ}\epsilon^{J} \ +\ H\ A_{a}\epsilon_{I} \ +\ iF^{+}_{ab}\gamma^{b}\varepsilon_{IJ}\epsilon^{J}\; .$$ This fermionic rule can be derived from the supersymmetry variations of minimal gauged $N=2$ $d=4$ supergravity, which has an anti-De Sitter type cosmological constant $\Lambda = 3g^{2}$, by Wick-rotating $g\rightarrow iH$. As eq. (\[eq:FM6a\]) looks like a Killing Spinor Equation but is not due to supersymmetry, we will refer to equations like it as fake-Killing Spinor Equations (fKSEs) [@Freedman:2003ax]. The KT solutions were subsequently generalised to higher dimensions by London in ref. [@London:1995ib], who also showed that his solutions solved a suitable fKSE, and generalised to spinning solutions in a stringy theory[^2] by Shiromizu [@Shiromizu:1999xj]. In ref. [@Behrndt:2003cx], Behrndt [&]{} Cvetič generalised the KT-solution to asymptotically DS solutions to $5$- and $4$-dimensional supergravities coupled to vector multiplets by observing the following substitution rule: as one can see form the expression for $\Omega$ in eq. (\[eq:KT1\]), the difference between the cosmological solution and the usual supersymmetric solutions is nothing but the linear $\tau$-dependence. As harmonic functions appear quite natural in supersymmetric solutions, the substitution rule is to add to these harmonic functions a piece linear in the time-coordinate. Furthermore, Behrndt [&]{} Cvetič showed that their solutions solved fKSEs that could be obtained from the KSEs of gauged supergravity coupled to vector-multiplets, by Wick-rotating the coupling constant, pointing out that this is equivalent to considering an $\mathbb{R}$-gauged symmetry. Indeed, the construction of [*e.g.*]{} gauged $N=2$ $d=4$ supergravity coupled to vector-multiplets calls for the inclusion of an $U(1)$ Fayet-Iliopoulos term, which as far as the Killing spinor is concerned means that it is gauged (see [*e.g.*]{} [@Andrianopoli:1996cm]), proportional to the coupling constant. Wick-rotating the coupling constant, then, is equivalent to Wick-rotating the gauge group, which becomes $\mathbb{R}$. It was recently realised by Grover [*et al.*]{} [@Grover:2008jr], that the techniques used to classify supersymmetric solutions to supergravity theories, could be used to construct solutions to theories admitting fKSEs; they applied the techniques of ref. [@Gillard:2004xq] to the classification of solutions to the time-like case of minimal ‘De Sitter $N=1$ $d=5$ supergravity’, which can be obtained by Wick-rotation from minimal gauged $N=1$ $d=5$ supergravity. Of special interest in these classification is the geometry of the 4-dimensional base-space, which in the DS case turns out to be hyperKähler-Torsion, whereas it is hyper-Kähler in the ungauged sugra [@Gauntlett:2002nw] and Kähler in the gauged sugra [@Gauntlett:2003fk]. In this article we will extend the results of K[&]{}T [@Kastor:1992nn] and B[&]{}C [@Behrndt:2003cx] by considering solutions to Wick’ed $N=2$ $d=4$ supergravity coupled to non-Abelian vector multiplets, by which we mean that we allow for gaugings of the isometries of the scalar manifold (see [*e.g.*]{} [@Andrianopoli:1996cm]). As indicated above, this theory can be obtained from gauged $N=2$ $d=4$ supergravity coupled to non-Abelian vector-multiplets by Wick-rotation, not of the coupling constant as we are allowing for non-Abelian couplings, but of the Fayet-Iliopoulos term responsible for gauging the $R$-symmetry; we shall refer to this theory as fake $N=2$ Einstein-Yang-Mills. For understandable reasons, the $N=2$ $d=4$ supergravity theories have attracted quite some interest in the last decades, and the theories for which the supersymmetric solutions have been fully classified/characterised are the minimal theory [@Tod:1983pm], the minimal theory coupled to vector- and hyper-multiplets [@Behrndt:1997ny; @Meessen:2006tu], the minimal theory coupled to non-Abelian vector-multiplets [@Hubscher:2008yz], minimal gauged theory [@Caldarelli:2003pb], and recently the minimal gauged theory coupled to vector-multiplets [@Cacciatori:2008ek]. Wick rotation of the coupling constant in gauged supergravity was also considered in ref. [@Skenderis:2007sm] in order to find a supergravity basis for the Domain Wall/Cosmology correspondence [@Skenderis:2006jq]. As ref. [@Skenderis:2007sm] focusses on proper supersymmetry, Wick rotation of the coupling constant has to be necessarily accompanied by a change of reality conditions on the spinors and, furthermore, a Wick rotation of the vector field: the result is a true De Sitter supergravity with its characteristic ghost-like vector field, [*i.e.*]{} the kinetic term for the vector field has the wrong sign (see [*e.g.*]{} [@Pilch:1984aw]). In our construction, however, we do not impose proper supersymmetry and do not change the reality conditions of the spinors: this avoids the problem of having ghost-like vector fields, implying that in the limit of vanishing FI-term we recover an ordinary supergravity theory. The outline of this paper is the following: in section (\[sec:DSVector\]) we shall set up the fake-Killing spinors equations we are going to solve and some information about special geometry and the gauging of isometries in special geometries, needed to understand the set up, are given in appendix (\[appsec:SpecGeom\]). In that section we will see that, as we are Wick-rotating the FI-term, the relations between the equations of motion one can derive from the integrability equation, are similar to the ones obtained in the supersymmetric case and that the implications as far as the checking of equations of motion are concerned are identical: this was to be expected as we are not changing the characteristics of the Killing spinors. Similar to the supersymmetric classifications, there are two cases to be considered, namely the ones depending on the norm of the vector one constructs as a bilinear of the fake-Killing spinors, and the time-like case, [*i.e.*]{} when the norm doesn’t vanish, will be treated in section (\[sec:VectBil\]). In section (\[sec:Null\]) we will have a go at the null case, [*i.e.*]{} when the norm of the vector vanishes identically. In that section, we shall ignore the possible non-Abelian couplings and furthermore will not obtain a complete characterisation; in stead we shall see that the solutions have infinitesimal holonomy contained in $\mathfrak{sim}(2)$ and discuss the general features such a solution should have. This will be illustrated by two solutions, namely the Nariai cosmos in the minimal theory in section (\[sec:NullSols\]) and in section (\[sec:Holomorphic\]) a general class of solutions with holomorphic scalars which can be seen as a, back-reacted, intersection of a cosmic string with a Robinson-Bertotti-Nariai solution. The reader might feel that the generic theories that can be treated in our setting are rather esoteric as their connection with supergravity theories or EYM-$\Lambda$ theories is rather weak: in section (\[sec:PotIsNul\]) we shall use the well-known fact that in gauged $N=2$ $d=4$ supergravity theories there are choices for the FI-terms for which the theory under consideration equals that of the bosonic part of an ungauged supergravity [@Cremmer:1984hj]. This in fact means that in those cases, our fake-supersymmetric solutions are nothing more than non-BPS solutions to an ordinary ungauged supergravity. The easiest model in which one can see this happen is the model which can be obtained by dimensionally reducing minimal $N=1$ $d=5$ supergravity, and we shall discuss some simple solutions to this model and also their uplift to five dimensions. Finally, in section (\[sec:Concl\]) we shall give our conclusions and a small outlook for related work in higher dimensions, and appendices (\[sec:Bil\]) and (\[sec:NullCurv\]) contains information about the normalisation of the bilinears and the curvatures for the null-case. Fake $N=2$ Einstein-Yang-Mills {#sec:DSVector} ============================== As was said in the introduction, the set-up that we want to consider can be obtained from ordinary $N=2$ $d=4$ gauged sugra coupled to vector multiples but no hyper-multiplets, by Wick-rotating the Fayet-Iliopoulos term: said differently, we Wick-rotate the constant tri-holomorphic map $\mathsf{P}^{x}_{\Lambda}\rightarrow i\mathtt{C}_{\Lambda}\delta^{x}_{2}$, where $\mathtt{C}_{\Lambda}$ are real constant. In supersymmetry the FY-term would gauge an $U(1)$ in the hyper-multiplets’ $SU(2)$, and the effect of the Wick-rotation is that we are gauging an $\mathbb{R}$-symmetry through the effective connection $\mathtt{C}_{\Lambda}A^{\Lambda}$ [@Behrndt:2003cx]. The presence of a FI-term is compatible with the gauging of non-Abelian isometries of the scalar manifold, as long as the action of the gauge group commutes with the FI-term (see [*e.g.*]{} [@Andrianopoli:1996cm]); taking the gauge algebra to have structure constants $f_{\Lambda\Sigma}{}^{\Gamma}$, then implies that we must impose the constraint $f_{\Lambda\Sigma}{}^{\Omega}\mathtt{C}_{\Omega}=0$. One result of the introduction of the $\mathtt{C}_{\Lambda}$ is that the dimension of the possible gauge-algebra is not $\bar{n}=n+1$, $n$ being the number of vector multiplets, but rather $n$, as ‘one‘ vector field is already used as the connection for the $\mathbb{R}$-symmetry. The gauging of isometries implies that field-strengths of the physical fields are given by $$\label{eq:Deriv1} \mathtt{D}Z^{i}\; \equiv\; dZ^{i}\ +\ gA_{\Lambda}^{\Lambda}\ \mathtt{K}^{i} \hspace{.5cm} ,\hspace{.5cm} F^{\Lambda} \; \equiv\; dA^{\Lambda} \ +\ \textstyle{g\over 2}\ f_{\Sigma\Gamma}{}^{\Lambda}\ A^{\Sigma}\wedge A^{\Gamma} \; .$$ where $\mathtt{K}_{\Lambda}^{i}$ is the holomorphic part of the Killing vector $\mathtt{K}_{\Lambda}$ (see appendix (\[appsec:SpecGeom\]) for the minimal information needed or refs. [@Andrianopoli:1996cm; @Hubscher:2008yz] for a fuller account). One implication of the above definition is that $\mathtt{C}_{\Lambda}F^{\Lambda} = d\left[ \mathtt{C}_{\Lambda}A^{\Lambda}\right]$, so that the linear combination $\mathtt{C}_{\Lambda}A^{\Lambda}$ is indeed an Abelian vector-field. As mentioned, we are introducing an $\mathbb{R}$-connection which together with the existent Kähler/$U(1)$-symmetry due to the vector coupling means that we should define the covariant derivative on the F-killing spinors as[^3] $$\begin{aligned} \label{eq:16} \mathbb{D}_{a}\epsilon_{I} & =& \nabla_{a}\epsilon_{I} \, +\, \textstyle{i\over 2}\mathcal{Q}_{a}\epsilon_{I} \, +\, \frac{ig}{2}\ A_{a}^{\Lambda}\ \left[\mathtt{P}_{\Lambda}\ +\ i \mathtt{C}_{\Lambda}\right]\epsilon_{I} \nonumber \\ & \equiv& \mathtt{D}_{a}\epsilon_{I} \, -\, \textstyle{g\over 2}\ \mathtt{C}_{\Lambda}A^{\Lambda}_{a}\epsilon_{I}\; ,\end{aligned}$$ where $\mathtt{P}_{\Lambda}$ is the momentum map corresponding to an isometry $\mathtt{K}_{\Lambda}$ of the special geometry. Using the above definitions we can write the fake Killing Spinor Equations as $$\begin{aligned} \label{eq:20} \mathbb{D}_{a}\epsilon_{I} & =& -\varepsilon_{IJ}\ \mathcal{T}_{ab}^{+}\gamma^{b}\ \epsilon^{J} \; -\; \textstyle{ig\over 4}\ \mathtt{C}_{\Lambda}\mathcal{L}^{\Lambda}\, \gamma_{a}\ \varepsilon_{IJ}\epsilon^{J} \; ,\\ \label{eq:20b} \mathbb{D}_{a}\epsilon^{I} & =& \varepsilon^{IJ}\ \overline{\mathcal{T}^{+}}_{ab}\gamma^{b}\ \epsilon^{J} \; -\; \textstyle{ig\over 4}\ \mathtt{C}_{\Lambda}\overline{\mathcal{L}}^{\Lambda}\, \gamma_{a}\ \varepsilon^{IJ}\epsilon_{J} \; ,\\ & & \nonumber \\ \label{eq:20c} i\slashed{\mathtt{D}}Z^{i}\ \epsilon^{I} & =& -\varepsilon^{IJ}\ \slashed{G}^{i+}\epsilon_{J} \; -\; \mathtt{W}^{i}\ \varepsilon^{IJ}\epsilon_{J} \; ,\\ \label{eq:20d} i\slashed{\mathtt{D}}\overline{Z}^{\bar{\imath}}\ \epsilon_{I} & =& -\varepsilon_{IJ}\ \slashed{\overline{G}}^{\bar{\imath}-}\epsilon^{J} \; -\; \overline{\mathtt{W}}^{\bar{\imath}}\ \varepsilon_{IJ}\epsilon^{J} \; , \end{aligned}$$ where for clarity we have given also the rules for $\mathbb{D}_{a}\epsilon^{I}$ and ${\scriptstyle\slashed{\mathtt{D}}\overline{Z}^{\bar{\imath}}}$ $\!\!\epsilon_{I}$ even though they can be obtained by complex conjugation from the other 2 rules. Furthermore, we introduced the abbreviation $$\label{eq:Vect1} \mathtt{W}^{i} \; =\; -\textstyle{ig\over 2}\ \bar{f}^{i\Lambda}\ \left[ \mathtt{P}_{\Lambda}\ +\ i\mathtt{C}_{\Lambda}\right] \hspace{.5cm},\hspace{.5cm} \overline{\mathtt{W}^{\bar{\imath}}} \; =\; \overline{\mathtt{W}^{i}} \; ,$$ and we used the standard $N=2$ $d=4$ sugra definitions [@Andrianopoli:1996cm] $$\label{eq:DefFStrength} \mathcal{T}^{+}\; \equiv\; 2i\mathcal{L}_{\Lambda}\ F^{\Lambda +} \;\;\; ,\;\;\; G^{i+} \; \equiv\; -\bar{f}^{i}_{\Lambda}\ F^{\Lambda\ +} \; .$$ The integrability conditions for the above system of equations can easily be calculated and give rise to $$\label{eq:VectInt1} \mathcal{B}_{ab}\ \gamma^{b}\epsilon_{I} \; =\; -2i\ \mathcal{L}^{\Lambda} \left[ \slashed{\mathcal{B}}_{\Lambda} \ -\ \mathcal{N}_{\Lambda\Sigma}\slashed{\mathcal{B}}^{\Sigma} \right]\ \varepsilon_{IJ}\gamma_{a}\epsilon^{J}\; ,$$ where we defined not only the Bianchi identity as $\star \mathcal{B}^{\Lambda} = \mathtt{D}F^{\Lambda}(=0)$ but also $$\begin{aligned} \label{eq:VectEOM1} \mathcal{B}_{ab} & =& R_{ab} \ +\ 2\mathcal{G}_{i\bar{\jmath}}\mathtt{D}_{(a}Z^{i}\mathtt{D}_{b)}\overline{Z}^{\bar{\jmath}} \ +\ 4\mathrm{Im}\left(\mathcal{N}\right)_{\Lambda\Sigma}\left[ F^{\Lambda}_{ac}F^{\Sigma}_{b}{}^{c} -\textstyle{1\over 4}\eta_{ab}F^{\Lambda}_{cd}F^{\Sigma cd} \right] \ -\ \textstyle{1\over 2}\eta_{ab}\ \mathtt{V} \; ,\\ & & \nonumber \\ \label{eq:VectEOM2} \star \mathcal{B}_{\Lambda} & =& \mathtt{D}\left[\ \mathcal{N}_{\Lambda\Sigma}\ F^{\Sigma -} + \overline{\mathcal{N}}_{\Lambda\Sigma}\ F^{\Sigma +} \right] - \textstyle{g\over 2}\mathrm{Re}\left( \mathtt{K}_{\Lambda \bar{\imath}} \star\mathtt{D}\overline{Z}^{\bar{\imath}} \right) \equiv \mathtt{D}F_{\Lambda} - \textstyle{g\over 2}\mathrm{Re}\left( \mathtt{K}_{\Lambda \bar{\imath}} \star\mathtt{D}\overline{Z}^{\bar{\imath}} \right)\; ,\\ & & \nonumber \\ \label{eq:VectPot} \mathtt{V} & =& \textstyle{g^{2}\over 2}\left[ 3\mathtt{C}_{\Lambda}\mathtt{C}_{\Sigma}\mathcal{L}^{\Lambda}\overline{\mathcal{L}}^{\Sigma} \, +\, f_{i}^{\Lambda}\ \bar{f}^{i\Sigma} \left(\mathtt{P}+i\mathtt{C}\right)_{\Lambda} \left(\mathtt{P}+i\mathtt{C}\right)_{\Sigma} \right] \; .\end{aligned}$$ The potential that follows from the integrability condition is not real, and imposing it to be real implies that we must satisfy the constraint $$\label{eq:GaugeConstr} 0\; =\; \mathrm{Im}\left(\mathcal{N}\right)^{-1|\Lambda\Sigma}\ \mathtt{P}_{\Lambda}\ \mathtt{C}_{\Sigma} \; ,$$ which is a gauge-invariant statement. For our choice of possible non-Abelian gaugings, this constraint is satisfied identically: by contracting the last equation in eq. (\[eq:SGK17\]) with $f_{i}^{\Sigma}$ and using identities (\[eq:SGImpId\]) and (\[eq:SGK10\]) one can obtain the identity $$\label{eq:GaugeConstr2} \mathrm{Im}\left(\mathcal{N}\right)^{-1|\Lambda\Sigma}\ \mathtt{P}_{\Sigma} \; =\; 4i\ \mathcal{L}^{\Sigma}\overline{\mathcal{L}}^{\Omega}\ f_{\Sigma\Omega}{}^{\Lambda}\; ,$$ which upon contracting with $\mathtt{C}_{\Lambda}$ and using its $G$-invariance gives the desired result. Therefore the potential $\mathtt{V}$ reads $$\begin{aligned} \label{eq:Potential} \mathtt{V}& =& \textstyle{g^{2}\over 2}\left[ 3\left|\mathtt{C}_{\Lambda}\mathcal{L}^{\Lambda}\right|^{2} \; +\; f_{i}^{\Lambda}\bar{f}^{i\Sigma}\left(\ \mathtt{P}_{\Lambda}\mathtt{P}_{\Sigma}\ -\ \mathtt{C}_{\Lambda}\mathtt{C}_{\Sigma} \right)\ \right]\; , \\ & & \nonumber \\ \label{eq:Potential1} & =& \textstyle{g^{2}\over 2}\left[ 4\left|\mathtt{C}_{\Lambda}\mathcal{L}^{\Lambda}\right|^{2} \; +\; \textstyle{1\over 2}\mathrm{Im}\left(\mathcal{N}\right)^{-1|\Lambda\Sigma} \left(\ \mathtt{C}_{\Lambda}\mathtt{C}_{\Sigma}\ -\ \mathtt{P}_{\Lambda}\mathtt{P}_{\Sigma} \right)\ \right]\; ,\end{aligned}$$ which is similar to the supersymmetric result in [@Andrianopoli:1996cm], upon Wick rotating the Fayet-Iliopoulos term. Likewise, the above equations of motion can then be obtained from the action $$\label{eq:VectAct} \int_{4}\sqrt{g}\left[ R + 2\mathcal{G}_{i\bar{\jmath}}\mathtt{D}_{a}Z^{i}\mathtt{D}^{a}\overline{Z}^{\bar{\jmath}} + 2\mathrm{Im}\left(\mathcal{N}\right)_{\Lambda\Sigma}F^{\Lambda}_{ab}F^{\Sigma ab} - 2\mathrm{Re}\left(\mathcal{N}\right)_{\Lambda\Sigma}F^{\Lambda}_{ab}\star F^{\Sigma ab} - \mathtt{V} \right] \; ,$$ which as stated in the introduction has correctly normalised kinetic terms. In sugra the integrability condition for the scalars relates the scalar e.o.m. with the Maxwell e.o.m.s, and the same happens here: a straightforward calculation results in $$\label{eq:VectInt2} \mathcal{B}^{i}\epsilon_{I} \; =\; -2i\ \bar{f}^{i\Lambda}\left[ \slashed{\mathcal{B}}_{\Lambda} \ -\ \mathcal{N}_{\Lambda\Sigma}\slashed{\mathcal{B}}^{\Sigma} \right]\ \varepsilon_{IJ}\epsilon^{J} \; ,$$ where we have introduced the equation of motion for the scalars $Z^{i}$ as $$\label{eq:VectEOM3} \mathcal{B}^{i} \, =\, \Box Z^{i} -i\partial^{i}\overline{\mathcal{N}}_{\Lambda\Sigma}F^{\Lambda +}_{ab}F^{\Sigma +\ ab} +i\partial^{i}\mathcal{N}_{\Lambda\Sigma}F^{\Lambda -}_{ab}F^{\Sigma -\ ab} +\textstyle{1\over 2}\partial^{i}\mathtt{V} \; .$$ In conclusion, the integrability conditions for the equations (\[eq:20\]–\[eq:20d\]) give relations between the equations of motion, which, forgetting about the changes in the form of the $\mathcal{B}$-tensors, are exactly the same as found in supersymmetry, which is hardly surprising. The implication of the relations (\[eq:VectInt1\]) and (\[eq:VectInt2\]) is then also the same [@Gauntlett:2002nw; @Bellorin:2005hy], namely that the independent number of equations of motion one has to check in order to be sure that a given solution to eqs. (\[eq:20\]–\[eq:20d\]) is also a solution to the equations of motion is greatly reduced.[^4] The minimal set of equations of motion one has to check depends on the norm of the vector bilinear $V_{a}=i\overline{\epsilon}^{I}\gamma_{a}\epsilon_{I}$: if the norm $V_{a}V^{a}$ is positive, referred to as the time-like case, we only need to solve the time-like direction of the Bianchi identity, [*i.e.*]{} $\imath_{V}\star\mathcal{B}^{\Lambda}=0$ and the Maxwell/YM equations, [*i.e.*]{} $\imath_{V}\star\mathcal{B}_{\Lambda}=0$. This case will be considered in section (\[sec:VectBil\]). If the norm of the bilinear is null, [*i.e.*]{} $V_{a}V^{a}=0$, then a convenient set of e.o.m.s is given by $N^{a}N^{b}\mathcal{B}_{ab}=0$, $N^{a}\mathcal{B}_{\Lambda a}=0$ and $N^{a}\mathcal{B}_{a}^{\Lambda}$, where $N$ is a vector normalised by $V^{a}N_{a}=1$: this case will be considered in section (\[sec:Null\]). Analysis of the Time-like case {#sec:VectBil} ============================== In this section we shall consider the time-like case and the strategy to be followed is the usual one: we analyse the differential constraints on the bilinears constructed out of the spinors $\epsilon_{I}$ defined in appendix (\[sec:Bil\]) coming from the fKSEs (\[eq:20\]–\[eq:20d\]), trying to solve these constraints as general as possible in as little unknowns as possible. After the constraints have been solved, we shall, following the comments made above, impose the Bianchi identity and the gauge-field equations of motion and to see what conditions they impose. After these steps we will be left with a minimal set of functions, structures and conditions they have to satisfy in order to construct fake-supersymmetric solutions: for the solutions to be constructed in this case, the algorithm will be outlined in section (\[sec:CosmMon\]). Let us start by discussing the differential constraints on the bilinears: using eq. (\[eq:20\]) and the definitions of the bilinears in appendix (\[sec:Bil\]), we can calculate $$\begin{aligned} \label{eq:VId1} \mathbb{D}X & =& \textstyle{g\over 4}\ \mathtt{C}_{\Lambda}\mathcal{L}^{\Lambda}\ V \; +\; i\ \imath_{V}\mathcal{T}^{+}\; ,\\ & & \nonumber\\ \label{eq:VId2} \mathbb{D}_{a}V_{b} & =& g|X|^{2}\ \mathtt{C}_{\Lambda}\mathcal{R}^{\Lambda}\ \eta_{ab} \ +\ 4\mathrm{Im}\left( \overline{X}\ \mathcal{T}^{+}_{ab} \right)\; ,\\ & & \nonumber \\ \label{eq:VId3} \mathbb{D}V^{x} & =& \textstyle{g\over 2}\mathtt{C}_{\Lambda}\mathcal{R}^{\Lambda}\ V\wedge V^{x} \ +\ \textstyle{g\over 2}\mathtt{C}_{\Lambda}\mathcal{I}^{\Lambda}\ \star\left[ V\wedge V^{x}\right] \; ,\end{aligned}$$ where following ref. [@Meessen:2006tu] we have introduced the real symplectic sections of Kähler weight zero, $$\label{eq:DefRandI} \mathcal{R} \; =\; \mathrm{Re}\left(\mathcal{V}/X\right) \;\;\; ,\;\;\; \mathcal{I} \; =\; \mathrm{Im}\left(\mathcal{V}/X\right) \;\longrightarrow\;\; \frac{1}{2|X|^{2}}\ =\ \langle\mathcal{R}|\mathcal{I}\rangle \; .$$ In the ungauged theory, as will also be the case here, the $2\bar{n}$ real functions $\mathcal{I}$ play a fundamental rôle in the construction of BPS solutions and the $2\bar{n}$ real functions $\mathcal{R}$ depend on $\mathcal{I}$: finding, given a Special Geometric model, the explicit $\mathcal{I}$-dependence of $\mathcal{R}$ is known as the [*stabilisation equation*]{}, and for many models solutions to it are known. A first difference with supersymmetric case lies in the character of the bilinear $V$: in that case it is always a Killing vector, which as one can see from eq. (\[eq:VId2\]) will not be the case here. We can still use it to introduce a time-like coordinate $\tau$ by choosing an adapted coordinate system through $V^{a}\partial_{a}=\sqrt{2}\partial_{\tau}$, but now the components of the metric will depend explicitly on $\tau$, as was to be expected from for instance the Kastor [&]{} Traschen solution [@Kastor:1993mj]. As the $V^{x}$ contain the information about the metric on the base-space, it is important to deduce its behaviour under translations along $V$; in order to investigate we calculate $$\label{eq:LieV} \pounds_{V}V^{x} \; =\; \imath_{V}dV^{x} +d\left(\imath_{V}V^{x}\right) \; =\; g \mathtt{C}_{\Lambda}\imath_{V}A^{\Lambda}\ V^{x} \ +\ 2g|X|^{2}\mathtt{C}_{\Lambda}\mathcal{R}^{\Lambda}\ V^{x} \; .$$ This implies that by choosing the gauge-fixing $$\label{eq:VId4} \imath_{V}A^{\Lambda} \; =\; -2|X|^{2}\ \mathcal{R}^{\Lambda} \; ,$$ we find that $\pounds_{V}V^{x}=0$. We would like to point out that the above gauge-fixing is the actual result one obtains when considering time-like supersymmetric solutions in $N=2$ $d=4$ supergravity theories [@Meessen:2006tu; @Hubscher:2008yz]. The above result has some nice implications, the first of which is derived by contracting eq. (\[eq:VId2\]) with $V^{a}V^{b}$, namely $$\label{eq:Res1} \langle\nabla_{V}\mathcal{R}|\mathcal{I}\rangle + \langle\mathcal{R}|\nabla_{V}\mathcal{I}\rangle \; =\; \nabla_{V}\frac{1}{2|X|^{2}} \; =\; g\mathtt{C}_{\Lambda}\ \mathcal{R}^{\Lambda} \; .$$ We can rewrite the above equation to a nicer form by observing that $$\begin{aligned} \langle \mathcal{V}/X| d\left(\mathcal{V}/X\right)\rangle & =& X^{-2}\ \langle\mathcal{V}|\mathfrak{D}\mathcal{V}\rangle \ -\ X^{-3}\mathfrak{D}X\; \langle\mathcal{V}|\mathcal{V}\rangle \; =\; 0\nonumber \\ & =& \langle\mathcal{R}|d\mathcal{R}\rangle \ -\ \langle\mathcal{I}|d\mathcal{I}\rangle \, +\, i\langle\mathcal{R}|d\mathcal{I}\rangle \, +\, i\langle\mathcal{I}|d\mathcal{R}\rangle \; ,\end{aligned}$$ which seeing the reality properties of the above expression implies[^5] $$\begin{aligned} \label{eq:StabId1} \langle\ d\mathcal{R}\ |\ \mathcal{I}\ \rangle & =& \langle\ \mathcal{R}\ |\ d\mathcal{I}\ \rangle \; ,\\ \label{eq:StabId2} \langle\ \mathcal{R}\ |\ d\mathcal{R}\ \rangle & =& \langle\ \mathcal{I}\ |\ d\mathcal{I}\ \rangle \; .\end{aligned}$$ If we then introduce the real symplectic section $\mathtt{C}^{T}=(0,\mathtt{C}_{\Lambda})$, we can rewrite eq. (\[eq:Res1\]) in the simple and suggestive form $$\label{eq:VId5a} 0\; =\; \langle\ \mathcal{R}\ |\ \nabla_{V}\mathcal{I}\ +\ \textstyle{g\over 2}\mathtt{C}\ \rangle \; .$$ The above equation could also have been obtained from the contraction of eq. (\[eq:VId1\]) with $V$, [*i.e.*]{} $$\label{eq:VId5b} \textstyle{\frac{1}{\overline{X}}}\ \mathtt{D}_{V}\textstyle{\frac{1}{X}} \; =\; -g\langle\mathcal{R}|\mathtt{C}\rangle \, +\, ig\langle\mathcal{I}|\mathtt{C}\rangle \; ,$$ and taking its real part. By taking the imaginary part and using the identity $$\mathrm{Im}\left(\ \textstyle{\frac{1}{\overline{X}}}\ \mathtt{D}\textstyle{\frac{1}{X}}\ \right) \, =\; -2\langle\mathcal{I}|\mathtt{D}\mathcal{I}\rangle \; ,$$ we find that apart form eq. (\[eq:VId5a\]), we also must have $$\label{eq:VId5c} 0\; =\; \langle\ \mathcal{I}\ |\ \nabla_{V}\mathcal{I}\ +\ \textstyle{g\over 2}\mathtt{C}\ \rangle \; .$$ By now, there are strong hints that the derivative of the symplectic section $\mathcal{I}$ in the direction $V$ should be constant and, in fact, the information needed to close the case is hidden in eqs. (\[eq:20c\]) and (\[eq:20d\]). From the contraction of (\[eq:20c\]) with $\bar{\epsilon}^{K}\gamma_{a}\varepsilon_{KI}$ we find $$\label{eq:VId6} 2\overline{X}\ \mathtt{D}Z^{i} \; =\; 4\ \imath_{V}G^{i+} \ -\ \mathtt{W}^{i}\ V\; ,$$ which upon contraction with $V$ leads to $$\label{eq:VId7} \mathtt{D}_{V}Z^{i} \; =\; -2\ X\ \mathtt{W}^{i} \; .$$ Using the gauge-fixing (\[eq:VId4\]), the identity $\bar{f}^{\Lambda i}\mathtt{P}_{\Lambda}=i\overline{\mathcal{L}}^{\Lambda}\mathtt{K}_{\Lambda}^{i}$ and the fact that for our choice of possible non-Abelian gauge groups we have $\mathcal{L}^{\Lambda}\mathtt{K}_{\Lambda}^{i}=0$, we see that the above equation is converted to $$\label{eq:VId7a} \nabla_{V}Z^{i} \; =\; -g\ X\ \bar{f}^{\Lambda i}\ \mathtt{C}_{\Lambda} \; .$$ Using then the special geometry identity $\langle\mathcal{U}_{i}|\overline{\mathcal{U}}_{\bar{\jmath}}\rangle = i\mathcal{G}_{i\bar{\jmath}}$, we can rewrite the above equation to $$\langle \nabla_{V}\mathcal{I}+g\mathtt{C}\, |\ \overline{\mathcal{U}}_{\bar{\jmath}}\rangle \; =\; i\langle\ \nabla_{V}\mathcal{R}|\overline{\mathcal{U}}_{\bar{\jmath}}\rangle \; ,$$ which can be manipulated by using the special geometry properties and a renewed call to eq. (\[eq:VId7\]) to give $$\label{eq:Vad6} \langle\ \nabla_{V}\mathcal{I}\ +\ \textstyle{g\over 2}\mathtt{C}\, |\ \overline{\mathcal{U}}_{\bar{\jmath}}\rangle \; =\; 0 \; .$$ The above equation plus eqs. (\[eq:VId5a\]) and (\[eq:VId5c\]) together with the completeness relation from special geometry, eq. (\[eq:SGSymplProj\]), then implies $$\label{eq:LaPolla} \nabla_{V}\mathcal{I} \; =\; -\textstyle{g\over 2}\ \mathtt{C} \; ,$$ which implies that the $\tau$-dependence of the functions $\mathcal{I}$ is at most linear, and in fact only half of them, namely the $\mathcal{I}_{\Lambda}$. At this point it is necessary to introduce a complete coordinate system ($\tau, y^{m}$), which we will take to be adapted to $V$ and compatible with the Fierz identities in appendix (\[sec:Bil\]), [*i.e.*]{} $$\label{eq:Coord} \begin{array}{lclclcl} V^{a}\partial_{a} & =& \sqrt{2}\partial_{\tau} &\hspace{.3cm},\hspace{.3cm}& V & =& 2\sqrt{2}|X|^{2}\left( d\tau \ +\ \omega\right) \\ & & & & & & \\ V^{xa}\partial_{a} & =& -2\sqrt{2}|X|^{2}V^{xm}\ \left(\partial_{m}-\omega_{m}\partial_{\tau}\right) & ,& V^{x} & =& \sqrt{2}\ V^{x}_{m}\ dy^{m} \; , \end{array}$$ where $\omega =\omega_{m}dy^{m}$ is a possibly $\tau$-dependent 1-form and we introduced $V^{xm}$ by $V^{xm}V^{y}_{m}=\delta^{xy}$; as the $V^{x}_{m}$ act as a Dreibein on a Riemannian space, the $x$-indices can be raised and lowered with $\delta^{xy}$, so that we won’t distinguish between co- and contravariant $x$-indices. Putting the Vierbein together with the Fierz identity (\[eq:Bil4\]) we find that the metric is takes on the conforma-stationary form $$\label{eq:37} ds^{2} \; =\; 2|X|^{2}\left( d\tau \ +\ \omega\right)^{2} \ -\ \textstyle{1\over 2|X|^{2}}\ h_{mn} dy^{m}dy^{n} \; ,$$ where $h_{mn}=V^{x}_{m}V^{x}_{n}$ is the metric on the 3-dimensional base-space. W.r.t. our choice of coordinates we have that $\pounds_{V}V^{x}=0$ equals $\partial_{\tau}V^{x}_{m}=0$; the $V^{x}$ are of course also constrained by eq. (\[eq:VId3\]), which in the chosen coordinate system and using the decomposition $$\label{eq:Dec1} A^{\Lambda} \; =\; -\textstyle{1\over 2}\mathcal{R}^{\Lambda}\ V \ +\ \tilde{A}^{\Lambda}_{m}\ dy^{m} \ \equiv\ -\textstyle{1\over 2}\mathcal{R}^{\Lambda}\ V \ +\ \tilde{A}^{\Lambda} \;\ \longrightarrow\;\ F^{\Lambda} \ =\ -\textstyle{1\over 2}\mathtt{D}\left[\mathcal{R}^{\Lambda}V\right] \ +\ \tilde{F}^{\Lambda} \; ,$$ reads $$\label{eq:ResVx} dV^{x} \; =\; g\mathtt{C}_{\Lambda}\ \tilde{A}^{\Lambda} \wedge V^{x} \ +\ \textstyle{g\over 4}\ \mathtt{C}_{\Lambda}\mathcal{I}^{\Lambda}\ \varepsilon^{xyz}\ V^{y}\wedge V^{z}\; .$$ A first remark to be made is that for consistency we must have $\mathtt{C}_{\Lambda}\partial_{\tau}\tilde{A}^{\Lambda}_{x}=0$. Further, we could use the residual gauge freedom $\mathtt{C}_{\Lambda}\tilde{A}^{\Lambda}\rightarrow \mathtt{C}_{\Lambda}\tilde{A}^{\Lambda} + d\phi (y)$, $V^{x}\rightarrow e^{g\phi}V^{x}$ to take $\mathtt{C}_{\Lambda}\mathcal{I}^{\Lambda}$ to be constant, a possibility we will not use. And lastly, the integrability condition $d^{2}V^{x}=0$ implies $$\label{eq:ResVxInt} 0\; =\; \textstyle{g\over 4} \left[\ \varepsilon^{xyz}\ \mathtt{C}_{\Lambda}\tilde{F}^{\Lambda}_{yz} \ +\ \sqrt{2}\ V_{x}^{m} \tilde{\mathbb{D}}_{m}\ \mathtt{C}_{\Lambda}\mathcal{I}^{\Lambda} \right] \; ,$$ where we have introduced $\tilde{F}^{\Lambda}_{xy}\equiv V_{x}^{m}V_{y}^{n}\tilde{F}^{\Lambda}_{mn}$ and $$\label{eq:DefTildeD} \tilde{\mathbb{D}}_{m}\mathcal{I} \; =\; \partial_{m} \mathcal{I} \ +\ g\mathtt{C}_{\Lambda}\tilde{A}^{\Lambda}_{m}\ \mathcal{I} \ +\ g\tilde{A}^{\Lambda}_{m}\ S_{\Lambda}\mathcal{I} \;\;\; ;\;\; \tilde{\mathbb{D}}_{x} \ \equiv\ V_{x}^{m}\ \tilde{\mathbb{D}}_{m} \; .$$ The system (\[eq:ResVx\]) was analysed by Gauduchon [&]{} Tod in ref. [@Gauduchon:1998], as it appeared in the discussion of 4-dimensional hyper-hermitian Riemannian metrics admitting a tri-holomorphic Killing vector. A first implication is that the geometry of the base-space belongs to a subclass of 3-dimensional Einstein-Weyl spaces, called hyper-CR or Gauduchon-Tod spaces: one of the extra constraints to be imposed on the EW-spaces is nothing more than the integrability condition (\[eq:ResVxInt\]) which is called the [*generalised Abelian monopole equation*]{}. As we will see later on, and can be expected from the similar discussion in [@Hubscher:2008yz], the equations determining the seed function $\mathcal{I}^{\Lambda}$, will be generalised non-Abelian monopole equation or, said differently, the straightforward generalisation of the standard Bogomol’nyi equation on $\mathbb{R}^{3}$ to GT-spaces; eq. (\[eq:ResVxInt\]) is of course implied by these upon contraction with $\mathtt{C}_{\Lambda}$. In ref. [@Behrndt:2003cx], Behrndt and Cvetič realised that their 5-dimensional cosmological solutions could be dimensionally reduced to 4-dimensional ones, which raises the question of what solution found by Grover [*et al.*]{} [@Grover:2008jr] can be reduced to solutions we are going ot find. As in this case we are dealing with a map between the 5-dimensional time-like case and the 4-dimensional time-like case, the dimensional reduction has to be over the 4-dimensional base-space, which was found to be hyperKähler-torsion [@Grover:2008jr]. The key to identifying the subclass of 5-dimensional solutions that can be reduced to ours, then lies in a further result of Gauduchon [&]{} Tod (see remark 2 in ref. [@Gauduchon:1998]), which states that the solutions to eqs. (\[eq:ResVx\]) and (\[eq:ResVxInt\]) are obtained by the reduction of a conformal hyper-Kähler space along a tri-holomorphic Killing vector. In fact, as is nicely discussed in [@Grover:2008jr sec. (3.2)], these spaces are particular instances of HKT-spaces. This inheritance of geometrical structures also ocurrs in ordinary supergravity theories in $6$, $5$ and $4$ dimensions and it is reasonable to suppose that this also holds for fake/Wick-rotated supergravities. As a final comment, let us mention that the 3-dimensional Killing spinor equation on a GT-manifold allows non-trivial solutions [@Buchholz:2000]. Before turning to the equations of motion, we deduce the following equation for $\omega$ from the anti-symmetrised version of eq. (\[eq:VId2\]) and the explicit coordinate expression in (\[eq:Coord\]). As the reader will observe, this calculation needs the explicit form for the 2-form $\mathcal{T}^{+}$, which can be obtained from eq. (\[eq:VId1\]) and the rule that an general imaginary self-dual 2-form $B^{+}$ is determined by its contraction with $V$ by means of (See refs. [@Caldarelli:2003pb] for more detail) $$\label{eq:Tplus} B^{+} \; =\; \frac{1}{4|X|^{2}}\left(\ V\wedge\ \imath_{V}B^{+} \ +\ i\star\left[\ V\wedge\ \imath_{V}B^{+} \right] \ \right)\; .$$ The result reads $$\label{eq:Rot} d\omega \ +\ g\mathtt{C}_{\Lambda}\tilde{A}^{\Lambda}\wedge (d\tau +\omega ) \; =\; \sqrt{2}\ \star\left[ V\wedge\ \langle\mathcal{I}|\ \mathtt{D}\mathcal{I}\rangle\right] \; .$$ Contracting the above equation with $V$ we find that $$\label{eq:OmegaV} \pounds_{V}\omega \; =\; g\sqrt{2}\mathtt{C}_{\Lambda}\tilde{A}^{\Lambda} \hspace{.5cm}\longrightarrow\hspace{.5cm} \omega \; =\; g\mathtt{C}_{\Lambda}\tilde{A}^{\Lambda}\ \tau \; +\; \varpi \; ,$$ where $\varpi = \varpi_{m}dy^{m}$ is $\tau$-independent. Substituting the above result into eq. (\[eq:Rot\]) and evaluating its r.h.s., we obtain $$\label{eq:Rot2} d\varpi \ +\ g\mathtt{C}_{\Lambda}\tilde{A}^{\Lambda}\wedge \varpi +g\mathtt{C}_{\Lambda}\tilde{F}^{\Lambda}\ \tau\; =\; \textstyle{1\over 2}\ \langle\mathcal{I}|\ \tilde{\mathtt{D}}_{m}\mathcal{I}\ -\ \omega_{m}\partial_{\tau}\mathcal{I}\rangle \ V^{xm}\varepsilon^{xyz}\ V^{y}\wedge V^{z}\; .$$ There is a possible inconsistency in this equation due to the possible $\tau$-dependence in the above equation; as the equation is at most linear in $\tau$, we can investigate the possible inconsistency by taking the $\tau$-derivative, only to find eq. (\[eq:ResVxInt\]). The equation determining $\varpi$ is then found by splitting off the $\tau$-dependent part and reads $$\label{eq:Rot3} \tilde{\mathbb{D}}\ \varpi \; =\; \textstyle{1\over 2}\ \varepsilon^{xyz}\ \langle \tilde{\mathcal{I}}\ |\ \tilde{\mathtt{D}}_{x}\tilde{\mathcal{I}} - \varpi_{x}\partial_{\tau}\mathcal{I}\rangle\ V^{y}\wedge V^{z} \; ,$$ where we introduced $\tilde{\mathcal{I}}=\mathcal{I}(\tau =0)$. The symplectic field strength $F^{T}=(F^{\Lambda},F_{\Lambda})$ then easily be deduced to give the standard supersymmetric result $$\begin{aligned} \label{eq:Fsympl} F & =& -\textstyle{1\over 2}\ \mathtt{D}\left( \mathcal{R}\ V\right) \; -\; \textstyle{1\over 2}\ \star\left[ V\wedge\ \mathbb{D}\mathcal{I} \right] \nonumber \\ & =& -\textstyle{1\over 2}\ \mathtt{D}\left( \mathcal{R}\ V\right) \; -\; \frac{\sqrt{2}}{8}\ \varepsilon^{xyz}\ V_{x}^{m}\left[\ \tilde{\mathbb{D}}_{m}\mathcal{I}\ -\ \omega_{m}\partial_{\tau}\mathcal{I} \ \right]\ V^{y}\wedge V^{z} \; ,\end{aligned}$$ which agrees completely with the imposed gauge-fixing (\[eq:VId4\]). At this point we would like to treat the Bianchi identity $\mathtt{D}F^{\Lambda}=0$, as it was treated in ref. [@Hubscher:2008yz], namely as leading to a Bogomol’nyi equation determining the pair $(\tilde{A}^{\Lambda},\mathcal{I}^{\Lambda})$; this approach boils down to stating that since we are given the potential in eq. (\[eq:Dec1\]), the Bianchi identity is solved identically. This does, however, not mean that any given $\tilde{A}^{\Lambda}$ leads to a field strength with the form prescribed by fake-supersymmetry in eq. (\[eq:Fsympl\]). If we then impose that a given $\tilde{A}^{\Lambda}$ leads to a field-strength with the prescribed form implies imposing the equation $$\label{eq:Bogo} \tilde{F}^{\Lambda}_{xy} \; =\; -\frac{1}{\sqrt{2}}\ \varepsilon^{xyz}\ \tilde{\mathbb{D}}_{z}\mathcal{I}^{\Lambda} \; ,$$ which due to eq. (\[eq:LaPolla\]) is manifestly $\tau$-independent. This equation is the generalisation of the standard Bogomol’nyi equation on $\mathbb{R}^{3}$ to a 3-dimensional Gauduchon-Tod space. Clearly, the above equation implies the constraint (\[eq:ResVxInt\]) upon contraction with $\mathtt{C}_{\Lambda}$. In order to show that the time-like solutions to the fKSEs we characterised are indeed solutions to the equations of motion, we need to impose the Maxwell-Yang-Mills equations of motion, [*i.e.*]{} eq. (\[eq:VectEOM2\]). This equation consists of 2 parts, namely one in the time-direction, [*e.g.*]{} $\mathcal{B}_{\Lambda}^{t}$, and one in the space-like directions, $\mathcal{B}_{\Lambda}^{x}$. A tedious but straightforward calculation shows that $\mathcal{B}_{\Lambda}^{t}=0$ identically, in full concordance with the discussion in section (\[sec:DSVector\]); the equations of motion in the $x$-direction, however, do not vanish identically. In stead, they impose the condition $$\label{eq:MYM1} \left(\tilde{\mathbb{D}}_{x}\ -\ \omega_{x}\partial_{\tau}\right)^{2}\mathcal{I}_{\Lambda} \; =\; \textstyle{g^{2}\over 2}\ f_{\Lambda (\Omega}{}^{\Gamma}f_{\Delta )\Gamma}{}^{\Sigma} \mathcal{I}^{\Omega}\mathcal{I}^{\Delta}\ \mathcal{I}_{\Sigma} \ -\ \textstyle{g^{2}\over 2}\ f_{\Lambda\Omega}{}^{\Sigma}\mathcal{I}^{\Omega}\mathcal{I}_{\Sigma}\ \mathtt{C}_{\Gamma}\mathcal{I}^{\Gamma}\; ,$$ which in the limit $\mathtt{C}\rightarrow 0$ coincides with the result obtained in ref. [@Hubscher:2008yz]. A simplification of the above equation can be obtained by observing that, due to eqs. (\[eq:DefTildeD\]) and (\[eq:OmegaV\]), $$\partial_{\tau}\left(\tilde{\mathbb{D}}_{m}\mathcal{I}_{\Lambda}\ -\ \omega_{m}\partial_{\tau}\mathcal{I}_{\Lambda}\right) \; =\; \partial_{\tau}\partial_{m}\mathcal{I}_{\Lambda} \; =\; 0\; .$$ Using the above identity and using the fact that $\mathcal{I}_{\Lambda}$ is linear in $\tau$, we can rewrite eq. (\[eq:MYM1\]) as $$\label{eq:MYM2} \tilde{\mathbb{D}}_{x}^{2}\ \tilde{\mathcal{I}}_{\Lambda} \ -\ \left(\tilde{\mathbb{D}}_{x}\varpi_{x}\right)\ \partial_{\tau}\mathcal{I}_{\Lambda} \ =\ \textstyle{g^{2}\over 2}\ f_{\Lambda (\Omega}{}^{\Gamma}f_{\Delta )\Gamma}{}^{\Sigma} \mathcal{I}^{\Omega}\mathcal{I}^{\Delta}\ \tilde{\mathcal{I}}_{\Sigma} \ -\ \textstyle{g^{2}\over 2}\ f_{\Lambda\Omega}{}^{\Sigma}\mathcal{I}^{\Omega}\tilde{\mathcal{I}}_{\Sigma}\ \mathtt{C}_{\Gamma}\mathcal{I}^{\Gamma}\; ,$$ which is a $\tau$-independent equation! Recapitulation and some comments {#sec:CosmMon} -------------------------------- Let us, before making some comments on the generic behaviour of the solutions, spell out the way how to construct solutions using the results obtained in the foregoing section: the first step is to decide which model to consider, [*i.e.*]{} one has to specify what special geometric manifold is to be used, what non-Abelian groups can and will be gauged, and furthermore the constants $\mathtt{C}_{\Lambda}$. Given the model, we must then decide what 3-dimensional hyperCR/GT space we are going to use to describe the geometry of the 3-dimensional base-space; this is equivalent to finding the triple $(V^{x} , \mathtt{C}_{\Lambda}\tilde{A}^{\Lambda},\mathtt{C}_{\Lambda}\mathcal{I}^{\Lambda})$ solving eq. (\[eq:ResVx\]). This decision, then, allows us in principle to solve the Bogomol’nyi equation (\[eq:Bogo\]) as to determine $(\tilde{A}^{\Lambda},\mathcal{I}^{\Lambda})$. The next step would be to determine the $\tau$-independent part of the seed functions $\mathcal{I}_{\Lambda}$, remember that their $\tau$-dependence is fixed by eq. (\[eq:LaPolla\]), using equation (\[eq:MYM2\]). As this equation contains not only the $\mathcal{I}_{\Lambda}$ but also $\varpi$, we are forced to determine both objects and make sure that eq. (\[eq:Rot3\]) is satisfied. Having gone through the above steps, all that needs to be done is to determine the field-strengths by means of eq. (\[eq:Fsympl\]), write down the physical scalars $Z^{i}=\mathcal{L}^{i}/\mathcal{L}^{0}$ and the metric by determining the stationarity 1-form $\omega$ by eq. (\[eq:OmegaV\]) and the metrical factor $|X|^{2}$ through eq. (\[eq:DefRandI\]). As usual the explicit construction of the last fields goes through the solution of the [*stabilisation equation*]{} which determines the symplectic section $\mathcal{R}$ in terms of the seed functions $\mathcal{I}$; for many models, solutions to the stabilisation equations are known. As mentioned in the previous section, the 3-dimensional Einstein-Weyl spaces that occur as the geometry of the base-space, can be obtained by reduction of hyper-Kähler spaces along a tri-holomorphic conformal Killing vector (see ref. [@Grover:2008jr sec. (3.2)] for detailed information), which would put us in a position to discuss the solutions to the Bogomol’nyi equation (\[eq:Bogo\]). However, knowing only explicit solutions to the non-Abelian Bogomol’nyi equation on $\mathbb{R}^{3}$,[^6] means that for the moment the only non-trivial non-Abelian solutions we can build are the ones that follow from the supersymmetric ones satisfying $\mathtt{C}_{\Lambda}\mathcal{I}^{\Lambda}=0$, which implies that $\mathtt{C}_{\Lambda}\tilde{A}^{\Lambda}$ is gauge trivial so that the base-space is $\mathbb{R}^{3}$, by substituting $\mathcal{I}_{\Lambda}\rightarrow \mathcal{I}_{\Lambda}-g\mathtt{C}_{\Lambda}\tau /(2\sqrt{2})$. As the base-space is $\mathbb{R}^{3}$, the equations determining the $\tau$-independent part of the $\mathcal{I}$, eqs. (\[eq:Bogo\]) and (\[eq:MYM2\]), reduce to the ones for $N=2$ EYM deduced in ref. [@Hubscher:2008yz]: indeed the only difference lies in the divergence of $\varpi$ occurring in eq. (\[eq:MYM2\]), and in the $\mathbb{R}^{3}$-case there is no obstruction to choosing it to vanish from the onset. At this point, then, the construction of fake-supersymmetric solutions boils down to the substitution principle put forward by Behrndt [&]{} Cvetič in ref. [@Behrndt:2003cx]: given a supersymmetric solution to $N=2$ $d=4$ EYM supergravity, Abelian [@Behrndt:1997ny] or non-Abelian [@Hubscher:2008yz; @Meessen:2008kb], substitute $\mathcal{I}_{\Lambda}\rightarrow \mathcal{I}_{\Lambda}-\mathtt{C}_{\Lambda}/(2\sqrt{2})\ \tau$ and impose the restriction $\mathtt{C}_{\Lambda}\mathcal{I}^{\Lambda}=0$. Of course, when dealing with non-Abelian gauge groups, not all choices for $\mathtt{C}_{\Lambda}$ are possible, as one must respect the constraint $f_{\Lambda\Sigma}{}^{\Gamma}\mathtt{C}_{\Gamma}=0$. The first observation is that generically the asymptotic form of the solution is not De Sitter but rather Kasner, [*i.e.*]{} the $\tau$-expansion of the base-space is power-like, making the definition of asymptotic mass even more cumbersome than in the De Sitter case.[^7] The second observation is that the metric has a curvature singularity at those events/points for which $|X|^{-2}=0$, which may be located outside our chosen coordinate system. This, however, raises the question of the possibility having an horizon, or said differently, how to decide in a practical manner when our solution describes a black hole. Observe that in the original Kastor [&]{} Traschen solution for one single black hole, this question is readily resolved by changing coordinates as to obtain the time-independent, spherically symmetric extreme RNDS black hole, for which the criteria to have an horizon are known: in the original coordinate system, the existence of a black hole can be expressed as the existence of a Killing horizon for a time-like Killing vector, covering the singularity. The last observation, then, is that in the general case no time-like Killing vector exists. To see this consider for instance the $\overline{\mathbb{CP}}^{1}$-model: this model has only one complex scalar field $Z$ living on the coset space $Sl(2;\mathbb{R})/SO(2)$ and associated Kähler potential $e^{\mathcal{K}}=1-|Z|^{2}$, so that we have the constraint $0\leq |Z|^{2}< 1$. Choosing $\mathtt{C}_{\Lambda}=(-2,0)$, the potential can be readily be calculated to be $$\label{eq:PotCPn} \mathtt{V} \; =\; 2g^{2}\ \left[ 1\ +\ 2\ e^{\mathcal{K}} \right]\; ,$$ which is manifestly positive. Imposing $\mathcal{I}^{0}=0$ in order to have $\mathbb{R}^{3}$ as the base space, and $\mathcal{I}_{1}=0$ in order to have a static solution ([*i.e.*]{} $\omega =0$), the equations of motion imply that a simple solution is $$\label{eq:PotCPn1} \mathcal{I}_{0} \; =\; \frac{g\tau}{\sqrt{2}} \hspace{.5cm}\mbox{and}\hspace{.5cm} \mathcal{I}^{1} \; =\; \sqrt{2}\ g\lambda \hspace{.5cm}\longrightarrow\hspace{.3cm} \frac{1}{2|X|^{2}} \ =\ g^{2}\left[ \tau^{2} \ -\ \lambda^{2}\right] \; ,$$ where $\lambda$ is a real constant. If $\lambda =0$ the above solution leads to $DS_{4}$, whereas if $\lambda\neq 0$ we can introduce a new coordinate $t$ through $\tau =\lambda\cosh\left( gt\right)$, such that the solution is given by $$\begin{aligned} \label{eq:PotCPn2} ds^{2} & =& dt^{2}\ -\ \sinh^{2}\left( gt\right)\ d\vec{x}^{2}_{(3)} \; ,\\ Z & =& -i\ \cosh^{-1}\left( gt\right) \; .\end{aligned}$$ At late times the metric is $DS_{4}$ but is singular when $t=0$; at that point in time also the scalar becomes problematic as $|Z(t=0)|^{2}=1$, violating the bound, which in its turn implies that the contribution of the scalars to the energy-momentum tensor blows up. Returning to the point we were going to make, it is paramount that in this case no time-like Killing vector exists. Had we on the other hand taken $\mathcal{I}^{1}= \sqrt{2}gp\ r^{-1}$, in which case a time-like Killing vector exists, the metric can be transformed to the static form $$\label{eq:PotCPn3} ds^{2} \ =\ \textstyle{p^{2} + R^{2}-g^{2}R^{4}\over R^{2}}\ dt^{2} \ -\ \textstyle{R^{4}\over ( R^{2}+p^{2})(p^{2} + R^{2}-g^{2}R^{4}/4)}\ dR^{2} \ -\ R^{2}dS^{2} \; .$$ This metric has one Killing horizon, identified with the cosmological horizon, for $R>0$ and is therefore a naked singularity, with the singularity being located at $R=0$. In the static coordinates, the scalar field reads $Z=-ip\ {\textstyle (p^{2}+R^{2})^{-1/2}}$, which explicitly breaks the bound $0\leq |Z|^{2}< 0$ at $R=0$, showing once again the link between the regularity of the metric and that of the scalars. A manageable prescription for deciding when a solution describes a black hole is clearly desirable. In this respect, we would like to mention the isolated horizon formalism (see [*e.g.*]{}[@Ashtekar:2004cn]) which attempt to give a local definition of horizons, without a reference to the existence of time-like Killing vectors. This formalism was recently applied to sugras in ref. [@Liko:2007mu] and similar work for fake sugras is in progress. Null case {#sec:Null} ========= In this section we shall characterise the fake-supersymmetric solutions in the so-called null-case, by which is meant the case when $V^{2}=0$: for simplicity we shall restrict ourselves to the theories with no YM-type couplings, a full analysis along the lines of ref. [@Hubscher:2008yz] being possible but, seeing the results obtained in that reference, not very rewarding. As in the time-like case, the difference with the supersymmetric case lies in the fact that the vector-bilinear $L$ to be introduced below, is not a Killing vector; introducing then an adapted coordinate $v$ through $L^{a}\partial_{a}=\partial_{v}$, we see that the metric will be explicitly $v$-dependent, unlike the supersymmetric case. The aim of this section, then, is to determine this $v$-dependence and give 2 minimal and simple, yet generic, solutions showing the changes brought about by the $\mathbb{R}$-gauging. In the Null case, the norm of the vector $V$ vanishes, whence $X=0$. This means that the 2 spinors $\epsilon_{I}$ are parallel, and following refs. [@Meessen:2006tu; @Hubscher:2008yz], we shall put $\epsilon_{I}=\phi_{I}\epsilon$, for some functions $\phi_{I}$ and the independent spinor $\epsilon$. The decomposition of $\epsilon^{I}$ follows from its definition as $\epsilon^{I}=(\epsilon_{I})^{*}$, which then implies $\epsilon^{I}=\phi^{I}\epsilon^{*}$, where we have defined $\overline{\phi_{I}}=\phi^{I}$. Furthermore, without loss of generality we can normalise the $\phi$’s such that $\phi_{I}\phi^{I}=1$. Once we take into account this normalisation, we can write down a completeness relation for the $I$-indices which is $$\label{eq:21} \Delta_{I}{}^{J} \; =\; \phi_{I}\phi^{J} \; +\; \varepsilon_{IK}\Phi^{K}\ \varepsilon^{JL}\Phi_{L} \; ,$$ which is such that $\Delta_{I}{}^{J}\phi_{J}=\phi_{I}$ and $\Delta_{I}{}^{J} \varepsilon_{JK}\phi^{K}=\varepsilon_{IK}\phi^{K}$. Moreover one can see that $\overline{\Delta_{I}{}^{J}}=\Delta_{J}{}^{I}$. Projecting, then, the fKSEs (\[eq:20\],…,\[eq:20d\]) onto the $\phi$’s we obtain $$\begin{aligned} \label{eq:22} 0 & =& \mathbb{D}_{a}\epsilon \; +\; \phi^{I}\nabla_{a}\phi_{I}\, \epsilon \; , \\ \label{eq:22b} 0 & =& \left( \mathcal{T}_{ab}^{+}\ +\ \textstyle{ig\over 4}\mathtt{C}_{\Lambda}\mathcal{L}^{\Lambda}\ \eta_{ab}\right)\ \gamma^{b}\epsilon^{*} \; -\; \varepsilon^{IJ}\phi_{I}\nabla_{a}\phi_{J}\ \epsilon \; ,\\ \label{eq:22c} 0 & =& i\slashed{\partial}Z^{i}\ \epsilon^{*} \; , \\ \label{eq:22d} 0 & =& \left[ \slashed{G}^{i+} \; +\: \mathtt{W}^{i}\right]\ \epsilon \; .\end{aligned}$$ In order to advance we will introduce an auxiliary spinor $\eta$, normalised by $\overline{\epsilon}\eta = \textstyle{1\over \sqrt{2}} = -\overline{\eta}\epsilon$; due to the introduction of $\eta$ we can introduce 4 null-vectors $$\label{eq:23} \begin{array}{lclclcl} L_{a} & =& i\overline{\epsilon}\gamma_{a}\epsilon^{*} &\hspace{.3cm},\hspace{.3cm}& N_{a} & =& i\overline{\eta}\gamma_{a}\eta^{*} \; ,\\ M_{a} & =& i\overline{\eta}\gamma_{a}\epsilon^{*} & ,& \overline{M}_{a} & =& i\overline{\epsilon}\gamma_{a}\eta^{*} \; , \end{array}$$ where $L$ and $N$ are real vectors and by construction $M^{*}=\overline{M}$, whence the notation. Observe that eq. (\[eq:Bil2a\]) implies that the vector $L$ is nothing but $V$, but we shall denote it by $L$(ightlike) in order to avoid confusion with the foregoing section. Given the above definitions it is a straightforward yet tedious calculation to show that they form an ordinary normalised null-tetrad, [*i.e.*]{} the only non-vanishing contractions are $$\label{eq:1} L^{a}\ N_{a} \; =\; 1 \; =\; -\ M^{a}\ \overline{M}_{a} \;\;\;\mbox{which implies}\;\;\; \eta_{ab} \ =\ 2\ L_{(a}N_{b)} \; -\; 2\ M_{(a}\overline{M}_{b)} \; .$$ Apart from the vectors one can also define imaginary-self-dual 2-forms, analogous to the ones defined in eq. (\[eq:Bil3\]), by $$\label{eq:2} \begin{array}{lclclcl} \Phi^{1}_{ab} & \equiv& \overline{\epsilon}\gamma_{ab}\epsilon &\hspace{.2cm},\hspace{.2cm}& \Phi^{1} & =& \sqrt{2}\ L\wedge \overline{M} \; , \\ \Phi^{2}_{ab} & \equiv& \overline{\epsilon}\gamma_{ab}\eta &\hspace{.2cm},\hspace{.2cm}& \Phi^{2} & =& \textstyle{1\over \sqrt{2}}\left[ L\wedge N \ +\ M\wedge \overline{M}\right] \; , \\ \Phi^{3}_{ab} & \equiv& \overline{\eta}\gamma_{ab}\eta &\hspace{.2cm},\hspace{.2cm}& \Phi^{3} & =& -\sqrt{2}\ N\wedge M \; , \\ \end{array}$$ where the identification on the r.h.s.  follows from the Fierz identities. The introduction of the above auxiliary spinor is not unique, and there still is some freedom left; first of all we have the freedom to rotate $\epsilon$ and $\eta$ by $\epsilon\rightarrow e^{i\theta}\epsilon$ and $\eta\rightarrow e^{-i\theta}\eta$. This rotation does not affect $L$ nor $N$, but rotates $M\rightarrow e^{-2i\theta}M$ and $\overline{M}\rightarrow e^{2i\theta}\overline{M}$: we will use this freedom to get rid of a phase-factor when introducing a coordinate expression for the tetrad. The second freedom arises, because a shift $\eta\rightarrow \eta +\delta\ \epsilon$, with $\delta$ a complex function, does not affect the normalisation condition. The effect of this shift on the vectors is $$\label{eq:3} L\ \rightarrow\ L \; ,\; M\ \rightarrow\ M \ +\ \delta\ L \; ,\; N\ \rightarrow\ N \ +\ |\delta|^{2}\ L \ +\ \delta\ \overline{M} \ +\ \bar{\delta}\ M \; ,$$ and this freedom can also be used to restrict the coordinate expressions of the tetrad. Let us start introducing a coordinate system by introducing a coordinate $v$ through $L^{\flat}\equiv L^{a}\partial_{a}=\partial_{v}$, and using eq. (\[eq:22\]) to derive $$\label{eq:5} \nabla_{a}L_{b} \; =\; g\mathtt{C}_{\Lambda}A^{\Lambda}_{a}\ L_{b} \; ,$$ whence $L$ is a recurrent null vector: this is the defining property of a space with holonomy $\mathrm{Sim}(2)$ (see ref. [@Gibbons:2007zu] for more information) and the combination $g\mathtt{C}_{\Lambda}A^{\Lambda}$ is called the recurrence 1-form. Anti-symmetrising this expression we see that $dL= g\mathtt{C}_{\Lambda}A^{\Lambda}\wedge L$, which implies not only $\mathtt{C}_{\Lambda}F^{\Lambda}\wedge L =0$, but also $L\wedge dL =0$. This last result states that the vector $L$ is hyper-surface orthogonal, which implies the local existence of functions $Y$ and $u$ such that $L = Ydu$. Seeing, however, that $L$ is charged under the $\mathbb{R}$-symmetry, we can always gauge-transform the function $Y$ away, leaving the statement that $L=du$, whence also that $\mathtt{C}_{\Lambda}A^{\Lambda} = \Upsilon\ L$, for some function $\Upsilon$. We can then write eq. (\[eq:5\]) as $$\label{eq:6} \nabla_{a}\ L_{b} \; =\; g\Upsilon\ L_{a}L_{b} \;\;\;\mbox{which immediately implies}\;\; \nabla_{L}L \ =\ 0 \; ,$$ so that $L$ is a geodesic null-vector. Given this information and the normalisation of the tetrad we can choose coordinates $u$, $v$, $z$ and $\bar{z}$ such that[^8] $$\label{eq:8} \begin{array}{lclclcl} L & =& du &\;\; ,\;\; & L^{\flat} & =& \partial_{v} \; , \\ N & =& dv \ +\ Hdu \ +\varpi dz\ + \overline{\varpi}d\bar{z} & ,& N^{\flat} & =& \partial_{u} \ -\ H\partial_{v} \; ,\\ M & =& e^{U}dz & ,& M^{\flat} & =& -e^{-U}\left( \partial_{\bar{z}} \ -\ \overline{\varpi}\partial_{v}\right) \; ,\\ \overline{M} & =& e^{U}d\bar{z} & ,& \overline{M}^{\flat} & =& -e^{-U}\left( \partial_{z} \ -\ \varpi\partial_{v}\right) \; , \end{array}$$ where we used the $U(1)$-rotation $M\rightarrow e^{-2i\theta}M$ to get rid of a possible phase in the expression of $M$ and $\overline{M}$. The spin-connection and curvatures for the tetrad is given in Appendix (\[sec:NullCurv\]). A last implication of the Fierz identities is that $$\label{eq:4} \varepsilon_{(4)} \;\equiv\; \textstyle{1\over 4!}\ \varepsilon_{abcd}\ e^{a}\wedge e^{b}\wedge e^{c}\wedge e^{d} \; =\; i\ L\wedge N\wedge M\wedge \overline{M} \; =\; i\ e^{+}\wedge e^{-}\wedge e^{\bullet}\wedge e^{\bar{\bullet}} \; ,$$ which implies that $\varepsilon^{+-\bullet\bar{\bullet}}=i$. Given the above expressions for the tetrad one can calculate the implications of the restriction (\[eq:6\]); one finds $$\label{eq:9} \partial_{v}H \; =\; g\Upsilon \hspace{.3cm}\mbox{and}\hspace{.3cm} 0\ =\ \partial_{v}U \ =\ \partial_{v}\varpi\ =\ \partial_{v}\overline{\varpi} \; ,$$ whence the only $v$-dependence of the metric resides in $H$; The resulting form of the metric is called a Walker metric, in honour of the late A.G. Walker, who was the first to give the general $d$-dimensional metric of a space with holonomy contained in $\mathrm{Sim}(d-2)$ in ref. [@art:walker1950]. In order to determine $\Upsilon$, we can use the identity $\mathtt{C}_{\Lambda}F^{\Lambda} = d\left( \mathtt{C}_{\Lambda}A^{\Lambda}\right) = d\Upsilon\wedge L$, which presupposes knowing $F^{\Lambda}$. The generic form of $F^{\Lambda}$ can be derived from the fKSEs (\[eq:22b\],\[eq:22d\]): consider first of all eq. (\[eq:22b\]). Contraction with $i\overline{\epsilon}$ and $i\overline{\eta}$ leads to $$\begin{aligned} \label{eq:10} \imath_{L}\mathcal{T}^{+} & =& \textstyle{ig\over 4}\ \mathtt{C}_{\Lambda}\mathcal{L}^{\Lambda}\ L \; ,\\ \label{eq:10a} \imath_{M}\mathcal{T}^{+} & =& \textstyle{ig\over 4}\ \mathtt{C}_{\Lambda}\mathcal{L}^{\Lambda}\ M \, +\, \textstyle{i\over \sqrt{2}} \phi_{I}\varepsilon^{IJ}\ d\phi_{J} \; .\end{aligned}$$ Coupling the above information to the fact that as $\mathcal{T}^{+}$ is an imaginary-self-dual 2-form it must be expressible in terms of the $\Phi$’s defined in eq. (\[eq:2\]), we see that $$\label{eq:11} \mathcal{T}^{+} \; =\; \aleph\ L\wedge\overline{M} \ -\ \textstyle{ig\over 4}\ \mathtt{C}_{\Lambda}\mathcal{L}^{\Lambda}\ \left[ L\wedge N + M\wedge \overline{M}\right] \;\;\mbox{with}\;\; \sqrt{2}\aleph \ =\ i\phi_{I}\varepsilon^{IJ}\nabla_{N}\phi_{J} \; ,$$ and furthermore $$\label{eq:13} \sqrt{2}\phi_{I}\varepsilon^{IJ}\nabla_{\overline{M}}\phi_{J} \; =\; g\mathtt{C}_{\Lambda}\mathcal{L}^{\Lambda} \;\;\; ,\;\;\; 0 \; =\; \phi_{I}\varepsilon^{IJ}\nabla_{L}\phi_{J} \; =\; \phi_{I}\varepsilon^{IJ}\nabla_{M}\phi_{J} \; .$$ Giving eq. (\[eq:22d\]) a similar treatment leads to $$\label{eq:14} G^{i+} \; =\; \aleph^{i}\ L\wedge\overline{M} \, -\, \textstyle{1\over 4}\ \mathtt{W}^{i} \left[ L\wedge N \, +\, M\wedge\overline{M}\right] \; ,$$ where $\aleph^{i}$ are, at this point, undetermined functions. Using the by-now-well-known rule $F^{\Lambda +}=i\overline{\mathcal{L}}^{\Lambda}\mathcal{T}^{+}+2f_{i}^{\Lambda}G^{i+}$, we find that $$\label{eq:15} F^{\Lambda +} \ =\ \varphi^{\Lambda}\ L\wedge\overline{M} \, +\, V^{\Lambda}\ \left[ L\wedge N + M\wedge\overline{M}\right] \; ,$$ where we introduced $$\label{eq:DefVLambda} V^{\Lambda} = \textstyle{g\over 8} \left(\ 4\overline{\mathcal{L}}^{\Lambda}\mathcal{L}^{\Sigma} + \mathrm{Im}(\mathcal{N})^{-1|\Lambda\Sigma}\ \right)\mathtt{C}_{\Sigma}$$ and $$\label{eq:17} \aleph = 2i\ \mathcal{L}_{\Lambda}\ \varphi^{\Lambda}\; ;\; \aleph^{i} \; =\; -\bar{f}_{\Lambda}^{i}\ \varphi^{\Lambda} \;\; \longleftrightarrow\;\; \varphi^{\Lambda} \; =\; i\aleph\ \overline{\mathcal{L}}^{\Lambda} \ +\ 2\aleph^{i}\ f_{i}^{\Lambda} \; .$$ Using then $F^{\Lambda} = F^{\Lambda +}+F^{\Lambda -}= 2\mathrm{Re}\left( F^{\Lambda +}\right)$, and doing the comparison $d\Upsilon\wedge L=\mathtt{C}_{\Lambda}F^{\Lambda}$, we obtain $$\begin{aligned} \label{eq:18} \nabla_{L}\Upsilon & =& -\mathtt{C}_{\Lambda}\left[ V \ +\ \overline{V}\right]^{\Lambda} \; ,\\ \label{eq:18b} \nabla_{M}\Upsilon & =& \mathtt{C}_{\Lambda}\ \varphi^{\Lambda} \; ,\\ \label{eq:18c} \nabla_{\overline{M}}\Upsilon & =& \mathtt{C}_{\Lambda}\ \overline{\varphi}^{\Lambda} \; .\end{aligned}$$ It is clear that eq. (\[eq:18\]) is the key to the possible $v$-dependence: in order to integrate it and obtain $H$ through eq. (\[eq:9\]), we need to know the coordinate dependence of the scalars $Z$. Information about said coordinate dependence can of course be obtained from eq. (\[eq:22c\]), by contracting it with the $i\overline{\epsilon}$ and $i\overline{\eta}$. The result is that $$\label{eq:19} 0\, =\, \nabla_{L} Z^{i} \, =\, \partial_{v}Z^{i} \;\;\;\mbox{and}\;\;\; 0\, =\, \nabla_{M} Z^{i} \, =\, e^{-U}\ \partial_{\bar{z}}Z^{i} \; ,$$ so that the $Z^{i}$ depend only on $u$ and $z$. Likewise, the $\overline{Z}^{\bar{\imath}}$ depend only on $u$ and $\bar{z}$. Using the fact that the scalars are $v$-independent, integration of eq. (\[eq:18\]) is straightforward and leads to $$\begin{aligned} \label{eq:7} \Upsilon & =& -\textstyle{g\over 4}\left[ 4\left|\mathtt{C}_{\Lambda}\mathcal{L}^{\Lambda}\right|^{2} \ +\ \mathrm{Im}\left(\mathcal{N}\right)^{-1|\Lambda\Sigma}\mathtt{C}_{\Lambda}\mathtt{C}_{\Sigma} \right]\ v \, +\, \Upsilon_{1}(u,z,\bar{z}) \; , \\ \label{eq:7a} H & =& -\textstyle{g^{2}\over 8}\left[ 4\left|\mathtt{C}_{\Lambda}\mathcal{L}^{\Lambda}\right|^{2} \ +\ \mathrm{Im}\left(\mathcal{N}\right)^{-1|\Lambda\Sigma}\mathtt{C}_{\Lambda}\mathtt{C}_{\Sigma} \right]\ v^{2} \, +\, \Upsilon_{1}\ v \ +\ \Upsilon_{0}(u,z,\bar{z}) \; .\end{aligned}$$ By doing a coordinate transformation $v\rightarrow v + f(u,z,\bar{z})$ we can take $\Upsilon_{1}=0$, but for the moment we shall ignore this possibility. $H$ can be written in terms of the potential $\mathtt{V}$ in eq. (\[eq:Potential\]), with $\mathtt{P}_{\Lambda}=0$ as we are ignoring possible non-Abelian couplings, as $$\label{eq:7b} H \ =\ \textstyle{1\over 2}\left[ g^{2}\left|\mathtt{C}_{\Lambda}\mathcal{L}^{\Lambda}\right|^{2} \ -\ \mathtt{V} \right]\ v^{2} \ +\ \Upsilon_{1}v \ +\ \Upsilon_{0} \; ,$$ which is calculationally advantageous when $\mathtt{V}$ is known. At this point we have nearly completely specified the $v$-dependence of the solution, the only field missing being the $A^{\Lambda}$; in order to determine its $v$-dependence it is worthwhile to impose the gauge-fixing $\imath_{L}A^{\Lambda}=0$, which is always possible and is furthermore consistent with the earlier result $\mathtt{C}_{\Lambda}A^{\Lambda}=\Upsilon\ L$. As a result of this gauge fixing we have that $$\label{eq:7c} \partial_{v}A^{\Lambda} \ =\ \pounds_{L}A^{\Lambda} \ =\ d\left(\imath_{L}F^{\Lambda}\right) \ =\ -\left(\ V\ +\ \overline{V}\ \right)^{\Lambda}\, L \; ,$$ so that $$\label{eq:7d} A^{\Lambda} \; =\; -\left(\ V\ +\ \overline{V}\ \right)^{\Lambda}\, v\ L \; +\; \tilde{A}^{\Lambda} \; =\; \textstyle{g\over 4}\ \mathsf{F}^{-1|\Lambda\Sigma}\mathtt{C}_{\Sigma}\ v\ L \; +\; \tilde{A}^{\Lambda} \; ,$$ where $\tilde{A}^{\Lambda}$ is a $v$-independent 1-form satisfying $\imath_{L}\tilde{A}^{\Lambda}=0$, and $\mathsf{F}$ is the imaginary part of the prepotential’s Hessian; see eq. (\[eq:PrePotImNinv\]) for why this ocurrs.. Given this expression for the vector potentials, the Bianchi identity is automatically satisfied, but, as in the time-like case, this does not necessarily mean that any $\tilde{A}^{\Lambda}$ leads to a field-strength of the desired form. Calculating the comparison we find that $$\begin{aligned} \label{eq:7e} d\tilde{A}^{\Lambda} & =& \left( V-\overline{V}\right)^{\Lambda}\ M\wedge\overline{M} \nonumber \\ & +& \left( \phi^{\Lambda}+\theta_{M}\left[ v\left( V+\overline{V}\right)^{\Lambda}\right]\right) \ L\wedge\overline{M} \ +\ \left( \overline{\phi}^{\Lambda}+\theta_{\overline{M}}\left[ v\left( V+\overline{V}\right)^{\Lambda}\right]\right) \ L\wedge M \, .\end{aligned}$$ Let us at this point return to the fKSEs, and evaluate eq. (\[eq:20c\]) using eqs. (\[eq:14\]) and (\[eq:19\]). This evaluation results in $$\label{eq:26} i\theta_{+}Z^{i}\ \gamma^{+}\epsilon^{I} \ +\ i\theta_{\bullet}Z^{i}\ \gamma^{\bullet}\epsilon^{I} \; =\; -\varepsilon^{IJ}\left[ \mathtt{W}^{i}\gamma^{-}\ -\ 2\alpha^{i}\gamma^{\bar{\bullet}}\right]\ \gamma^{+}\epsilon_{J} \; .$$ The above equation is readily seen to be solved by observing that the constraint $\gamma^{+}\epsilon_{J}=0$ not only leads to $\gamma^{+}\epsilon^{I}=0$ under complex conjugation, but also to $\gamma^{\bar{\bullet}}\epsilon_{I}=0$ and $\gamma^{\bullet}\epsilon^{I}=0$; these last implications are due to the fact that we dealing with chiral spinors and the normalisation in eq. (\[eq:4\]). Doing a similar analysis on the fKSE (\[eq:20\]) in the $v$-direction shows that the spinor $\epsilon_{I}$, whence also $\epsilon^{I}$, is $v$-independent. The other equations become $$\begin{aligned} \label{eq:27} \mathbb{D}_{\bar{\bullet}}\epsilon_{I} & =& 0\; ,\\ \label{eq:27a} \mathbb{D}_{\bullet}\epsilon_{I} & =& \textstyle{ig\over 2}\ \mathtt{C}_{\Lambda}\mathcal{\Lambda}\ \varepsilon_{IJ}\gamma^{\bar{\bullet}}\epsilon^{J} \; , \\ \label{eq:27c} \mathbb{D}_{+}\epsilon_{I} & =& -\aleph\ \varepsilon_{IJ}\gamma^{\bar{\bullet}}\epsilon^{J} \; .\end{aligned}$$ Using the definition (\[eq:16\]) and the spin-connection in eq. (\[eq:NCspincon\]), we can expand eqs. (\[eq:27\]) and (\[eq:27a\]) as $$\begin{aligned} \label{eq:28} 0 & =& \theta_{\bar{\bullet}}\epsilon_{I} \ -\ \textstyle{1\over 2}\theta_{\bar{\bullet}}\left( U +\textstyle{1\over 2}\mathcal{K}\right)\ \epsilon_{I} \; , \\ \label{eq:28a} 0 & =& \theta_{\bullet}\epsilon_{I} \ +\ \textstyle{1\over 2}\theta_{\bullet}\left( U +\textstyle{1\over 2}\mathcal{K}\right)\ \epsilon_{I} \ -\ \textstyle{ig\over 2}\mathtt{C}_{\Lambda}\mathcal{L}^{\Lambda}\ \varepsilon_{IJ}\gamma^{\bar{\bullet}}\epsilon^{J} \; ,\end{aligned}$$ The first equation is easily integrated by putting $$\label{eq:29} \epsilon_{I} \; =\; \exp\left( \textstyle{1\over 2}\ S\right)\ \chi_{I} (u,z) \hspace{.5cm}\mbox{with}\hspace{.5cm} S \ \equiv\ U \ +\ \textstyle{1\over 2}\mathcal{K} \; ,$$ which upon substitution into eq. (\[eq:28a\]) leads to $$\label{eq:30} \partial_{z}\chi_{I} \ +\ \left(\partial_{z}S\right)\chi_{I} \; =\; \textstyle{ig\over 2}\ \mathtt{C}_{\Lambda}\mathcal{X}^{\Lambda}\ \varepsilon_{IJ}\gamma^{\bar{\bullet}}\ e^{S}\chi^{J} \; .$$ This last equation is potentially dangerous as it has a residual $\bar{z}$-dependence, even though $\eta$ and $\mathcal{X}^{\Lambda}$ are $\bar{z}$-independent; it is this possible inconsistency that fixes $S$, as can be seen by deriving eq. (\[eq:30\]) w.r.t. $\bar{z}$ and using the complex conjugated version of eq. (\[eq:30\]) to get rid of $\eta^{I}$ in the resulting equations. The result is that $S$ has to satisfy $$\label{eq:31} \partial_{z}\partial_{\bar{z}}S \; =\; -\textstyle{g^{2}\over 2}\ e^{2S}\ \left|\mathtt{C}_{\Lambda}\mathcal{X}^{\Lambda}\right|^{2} \;\;\;\longrightarrow\;\;\; e^{-2S} \, =\, \textstyle{g^{2}\over 2}\ \left|\mathtt{C}_{\Lambda}\mathcal{X}^{\Lambda}\right|^{2}\ \left( 1+|z|^{2}\right)^{2} \; .$$ This unique choice for $S$ is a necessary condition for the eqs. (\[eq:28\]) and (\[eq:28a\]) to admit a solution, but it may not be sufficient; in the next section we shall discuss the simplest null-case solution to the minimal theory, and show that the system can be solved completely. The lesson to be learned from that section is that the system (\[eq:28\],\[eq:28a\]) once we introduce $S$, corresponds to an equation determining spinors on a 2-sphere, and has solutions even though this is hard to see. The electrically charged Nariai cosmos belongs to the Null case {#sec:NullSols} --------------------------------------------------------------- The minimal theory is obtained by putting $\mathcal{V}^{T}=(1,-i/2)$, which leads to the monodromy matrix $\mathcal{N}=-i/2$, so that $\mathrm{Re}(\mathcal{N})=0$. if we then further fix $\mathtt{C}_{0}=2$, we see that the minimal De Sitter theory is given by $$\label{eq:MinDSaction} \int_{4}\sqrt{g}\left( R\ -\ F^{2}\ -\ 6g^{2}\right) \; .$$ Using the general results obtained thus far, we can write down the following solution $$\begin{aligned} \label{eq:NullNariai} ds^{2} & =& 2du\left( dv -g^{2}v^{2}du\right) \, -\, \frac{dzd\bar{z}}{g^{2}(1+|z|^{2})^{2}} \; .\nonumber \\ A & =& -gv\ du\; ,\end{aligned}$$ A small analysis shows that the metric is nothing more than $DS_{2}\times S^{2}$, albeit in a non-standard coordinate system, and the solutions is known to the literature as the electrically charged Nariai solution [@art:nariai]. Observe that the local holonomy of the Nariai solution is not the full $\mathfrak{sim}(2)$, but rather $\mathfrak{so}(1,1)\oplus\mathfrak{so}(2)\subset \mathfrak{sim}(2)$ [@Gibbons:2007zu]. In order to discuss the preserved fake-supersymmetries it is easier to write the metric as $$\label{eq:NullNariai1} ds^{2} \; =\; 2du\left( dv -g^{2}v^{2}du\right) \, -\, \frac{1}{4g^{2}}\left[ d\theta^{2}\ +\ \sin^{2}(\theta )d\varphi^{2}\right] \; ,$$ and consider the fake-supergravity equations in terms of a 2-component vector of Majorana spinors, also denoted by $\epsilon$, namely $$\label{eq:NullNariai2} \nabla_{a}\epsilon \ -\ gA_{a}\epsilon \; =\; -\textstyle{1\over 4}\slashed{F}\gamma_{a}\sigma^{2}\epsilon -\textstyle{g\over 2}\gamma_{a}\sigma^{2}\epsilon \; .$$ The solution to the above equation is then seen to be $$\label{eq:NullNariai3} \epsilon \; =\; \exp\left( \textstyle{\theta\over 2}\gamma^{3}\sigma^{2}\right)\ \exp\left( -\textstyle{\varphi\over 2}\gamma^{34}\right)\ \epsilon_{0} \hspace{.4cm}\mbox{with}\hspace{.4cm} \gamma^{+}\epsilon_{0}=0\; ,$$ where $\epsilon_{0}$ is a 2-vector of constant spinors. Some remarks are in order: in supersymmetry one can associate a Lie superalgebra to a given supersymmetric solution [@Gauntlett:1998kc], and for the supersymmetric $aDS_{2}\times S^{2}$ maximally supersymmetric solutions in minimal $N=2$ $d=4$, this algebra is $\mathfrak{su}(1,1|2)$. In the fake-supersymmetric case, however, one cannot assign a Lie superalgebra to the solution, as the vector bilinears which would represent the supertranslation part, do not lead to Killing vectors; this fact is already illustrated by eq. (\[eq:5\]). A perhaps worrisome point is the action of the De Sitter’s Killing vectors on the preserved fake-supersymmetry, especially since the Killing spinors are $u$- and $v$-independent. Taking into account that the spinors are gauge-dependent objects means that this action is defined using the $\mathbb{R}$-covariant Lie derivative on spinors [@Ortin:2002qb]; this derivative is defined for Killing vectors $X$ and $Y$ as $$\label{eq:Kosmann} \mathbb{L}_{X}\epsilon \; =\; \nabla_{X}\epsilon \ +\ \textstyle{1\over 4}\left(\partial_{a}X_{b}\right)\ \gamma^{ab}\epsilon \ -\ g\xi_{X}\ \epsilon \;\;\;\mbox{with}\;\;\; \left\{ \begin{array}{rcl} d\xi_{X} & =& \pounds_{X}A \\ \xi_{[X,Y]} & =& \pounds_{X}\xi_{Y}-\pounds_{Y}\xi_{X} \end{array} \right.$$ Using this Lie derivative, one can see that $\mathbb{L}_{X}\epsilon =0$ for any $X\in\mathrm{Isom}(DS_{2})$. Holomorphic scalars and deformations of the Nariai cosmos {#sec:Holomorphic} --------------------------------------------------------- In the supersymmetric case, there are 2 generic classes of solutions in the null case whose supersymmetry is straightforward to see: the first are the pp-waves which are characterised by the fact that the scalars depend only on $u$, and the [*cosmic strings*]{} which are characterised by vanishing vector potentials $A^{\Lambda}$, vanishing Sagnac connection, $\varpi =0$, and a holomorphic spacetime dependence of the scalars, [*i.e.*]{} $Z^{i}=Z^{i}(z)$ [@Tod:1995jf; @Meessen:2006tu]. In this section we will consider the analogue of the latter case and impose $\varpi =0$ and that $Z^{i}$ is a function of $z$ only. Due to eq. (\[eq:7d\]), however, the vector potentials cannot vanish and we will look for the minimal expression for $\tilde{A}^{\Lambda}$ for which the Bianchi identity, eq. (\[eq:7e\]) is solved: minimality implies that $\phi^{\Lambda} = v e^{-U}\partial_{\bar{z}}\left( V+\overline{V}\right)^{\Lambda}$ and the Bianchi identity reduces to $$\label{eq:Hol1} d\tilde{A}^{\Lambda} \; =\; 2i\ \mathrm{Im}\left(\frac{\mathcal{X}^{\Lambda}}{g\mathtt{C}_{\Sigma}\mathcal{X}^{\Sigma}}\right) \, \frac{dz\wedge d\bar{z}}{(1+|z|^{2})^{2}} \; ,$$ a solution to which exists locally and determines $\tilde{A}_{u}=0$ and $\tilde{A}_{z}^{\Lambda}$ and $\tilde{A}^{\Lambda}_{\bar{z}}$ as functions of $z$ and $\bar{z}$. Given the above identifications we can use eq. (\[eq:15\]) to calculate the constraints imposed by the Maxwell e.o.m.s, [*i.e.*]{} $\mathcal{B}_{\Lambda}=0$ in eq. (\[eq:VectEOM2\]), which leads to $$\begin{aligned} \label{eq:Hol2a} \mathcal{N}_{\Lambda\Sigma}\ \partial_{z}\left( V +\overline{V}\right)^{\Sigma} & =& \partial_{z}\left[ \overline{\mathcal{N}}_{\Lambda\Sigma}V^{\Sigma} \ +\ \mathcal{N}_{\Lambda\Sigma}\overline{V}^{\Sigma} \right] \; ,\\ \label{eq:Hol2b} \overline{\mathcal{N}}_{\Lambda\Sigma}\ \partial_{\bar{z}}\left( V +\overline{V}\right)^{\Sigma} & =& \partial_{\bar{z}}\left[ \overline{\mathcal{N}}_{\Lambda\Sigma}V^{\Sigma} \ +\ \mathcal{N}_{\Lambda\Sigma}\overline{V}^{\Sigma} \right] \; , \\ \label{eq:Hol2c} \partial_{z}\left[ \overline{\mathcal{N}}_{\Lambda\Sigma}\ \partial_{\bar{z}}\left( V +\overline{V}\right)^{\Sigma} \right] & =& \partial_{z}\left[ \mathcal{N}_{\Lambda\Sigma}\ \partial_{z}\left( V +\overline{V}\right)^{\Sigma} \right] \; ,\end{aligned}$$ the contribution due to $\tilde{A}^{\Lambda}$ dropping out identically. As eq. (\[eq:Hol2b\]) is the complex conjugated version of (\[eq:Hol2a\]), and eq. (\[eq:Hol2c\]) is the integrability condition for eqs. (\[eq:Hol2a\]) and (\[eq:Hol2b\]), we only need to see that eq. (\[eq:Hol2a\]) holds. Using the holomorphicity of the scalars in order to write $\partial_{z}=\partial_{z}Z^{i}\ \partial_{i}$, one can rewrite eq. (\[eq:Hol2a\]) as an equation in Special Geometry, namely $$\begin{aligned} \label{eq:Hol3a} \partial_{i}\overline{\mathcal{N}}_{\Lambda\Sigma}\ V^{\Sigma} \ +\ \partial_{i}\mathcal{N}_{\Lambda\Sigma}\ \overline{V}^{\Sigma} & =& 2i\mathrm{Im}\left(\mathcal{N}\right)_{\Lambda\Sigma}\ \partial_{i}V^{\Sigma} \nonumber \\ & =& gi\overline{\mathcal{L}}_{\Lambda}\ \mathtt{C}_{\Gamma}f_{i}^{\Gamma} \ -\ \textstyle{gi\over 4}\partial_{i}\mathrm{Im}\left(\mathcal{N}\right)_{\Lambda\Sigma}\, \mathrm{Im}\left(\mathcal{N}\right)^{-1|\Sigma\Gamma}\mathtt{C}_{\Gamma} \; .\end{aligned}$$ Some straightforward algebra using the expressions (\[eq:SGdifiN\]) and (\[eq:SGdifbiN\]) shows that the above equation holds, whence the Maxwell equations are solved for arbitrary scalar functions $Z^{i}(z)$. Had we been sure of the fact that the generic expressions for the fields we are using solve the fKSEs, we would have deduced from the KSIs that we only need to verify $\mathcal{B}_{++}=0$ as to be sure that the proposed configuration solves the equations of motion. As we are not 100[%]{} sure of this fact, however, we checked that all of the equations of motion are indeed satisfied. As was to be expected from the discussion of the Maxwell equations, all the e.o.m.s reduce to Special Geometry calculations. In conclusion then, given an expression for $Z^{i}=Z^{i}(z)$, we need to find the local expression for $\tilde{A}^{\Lambda}$ from eq. (\[eq:Hol1\]), and the solution is given by $$\begin{aligned} \label{eq:NarGen1} ds^{2} & =& 2du\left( dv - \textstyle{1\over 2}H_{0}v^{2}\ du\right) \ -\ \frac{4}{g^{2}\left|\mathtt{C}_{\Lambda}\mathcal{L}^{\Lambda}\right|^{2}} \frac{dzd\bar{z}}{(1+|z|^{2})^{2}} \; , \\ & & \nonumber \\ \label{eq:NarGen2} A^{\Lambda} & =& \textstyle{g\over 4}\mathsf{F}^{-1|\Lambda\Sigma}\mathtt{C}_{\Sigma}\ v\ du \ +\ \tilde{A}^{\Lambda} \; ,\end{aligned}$$ where $$\label{eq:NarGen3} H_{0} \; =\; \mathtt{V} \ -\ g^{2}\left| \mathtt{C}_{\Lambda}\mathcal{L}^{\Lambda}\right|^{2} \; .$$ Nariai-like solutions can be obtained by taking the scalars $Z^{i}$ to be constants, in which case the $z\bar{z}$-part of the metric describes a 2-sphere of radius $g|\mathtt{C}_{\Lambda}\mathcal{L}|$. Depending on $H_{0}$, the $uv$-part of the metric describes $DS_{2}$ ($H_{0}>0$), 2-dimensional Minkowski space ($H_{0}=0$) or $aDS_{2}$ ($H_{0}<0$). As before, these spaces have local holonomy contained in $\mathrm{sim}(2)$; the solution for generic $Z^{i}(z)$, however, has proper $\mathrm{sim}(2)$ holonomy. Non-BPS solutions to $N=2$ sugra from fEYM {#sec:PotIsNul} ========================================== As is well-known, there are models in $N=2$ $d=4$ sugra coupled to vector-multiplets for which one can choose the Fayet-Iliopoulos terms such that the hyper-multiplet contribution to the potential vanishes (see [*e.g.*]{} [@Cremmer:1984hj] or [@Andrianopoli:1996cm sec. 9] for a discussion of this point). As we are basically dealing with a Wick-rotated version of the general supersymmetric set-up, this implies that there are fake-supersymmetric models in which the only contribution to the potential comes from the gauging of the isometries, as the FI-contributions cancel. In that case the bosonic action (\[eq:VectAct\]) coincides with that of an ordinary YMH-type of supergravity theory, and we must conclude that for those specific models the solutions we obtained are in fact non-BPS solutions of a regular supergravity theory.[^9] Let us illustrate this fact with an example: the dimensional reduction of minimal 5-dimensional sugra. The dimensional reduction of minimal 5-dimensional sugra leads to a specific $N=2$ $d=4$ sugra, namely minimal sugra coupled to one vector-multiplet with a prepotential given by $$\label{eq:24} \mathcal{F}\left( \mathcal{X}\right) \; =\; -\textstyle{1\over 8}\ \frac{\left(\mathcal{X}^{1}\right)^{3}}{\mathcal{X}^{0}} \; .$$ With the usual choice $Z=\mathcal{X}^{1}/\mathcal{X}^{0}$, one finds that the scalar-manifold is $\mathrm{Sl}(2;\mathbb{R})/\mathrm{U}(1)$ with the corresponding Kähler potential $e^{\mathcal{K}}\ =\ \mathrm{Im}^{3}\left( Z\right)$; observe that this implies the constraint $\mathrm{Im}\left( Z\right) > 0$. Ignoring the possibility of gauging isometries of the resulting scalar-manifold, so that $\mathtt{P}=0$, we can calculate the potential in eq. (\[eq:Potential\]) only to find $$\label{eq:25} \mathtt{V} \; =\; \textstyle{2g^{2}\over 3}\left[\ \mathtt{C}_{1}^{2}\ \mathrm{Im}^{-1}(Z) \; +\; 6\mathtt{C}_{0}\mathtt{C}_{1}\ \mathrm{Re}(Z)\mathrm{Im}^{-3}(Z) \ \right] \; .$$ There are two interesting sub-classes to be considered, the first one being $\mathtt{C}_{\Lambda}=(0,\mathtt{C}_{1})$ for which the potential is of the correct form to correspond to the dimensionally reduced version of the theory considered in [@Grover:2008jr]. The second case is $\mathtt{C}_{\Lambda}=(\mathtt{C}_{0},0)$, which seeing that the potential is linear in $\mathtt{C}_{0}$ means that the potential vanishes. By construction this not only means that we can construct non-BPS solutions to the 4-dimensional supergravity theory, but also that it can be oxidised to minimal 5-dimensional sugra. A simple time-like static solution for this latter case can be found by putting $\mathcal{I}^{0}=0$, so that we can take the base-space to be $\mathbb{R}^{3}$, and $\mathcal{I}_{1}=0$ as to ensure staticity, [*i.e.*]{} $\omega = 0$; the regularity of the solution to the stabilisation equations, or equivalently the consistency of the metrical factor $|X|^{2}$, imposes the constraint $\mathcal{I}_{0}\left(\mathcal{I}^{1}\right)^{3}<0$. With this information the solution is determined by $$\label{eq:32} \frac{1}{2|X|^{2}} \ =\ \sqrt{\ 2\ \left| \mathcal{I}_{0}\left(\mathcal{I}^{1}\right)^{3}\right|\ } \hspace{.4cm} ,\hspace{.4cm} Z \, =\, 2i\sqrt{\ \left|\frac{\mathcal{I}_{0}}{\mathcal{I}^{1}}\right|\ } \; ,$$ so that the solution is asymptotically Kasner. As the effective radius of the compactified fifth direction is proportional to $\mathrm{Im}(Z)$ which grows linear in $\tau$, this solutions is asymptotically decompactifying; the resulting 5-dimensional metric is readily found to be (shifting $\mathcal{I}^{1}\rightarrow \sqrt{2}\ H$) $$\label{eq:33} ds_{(5)}^{2} \ =\ 2H^{-1}\ dy\left(\ d\tau \ -\ 2\sqrt{2}\ \left|\mathcal{I}_{0}\right|\ dy\right) \ -\ H^{2}\ d\vec{x}^{2}\; .$$ which can be transformed to a Walker metric for a space of holonomy $\mathrm{Sim}(3)$ [@art:walker1950]. Observe that the relation between $d+1$ dimensional spaces of holonomy in $\mathrm{Sim}(d-1)$ and time-dependent black holes, of which the foregoing is one example, was first introduced and used in ref. [@Gibbons:2007zu]. The generic solution in section (\[sec:Holomorphic\]) can readily be adapted to the model at hand and reads $$\label{eq:34} ds^{2} \; =\; 2du\left( dv \ +\ \lambda^{2}\ v^{2}\ \mathcal{Z}^{-3}\ du\right) \; -\; \textstyle{2\over \lambda^{2}}\ \mathcal{Z}^{3}\ \frac{dzd\bar{z}}{(1-|z|^{2})^{2}} \; ,$$ where we introduced the abbreviations $\sqrt{2}\lambda = g\mathtt{C}_{0}$ and $\mathcal{Z}=\mathrm{Im}(Z)$. The vector fields are given by the expression (\[eq:7d\]), with $\tilde{A}^{0}=0$ and $\tilde{A}^{1}$ needs to satisfy $$\label{eq:35} d\tilde{A}^{1} \; =\; \textstyle{\sqrt{2}i}{\lambda}\ \mathcal{Z}\ \frac{dz\wedge d\bar{z}}{(1+|z|^{2})^{2}} \; ,$$ which presupposes knowing the explicit dependence of $Z$ on $z$. Lifting this solution up to 5 dimensions we obtain, after the coordinate transformations $v\rightarrow e^{\sqrt{2}\lambda y}w$ where $y$ is the 5$^{th}$ direction, the following solution $$\begin{aligned} \label{eq:36} ds^{2}_{(5)} & =& 2\mathcal{Z}^{-1}\ e^{\sqrt{2}\lambda y}\ dudw \; -\; \mathcal{Z}^{2}\left[ dy^{2} \ +\ \textstyle{2\over \lambda^{2}}\ \frac{dzd\bar{z}}{(1-|z|^{2})^{2}} \right] \; , \\ \hat{A} & =& \sqrt{3}\ \mathrm{Re}\left( Z\right)\ \left[ dy \ +\ 2\sqrt{2}\lambda\ \mathcal{Z}^{-3}\ vdu \right] \, -\, \sqrt{3}\ \tilde{A}^{1} \; ,\end{aligned}$$ where $\tilde{A}^{1}$ is determined by the condition (\[eq:36\]): this solution is a deformation of the maximally supersymmetric $aDS_{3}\times S^{2}$ solution, and deformations of the other maximally supersymmetric 5-dimensional solutions can be obtained by using the $Sp(2;\mathbb{R})$-duality transformations before oxidation, similar to how the 4- and 5-dimensional vacua are related (see [*e.g.*]{} ref. [@LozanoTellechea:2002pn]). Let us end this section by pointing out that there are more models for which the FI-contribution to the potential vanishes [@Andrianopoli:1996cm]. One of them is the $\mathcal{ST}[2,m]$-model, which in the ungauged supergravity model, allows for the embeddings of monopoles and the construction of non-Abelian black holes [@Hubscher:2008yz], and we will briefly talk about the solutions. A convenient parameterisation of the model is given by the symplectic section $$\label{eq:38} \mathcal{V} \, =\, \left(\begin{array}{c} \mathrm{L}^{\Lambda} \\ \eta_{\Lambda\Sigma}\mathrm{S}\ \mathrm{L}^{\Sigma}\end{array}\right) \;\;\mbox{with}\;\; \left\{\begin{array}{lcl} \eta & =& \mathrm{diag}([+]^{2},[-]^{m}) \\ & & \\ 0 & =& \eta_{\Lambda\Sigma} \mathrm{L}^{\Lambda}\mathrm{L}^{\Sigma} \end{array}\right. \; .$$ The FI-part of the potential is easily calculated and gives [@Cremmer:1984hj; @Andrianopoli:1996cm] $$\label{eq:39} \mathtt{V}_{FI} \; =\; -\textstyle{g^{2}\over 4}\ \mathrm{Im}^{-1}\left(\mathrm{S}\right)\ \mathtt{C}_{\Lambda}\eta^{\Lambda\Sigma}\mathtt{C}_{\Sigma} \; ,$$ so that $\mathtt{V}_{FI}=0$ whenever $\mathtt{C}$ is a null-vector w.r.t. $\eta$. Taking $\mathcal{ST}[2,4]$ as the model to work with and $\mathtt{C}$ to be a null-vector, we can gauge an $SU(2)$-gauge group, and by further taking $\mathtt{C}_{\Lambda}\mathcal{I}^{\Lambda}=0$, implying that the base-space is $\mathbb{R}^{3}$, we can generalise the solutions found in ref. [@Meessen:2008kb] to cosmological solutions. For that take the indices $\Lambda$ to run over $(0,+,-, i)$ (with $0$ a time-like direction, $\pm$ the null directions and $i=1,2,3$) and let $\mathtt{C}_{+}$ be the only non-vanishing element of the $\mathtt{C}$s. By taking then $\mathcal{I}^{\pm}=\mathcal{I}_{0}=\mathcal{I}_{i}=0$ we find a static solution, [*i.e.*]{} $\omega =0$, which allows for the embedding of an ’t Hooft-Polyakov monopole, say, in the $\mathcal{I}^{i}$s. If we then further take $\tilde{\mathcal{I}}_{+}=0$ and normalise the metric on constant-$\tau$ slices to be asymptotically $\mathbb{R}^{3}$, which is equivalent to taking $\tilde{\mathcal{I}}_{-}$ and $\mathcal{I}^{0}$ to be suitable constants, we see that the metric is determined through eq. (\[eq:37\]) and $$\label{eq:40} \frac{1}{2|X|^{2}}\; =\; \sqrt{\tau}\ \sqrt{ 1\, +\ \textstyle{\mu^{2}\over g^{2}}\left[ 1-\overline{H}^{2}\right] } \; ,$$ where $\overline{H}$ is a completely regular function of $r\in\mathbb{R}$ coming from the ’t Hooft-Polyakov monopole: it reads $$\label{eq:41} \overline{H} \; =\; \coth \left(\mu r\right) \ -\ \frac{1}{\mu r} \; ,$$ and is a monotonic functions with $\overline{H}(r=0)=0$ and asymptoting to $\overline{H}(r\rightarrow\infty )=1$. This means that the constant-$\tau$ slices are complete: the full metric, however, suffers from an initial singularity at $\tau =0$ and also from Kasner expansion. More general solutions can of course be constructed by considering the hairy or coloured solutions in refs. [@Meessen:2008kb; @Hubscher:2008yz], in case one is interested in non-Abelian solutions, or the general Abelian solutions of ref. [@Behrndt:1997ny]; to these solutions the general the comments made in section (\[sec:CosmMon\]) apply. Conclusions... {#sec:Concl} ============== In this article we studied the fake-supersymmetric solution that can be obtained from $N=2$ $d=4$ gauged supergravity coupled to (non-Abelian) vector multiplets, by Wick-rotating the FI-term needed in order to obtain gauged supergravity. As is usual in the classification of (fake-)supersymmetric solutions, the solutions are divided into two classes, denoted the time-like- and the null-case, which are distinguished by the norm of the vector built out of the preserved Killing spinor. In the time-like case we find that the metric is of the standard conformastationary form, appearing naturally in the supersymmetric time-like solutions, with the difference that the metric is to have a specific time dependence; this time dependence is such that there is a natural substitution principle, as first pointed out by Behrndt and Cvetič [@Behrndt:2003cx], of creating solutions from the known supersymmetric solutions to $N=2$ $d=4$ supergravity coupled to (non-Abelian) vector multiplets. Apart from this time-dependence, we find that the base-space must be a subclass of 3-dimensional Einstein-Weyl spaces known as hyperCR- or Gauduchon-Tod spaces [@Gauduchon:1998], and that half of the seed functions, namely the $\mathcal{I}^{\Lambda}$, must obey the Bogomol’nyi equation generalised to GT-spaces. In the null-case we find that the solutions must have a holonomy contained in $\mathrm{Sim}(2)$, which arguably can be considered to be a minor detail: it was, however, shown in ref. [@art:coley] that the purely gravitational solutions of this kind have rather special properties with respect to quantum corrections, and it is not unconceivable that this holds for the more general class of solutions with $\mathrm{Sim}(2)$-holonomy in supergravity theories, such as the one presented in section (\[sec:PotIsNul\]). We did not develop a full-fledged characterisation of the solutions in the null-case, but instead focussed on the new characteristics induced by the interplay between $\mathrm{Sim}(2)$-holonomy and Special Geometry. The end result is what can be considered to be a back-reacted solution describing the intersection of a Nariai/Robinson-Bertotti space with a generic (stringy) cosmic string [@Meessen:2006tu]. The fact that the holonomy is contained in $\mathrm{Sim}(2)$ is caused by the fact that we are gauging an $\mathbb{R}$-symmetry, where-from one deduces that the null-vector one constructs as a bilinear of the preserved Killing spinor is gauge-covariantly constant null-vector; said differently it is a recurrent null-vector, whence the 4-dimensional space has holonomy $\mathrm{Sim}(2)$ [@Gibbons:2007zu]. As the Wick-rotation needed to create fake supergravities from ordinary gauged supergravities will always introduce an $\mathbb{R}$-gauging, one might be inclined to think that fake supersymmetric solutions in the null case always have infinitesimal holonomy in $\mathfrak{sim}(d-2)$. This is, however, only partially true. Consider for instance the theory studied by Grover [*et al*]{} [@Grover:2008jr]: in that case one can see that the recurrency condition (\[eq:5\]) still holds [*but*]{} with the Levi-Cività connection replaced with a metric compatible, torsionful connection, where the torsion is completely anti-symmetric and proportional to the Hodge dual of the graviphoton field strength. As the connection is metric, the link between the recurrency relation and $\mathfrak{sim}$-holonomy going through [*mutatis mutandis*]{}, we see that in fake $N=1$ $d=5$ gauged supergravity theories, there is a $\mathrm{Sim}(3)$ holonomy even though in general it is not associated to the Levi-Cività connection. As was shown by Gibbons [&]{} Pope in ref. [@Gibbons:2007zu], and illustrated in section (\[sec:PotIsNul\]), time-dependent solution of the kind found in the time-like case can be obtained by dimensional reduction of spaces with $\mathrm{Sim}$-holonomy; the solutions in the time-like case can also be obtained from the solutions in the 5-dimensional time-like case. This strongly suggest that the ordinary hierarchy of supersymmetric solutions, and the geometric structures appearing in them, to theories in $d=6$, $5$ and $4$ with eight supercharges has a fake analogue.\ [**Note added:**]{} Shortly after this paper appeared, Gutowski [&]{} Sabra [@Gutowski:2009vb] published the classification of the fake supersymmetric solutions to the minimal theory. Let us for completeness point out that the general solution to the null-case is the Nariai solution in eq. (\[eq:NullNariai\]) with the substitution $g_{uu}=-2g^{2}v^{2}\rightarrow -2g^{2}v^{2} +2\Upsilon_{0}(z,\bar{z})$, with $\partial_{z}\partial_{\bar{z}}\Upsilon_{0}=0$ Acknowledgements {#acknowledgements .unnumbered} ================ This work has been supported in part by a C.S.I.C. scholarship JAEPre-07-00176 (AP), the Comunidad de Madrid grant HEPHACOS P-ESP-00346, by the EU Research Training Network *Constituents, Fundamental Forces and Symmetries of the Universe* MRTN-CT-2004-005104, the Spanish Consolider-Ingenio 2010 program CPAN CSD2007-00042 and by the [*Fondo Social Europeo*]{} through an I3P-doctores scholarship (PM). PM wishes to thank J. Hartong, C. Herdeiro, D. Klemm, T. Ortín and S. Vaulà for fruitful discussions. [9]{} D. Kastor and J. H. Traschen, Phys. Rev.  D [**47**]{} (1993) 5370 \[hep-th/9212035\]. K.P. Tod, Phys. Lett.  B [**121**]{} (1983) 241. D. Kastor and J.H. Traschen, Class. Quant. Grav.  [**13**]{} (1996) 2753 \[gr-qc/9311025\]. M. Hübscher, P. Meessen, T. Ortín and S. Vaulà, JHEP [**0809**]{} (2008) 099 \[arXiv:0806.1477\]; Phys. Rev.  D [**78**]{} (2008) 065031 \[arXiv:0712.1530\]. L. Andrianopoli, M. Bertolini, A. Ceresole, R. D’Auria, S. Ferrara, P. Frè and T. Magri, J. Geom. Phys.  [**23**]{} (1997) 111 \[hep-th/9605032\]. D.Z. Freedman, C. Nunez, M. Schnabl and K. Skenderis, Phys. Rev.  D [**69**]{} (2004) 104027 \[hep-th/0312055\]. L.A.J. London, Nucl. Phys.  B [**434**]{} (1995) 709. T. Shiromizu, Prog. Theor. Phys.  [**102**]{} (1999) 1207 \[hep-th/9910176\]. K. Behrndt and M. Cvetič, Class. Quant. Grav.  [**20**]{} (2003) 4177 \[hep-th/0303266\]. J. Grover, J.B. Gutowski, C.A.R. Herdeiro and W. Sabra, Nucl. Phys.  B [**809**]{} (2009) 406 \[arXiv:0806.2626\]. J. Gillard, U. Gran and G. Papadopoulos, Class. Quant. Grav.  [**22**]{} (2005) 1033 \[hep-th/0410155\]. J.P. Gauntlett, J B. Gutowski, C.M. Hull, S. Pakis and H.S. Reall, Class. Quant. Grav.  [**20**]{} (2003) 4587 \[hep-th/0209114\]. J.P. Gauntlett and J.B. Gutowski, Phys. Rev.  D [**68**]{} (2003) 105009 \[Erratum [*ibid.*]{}  D [**70**]{} (2004) 089901\] \[hep-th/0304064\]. K. Behrndt, D. Lüst and W.A. Sabra, Nucl. Phys. B [**510**]{} (1998) 264 \[hep-th/9705169\]; G. Lopes Cardoso, B. de Wit, J. Kappeli and T. Mohaupt, JHEP [**0012**]{} (2000) 019 \[hep-th/0009234\]; P. Meessen and T. Ortín, Nucl. Phys.  B [**749**]{} (2006) 291 \[hep-th/0603099\]; M. Hübscher, P. Meessen and T. Ortín, Nucl. Phys.  B [**759**]{} (2006) 228 \[hep-th/0606281\]. M.M. Caldarelli and D. Klemm, JHEP [**0309**]{} (2003) 019 \[hep-th/0307022\]; S.L. Cacciatori, M.M. Caldarelli, D. Klemm and D.S. Mansi, JHEP [**0407**]{} (2004) 061 \[hep-th/0406238\]; S.L. Cacciatori, M.M. Caldarelli, D. Klemm, D.S. Mansi and D. Roest, JHEP [**0707**]{} (2007) 046 \[arXiv:0704.0247\]. S.L. Cacciatori, D. Klemm, D.S. Mansi and E. Zorzan, JHEP [**0805**]{} (2008) 097 \[arXiv:0804.0009\]; D. Klemm and E. Zorzan, arXiv:0902.4186. K. Skenderis, P.K. Townsend and A. Van Proeyen, JHEP [**0708**]{} (2007) 036 \[arXiv:0704.3918\]. K. Skenderis and P.K. Townsend, Phys. Rev. Lett.  [**96**]{} (2006) 191301 \[hep-th/0602260\]; J. Phys. A [**40**]{} (2007) 6733 \[hep-th/0610253\]. K. Pilch, P. van Nieuwenhuizen and M.F. Sohnius, Commun. Math. Phys.  [**98**]{} (1985) 105. E. Cremmer, C. Kounnas, A. Van Proeyen, J.P. Derendinger, S. Ferrara, B. de Wit and L. Girardello, Nucl. Phys.  B [**250**]{} (1985) 385. J. Bellorin and T. Ortin, Phys. Lett.  B [**616**]{} (2005) 118 \[hep-th/0501246\]. J. Bellorín, P. Meessen and T. Ortín, Nucl. Phys.  B [**762**]{} (2007) 229 \[hep-th/0606201\]. P. Gauduchon and K.P. Tod, J. Geom. Phys. [**25**]{}(1998), 291-304. V. Buchholz. J. Geom. Phys. [**35**]{}(2000), 93-98 \[math.DG/9912115\]; Suppl. Rend. Circ. Mat. di Palermo [**63**]{}(2000), 63 \[math.DG/9901125\]. M. Dunajski and P. Tod, Differ. Geom. Appl.  [**14**]{} (2001) 39 \[math/9907146\]. D.M.J. Calderbank and P. Tod, Differ. Geom. Appl.  [**14**]{} (2001) 199–208 \[math/9911121\]. P. Meessen, Phys. Lett.  B [**665**]{} (2008) 388 \[arXiv:0803.0684\]. D. Kastor and J.H. Traschen, Class. Quant. Grav.  [**19**]{} (2002) 5901 \[hep-th/0206105\]. A. Ashtekar and B. Krishnan, Living Rev. Rel.  [**7**]{} (2004) 10 \[gr-qc/0407042\]. T. Liko and I. Booth, Class. Quant. Grav.  [**25**]{} (2008) 105020 \[arXiv:0712.3308\]; I. Booth and T. Liko, Phys. Lett.  B [**670**]{} (2008) 61 \[arXiv:0808.0905\]. G.W. Gibbons and C.N. Pope, Class. Quant. Grav.  [**25**]{} (2008) 125015 \[arXiv:0709.2440\]. A.G. Walker, Quart. J. Math. Oxford [**1**]{}(1950), 69–79. H. Nariai, Sci. Rept. Tohoku Univ. [**34**]{}(1950), 160; Sci. Rept. Tohoku Univ. [**35**]{}(1951), 62. J.P. Gauntlett, R.C. Myers and P.K. Townsend, Phys. Rev.  D [**59**]{} (1999) 025001 \[hep-th/9809065\]; J.M. Figueroa-O’Farrill, Class. Quant. Grav.  [**16**]{} (1999) 2043 \[hep-th/9902066\]. Y. Kosmann-Schwarzbach, Annali di Mat. Pura Appl. [**91**]{}(1972), 317–395; T. Ortín, Class. Quant. Grav.  [**19**]{} (2002) L143 \[hep-th/0206159\]. K.P. Tod, Class. Quant. Grav.  [**12**]{} (1995) 1801. E. Lozano-Tellechea, P. Meessen and T. Ortín, Class. Quant. Grav.  [**19**]{} (2002) 5921 \[hep-th/0206200\]. A.A. Coley, G.W. Gibbons, S. Hervik and C.N. Pope, Class. Quant. Grav.  [**25**]{} (2008) 145017 \[arXiv:0803.2438\]. J.B. Gutowski and W.A. Sabra, arXiv:0903.0179. [^1]: In this article we will be following the conventions of ref. [@Hubscher:2008yz], which in its turn are adapted from those of ref. [@Andrianopoli:1996cm]. Specifically this means that the metric is mostly-minus, the $\gamma$-matrices are purely imaginary and the spinors are chiral, with $\gamma_{5}\epsilon^{I}=\epsilon^{I}$ and $\gamma_{5}\epsilon_{I}=-\epsilon_{I}$. As $\gamma_{5}=-i\gamma^{0123}$ is purely imaginary, the above chirality assignment is compatible with the convention of raising and lowering $I$-indices by complex conjugation. [^2]: The model used by Shiromizu can be seen as a truncation of a model with prepotential $\mathcal{F}=-i/2\mathcal{X}^{0}\mathcal{X}^{1}$ and $\mathtt{C}_{1}=0$, the meaning of which will be explained in section (\[sec:DSVector\]). His solutions can be obtained from the results in section (\[sec:VectBil\]). [^3]: In the notation that we will follow throughout this article, $\mathbb{D}$ will be the total connection, whereas we will reserve $\mathtt{D}$ for the connection without the $\mathbb{R}$-part and $\mathfrak{D}$ for the Kähler-connection, [*i.e.*]{} the connection appearing in ungauged supergravity. [^4]: As we are using the same conventions as ref. [@Meessen:2006tu], we can copy their arguments as they stand. [^5]: These expressions were derived in ref. [@Bellorin:2006xr] starting from a prepotential and using the obvious homogeneity of the symplectic section $\mathcal{R}$. The derivation presented here is far less involved and also holds in situations where no prepotential exists. [^6]: Observe that this is a purely non-Abelian restriction as hyperCR/GT-metrics are known, see [*e.g.*]{} [@Dunajski:1999qs] [^7]: Let us in passing point out that in the resulting Kasner spaces there is a time-like conformal isometry of the kind used in ref. [@Kastor:2002fu] to define a conformal energy. [^8]: See appendix (\[sec:NullCurv\]) for the spin-connection and curvatures for this tetrad. [^9]: Needless to say, this reasoning also holds for the ordinary gauged $N=2$ $d=4$ supergravities with potentials whose FI-contribution vanishes.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We show that the discrete Sinkhorn algorithm, as applied in the setting of Optimal Transport on a compact manifold, converges, in a large-scale limit, to the solution of a parabolic PDE of Monge-Ampère type. The latter evolution equation has previously appeared in different contexts (in particular, on the torus it can be be identified with the Ricci flow). This leads to an algorithmic approximation of the potential of the optimal transport map (or equivalently, certain geometric Monge-Ampère equations) with explicit bounds on the time-complexity of the construction and the approximation errors. As applications we obtain explicit schemes of nearly linear complexity for optimal transport on the torus and the two-sphere, as well as the far-field antenna problem. Connections to Quasi-Monte Carlo methods are also pointed out.' author: - 'Robert J. Berman' title: 'The Sinkhorn algorithm, parabolic optimal transport and geometric Monge-Ampère equations' --- Introduction ============ The theory of Optimal Transport [@v1; @v2] is used in a multitude of applications ranging from economy, statistics, cosmology, geometric optics and meterology to more recent applications in data processing (see the surveys [@s-d---; @so]). In the last years there has been a flurry of numerical work applying the Sinkhorn algorithm [@s] (aka the Iterative Proportional Fitting Procedure [@r]) as a fast and efficient way of computing approximations to optimal transport maps, or equivalently, solutions to certain geometric Monge-Ampère type equations. This is motivated by applications to machine learning [@cu] (concerning optimal transport in Euclidean $\R^{n})$ and computer graphics and image processing [@s-d---] (where the general setting of optimal transport on Riemannian manifolds is considered). The key advantage of the Sinkhorn algorithm in this context is its favorable large-scale computational properties (parallelization, quadratic complexity, linear time convergence, etc [@b-g-c-N]). The main aim of the present paper is to show that, in the large-scale limit, the Sinkhorn algorithm converges towards the solution of a parabolic PDE of Monge-Ampère type, which, incidentally, previously has appeared in [@s-s; @k-s-w; @ki] and is called the *parabolic optimal transport equation* in [@k-s-w]. The convergence is shown with explicit error estimates. This leads, in particular, to an algorithmic approximation of the potential of the optimal transport map with explicit bounds on the time-complexity of the construction and the approximation errors introduced by the discretization. Background and setup -------------------- ### \[subsec:The-Sinkhorn-algorithm\]The Sinkhorn algorithm Let $p$ and $q$ be two vectors in $\R_{+}^{n}$ whose entries sum to one. Given any matrix $K\in\R_{+}^{N}\times\R_{+}^{N}$ there exists, by Sinkhorn’s theorem [@s], two diagonal positive matrices $D_{a}$ and $D_{b}$ with diagonal vectors $a$ and $b$ in $\R_{+}^{N}$ such that the matrix $$B:=D_{b}KD_{a}$$ has the property that the rows sum to $p$ and the columns sum to $q.$ The construction is essentially unique (i.e. up to scaling $D_{a}$ and $D_{b}^{-1}$ by a positive number) and $B$ can be obtained as the limit of the process of alternately normalizing the rows and columns of the matrix. In other words, $$B=\lim_{m\rightarrow\infty}B(m),\,\,\,B(m)=D_{a(m)}KD_{b(m)},$$ where the pair of positive vectors $(a(m),b(m))$ are defined by the following recursion, formulated in terms of matrix vector multiplications and component-wise division of vectors: $$a(m+1)=\frac{q}{K^{T}b(m)}$$ $$b(m)=\frac{p}{Ka(m)}$$ with initial data $a(0)$ taken as the vector with entries $1.$ In fact, any initial positive vector $a(0)$ will do and $a(m)$ and $b(m)$ even converge with any need of scaling, as follows from Theorem \[thm:conv of u m in general setting\]). The same algorithm has appeared in various fields. In its most general (infinite dimensional) form, known as the Iterative Proportional Fitting Procedure in the statistics literature, the roles of $p$ and $q$ are played by probability measures on two (possible non-finite) topological spaces. In this setting the corresponding convergence of $B(m)$ towards a limit $B$ was established in [@r] using a maximum entropy characterization of $B,$ which in the finite setting above says that $B$ is the unique element realizing the infimum $$\inf_{\gamma\in\Pi(p,q)}\mathcal{I}(\gamma|K)$$ where $\mathcal{I}$ denotes the Kullback-Leibler divergence of $\gamma$ relative $K,$ when $\gamma$ and $K$ are identified with measures on the discrete product $\{1,...,N\}^{2}$ and $\Pi(p,q)$ denotes the set of all matrices $\gamma$ in $\R_{+}^{N}\times\R_{+}^{N}$ with row sum $p$ and column sum $q,$ i.e. the corresponding measures on $\{1,...,N\}^{2}$ have marginals $p$ and $q,$ respectively ($-\mathcal{I}(\gamma|K)$ is the corresponding “physical” entropy). An alternative proof of the convergence follows from Theorem \[thm:conv of u m in general setting\] below, which also shows that $a(m)$ and $b(m)$ have unique limits (determined by the initial value $a(0)).$ ### \[subsec:Discrete-optimal-transport\]Discrete optimal transport {#subsec:Now-replace-} Now replace $K$ with a family of matrices $K_{\epsilon}$ of the form $$(K_{\epsilon})_{ij}=e^{-\epsilon^{-1}C_{ij}},$$ for a given matrix $C_{ij}$ and parametrized by a positive number $\epsilon.$ Then the corresponding matrix $B_{\epsilon}\in\Pi(p,q)$ furnished by Sinkhorn’s theorem converges, as $\epsilon\rightarrow0$ to a matrix $B_{0}$ realizing the infimum $$\mathcal{C}:=\inf_{\gamma\in\Pi(p,q)}\left\langle C,\gamma\right\rangle$$ In the terminology of discrete optimal transport theory [@v1; @m-p] this means that $B_{0}$ is an *optimal transport plan (coupling) between $p$ and $q,$ with respect to the cost matrix $C.$* The convergence follows from noting that, the maximum entropy characterization of $B_{\epsilon}$ shows that $B_{\epsilon}$ realizes the perturbed minimum $$\mathcal{C}_{\epsilon}:=\inf_{\gamma\in\Pi(p,q)}\left\langle c,\gamma\right\rangle +\epsilon\mathcal{I}(\gamma|I).$$ While the matrix $B_{0}$ is sparse (and is typically supported on the graph of a transport map) the approximation $B_{\epsilon}$ always has full support (by Sinkhorn’s theorem) and is thus more “regular” than $B_{0}.$ Accordingly, the small parameter $\epsilon$ is sometimes referred to as the *entropic regularization parameter.* This is illustrated by the simulations in [@b-g-c-N] for the case when $p$ and $q$ represent the discretization of two probability measures on the unit-interval in $\R,$ using a large number $N$ of points and with $C_{ij}$ the cost matrix defined by to the squared distance function on $\R.$ When $\epsilon$ is taken to be of the order $1/N$ [@b-g-c-N Fig 1] shows how the discrete probability measures on $\R\times\R$ appear as smoothed out versions of the graph of the corresponding optimal transport map. It should also be pointed out that the entropy minimization problem above goes back to the work by Schrödinger on Quantum Mechanics in the 30s [@schr] (see the survey [@le], where the connection to optimal transport is emphasized). ### \[subsec:Discretization-of-Optimal\]Discretization of Optimal Transport on the torus Let now $X$ be a compact manifold (without boundary) endowed with a Riemannian distance function $d(x,y).$ In applications one typically discretizes the manifold by a fixing $N$ points on $X,$ i.e. a “point cloud”. One then fixes an entropic regularization parameter $\epsilon,$ whose size depends on $N$ (and is typically taken to be of the order of the “spatial resolution” of $X$ [@s-d---]). To keep things as simple as possible we will start by taking the manifold $X$ to be the $n-$dimensional torus $$T^{n}:=(\frac{\R}{\Z})^{n}$$ endowed with the standard distance function $d(x,y),$ induced from the Euclidean distance function on $\R^{n}.$ Let $\mu$ and $\nu$ be two probability measure on $T^{n}$ (which thus correspond to two periodic measures on $\R^{n})$ with $C^{2}-$smooth and strictly positive densities $e^{-f}$ and $e^{-g},$ respectively [^1]: $$\mu=e^{-f}dV,\,\,\,\nu=e^{-g}dV\label{eq:mu and nu in terms of f and g intro}$$ where $dV$ is the Riemannian normalized volume form on $T^{n}.$ As is well-known, a map $F$ transporting (i.e. pushing forward) $\mu$ to $\nu$ is *optimal* with respect to the cost function $d(x,y)^{2}$ iff it can be expressed in terms of a potential $u\in C^{4}(T^{n}):$ $$F(x):=x+\nabla u(x),\,\,\,T^{n}\rightarrow T^{n},$$ which is strictly *quasi-convex* in the sense that symmetric matrix $\nabla^{2}u+I$ is positive definite: $$\nabla^{2}u+I>0$$ (we identify $u$ with a $\Z^{n}-$periodic function on $\R^{n}$ so that the gradient $\nabla u$ on $\R^{n}$ descends to define a self-map of $T^{n}).$ The function $u$ is uniquely determined, up to an additive constant, by the following Monge-Ampère equation $$\exp(-g(x+\nabla u(x))\det(I+\nabla^{2}u(x))=\exp(-f(x))\label{eq:MA eq intro}$$ A standard way to discretize this setting is to replace the torus $T^{n}$ with a regular point set $\Lambda_{k}$ in $T^{n}$ such that the corresponding “grid” has edge lengths $k^{-1},$ where $k$ is a positive integer, i.e. $\Lambda_{k}$ is the discrete torus defined by $$\Lambda_{k}:=(\frac{(k^{-1}\Z)}{\Z})^{n}\subset T^{n}$$ Let $\{x_{i}^{(k)}\}_{i=1}^{N_{k}}$ be an enumeration of the points in $\Lambda_{k}$ The total number $N_{k}$ of points is thus given by $$N_{k}=k^{n}$$ Denote by $p^{(k)}$ and $q^{(k)}$ the corresponding discrete approximations in $\R^{N_{k}}$ of $\mu$ and $\nu,$ defined by the normalized values of the densities of $\mu$ and $\nu$ at the points in $\Lambda_{k}.$ Defining a sequence of $N_{k}\times N_{k}$ matrices $K^{(k)}$ by $$K_{ij}^{(k)}:=\mathcal{K}^{(k)}(x_{i}^{(k)},x_{j}^{(k)}),\,\,\,\mathcal{K}^{(k)}(x,y):=e^{-kd(x,y)^{2}/2}\label{eq:def of A k ij intro}$$ and applying Sinkhorn’s theorem to the triple $(K^{(k)},p^{(k)},q^{(k)})$ furnishes two positive vectors $a^{(k)}$ and $b^{(k)}$ in $\R^{N_{k}},$ uniquely determined by the normalization condition that $a_{i_{k}}^{(k)}=0$ for the index $i_{k}$ corresponding to the point $x_{i_{k}}=0$ in $\Lambda_{k}.$ \[subsec:Main-results-in-the torus\]Main results in the torus setting --------------------------------------------------------------------- We will investigate the joint limit where the entropic regularization parameter $\epsilon$ coincides with the edge length in the discretization scheme above: $$\epsilon:=k^{-1}$$ Our first result shows that the potential $u$ for the optimal transport problem between $\mu$ and $\nu$ can be recovered from the positive vectors $a^{(k)}$ and $b^{(k)},$ furnished by the Sinkhorn theorem: \[thm:conv in static torus setting intr\](Static case) If $x_{i_{k}}^{(k)}$ is a sequence of points in the discrete torus $\Lambda_{k}$ converging to the point $x$ in the torus $T^{n},$ as $k\rightarrow\infty,$ then $$-\lim_{k\rightarrow\infty}k^{-1}\log a_{i_{k}}^{(k)}=u(x)$$ where $u$ is the unique optimal transport potential solving the Monge-Ampère equation \[eq:MA eq intro\] and normalized by $u(0)=0.$ Moreover, the convergence is uniform with respect to $k.$ This convergence result should come at no surprise and it holds in a very general setting (see Theorems \[Thm:weak conv of fixed points towards optimal transport plans\]\[thm:Static conv in Riem setting\]). But the main point of the present paper is that the Sinkhorn algorithm itself, when viewed as a discrete dynamical system for the positive vectors $a_{i_{k}}^{(k)},$ also admits a continuous large-scale limit $u_{t}(x),$ evolving according to the following parabolic PDE: $$\frac{\partial u_{t}(x)}{\partial t}=\log\det(I+\nabla^{2}u_{t}(x))-g(x+\nabla u_{t}(x))+f(x).\label{eq:parabolic eq intro}$$ The existence of a $C^{4}-$smooth solution $u_{t},$ given strictly quasi-convex $C^{2}-$smooth initial data $u_{0}$ follows from the results in [@s-s]. In order to prove the convergence we first observe that the function $$u_{m}^{(k)}(x_{i_{k}})=-k^{-1}\log a_{i_{k}}^{(k)}(m)$$ on the discrete torus $\Lambda_{k}$ admits a canonical extension, defining a quasi-convex function on $X:$ $$u_{m}^{(k)}(x):=k^{-1}\log\sum_{i=1}^{N_{k}}e^{-kd(x,y_{i}^{(k)})^{2}/2}b_{i}^{(k)}(m-1)q_{i}^{(k)}$$ expressed in terms of a Fourier/Gauss type sum, with $k$ playing the role of the band-with. \[thm:conv in dynamic torus setting intro\](Dynamic case) For any sequence of discrete times $m_{k}$ such that $m_{k}/k\rightarrow t$ we have $$\lim_{k\rightarrow\infty}u_{m}^{(k)}=u_{t}$$ uniformly on $T^{n},$ where $u_{t}$ is the smooth and strictly quasi-convex solution of the parabolic PDE \[eq:parabolic eq intro\] with initial data $u_{0}=0.$ More precisely, $$\sup_{T^{n}}\left|u_{m}^{(k)}-u_{m/k}\right|\leq C\frac{m}{k}k^{-1},$$ where the constant $C$ depends explicitly on upper bounds on the $C^{2}-$norms of $f$ and $g$ and a strict positive lower bound on $f$ and $g.$ More generally, if at the initial discrete time $m=0$ $$u_{0}^{(k)}=u_{0},$$ where $u_{0}$ is the restriction to the discrete torus $\Lambda_{k}$ of a fixed $C^{4}-$smooth and strictly quasi-convex function on the torus $T^{n},$ then the corresponding result still holds. Using that $u_{t}$ converges exponentially, as $t\rightarrow\infty,$ to a potential $u$ for the optimal transport problem [@k-s-w] we thus arrive at the following \[cor:conv of explicit apprl intro\] (Constructive approximation) There exists a positive constant $A$ such that at the discrete time $m_{k}=[Ak\log k]$ the functions $u_{k}(x):=u_{m_{k}}^{(k)}(x_{i_{k}})$ converge uniformly to the optimal transport potential $u(x)$ as $k\rightarrow\infty.$ More precisely, $$\sup_{T^{n}}\left|u_{k}-u\right|\leq Ck^{-1}\log k\label{eq:estimate in cor torus intro}$$ Moreover, the discrete probability measures $\gamma_{k}$ on $T^{n}\times T^{n},$ determined by the Sinkhorn algorithm, converge weakly towards the corresponding optimal transport plan $(I\times(\nabla u+I))_{*}\mu$ concentrating exponentially on the graph $\Gamma$ of the transport map $\nabla u+I:$ $$\gamma_{k}\leq k^{p}e^{-kd_{\Gamma}^{2}/p}dx\otimes dx,\label{eq:estimate for gamma k cor torus intro}$$ for some positive constant $p,$ where $d_{\Gamma}$ denotes the distance to the graph $\Gamma$ in $T^{n}\times T^{n}.$ Since each iteration in the Sinkhorn algorithm may be formulated in terms of matrix-vector multiplication, which requires $O(N^{2})$ arithmetic operators, and since $k=N^{1/n},$ the direct construction of the function $u_{k}$ uses, in general, $O(N^{2+1/n})\log N$ elementary arithmetic operations. Moreover, in the present case of the torus the matrix-vector operations in question are discrete convolutions and can thus be performed using merely $O(N)(\log N)$ arithmetic operations, by using the Fast Fourier Transform (see Section \[subsec:Optimal-transport-on linear\]). By symmetry, Theorem \[thm:conv in dynamic torus setting intro\] also shows that, on the one hand, the functions $$v_{m_{k}}^{(k)}:=-k^{-1}\log b^{(k)}(m_{k})$$ converge, as $k\rightarrow\infty,$ towards the solution $v_{t}$ of the parabolic equation obtained by interchanging the roles of $\mu$ and $\nu.$ On the other hand, by Lemma \[lem:dens prop\], the function $v_{m_{k}}^{(k)}$ is equal to the Legendre transform (in the space variable) of $u_{m_{k}}^{(k)},$ up to negligable $O(\log k)k$ error term. Thus Theorem \[thm:conv in dynamic torus setting intro\] is consistent, as it must, with the fact that the Legendre transform of the solution $u_{t}$ of equation \[eq:parabolic eq intro\] solves the parabolic equation obtained by interchanging the roles of $\mu$ and $\nu$ (as can be checked by a direct calculation). Generalizations to compact manifolds ------------------------------------- The result in the static setting is shown to hold in a very general setting of optimal transport between two probability measures $\mu$ and $\nu$ defined on compact topological spaces $X$ and $Y,$ respectively (see Theorem \[Thm:weak conv of fixed points towards optimal transport plans\] and also Theorem \[thm:non-cpt static\] which holds in the classical (non-compact) Euclidean setting where $X$ is equal to $\R^{n}).$ Then the roles of the positive vectors $p^{(k)}$ and $q^{(k)},$ discretizing $\mu$ and $\nu,$ are played by two sequences $\mu^{(k)}$ and $\nu^{(k)},$ satisfying certain density properties with respect to $\mu$ and $\nu,$ which are almost always satisfied in practice. Moreover, the cost function $c$ is merely assumed to be continuous and can even be replaced by any sequence $c_{k}$ converging uniformly to $c,$ as the inverse $k$ of the entropic regularization parameter tends to infinity (which applies, in particular, to the convolutional Wasserstein distances introduced in [@s-d---], where the matrix in formula \[eq:def of A k ij intro\] is replaced by a heat kernel; see Section \[subsec:Application-to-convolutional\]). The corresponding result in the dynamic setting (see Theorem \[thm:dynamic general\]) requires that $X$ and $Y$ be compact Riemannian manifold (without boundary) and further regularity assumptions on $c,$ $\mu$ and $\nu.$ In particular, a “local density property” on the approximations $\mu^{(k)}$ and $\nu^{(k)}$ is required, roughly saying that the approximations of $\mu$ and $\nu$ hold up to length scales of the order $k^{-1/2}$ point, with an $O(1/k)-$error term. Interestingly, the local density properties turn out to be satisfied when the approximations $\mu^{(k)}$ and $\nu^{(k)}$ are defined by (weighted) point clouds, generated using suitable *Quasi-Monte Carlo methods* in numerical analysis [@b-e-g; @b-s-s-w; @b-c-c-g-s-t] (as explained in Section \[subsec:Sequences-of-measures\]). Then the entropic regularization parameter $k^{-1}$ is comparable to the covering radius (mesh norm) of the corresponding point clouds [@b-e-g]. This provides a theoretic justification for taking the entropic regularization parameter in the Sinkhorn algorithm to be comparable to the “spatial resolution” of the corresponding discretization scheme (as was done in the numerical simulations on Riemannian manifolds in [@s-d---]). The general results are applied to the case of the round sphere endowed with the two different cost functions: (1) $d(x,y)^{2}$ and (2) $-\log|x-y|.$ These two cases appear, for example, in applications to (1) computer graphics (texture mapping), medical imaging [@d-t; @z; @et; @al] and mesh adaptation for global whether and climate prediction [@w-b-b-c] and (2) the reflector antenna problem in geometric optics [@waII; @g-o]. Nearly linear complexity of the corresponding Sinkhorn iteration is achieved in both cases, using fast transforms. Relation to previous results ---------------------------- To the best of the authors knowledge these are the first convergence results concerning the Sinkhorn algorithm (and its fixed points) in the limit when the number $N$ of points and the inverse of the regularization parameter $\epsilon$ *jointly* tend to infinity (see [@c-d-p-s] and references therein for the static case when only $\epsilon^{-1}$ tends to infinity in the Euclidean $\R^{n}-$ setting and [@le1; @le] for a very general setting). This kind of joint limit is, in practice, what is studied in numerical simulations (see for example [@s-d---]) and the convergence analysis in the present paper thus provides a theoretical bases for the simulations. We do not adress questions of numerical stability (in connection to floating point-arithmetic) and stabilization and instead refer to [@c-d-p-s Section 4] and the in-depth study in [@sc] (where various numerical speed-ups are explored). The convergence in Corollary \[cor:conv of explicit apprl intro\] should be compared with previous results in the rapidly growing literature on numerical approximations schemes for solutions to Optimal Transport problems and (generalized) Monge-Ampère equations (satifying a second boundary value condition). However, the author is not aware of any previous results providing both complexity bounds (in terms of $N)$ *and* a quantified rate of convergence of the error of the approximate solution, as $N\rightarrow\infty.$ We recall that the oldest approach is to approximate the optimal transport potential $u$ by solving the linear program which is dual to the discrete optimal transport problem. This can be done in about $O(N^{3})$ time-complexity (see, for example, the exposition in [@Br-F-H-L...], where applications to cosmology are given). Another influential approach is the Benamou-Brenier approach, using computational fluid mechanics [@b-b] (numerical experiments suggests that it has $O(N^{3})$ time-complexity, as pointed out in [@b-d]). There is also a rapidly expanding literature on other discretization schemes in the PDE and numerics literature (mainly in the flat case of $\R^{n}$ and the torus). We refer the reader to the introduction of the very recent work [@b-d] where references to the main two numerical branches are provided (the semi-discrete approach e.g. [@k-m-t] and finite difference methods e.g. [@s-a-k]), which, experimentally, show nearly linear complexity. A survey of numerical methods for general fully non-linear partial differential equations, with an emphasis on Monge-Ampère equations, is given in [@f-g-n]. See also the recent work [@l-r] for a new convex optimization approach and [@s-w] for numerical simulation (based on finite difference schemes) of the analog of the parabolic equation in the case of bounded convex Euclidean domains. One advantage of the present framework over finite difference and finite element type approaches, when applied to general manifolds, is that it is meshfree. In other words, it does not require generating a grid or a polyhedral tessellation of the manifolds, but only a suitable point cloud, which can be efficently generated using Quasi-Monte Carlo methods. In the case of the round sphere various different numerical algorithms have previously been explored in the literature: see [@d-t; @z; @et; @al; @w-b-b-c] for experimental work on the case of the cost function $d(x,y)^{2}$ and [@c-k-o; @de-m-t] for the case of the cost function $-\log|x-y|$, as applied to the reflector antenna problem in geometric optics. The present results are very much inspired by an analogous setup which appears in complex (Kähler) geometry. Briefly, the role of the Sinkhorn algorithm is then played by Donaldson’s iteration [@do], whose fixed points are called balanced metrics and $k$ appears as the power of a given ample line bundle over $X.$ From this point of view Theorem \[thm:conv in static torus setting intr\] (and its generalization Theorem \[Thm:weak conv of fixed points towards optimal transport plans\]) is the analog of [@bbgz Thm B] and Theorem \[thm:conv in dynamic torus setting intro\] is the analog of the result in [@Ber] showing that Donaldson’s iteration converges to the Kähler-Ricci flow [@ca]. In fact, identifying the real torus $T^{n}$ with a reduction of the complex torus $X:=\C^{n}/(\Z^{n}+i\Z^{n})$ (or more generally, an abelian variety) the parabolic flow \[eq:parabolic eq intro\] can, in the case when $g=0,$ be identified with a twisted Kähler-Ricci flow whose stationary solutions are Kähler potentials solving the corresponding complex Monge-Ampère equation (known as the Calabi-Yau equation in this context).[^2] A new feature of the theoretical analysis in the present paper, compared to the usual situation in Kähler geometry, is that the target measure $\mu$ is taken to be discrete and depend on $k$ (in practise, such discretizations are used in implementations of Donaldson’s iteration, such as the experimental work [@d-k-l-r], motivated by String Theory). Interestingly, the density condition on the the sequence $\mu^{(k)}$ appearing in Lemma \[lem:dens prop\] can be viewed as a real analog of the Bernstein-Markov property for a sequence $\mu_{k},$ as studied in the complex geometric and pluripotential theoretic setting (see the discussion on page 8 in [@b-b-w]). The relations between the real and complex settings will be expanded on elsewhere. Acknowledgements ---------------- I am grateful to Klas Modin for many discussions and, in particular, for drawing my attention to the recent numerical work on the applications of the Sinkhorn algorithm to Optimal Transport, which was the starting point of the present paper. Also thanks to Gabriel Peyré for providing me with further references. Lacking background in Numerics and Optimal Transport I would like to apologize for any omission in accrediting results properly. This work was supported by grants from the ERC and the KAW foundation. Organization ------------ In section \[sec:General-setup-and\] a general setting for iterations on $C(X),$ generalizing the Sinkhorn algorithm, is introduced. The iteration in question, which is determined by a triple $(\mu,\nu,c),$ is essentially equivalent to the Iterative Proportional Fitting Procedure and the results in Section \[sec:General-setup-and\] are probably more or less well-known (expcept perhaps Theorem \[thm:conv of u m in general setting\]). But one point of the presentation is to exploit the variational structure. It can be viewed as a real analogue of the formalism introduced in [@Ber], in the setting of Donaldson’s iteration [@do] and it lends itself to various generalizations of the optimal transport problem (such as Monge-Ampère equations with exponential non-linearities). The main results in the torus setting, stated in the introduction above, are proved in Section \[sec:Proofs-of-the\]. In fact, the general form of the static result given in Theorem \[thm:conv in static torus setting intr\] is proved, as it requires not much more work then in the torus setting. Applications to convolutional Wasserstein distances are also given. Then, in the following Section \[sec:Generalizations-and-outlook\], the dynamic result in Theorem \[thm:conv in dynamic torus setting intro\] is generalized from the torus setting to a rather general setting of optimal transport on compact manifolds, where the connection to the general parabolic transport equation introduced in [@k-s-w] is made. In Section \[sec:Nearly-linear-complexity\] it is shown that nearly linear complexity can be achieved in the case of optimal transport on the torus and the sphere (which applies, in particular, to the reflector antenna problem). Section \[sec:Outlook\] gives an outlook on relations to singularity formation in the parabolic optimal transport equations. In the appendix a proof of a discrete version of the classical stationary phase approximation is provided. \[sec:General-setup-and\]General setup and preliminaries ======================================================== If $Z$ is a compact topological space then we will denote by $C(Z)$ the space of continuous functions on $Z$ endowed with the sup-norm and by $\mathcal{P}(Z)$ the space of all (Borel) probability measures on $Z,$ endowed with the weak topology. Given a subset $S$ of $Z$ we will denote by $\chi{}_{S}$ the function which is equal to $1$ on $S$ and infinity otherwise. Throughout the paper we will assume given a triple $(\mu,\nu,c)$ where $\mu$ and $\nu$ are probability measures on compact topological spaces $X$ and $Y,$ respectively and $c$ is a continuous function on $X\times Y.$ The support of $\mu$ and $\nu$ will be denote by $X_{\mu}$ and $Y_{\nu},$ respectively. Given $u\in C(X)$ and $v\in C(Y)$ we will, abusing notation slightly, identify $u$ and $v$ with their pull-backs to $X\times Y.$ \[subsec:Recap-of-Optimal\]Recap of Optimal Transport and the $c-$Legendre transform ------------------------------------------------------------------------------------- Let us start by recalling the standard setup for optimal transport (see the book [@v1] for further background). A probability measure $\gamma\in\mathcal{P}(X\times Y)$ is said to be a *transport plan (or coupling) between $\mu$ and $\nu$* if the the push forwards of $\gamma$ to $X$ and $Y$ are equal to $\mu$ and $\nu,$ respectively. The subspace of all such probability measures in $\mathcal{P}(X\times Y)$ will be denote by $\Pi(\mu,\nu).$ A transport plan in $\Pi(\mu,\nu)$ is said to be *optimal wrt the cost function $c,$* if it realizes the following infimum: $$\inf_{\gamma\in\Pi(\mu,\nu)}\int_{X\times Y}c\gamma$$ By weak compactness such an optimal transport plan always exists. The *$c-$Legendre transform $u^{c}$* of a function $u\in C(X)$ is defined as the following function in $C(Y)$ $$u^{c}(y):=\sup_{x\in X}-c(x,y)-u(x).$$ A function $u\in C(X)$ is said to be *$c-$convex* if $$(u^{c})^{c}=u$$ Equivalently, $u$ is $c-$convex iff there exists some $v\in C(Y)$ such that $u=v^{c}.$ Indeed, this follows from the observation that $u^{ccc}=u$ for any $u\in C(X),$ which in turn follows from $u^{cc}\leq u.$ \[prop:optim crit\](“Optimality criterion”) A transport plan $\gamma\in\Pi(\mu,\nu)$ is optimal iff there exists $u\in C(X)$ which is $c$- convex and such that $\gamma$ is supported in $$\Gamma_{u}:=\left\{ (x,y)\in X\times Y:\,u(x)+u^{c}(y)+c(x,y)=0\right\} \label{eq:def of Gamma u}$$ This is standard and known as the Knott-Smith optimality criterion (in the Euclidean setting) [@v1]. For completeness we provide the simple proof of the direction that we shall use later on. Consider the following functional on $C(X)$ (often called the *Kantorovich functional*): $$J(u):=\int u\mu+\int u^{c}\mu.\label{eq:Kant functional}$$ Since $u+u^{c}+c\geq0$ on $X\times Y$ the following lower bound holds for any given $\gamma\in\Pi(\mu,\nu)$ $$\int_{X\times Y}c\gamma\geq-\inf_{u\in C(X)}J(u)\label{eq:lower bd in pf prop opt crit}$$ Now, if $\gamma$ is supported in $\Gamma_{u}$ it follows directly that $\int_{X\times Y}c\gamma=-J(u)$ and hence $\gamma$ attains the lower bound above, i.e. $\gamma$ is optimal. A byproduct of Theorem \[Thm:weak conv of fixed points towards optimal transport plans\] below (applied to the case when $\mu_{k}=\mu$ and $\nu_{k}=\nu$ for all $k)$ is a proof that there always exists a transport plan $\gamma_{*}$ with support in $\Gamma_{u}$ for some $c-$convex function. Since $\gamma_{*}$ saturates the lower bound \[eq:lower bd in pf prop opt crit\] it follows that taking the infimum over all $\gamma$ in $\Pi(\mu,\nu)$ yields equality in \[eq:lower bd in pf prop opt crit\] . As a consequence, $$\inf_{\gamma\in}\int_{X\times Y}c\gamma=\sup_{\Phi_{c}}\int u\mu+\int v\nu,\,\,\,\Phi_{c}:=\left\{ (u,v)\in C(X)\times C(Y):\,u+v\leq c\right\} \label{eq:Kant duality}$$ This is the content of Kantorovich duality which is usually shown using Rockafeller-Fenchel duality in topological vector spaces [@v1]. ### The torus setting In the case when $X=Y=T^{n}:=(\frac{\R}{\Z})^{n}$ we will take $c(x,y)$ to be half the squared standard distance function on $T^{n}:$ $$c(x,y):=\frac{1}{2}d(x,y)^{2}:=\frac{1}{2}\inf_{m\in\Z^{n}}|x+m-y|^{2}$$ \[lem:quasi conv\]Let $u(x)$ be a smooth function on $T^{n}.$ Then the following conditions are equivalent and will be referred to as strict *quasi-convexity:* - $u(x)$ is strictly $c-$convex - $u$ satisfies the following inequality on $T^{n}:$ $$\nabla^{2}u+I>0$$ - The map $$x\mapsto y_{x}:=x+(\nabla u)(x)\label{eq:c gradient map torus}$$ is a diffeomorphism of $T^{n}.$ Moreover, the previous conditions on $u$ are equivalent to the corresponding conditions for $u^{c}.$ The corresponding map $$y\mapsto x_{y}:=y+(\nabla u^{c})(y)\label{eq:def of x y}$$ is the inverse of the map \[eq:c gradient map torus\] and the following matrix relation holds $$(\nabla^{2}u+I)(x)^{-1}=(\nabla^{2}u^{c}+I)(y_{x})\label{eq:matrix relation for u}$$ This is well-known and can be deduced from the corresponding classical properties of the ordinary Legendre transform $\phi\mapsto\phi^{*}$ on $\R^{n}$ corresponding to $c(x,y):=-x\cdot y$ (see for example [@v1 Section 2.1.3]). Indeed, setting $\phi(x):=u(x)+|x|^{2}/2$ it follows directly from the definitions that $\phi^{*}(x)=u^{c}(x)+|x|^{2}/2$ (compare formula \[eq:inf u equal to inf u\] below). We will also have use for the following \[lem:smooth\]Assume that $u$ is $C^{1}-$smooth and strictly quasi-convex. Then, for any fixed $y\in T^{n},$ the unique infimum of the function $x\mapsto d(x,y)^{2}/2+u(x)$ is attained at $x=x_{y}$ (defined by formula \[eq:def of x y\]). Moreover, the function $x\mapsto d(x,y)^{2}/2$ is smooth on some neighborhood of $x_{y}$ and its Hessian is equal to the identity there. Given two points $\bar{x}$ and $\bar{y}$ in $T^{n}$ we can identify them with two points in $\R^{n}$ in the fundamental domain $[0,1]^{n}$ for the $\Z^{n}-$action such that $$d(\bar{x},\bar{y})^{2}:=\inf_{m\in\Z^{n}}|\bar{x}+m-\bar{y}|^{2}=|\bar{x}-\bar{y}|^{2}$$ We claim that, under the assumptions of the lemma, the inf above is uniquely attained at $m=0$ when $\bar{x}=\bar{x}_{\bar{y}}:$ $$m\neq0\implies|\bar{x}+m-\bar{y}|^{2}>|\bar{x}-\bar{y}|^{2}.$$ To see this we identify $u$ with a periodic function on $\R^{n}$ and note that $$\inf_{\bar{x}\in T^{n}}d(\bar{x},\bar{y})^{2}/2+u(\bar{x})=\inf_{x\in\R^{n}}|x-y|^{2}/2+u(x)\label{eq:inf u equal to inf u}$$ The inf in the left hand side above is attained at $\bar{x}_{\bar{y}}$ and hence so is the inf in the right hand side. Now assume, to get a contradiction, that the claim above does not hold, i.e. there exists a non-zero $m\in\Z^{n}$ such that $|\bar{x}_{\bar{y}}+m-\bar{y}|=|\bar{x}_{\bar{y}}-\bar{y}|.$ This implies that the inf in the right hand side in formula \[eq:inf u equal to inf u\] is attained both at $\bar{x}_{\bar{y}}$ and at $\bar{x}_{\bar{y}}+m$ (since $u$ is periodic). But this contradicts the fact that the function $x\mapsto|x-y|^{2}/2+u(x)$ is strictly convex on $\R^{n}$ (by the quasi-convexity of $u$ on $T^{n}).$ Finally, the claim shows, since the inequality in the claim is preserved when $\bar{x}$ is perturbed slighly, that $d(\bar{x},\bar{y})^{2}=|\bar{x}-\bar{y}|^{2}$ for all $\bar{x}$ sufficently close to $\bar{x}_{\bar{y}}.$ Hence, $x\mapsto d(x,y)^{2}/2$ is smooth there and its Hessian is constant. The iteration on $C(X)$ ----------------------- In this section we will consider an iteration on $C(X),$ which can be viewed as a generalization of the Sinkhorn algorithm (see Section\[subsec:Discretization-and-the\]). Given data $(\mu,\nu,c),$ as in Section \[subsec:Recap-of-Optimal\], we first introduce the following maps $$T_{\mu}:\,C(X)\rightarrow C(Y),\,\,\,u\mapsto v[u]:=\log\int e^{-c(x,\cdot)-u(x)}\mu(x)$$ and $$T_{\nu}:\,C(Y)\rightarrow C(X),\,\,\,v\mapsto u[v]:=\log\int e^{-c(\cdot,y)-v(y)}\nu(y)$$ (abusing notation slightly we will write $T_{\nu}(u)=v[u]$ etc). This yields an iteration on $C(X)$ defined by $$u_{m+1}:=S[u_{m}],\label{eq:iteration for u}$$ where $S$ is defined as the the composed operator $T_{\nu}\circ T_{\mu}$ on $C(X):$ $$S:\,C(X)\rightarrow C(X),\,\,\,u\mapsto u[v[u]]$$ It will be convenient to rewrite the iteration \[eq:iteration for u\] for $u_{m}$ as the following difference equation: $$u_{m+1}-u_{m}=\log(\rho_{u_{m}}),\label{eq:difference eq}$$ where $\rho_{u}$ is defined by $$\rho_{u}:=e^{S[u]-u}\label{eq:def of pho u in terms of S}$$ and has the property that $\rho_{u}\mu$ is a probability measure on $X$ (as follows directly from the definitions). ### Existence and uniqueness of fixed points Consider the following functional $\mathcal{F}$ on $C(X):$ $$\mathcal{F}:=I_{\mu}-\mathcal{L}_{\nu},\,\,\,I_{\mu}(u)=\int_{X}u\mu,\,\,\,\mathcal{L}(u):=-\int_{Y}v[u]\nu\label{eq:def of F and L}$$ Note that $I_{\mu}$ and $\mathcal{L}_{\mu}$ are equivariant under the additive action of $\R$ and hence $\mathcal{F}$ is invariant. The following is equivalent: - $u$ is a critical point for the functional $\mathcal{F}$ on $C(X)$ - $\rho_{u}=1$ a.e. with respect to $\mu$ Moreover, if $u$ is a critical point, then $u_{*}:=S(u)$ is a fixed point for the operator $S$ on $C(X)$ First observe that that the differential of the functional $\mathcal{L}$ defined in formula \[eq:def of F and L\], at an element $u\in C(X),$ is represented by the probability measure $\rho_{u}\mu,$ where $\rho_{u}$ is defined by formula \[eq:def of pho u in terms of S\]. This means that for any $\dot{u}\in C(X)$ $$\frac{d(\mathcal{L}(u+t\dot{u}))}{dt}|_{t=0}=\int\dot{u}\rho_{u}\mu$$ This follows readily from the definitions by differentiating $t\mapsto v[(u+t\dot{u})]$ to get an integral over $(X,\mu)$ and then switching the order of integration. As a consequence, $u$ is a critical point of the functional $\mathcal{F}$ on $C^{0}(X)$ iff $\rho_{u}\mu=\mu,$ i.e. iff $\rho_{u}=1$ a.e. with respect to $\mu.$ Finally, if this is the case then $S(u)=u$ a.e wrt $\mu$ and hence $S(S(u))=S(u)$ (since $S(f)$ only depends on $f$ viewed as an element in $L^{1}(X,\mu)).$ The following basic compactness property holds: \[lem:compactness of function spaces\]Given a point $x\in X$ the subset $\mathcal{K}_{x_{0}}$ of $C(X)$ defined as all elements $u$ in the image of $S$ satisfying $u(x_{0})=0$ is compact in $C(X).$ First observe that, since $X\times Y$ is assumed compact, the continuous function $c$ is, in fact, uniformly continuous on $X.$ Hence, it follows from the very definition of $S$ that $S(C(X))$ is an equicontinuous family of continuous functions on $X.$ By Arzela-Ascoli theorem that the set $\mathcal{K}_{x_{0}}$ is thus compact in $C(X).$ Using the previous two lemmas gives the following \[prop:exist and uniq of fix point\]The operator $S$ has a fixed point $u_{*}$ in $C(X).$ Moreover, $u_{*}$ is uniquely determined a.e. wrt $\mu$ up to an additive constant. More precisely, there exists a unique fixed point in $S(C(X))/\R.$ We start by noting that $$\mathcal{F}(Su)\leq\mathcal{F}(u)$$ (this is shown in the first step of Theorem \[thm:conv of u m in general setting\] below). Since $\mathcal{F}$ is invariant under the natural $\R-$action we conclude that $$\inf_{C(X)}\mathcal{F}=\inf_{\mathcal{K}_{0}}\mathcal{F},$$ where $\mathcal{K}_{0}$ denotes the compact subset of $C(X)$ appearing in Lemma \[lem:compactness of function spaces\]. Since $\mathcal{F}$ is clearly continuous on $C(X)$ this implies the existence of a minimizer of $\mathcal{F}$ which is moreover in $\mathcal{K}_{0}.$ Next observe that $\mathcal{F}$ is convex on $C(X).$ Indeed, for any fixed $y\in Y,$ $v\mapsto v[u](y)$ is convex on $C(Y),$ as follows directly from Jensen’s inequality. Hence, $-\mathcal{L}_{\nu}$ is convex and since $I_{\mu}$ is affine we conclude that $\mathcal{F}$ is convex. More precisely, Jensen’s (or Hölder’s) inequality implies that $\mathcal{F}$ is strictly convex on $C(X)/\R$ viewed as a subset of $L^{1}(\mu)/\R.$ Hence, if $u_{0}$ and $u_{1}$ are two minimizers, then there exists a constant $C$ such that $u_{0}=u_{1}+C$ a.e. wrt $\mu.$ In particular, if $C=0$ then $u_{*}:=S(u_{0})=S(u_{1})$ gives the same fixed point of $S.$ ### Monotonicity and convergence properties of the iteration We next establish the following result, which can be seen as a refinement, in the present setting, of the convergence of the general Iterative Proportional Fitting Procedure established in [@r]. The result will be used in the proof of Proposition \[prop:generalized parabolic\]. \[thm:conv of u m in general setting\]Given $u_{0}\in C(X)$ the corresponding iteration $u_{m}:=S^{m}u_{0}$ converges uniformly to a fixed point $u_{\infty}$ of $S.$ *Step 1: $I_{\mu}$ and $-\mathcal{L}$ are decreasing along the iteration and hence $\mathcal{F}$ is also decreasing. The functionals are strictly decreasing at $u_{m}$ unless $S(u_{*})=u_{*}$ for $u_{*}:=S(u_{m}).$* Using the difference equation \[eq:difference equation\] for $u_{m}$ and Jensen’s inequality, we have $$I_{\mu}(u_{m+1})-I_{\mu}(u_{m})=\int\log\rho_{u_{m}}\mu\leq\log\int\rho_{u_{m}}\mu=\log1=0$$ Moreover, equality holds unless $\rho_{u_{m}}=1$ a.e wrt $\mu$ i.e. $S(u_{m})=u_{m}$ and $S(u_{*})=u_{*}$ everywhere on $X.$ Similarly, by symmetry, $$\mathcal{L}(u_{m})-\mathcal{L}(u_{m+1})=\int\log\rho_{v_{m}}\nu\leq\log\int\rho_{v_{m}}\nu=\log1=0,$$ where now $\rho_{v},$ for $v\in C(Y),$ denotes the probability measure on $Y$ defined as in formula \[eq:def of pho u in terms of S\], but with the roles of $\mu$ and $\nu$ interchanged. *Step 2: Convergence in $C(X)/\R.$* Given the initial data $u_{0}$ we denote by $\mathcal{K}_{u_{0}}$ the closure of the orbit of $u_{0}$ in $C(X).$ By Lemma \[lem:compactness of function spaces\] $\mathcal{K}_{u_{0}}/\R$ is compact in $C(X)/\R.$ Hence, after perhaps passing to a subsequence, $u_{m}\rightarrow u_{\infty}$ in $C(X)/\R.$ Now, since $\mathcal{F}$ is decreasing along the orbit we have $$\mathcal{F}(u_{\infty})=\inf_{\mathcal{K}_{0}}\mathcal{F}.$$ Hence, by the condition for strict monotonicity it must be that $Tu_{\infty}=u_{\infty}$ a.e. wrt $\mu$ and hence, since $u_{\infty}$ is the image of $S,$ it follows that $Tu_{\infty}=u_{\infty}$ on all of $X.$ It then follows from Proposition \[prop:exist and uniq of fix point\] that $u_{\infty}$ is uniquely determined in $C(X)/\R$ (by the initial data $u_{0}),$ i.e. the whole sequence converges in $C(X)/\R.$ *Step $3:$ Convergence in $C(X)$* Let us first show that there exists a number $\lambda\in\R$ such that $$\lim_{m\rightarrow\infty}I_{\mu}(u_{m})=\lambda.\label{eq:limit is lambda}$$ By Step $1$ $I_{\mu}$ is decreasing and hence it is enough to show that $I_{\mu}(u_{m})$ is bounded from below. But $I_{\mu}=\mathcal{F}+\mathcal{L}_{\mu},$ where, by Prop \[prop:exist and uniq of fix point\] (or the previous step) $\mathcal{F}$ is bounded from below (by $\mathcal{F}(u_{\infty}))$. Moreover, by the first step $\mathcal{L}_{\mu}(u_{m})\geq\mathcal{L}_{\mu}(u_{0}),$ which concludes the proof of \[eq:limit is lambda\]. Next, decompose $$u_{m}=\tilde{u}_{m}+u_{m}(x_{0}),\,\,\,$$ By Lemma \[lem:compactness of function spaces\] the sequence $(\tilde{u}_{m})$ is relatively compact in $C(X)$ and we claim that $|u_{m}(x_{0})|\leq C$ for some constant $C.$ Indeed, if this is not the case then there is a subsequence $u_{m_{j}}$ such that $|u_{m_{j}}|\rightarrow\infty$ uniformly on $X.$ But this contradicts that $I_{\mu}(u_{m})$ is uniformly bounded (by \[eq:limit is lambda\]). It follows that the sequence $(u_{m})$ is also relatively compact. Hence, by the previous step the whole sequence $u_{m}$ converges to the unique minimizer $u_{*}$ of $\mathcal{F}$ in $S(C(X))$ satisfying $I_{\mu}(u_{*})=\lambda.$ The convergence result in [@r] is, in the present setting, equivalent to convergence of the iteration on $C(X)/\R$ induced by $u_{n}.$ In fact, the latter convergence holds at linear rate, i.e. there exists a metric $d$ on $C(X)/\R$ and a positive number $\delta$ such that $d(u_{m},u_{\infty})\leq e^{-\delta m}.$ Indeed, setting $d(u,u'):=\left\Vert \sup(u-u')-\inf(u-u')\right\Vert _{C(X)}$ (which correponds, under $u\mapsto e^{-u},$ to the Hilbert metric on the cone of positive functions in $C(X))$ this follows from Birkhoff’s theorem about positive operators on cones, precisely as in the finite dimensional situation of the Sinkhorn iteration considered in [@f-l]; see also [@m-p Thm 4.2]. ### The induced discrete evolution on $C(X)\times C(Y)$ Fixing an initial data $u_{0}\in C(X)$ the corresponding evolution $m\mapsto u_{m}$ induces a sequence of pairs $(u_{m},v_{m})\in C(X)\times C(Y)$ defined by the following recursion: $$(u_{m+1},v_{m+1}):=(u[v_{m+1}],v[u_{m}])$$ ### The induced discrete evolution on $\mathcal{P}(X\times Y)$ and entropy Let us briefly explain the dual point of view involving the space $\mathcal{M}(X\times Y)$ of measures on $X\times Y$ (which, however, is not needed for the proofs of the main results). The data $(\mu,\nu,c)$ induces the following element $\gamma_{c}\in\mathcal{M}(X\times Y):$ $$\gamma_{c}:=e^{-c}\mu\otimes\nu$$ Given a function $u\in C(X)$ we will write $$\gamma_{u}:=e^{-(u+v[u])}\gamma_{c}\label{eq:def of gamma u}$$ (so that $\gamma_{0}=\gamma_{c}).$ \[lem:-u fixed iff gamma u is in Pi\]$u$ satisfies $S(u)=u$ a.e. wrt $\mu$ iff $\gamma_{u}\in\Pi(\mu,\nu).$ This follows immediately from observing $$\int_{X}e^{-(u+v[u])}\gamma_{c}=\nu,\,\,\,\,\int_{Y}e^{-(u+v[u])}\gamma_{c}=\rho_{u}\mu$$ The discrete dynamical system $u_{m}$ induces a sequence $$\gamma_{m}:=\gamma_{u_{m}}(=e^{-u_{m}(x)}e^{-v_{m}(x)}\gamma_{c})\in\mathcal{P}(X\times Y)$$ \[prop:The-unique-minimizer\]The unique minimizer $\gamma_{*}$ of the functional $\mathcal{I}(\cdot|\gamma_{c})$ on $\Pi(\mu,\nu)$ is characterized by the property that it has the form $$\gamma_{_{*}}=e^{-\Phi}\gamma_{c}$$ for some $\Phi\in C(X)+C(Y).$ Moreover, $\gamma_{_{*}}=\gamma_{u_{*}}$, where $u_{*}$ is a fixed point for $S$ on $C(X)$ (or more generally, on $L^{1}(X,\mu))$ and $$\inf_{\Pi(\mu,\nu)}\mathcal{I}(\cdot|\gamma_{c})=\inf_{C(X)\times C(Y)}\mathcal{F}\label{eq:inf of I is inf of F}$$ and given any function $u_{0}\in C(X),$ the corresponding sequence $\gamma_{m}$ converges in $L^{1}$ (i.e. in variation norm) towards $\gamma_{*}$ (and moreover $\mathcal{I}(\gamma_{m}|\gamma_{*})\rightarrow0).$ By construction $\gamma_{*}:=\gamma_{u_{*}}$ has the property that $$\gamma_{_{*}}=e^{-\Phi}\gamma_{c},\,\,\,\gamma_{_{*}}\in\Pi(\mu,\nu)$$ for some $\Phi\in L^{\infty}(X)+L^{\infty}(Y).$ But a standard calculus argument reveals that any such $\gamma_{*}$ is the unique minimizer of the restriction of $\mathcal{I}$ to the affine subspace $\Pi(\mu,\nu)$ of $\mathcal{P}(X\times Y$) (using that $\mathcal{I}$ is strictly convex). The last convergence statement then follows directly from Theorem \[thm:conv of u m in general setting\] (only the easier convergence in $C(X)/\R$ is needed). Rewriting $$k^{-1}\mathcal{I}(\gamma|\gamma_{kc})=\int c\gamma+k^{-1}\mathcal{I}(\gamma|\mu\otimes\nu),$$ the equality \[eq:inf of I is inf of F\] can be viewed as an entropic variant of Kantorovich duality \[eq:Kant duality\] in the limit when $c$ is replaced by $kc$ for a large positive number $k.$ In fact, it follows from Theorem \[Thm:weak conv of fixed points towards optimal transport plans\] applied to $\mu_{k}=\mu$ and $\nu_{k}=\nu$ that $$\inf_{\gamma\in\Pi(\mu,\nu)}k^{-1}\mathcal{I}(\gamma|\gamma_{kc})=\inf_{\gamma\in\Pi(\mu,\nu)}\int c\gamma=\sup_{\Phi_{c}}\int u\mu+\int v\nu,$$ as in the Kantorovich duality \[eq:Kant duality\]. In the next section we will consider the setting where $\mu$ and $\nu$ also change with $k.$ \[subsec:The-parametrized-setting\]The $k-$parametrized setting and discretization ---------------------------------------------------------------------------------- Let us next consider the following variant of the previous setting, parametrized by a parameter $k$ (which is the parameter that will later on tend to infinity and which corresponds to the entropic regularization parameter $\epsilon:=k^{-1}$). This means that we replace the triple $(\mu,\nu,c)$ with a sequence $(\mu^{(k)},\nu^{(k)},kc).$ As explained in Section \[subsec:Discrete-optimal-transport\] replacing $c$ with $kc$ corresponds to introducing the entropic regularization parameter $\epsilon=k^{-1}.$ We then rescale the functions in $C(X)$ and $C(Y)$ by $k$ and consider the corresponding rescaled operators: $$v^{(k)}[u]:=k^{-1}v[ku]=k^{-1}\log\int e^{-kc(x,\cdot)-ku(x)}\mu^{(k)}(x)$$ $$u^{(k)}[v]:=k^{-1}u[kv]=k^{-1}\log\int e^{-kc(\cdot,y)-kv(y)}\nu^{(k)}(y)$$ $$S^{(k)}(u):=k^{-1}S(ku)$$ etc. The corresponding rescaled iteration is thus defined by the iteration $$u_{m}^{(k)}:=S^{(k)}u_{0}^{(k)}\in C(X),\label{eq:scaled iteration}$$ given the initial value $u_{0}^{(k)}\in C(X).$ Equivalently, $$u_{m+1}^{(k)}-u_{m}^{(k)}=k^{-1}\log(\rho_{ku_{m}^{(k)}}),\label{eq:difference equation}$$ where $$\rho_{ku}(x)=\int_{Y}\frac{e^{-kc(x,y)-ku(x)}}{\int_{X}e^{-kc(x',y)-ku(x')}\mu^{(k)}(x')}v^{(k)}(y)\label{eq:explicit expres for pho}$$ By Theorem \[thm:conv of u m in general setting\], as $m\rightarrow\infty$ the iteration $u_{m}^{(k)}$ converges in $C(X)$ to a fixed point $u^{(k)}$ of the operator $S^{(k)}$ (uniquely determined by the initial value $u_{0}^{(m)})).$ We observe that the following compactness property holds (and is proved exactly as in Lemma \[lem:compactness of function spaces\]): \[lem:compactness of function with k\]The union $\bigcup_{k\geq0}S^{(k)}$ is relatively compact in $C(X)/\R$ (identifying $C(X)/\R$ is identified with the set of all continuous functions vanishing at a given point $x_{0})$ ### \[subsec:Discretization-and-the\]Discretization and the Sinkhorn algorithm Now assume that $\mu^{(k)}$ and $\nu^{(k)}$ are discrete probability measures whose supports are finite sets $$X^{(k)}:=\{x_{i}^{(k)}\}_{i=1}^{N_{k}},\,\,\,Y^{(k)}:=\{y_{i}^{(k)}\}_{i=1}^{N_{k}}$$ of the same number $N_{k}$ of points in $X$ and $Y,$ respectively. This means that there exist vectors $p^{(k)}$ and $q^{(k)}$ in $\R^{N_{k}}$ such that $$\mu^{(k)}=\sum_{i=1}^{N_{k}}\delta_{x_{i}^{(k)}}p^{(k)},\,\,\,\nu^{(k)}=\sum_{i=1}^{N_{k}}\delta_{x_{i}^{(k)}}q^{(k)}.$$ Moreover, since $\mu^{(k)}$ and $\nu^{(k)}$ are probability measures the vectors $p^{(k)}$ and $q^{(k)}$ are elements in the simplex $\Sigma_{N_{k}}$ in $\R^{N_{k}}$ defined by $$\Sigma_{N}:=\left\{ v\in\R^{N}:v_{i}\geq0,\,\,\,\sum_{i=1}^{N}v_{i}=1\right\} ,$$ which we identify with $\mathcal{P}(\{1,...,N\}).$ Similarly, we identify the discrete measure $$\gamma_{c}^{(k)}:=e^{-kc}\mu^{(k)}\otimes\nu^{(k)}$$ on $X\times Y$ with the matrix $\tilde{A}\in\R^{N_{k}}\times\R^{N_{k}}$ defined by $$\tilde{K}_{ij}:=K_{ij}^{(k)}p_{i}^{(k)}q_{j}^{(k)},\,\,\,K_{ij}^{(k)}:=\exp(-kC_{ij}),\,\,\,C_{ij}:=c(x_{i}^{(k)},x_{j}^{(k)}),$$ where $C_{ij}$ is viewed as a cost function on $\{1,...,N\}^{2}.$ Under the identifications $$C(X^{(k)})\leftrightarrow\R_{+}^{N_{k}},\,\,\,u\mapsto a,\,\,\,a_{i}:=e^{-ku}p_{i}^{(k)}$$ and $$C(Y^{(k)})\leftrightarrow\R_{+}^{N_{k}},\,\,\,v\mapsto b,\,\,\,b_{i}:=e^{-kv}q_{i}^{(k)}$$ the scaled iteration \[eq:scaled iteration\] gets identified with the recursion $a^{(k)}(m)$ defined by the Sinkhorn algorithm determined by the matrix $K^{(k)}$ and the positive vectors $p^{(k)}$ and $q^{(k)}$ (see Section \[subsec:The-Sinkhorn-algorithm\]). Given an initial positive vector $a^{(k)}(0)$ Theorem \[thm:conv of u m in general setting\] thus shows that $(a^{(k)}(m),b^{(k)}(m))$ converges, as $m\rightarrow\infty,$ to a pair of positive vectors $(a^{(k)},b^{(k)})$ such that the scaled matrix $D_{b}K^{(k)}D_{a}$ has the property that the rows sum to $p^{(k)}$ and the columns sum to $q^{(k)}.$ \[rem:Fourier\]By construction, the functions $u_{k}^{(m)}(x)$ on $X$ can be expressed in terms of a Fourier type sum: $$u_{k}^{(m)}(x)=k^{-1}\log\sum_{i=1}^{N_{k}}e^{-kc(x,y_{i}^{(k)})}b_{i}^{(k)}(m-1)q_{i}^{(k)}$$ where the “Fourier coefficients” $b_{i}^{(k)}(m-1)q_{i}^{(k)}$ are given by the Sinkhorn algorithm. In the non-compact setting when $X=Y=\R^{n},$ with $c(x,y)=-x\cdot y,$ this is the analytic continuation to $i\R^{n}$ of a bona fide Fourier sum with fourier coefficients in $k$ times the support of $\nu^{(k)}.$ Hence, $k$ plays the role of the “band-with”. \[sec:Proofs-of-the\]Proofs of the main results in the torus setting ==================================================================== Proof of Theorem \[thm:conv in static torus setting intr\] (and generalizations) -------------------------------------------------------------------------------- Throughout the section we will consider the parametrized setting in Section \[subsec:The-parametrized-setting\] and assume that the sequences $\mu^{(k)}$ and $\nu^{(k)}$ converge to $\mu$ and $\nu$ in $\mathcal{P}(X)$ and $\mathcal{P}(Y),$ respectively. We will denote by $u^{(k)}$ the fixed point of the corresponding operator $S^{(k)}$ on $C(X),$ uniquely determined by the normalization condition $u^{(k)}(x_{0})=0,$ at a given point $x_{0}$ in $X.$ Similarly, $v^{(k)}$ denotes the corresponding fixed point obtained by reversing the roles of $\mu$ and $\nu.$ We start by giving a density condition on $\mu^{(k)}$ ensuring that $v^{(k)}[u]$ converges uniformly to the $c-$Legendre transform $u^{c}$ of $u,$ when $\mu$ has full support: \[lem:dens prop\]Assume that the sequence $\mu^{(k)}$ converging to $\mu$ in $\mathcal{P}(X)$ has the following “density property”: for any given open subset $U$ intersecting the support $X_{\mu}$ of $\mu$ $$\liminf_{k\rightarrow\infty}k^{-1}\log\mu^{(k)}(U)\geq0\label{eq:density property}$$ Then, for any given $u\in C(X),$ the sequence $v^{(k)}[u]$ converges uniformly to $(\chi_{X_{\mu}}u)^{*,c}$ in $C(Y).$ Replacing the integral over $\mu^{(k)}$ with a sup directly gives $$v^{(k)}[u](y)\leq(\chi_{X_{\mu}}u)^{c}(y)$$ for any $y\in Y.$ To prove a reversed inequality let $x_{y}$ be a point in $X_{\mu}$ where the sup defining $(1_{X_{\mu}}u)^{c}(y)$ is attained and $U_{\delta}$ a neighborhood of $x_{y}$ where the oscillation of $c(\cdot,y)+u$ is bounded from above by $\delta$ (by compactness/continuity $U_{\delta}$ can be taken to be independent of $y).$ Then $$v^{(k)}[u](y)\geq k^{-1}\log\int_{U_{\delta}}e^{-k(c(x,y)+u(x))}\mu^{(k)}(x)\geq k^{-1}\log\mu^{(k)}(U_{\delta})+u^{c}(y)-\delta$$ Hence, as $k\rightarrow\infty,$ $v^{(k)}[u](y)\rightarrow u^{c}(y)$ and since $v^{(k)}[u]$ is equicontinuous (by the assumed compactness of $X\times Y$ and the continuity of $c)$ this implies the desired uniform convergence. \[exa:The-density-property\](Weighed point clouds). If $\mu_{k}=\mu$ for any $k$ then the density property is trivially satisfied. More generally, the density property \[eq:density property\] is satisfied by any reasonable approximation $\mu^{(k)}.$ For example, in the discrete case where $\mu^{(k)}=\sum_{i=1}^{N_{k}}w_{i}^{(k)}\delta_{x_{i}^{(k)}}$ the property in question holds as long as $\sup_{i}w_{i}^{(k)}$ and the inverse of the number of points $x_{i}^{(k)}$ in any given open set $U$ intersecting $X_{\mu}$ have sub-exponential growth in $k.$ \[Thm:weak conv of fixed points towards optimal transport plans\]Suppose that $\mu^{(k)}\rightarrow\mu$ and $\nu^{(k)}\rightarrow\mu$ in $\mathcal{P}(X)$ and $\mathcal{P}(Y),$ respectively and assume that $\mu^{(k)}$ and $\nu^{(k)}$ satisfy the density property \[eq:density property\]. Let $u^{(k)}$ be a fixed point for the scaled operator $S^{(k)}$ on $C(X).$ Then, after perhaps passing to a subsequence, the following holds: $$u^{(k)}\rightarrow u$$ uniformly on $X,$ where $u$ is a $c-$convex minimizer of the Kantorovich functional $J$ (formula \[eq:Kant functional\]) satisfying $$u=(\chi_{Y_{\nu}}(\chi_{X_{\mu}}u)^{c})^{c}$$ As a consequence, the corresponding probability measures $$\gamma^{(k)}:=e^{-k(u^{(k)}+v^{(k)})}\mu^{(k)}\otimes\nu^{(k)}\in\mathcal{P}(X\times Y)$$ converge weakly to a transport plan $\gamma$ between $\mu$ and $\nu,$ which is optimal wrt the cost function $c.$ *Step 1: Convergence of a subsequence of $u^{(k)}$* In the following all functions will be normalized by demanding that the values vanish at a given point. By Lemma \[lem:compactness of function with k\] we may assume that $u^{(k)}\rightarrow u^{(\infty)}$ uniformly on $X,$ for some element $u^{(\infty)}$ in $C(X).$ Now, by the previous lemma, for any given $u\in C(X)$ we have, with $J(u)$ defined by formula \[eq:Kant functional\], $$\mathcal{F}^{(k)}(u)=J(u)+o(1).\label{eq:F k asympt to J}$$ Now take a sequence $\epsilon_{k}$ of positive numbers tending to zero such that $$u^{(\infty)}-\epsilon_{k}\leq u^{(k)}\leq u^{(\infty)}+\epsilon_{k}\label{eq:u k compard to u infy}$$ Since $u\mapsto v^{(k)}[u]$ is decreasing it follows that $$\mathcal{F}^{(k)}(u^{(\infty)})-2\epsilon_{k}\leq\mathcal{F}^{(k)}(u^{(k)})+2\epsilon_{k}\label{eq:ineq for F k}$$ By definition, $\mathcal{F}(u^{(k)})\leq\mathcal{F}(u)$ for any given $u$ in $C(X).$ Hence, combining \[eq:ineq for F k\] and \[eq:F k asympt to J\] and letting $k\rightarrow\infty$ gives $$J(u^{(\infty)})\leq\inf_{u\in C(X)}J(u),$$ showing that $u^{(\infty)}$ minimizes $J$ on $C(X).$ To see that $u^{(\infty)}$ is $c-$convex first recall that, by definition, $u^{(k)}$ satisfies $$u^{(k)}=u^{(k)}[v^{(k)}[u^{(k)}]].$$ Hence, combing \[eq:u k compard to u infy\] with the previous lemma, applied twice, gives $$u^{(k)}=u^{(k)}[(\chi_{X_{\mu}}u^{(\infty)})^{c}]+o(1)=((\chi_{Y_{\mu}}(\chi_{X_{\mu}}u^{(\infty)})^{c})^{c}+o(1)$$ This shows that $u^{(\infty)}=((\chi_{Y_{\mu}}(\chi_{X_{\mu}}u^{(\infty)})^{c})^{c},$ proving that $u^{(\infty)}=f^{c}$ for some $f\in C(Y).$ Hence $u^{(\infty)}$ is $c-$convex. *Step 2: Convergence of $\gamma^{(k)}$ towards an optimizer* By Lemma \[lem:-u fixed iff gamma u is in Pi\] *$\gamma^{(k)}$ is in $\Pi(\mu,\nu).$* Hence, by weak compactness, we may assume that *$\gamma^{(k)}$* converges towards an element $\gamma^{(\infty)}$ in $\mathcal{P}(X\times Y).$ By Prop \[prop:optim crit\] it will thus be enough to show that $\gamma^{(\infty)}$ is supported in $\Gamma_{u^{(\infty)}}.$ To this end let $\Gamma_{\delta}$ be the closed subset of $X\times Y$ where $u+u^{c}\geq\delta>0$ for $u:=u^{(\infty)}.$ By the previous lemma $\gamma^{(k)}\leq e^{-k\delta/2}\mu^{(k)}\otimes\nu^{(k)}$ on $\Gamma_{\delta},$ when $k$ is sufficiently large and hence the limit $\gamma^{(\infty)}$ is indeed supported on $\Gamma_{u^{(\infty)}}.$ Taking the cost function $c(x,y)$ to be the squared distance function on a compact Riemannian manifold we arrive at the following generalization of Theorem \[thm:conv in static torus setting intr\] stated in the introduction (using that the density property is satisfied for the discretization scheme used in Theorem \[thm:conv in static torus setting intr\] , as explained in the example above): \[thm:Static conv in Riem setting\]Let $X$ be a compact Riemannian manifold (possibly with boundary) and set $c(x,y):=d(x,y)^{2}/2,$ where $d$ is the Riemannian distance function. Suppose that $\mu^{(k)}\rightarrow\mu$ and $\nu^{(k)}\rightarrow\mu$ in $\mathcal{P}(X)$ and assume that $\nu$ is absolutely continuous wrt the Riemannian volume form on $X$ and has full support. If $\nu^{(k)}$ satisfies the density property \[eq:density property\], then $v^{(k)}$ converges uniformly in $Y$ to the normalized potential for the unique optimal Borel map transporting $\nu$ to $\mu,$ i.e. the map that can be expressed as $$y\mapsto x_{y}:=\text{exp}_{y}(\nabla v),\label{eq:y mapsto x as exp}$$ (which means that $y_{x}$ is obtained by transporting $x$ along a unit-length geodesic in the direction of $(\nabla v)(y)).$ Moreover, $u^{(k)}$ converges uniformly on $X$ towards the $d^{2}/2-$Legendre transform of $v$ and $x\mapsto\exp_{y}(\nabla u)$ defines the optimal transport of $\mu$ to $\nu.$ This follows from the previous proposition combined with the results in [@mc]. Indeed, it is shown in [@mc] that when $v$ is taken as a minimizer of the Kantorovich functional, then the the map \[eq:y mapsto x as exp\] is well-defined a.e. wrt $\nu$ and the corresponding transport plan is the unique optimal transport plan (the result in [@mc] is a Riemannian version of Brenier’s theorem in $\R^{n}$ [@br]; see also[@c-e] for a direct proof in the torus setting). Since $\nu$ has full support it follows that the minimizer $v$ is uniquely determined modulo additive constants and since $u=v^{*}$ so is the limit of $u^{(k)}.$ The previous theorem applies more generally as soon as a unique Borel optimal map exists (see for example [@v2 Thm 10.38] for conditions on $c$ ensuring that this is the case). ### \[subsec:The-non-compact-setting\]The Euclidean setting in $\R^{n}$ Theorem \[thm:Static conv in Riem setting\] applies, in particular, to the Euclidean setting when $\mu$ and $\nu$ are probability measures on $\R^{n},$ assuming that their supports $X$ and $Y$ are bounded domains (i.e. the closure of open connected subsets). This yields an analog of the recent convergence result [@l-r Thm 1.9], where the role of $u^{(k)}$ is played by the solution to a convex optimization problem formulated using “almost triangulations” of $X$ and $Y$ (assuming that $Y$ is be convex). Moreover, the present result also applies in a non-compact Euclidean setting when $X=\R^{n}$ and gives the following result formulated in terms of the Monge-Ampère measure $MA$ (in the sense of Alexandrov) and the sub-gradient maping $\partial u$ [@v1]: \[thm:non-cpt static\]Assume that $\mu$ and $\nu$ are probability measures on Euclidean $\R^{n}$ such that $\mu$ and $\nu$ have compact support and $\nu$ has convex support $Y$ and a bounded density $g.$ If $\mu^{(k)}$ and $\nu^{(k)}$ satisfy the density property \[eq:density property\], then the (normalized) fixed points $u^{(k)}$ of the corresponding iteration converge uniformly on compacts to a convex function $u$ on $\R^{n}$ satisfying $$(i)\,g(\nabla u)MA(u)=\mu,\,\,\,(i)\,\overline{(\partial u)(\R^{n})}=Y\label{eq:sec bv problem}$$ (where $g(\nabla u)$ is viewed as an $L^{\infty}-$function on $\R^{n})$ In the following $v^{*}$ will denote the classical Legendre transform of $v.$ Applying Theorem \[Thm:weak conv of fixed points towards optimal transport plans\] to a ball $X_{R}$ of radius $R$ containing the support of $\mu$ and the support $Y$ of $\nu$ gives that, after perhaps passing to a subsequence, $v^{(k)}\rightarrow v$ in $C(Y)$ where $v$ is such that the that $L^{\infty}-$map $\nabla v$ has the transport property $(\nabla v)_{*}\mu=\nu$ (note that $v^{(k)}$ and hence $v$ are independent of $R$ since they only depend on the support of $\mu).$ As a consequence, by Brenier’s theorem [@br] (or more generally, by [@mc]) $v$ is uniquely determined in the interior of $Y$ and hence the whole sequence $v^{(k)}$ converges towards $v$ on $Y.$ Next, observe that, since $Y$ is convex the function $\chi_{Y}v,$ is convex on $\R^{n}$ and hence $u:=(\chi_{Y}v)^{*}$ is a convex function on $\R^{n}$ such that $u^{*}=\chi_{Y}v.$ But then it follows from basic convex analysis that conditions $(i)$ and $(ii)$ hold. Moreover, since, by definition, $u_{|X_{R}}=(\chi_{Y}v)_{|X_{R}}^{*}$ it follows from Theorem \[Thm:weak conv of fixed points towards optimal transport plans\] applied to $(X_{R},Y)$ that $u^{(k)}\rightarrow u$ uniformly on $X_{R}.$ Since $R$ can be taken arbitrarily large this concludes the proof of the theorem. The equation \[eq:sec bv problem\] is usually refered to the *second boundary value problem* for the Monge-Ampère operator in $\R^{n}.$ It admits a unique normalized convex solution $u$ (as follows, for example, from the argument using Legendre transforms in the proof of the previous theorem). The previous theorem can be viewed as an analog of [@b-d Thm 5.5], where the role of $u^{(k)}$ is played by the solution to a finite difference type discretization of equation \[eq:sec bv problem\] (in [@b-d] it is assumed that $g$ is constant in order to ensure the existence of a discrete solution). ### \[subsec:Application-to-convolutional\]Application to convolutional Wasserstein distances Theorem \[Thm:weak conv of fixed points towards optimal transport plans\] holds more generally (with essentially the same proof) when the function $c$ is replaced by a sequence $c_{k}$ such that $$\left\Vert c_{k}-c\right\Vert _{L^{\infty}(X\times Y)}\rightarrow0\label{eq:unif conv of ck}$$ For example, in the Riemannian setting of Theorem \[thm:Static conv in Riem setting\]. denoting by $\mathcal{K}_{t}(x,y)$ the corresponding heat kernel and setting $t:=2k^{-1},$ the sequence $$c_{k}:=-t^{-1}\log\mathcal{K}_{t}(x,y)\label{eq:c k as heat}$$ satisfies \[eq:unif conv of ck\], by Varadhan’s formula (which holds more generally on Lipschitz Riemannian manifolds [@no]). Replacing $c$ by $c_{k}$ in this setting thus has the effect of replacing the matrix $A_{ij}:=e^{-kd^{2}(x_{i},x_{j})/2}$ appearing in the corresponding Sinkhorn algorithm with the heat kernel matrix $\mathcal{K}_{2k^{-1}}(x_{i},x_{j})$ which, as emphasized in [@s-d---], has computational advantages. Following [@s-d---] we consider the squared *convolutional Wasserstein distance* between $\mu$ and $\nu:$ $$\mathcal{W}_{(k)}^{2}(\mu,\nu):=k^{-1}\inf_{\gamma\in\Pi(\mu^{(k)},\nu^{(k})}\mathcal{I}(\gamma,\mathcal{K}_{2k^{-1}}\mu^{(k)}\otimes\nu^{(k)}),$$ definied wrt approximations $\mu^{(k)}$ and $\nu^{(k)},$ for example given by weighed point clouds, as in Example \[exa:The-density-property\]. In [@s-d--- Page 3], the problem of developing conditions for the convergence of $\mathcal{W}_{(k)}^{2}(\mu,\nu)$ was posed. The following result provides an answer: \[thm:conv wasserst\]Let $X$ be a compact Riemannian manifold (possibly with boundary) and set $c(x,y):=d(x,y)^{2}/2,$ where $d$ is the Riemannian distance function. Suppose that $\mu^{(k)}\rightarrow\mu$ and $\nu^{(k)}\rightarrow\mu$ in $\mathcal{P}(X)$ and that $\mu^{(k)}$ and $\nu^{(k)}$ satisfy the density property \[eq:density property\]. Then $$\lim_{k\rightarrow\infty}\mathcal{W}_{(k)}^{2}(\mu,\nu)=\mathcal{W}^{2}(\mu,\nu),$$ where $\mathcal{W}^{2}(\mu,\nu)$ denotes the squared $L^{2}-$Wasserstein distance between $\mu$ and $\nu.$ Repeating the argument in the proof of Theorem \[Thm:weak conv of fixed points towards optimal transport plans\], with $c$ replaced by $c_{k}$ as above, gives $$\lim_{k\rightarrow\infty}\inf_{u\in C^{0}(X)}\mathcal{F}^{(k)}=\inf_{u\in C(X)}J(u)$$ According to formula \[eq:inf u equal to inf u\] the infimum appearing in the left hand side above is precisely $\mathcal{W}_{(k)}^{2}(\mu,\nu).$ Since the infimum in the right hand side above is equal to $\mathcal{W}^{2}(\mu,\nu),$ by Kantorovich duality (formula \[eq:Kant duality\]), the result follows. Proof of Theorem \[thm:conv in dynamic torus setting intro\] ------------------------------------------------------------ We will denote by $\delta_{\Lambda_{k}}$ the uniform discrete measure supported on the discrete torus $\Lambda_{k}$ with edge-length $1/k$ and by $\mu^{(k)}$ and $\nu^{(k)}$ the corresponding discretizations of $\mu=e^{-f}dx$ and $\nu=e^{-g}dy.$ We start with the following discrete version of the classical stationary phase approximation, proved in the appendix. \[lem:discrete stat phase\]Let $\alpha$ be a $C^{4}-$smooth function on $T^{n}$ with a unique minimum at $x_{0}$ which is non-degenerate, i.e. $\nabla^{2}\alpha(x_{0})>0.$ Then, if $h$ is $C^{2}-$smooth $$k^{-n/2}\int e^{-k\alpha}f\delta_{\Lambda_{k}}=(2\pi)^{n/2}e^{-k\alpha(x_{0})}\frac{h(x_{0})}{\sqrt{\det(\nabla^{2}\alpha(x_{0}))}}(1+Ck^{-1})$$ where the constant $C$ only depends on an upper bounds on the $C^{4}-$norm of $\alpha,$ the $C^{2}-$norm of $h$ and a strict lower bound on the smallest eigenvalue of $\nabla^{2}\alpha$ close to $x_{0}.$ We next prove the key asymptotic result that will be used in the proof of Theorem \[thm:conv in dynamic torus setting intro\] giving the asymptotics of the function $\rho_{ku}(x),$ explicitly defined by formula \[eq:explicit expres for pho\] (the result can be viewed as a refinement of Lemma \[lem:dens prop\]). \[prop:asympt of pho\]Let $u$ be a smooth and strictly quasi-convex function on $T^{n}.$ Then the following asymptotics hold $$\rho_{ku}(x)=\det(I+\nabla^{2}u(x))e^{f(x)-g(x+\nabla u(x))}(1+Ck^{-1})$$ where the constant $C$ only depends on upper bounds on the $C^{4}-$norm of $u$ and the $C^{2}-$norms of $f$ and $g$ and a strict positive lower bounds on the matrix $(I+\nabla^{2}u(x).$ In the proof we will denote by $\epsilon_{k}$ any sequence of functions satisfying $|\epsilon_{k}|\leq Ck^{-1}.$ First observe that the integral over $\mu^{(k)}(x)$ defining $v^{(k)}[u](y),$ i.e. $$e^{v^{(k)}[u](y))}=\int e^{-k(d(x,y)^{2}/2+u(x))}\mu^{(k)}(x),$$ localizes around a small neighborhood $U$ of $x=x_{y}$ (defined as in Lemma \[lem:quasi conv\]) i.e. the integration may be restricted to $U$ up to introducing an exponentially small error term. Indeed, the unique infimum of the function $x\mapsto d(x,y)^{2}/2+u(x)$ is attained at $x=x_{y}$ and is equal to $u^{c}(y_{x}).$ Hence, we can take $U$ to be defined by all $y$ satisfying $$u^{c}(y)+u(x_{y})<\delta$$ Since the function $x\mapsto d(x,y)^{2}/2+u(x)$ is $C^{4}-$smooth on $U$ (by Lemma \[lem:smooth\]) applying Lemma \[lem:discrete stat phase\] gives $$k^{-n/2}e^{kv[u](y))}=e^{ku^{c}(y)}h(y)(1+\epsilon_{k}(y)),\,\,\,h(y):=(2\pi)^{n/2}\frac{\exp(-f(x_{y}))}{\sqrt{\det(I+\nabla^{2}u(x_{y}))}}.$$ Hence, $$\rho_{ku}(x):=k^{n/2}\int e^{-k(d(x,y)^{2}/2+u(x)+u^{c}(y))}h(y)^{-1}\nu^{(k)}(y)+R_{k}(y),\label{eq:rho in pf prop rhp}$$ where $$R_{k}(y)\leq C\sup\epsilon_{k}(y)k^{n/2}\int e^{-k(d(x,y)^{2}/2+u(x)+u^{c}(y))}\nu^{(k)}(y)$$ Since, by Lemma \[lem:discrete stat phase\], we have $$k^{n/2}\int e^{-k(d(x,y)^{2}/2+u(x)+u^{c}(y))}\nu^{(k)}(y)\leq C$$ it follows that $$R_{k}(y)\leq C^{2}k^{-1}$$ Now, the same localization argument as above shows that the integral over $\nu^{(k)}(y)$ in formula \[eq:rho in pf prop rhp\] localizes around a small neighborhood $V$ of $y=y_{x}.$ Hence, applying Lemma \[lem:discrete stat phase\] again gives $$k^{n/2}\int e^{-k(d(x,y)^{2}/2+u(x)+u^{c}(y))}h(y)^{-1}\nu^{(k)}(y)=h^{-1}(y_{x})\frac{(2\pi)^{n/2}\exp(-g(y_{x}))}{\sqrt{\det(I+\nabla^{2}u^{c}(y_{x}))}}(1+\epsilon_{k}(x)).$$ The proof is now concluded using the inverse properties of the Hessians in Lemma \[lem:quasi conv\]. \[lem:general conv towards parab\]Let $X$ be a compact topological space and consider the following family of difference equations on $C(X),$ parametrized by a positive number $k$ and a discrete time $m:$ $$u_{m+1}^{(k)}-u_{m}^{(k)}=k^{-1}D^{(k)}(u_{m}^{(k)}),\label{eq:differ eq abs}$$ where $D^{(k)}$ is an operator on $C(X),$ which descends to $C(X)/\R$ and with the property that $I+k^{-1}D^{(k)}$ is an increasing operator (wrt the usual order relation on $C(X)).$ Assume that there exists a subset $\mathcal{H}$ of $C(X)$ and an operator $D$ on $\mathcal{H}$ such that for any $u$ in the class $$\left|D^{(k)}(u)-D(u)\right|\leq C_{u}\epsilon_{k},\label{eq:conv of D k}$$ where $C_{u}$ is a positive constant depending on $u$ and $\epsilon_{k}$ is a sequence of positive numbers converging towards zero. Assume that $u_{0|}^{(k)}=u_{0}$ where $u_{0}$ is a fixed function in $\mathcal{H}$ and that there exists a family $u_{t}\in\mathcal{H},$ which is two times differentiable wrt $t$ and solving $$\frac{\partial u_{t}}{\partial t}=D(u_{t}),\,\,\,\,(u_{t})_{|t=0}=u_{0},\label{eq:evol eq for u t abs}$$ for $t\in[0,T].$ Then, for any $(k,m)$ such that $m/k\in[0,T],$ $$\sup_{T^{n}}\left|u_{m}^{(k)}-u_{m/k}\right|\leq C_{T}\frac{m}{k}\max\{\epsilon_{k},k^{-1}\},\,\,\,C_{T}:=\max\{\sup_{t\in[0,T]}C_{u_{t}},\sup_{X\times[0,T]}\frac{\partial^{2}u_{t}}{\partial^{2}t}\}$$ We will write $\psi_{k,m}=u_{m/k}.$ *Step 1: The following holds for all $(k,m)$* $$\sup\left|\psi_{k,m+1}-\psi_{k,m}-k^{-1}D^{(k)}(\psi_{k,m})\right|\leq k^{-1}C_{T}\epsilon'_{k},\,\,\,\epsilon'_{k}:=\max\{\epsilon_{k},k^{-1}\}$$ Indeed, using the mean value theorem we can write $$\psi_{k,m+1}-\psi_{k,m}=\frac{1}{k}(\frac{u_{m/k+1/k}-u_{m/k}}{1/k})=\frac{1}{k}(\frac{\partial u_{t}}{\partial t}_{t=m/k}+O(k^{-1})),$$ where the term $O(k^{-1})$ may be estimated as $|O(k^{-1})|\leq A_{T}k^{-1},$ where $A=\sup_{X\times[0,T]}|\frac{\partial^{2}u_{t}}{\partial^{2}t}|.$ Using the evolution equation for $u_{t}$ and applying formula \[eq:conv of D k\] thus proves Step 1. *Step 2: The discrete evolution on $C(X)$ defined by the difference equation \[eq:differ eq abs\] contracts the sup-norm*. Set $C:=\sup|\phi_{m}-\psi_{m}|$ where $\phi_{m}$ and $\psi_{m}$ satisfy the difference equation \[eq:differ eq abs\] for a fixed $k.$ Then, $\phi_{m}\leq\psi_{m}+C$ and hence, since $I+k^{-1}D^{(k)}$ is assumed to be increasing, $$\phi_{m+1}=\phi_{m}+k^{-1}D^{(k)}(\phi_{m})\leq\psi_{m}+C+k^{-1}D^{(k)}(\psi_{m}+C)=\psi_{m+1}+C$$ In particular, $\sup(\phi_{m+1}-\psi_{m+1})\leq C:=|\sup(\phi_{m+1}-\psi_{m+1})|.$ Applying the same argument with the roles of $\phi$ and $\psi$ interchanged concludes the proof. *Step 3: Conclusion: $$\sup_{X}|u_{m}^{(k)}-\psi_{k,m}|\leq C_{T}\frac{m}{k}\epsilon'_{k}\label{eq:induction hyp}$$* We will prove this by induction over $m$ (for $k$ fixed) the statement being trivially true for $m=0.$ We fix the integer $k$ and assume as an induction hypothesis that \[eq:induction hyp\] holds for $m$ with $C$ the constant in the previous inequality. Applying first Step 2 and then the induction hypothesis thus gives $$\sup_{X}|(\psi_{k,m}+k^{-1}D^{(k)}(\psi_{k,m}))-(u_{m}^{(k)}+k^{-1}D^{(k)}(u_{m}^{(k)}))|\leq\sup_{X}|\psi_{k,m}-u_{m}^{(k)}|\leq C_{T}\frac{m}{k}\epsilon'_{k},$$ Now, by Step 1, $$\sup_{X}|\psi_{k,m+1}-(\psi_{k,m}+k^{-1}D^{(k)}(\psi_{k,m})|\leq\frac{C_{T}}{k}\epsilon'_{k}$$ for all $(m,k).$ Hence, $$\sup_{X}|\psi_{k,m+1}-u_{m+1}^{(k)}|\leq C_{T}\frac{m}{k}\epsilon'_{k}+C_{T}\frac{1}{k}\epsilon'_{k}=C_{T}\frac{(m+1)}{k}\epsilon'_{k}$$ proving the induction step and hence the final Step 3. In the present setting $\mathcal{H}$ will be taken as subspace of $C^{4}(T^{n})$ consisting of all strictly quasi-convex functions $u$ and $$D(u)(x):=\log(\det(\nabla^{2}u(x)+I))-g(\nabla u(x)+x)+f(x)$$ The existence and large time-convergence properties of the corresponding flow follows from the results in [@s-s; @k-s-w]: \[prop:existence and conv parab torus\]Let $f$ and $g$ be two functions in $C^{2}(T^{4})$. Then, for initial data $u_{0}\in C^{2}(T^{n})$ which is strictly quasi-convex there exists a solution $u(x,t)$ to the corresponding parabolic PDE in $C^{2}(]0,\infty[\times T^{n})$ such that $u_{t}\in C^{4}(X)$ and $u_{t}$ is strictly quasi-convex (and $u(x,t)\in C^{4}([0,\infty[\times T^{n})$ if $u_{0}\in C^{4}(X)).$ Moreover, as $t\rightarrow\infty$ the functions $u_{t}$ converge in the $C^{4}-$topology to a static solution $u_{\infty}$ (i.e. a potential to the corresponding optimal transport problem) and there exists a positive constant $A$ such that $$\sup_{T^{n}}|u_{t}-u_{\infty}|\leq Ae^{-t/A}\label{eq:exp conv}$$ (if moreover $f$ and $g$ are smooth then so is $u_{t}$ and the convergence holds in the $C^{\infty}-$topology). We first observe that the existence of a solution $u_{t}$ as above, follows from the results in [@s-s], as do the large $t-$ convergence. Strictly speaking the setting in [@s-s] is a bit different as the role of $T^{n}$ is played by a bounded smooth domain in $\R^{n}$ and some boundary conditions are imposed on $u_{t}.$ But the proof in the case of $T^{n}$ is actually easier than the ones in [@s-s] since $T^{n}$ is compact, without boundary. Indeed, the key point is the interior $C^{2}-$estimate in [@s-s Lemma 4.6] which directly apply to the present setting. Then parabolic Krylov- Safonov theory and (Shauder) boot strapping can be applied in the standard way. As for the exponential convergence it follows from the Li-Yau type Harnack inequality established in [@k-s-w]; see Section 5-7 in [@k-s-w] (in the case when $g=0$ the usual Li-Yau Harnack inequality can be applied, precisely as in [@ca]). ### Conclusion of proof of Theorem \[thm:conv in dynamic torus setting intro\] and Corollary \[cor:conv of explicit apprl intro\] The proof of Theorem \[thm:conv in dynamic torus setting intro\] now follows directly from combining Lemma \[lem:general conv towards parab\] with Propositions \[prop:asympt of pho\], \[prop:existence and conv parab torus\] and also using that the corresponding operator $S^{(k)}:=I+k^{-1}D^{(k)}$ is clearly increasing and invariant under the additive $\R-$action. Finally, Corollary \[cor:conv of explicit apprl intro\] follows directly by combining Theorem \[thm:conv in dynamic torus setting intro\] with the exponential convergence in formula \[eq:exp conv\]. Indeed, setting $m_{k}=kt_{k}$ with $t_{k}:=A\log k,$ where $A$ is the constant appearing formula \[eq:exp conv\]., gives $$\left\Vert u_{m_{k}}^{(k)}-u_{\infty}\right\Vert _{C(X)}\leq\left\Vert u_{m_{k}}^{(k)}-u_{t_{k}}\right\Vert _{C(X)}+\left\Vert u_{t_{k}}-u_{\infty}\right\Vert _{C(X)}\leq ACk^{-1}\log k+Ak^{-1}$$ as desired (also using the $C^{\infty}-$convergence of $u_{t}$ which ensures that $C$ is bounded from above (independently of $t).$ Setting $u_{k}:=u_{m_{k}}^{(k)}$ this proves the estimate \[eq:estimate in cor torus intro\], which also implies the estimate \[eq:estimate for gamma k cor torus intro\] for $\gamma_{k}.$ Indeed, by definition, $\gamma_{k}:=e^{-kd^{2}(x,y)^{2}/2}e^{-ku_{k}(x)}e^{-kv_{k}(y)}\mu^{(k)}\otimes\nu^{(k)},$ where $v_{k}:=v[u_{k}].$ By the estimate \[eq:estimate in cor torus intro\] and Lemma \[lem:local dens prop in model case\] there exists a positive number $C$ such that $$u_{k}(x)+v_{k}(y)\geq u(x)+u^{c}(y)+Ck^{-1}\log k.$$ The proof is thus concluded by invoking the following elementary inequality, which we claim holds for any smooth quasi-convex function $u$ on $T^{n}:$ $$d^{2}(x,y)^{2}/2+u(x)+u^{c}(y)\geq\frac{\delta}{2}d(x,F_{u}(x))^{2},\,\,\,F_{u}(x)=x+(\nabla u)(x).\label{eq:ineq using strict conv}$$ To see this we identify $u$ and $u^{c}$ with $\Z^{n}-$periodic functions on $\R^{n}$ and set $\phi(x):=u(x)+|x|^{2}/2.$ Then $\phi^{*}(y)=u^{c}(y)+|y|^{2}/2,$ where $\phi^{*}$ is the classical Legendre-Fenchel transform on $\R^{n}$ and it will be enough to show that $$|x-y|^{2}/2+\phi(x)+\phi^{*}(y)\geq\frac{\delta}{2}|y-\nabla\phi(x)|^{2}.\label{eq:ineq using strict ineq in phi}$$ Indeed, the claimed inequality \[eq:ineq using strict conv\] follows from the latter one after replacing $x$ with $x+m$ and taking the infimum over all $m\in\Z^{n}.$ Now, by assumption $\nabla^{2}\phi\leq\delta^{-1}I$ and hence $\nabla^{2}\phi^{*}\geq\delta I$ (by \[eq:matrix relation for u\]). As a consequence, $\phi^{*}(y)\geq\phi^{*}(y-t)+t\cdot\nabla\phi^{*}(y-t)+\delta|t|^{2}/2$ for any $t\in\R^{n}.$ Setting $t:=y-\nabla\phi(x)$ and using that $\phi^{*}(\nabla\phi(x))=\nabla\phi(x)\cdot x-\phi(x)$ and $(\nabla\phi^{*})(\nabla\phi(x))=x$ this implies the desired inequality \[eq:ineq using strict ineq in phi\]. \[sec:Generalizations-and-outlook\]Convergence towards parabolic optimal transport equations on compact manifolds ================================================================================================================= Let $X$ and $Y$ be compact smooth manifolds (without boundary) and $c$ a lsc function on $X\times Y,$ taking values in $]-\infty,\infty]$ which is smooth on the complement of a closed proper subset, denoted by $\text{sing \ensuremath{(c)}.}$ We will denote by $\partial_{x}$ the vector of partial derivatives defined wrt a choice of local coordinates around a fixed point $x\in X.$ Given two normalized volume forms $\mu$ and $\nu$ in $\mathcal{P}(X)$ and $\mathcal{P}(Y)$ we locally express $$\mu=e^{-f}dx,\,\,\,\nu=e^{-g}dy$$ in terms of the local volume forms $dx$ and $dy$ determined by a choice of local coordinates. Following standard practice we will assume that the cost function satisfies the following assumptions - *(A1) (“Twist condition”)* The map $y\mapsto\partial_{x}c(x,y)$ is injective for any $(x,y)\in X\times Y-\text{sing \ensuremath{(c)}}$ - *(A2) (“Non-degeneracy”)* det $(\partial_{x_{i}}\partial_{y_{j}}c)(x,y)\neq0$ for any $(x,y)\in X\times Y-\text{sing \ensuremath{(c)}}$ See [@v2 pages 246, 313] for an in depth discussion of various assumption on cost functions. In [@v2] A1+A2 is called the strong twist condition and as pointed out in [@v2 Remark 12.23] it holds for the cost function derived from any well-behaved Lagrangian, including the Riemannian setting where $c=d^{2}/2).$ The space $\mathcal{H}(X)$ (or $\mathcal{H}$ for short) of all *smooth potentials* on $X$ is defined as the subspace of $C^{\infty}(X)$ consisting of all $c-$convex (i.e. such that $(u^{c})^{c}=u)$ smooth functions $u$ on $X$ such the subset $\Gamma_{u}$ of $X\times Y$ defined by formula \[eq:def of Gamma u\] is the graph of a diffeomorphism, denoted by $F_{u}$ and $c$ is smooth in a neighborhood of $\Gamma_{u}.$ The definition has been made so that, if $u\in\mathcal{H}$ and $(F_{u})_{*}\mu=\nu,$ then $F_{u}$ is an optimal map (diffeomorphism) wrt the cost function $c$ (by Prop \[prop:optim crit\]). Accordingly, we will call $u$ the *potential* of the map $F_{u}.$ We note that it follows immediately from the assumption A1 that, for a given $x\in X,$ $$y:=F_{u}(x)\Longleftrightarrow(\partial_{x}c)(x,y)+(\partial_{x}u)(x)=0\label{eq:eq for F u}$$ Assume that $X=Y$ and that $c(x,y)\geq0$ with equality iff $x=y$ and that $c$ is smooth in a neighborhood of the diagonal. Then $u=0$ is in $\mathcal{H}$ (with $F_{0}$ given by the identity) and more examples of potentials are obtained by using that $\mathcal{H}$ is open in the $C^{\infty}-$topology, in general. In particular, this applies in the “Riemannian setting”, where $c=d^{2}/2$ on a compact Riemannian manifold $X.$ Parabolic optimal transport equations ------------------------------------- Consider now the following parabolic PDE, introduced in [@k-s-w]: $$\frac{\partial u_{t}(x)}{\partial t}=\log\det\left(\partial_{x}F_{u_{t}}\right)-g(F_{u_{t}}(x))+f(x)\label{eq:parabolic pde general setting}$$ expressed in terms of a choice of local coordinates, where $\det\left(\partial_{x}F_{u}\right)$ denotes the local Jacobian of the map $x\mapsto F_{u}(x).$ We note that - The right hand side in the equation \[eq:parabolic pde general setting\] is globally well-defined (i.e. independent of the choice of local coordinates around $(x,F_{u}(x))$ in $X\times Y).$ Indeed, it is equal to the logarithm of the quotient of $(F_{u})_{*}\mu$ and $\nu.$ Accordingly, $u$ is a stationary solution iff it is the potential of an optimal transport map. - Differentiating the equation \[eq:eq for F u\] reveals that $$\det(\partial_{x}F_{u})=\frac{\det\left((\partial_{x}^{2}c)(x,F_{u}(x))+(\partial_{x}^{2}u)(x)\right)}{\det\left((\partial_{x}\partial_{y}c)(x,F_{u}(x))\right)}\label{eq:form for det partial F}$$ \[subsec:Sequences-of-measures\]Sequences of measures satisfying a local density property and Quasi-Monte Carlo methods ----------------------------------------------------------------------------------------------------------------------- We next introduce the following stronger local form of the “global” density property appearing in Lemma \[lem:dens prop\]. \[def:local dens\]Given a positive integer $s,$ a sequence of probability measures $\mu^{(k)}$ on $X,$ converging weakly towards a measure $\mu,$ is said to have the *local density property (at the length scale $k^{-1/2})$* *of order $s$* if for any fixed $x_{0}\in X$ there exists local coordinates $\xi:=(\xi_{1},...,\xi_{n})$ centered at $x_{0}$ such that for any sequence $h_{k}$ defined on the polydisc $D_{k}$ of radius $\log k,$ centered at $0,$ and satisfying $\left\Vert h_{k}\right\Vert _{C^{s}(D_{k})}\leq Ce^{-x^{2}/C}$ the following holds: $$|\left\langle k^{n/2}(F_{x_{0}}^{(k)})_{*}(\mu^{(k)}-\mu),h_{k}\right\rangle |\leq C'k^{-s/2}$$ where $F_{x_{0}}^{(k)}$ is the scaled coordinate map from a neighborhood of $x_{0}$ in $X$ into $\R^{n}$ defined by $F_{x_{0}}^{(k)}(x):=(k^{1/2}\xi(x))$ and $d\lambda$ denotes Lebesgue measure. The model case for the local density property introduced above (of order two) is the case when $X$ is the torus and $\mu^{(k)}$ is obtained from a volume form $\mu$ using the standard discretization scheme described in Section \[subsec:Discretization-of-Optimal\], based on a regular grid. This means that $\mu^{(k)}=f\delta_{\Lambda_{k}}/Z_{k},$ where $Z_{k}$ is the normalization constant and $\Lambda_{k}$ is the discrete torus. In practice, for a general manifold $X$ it is convenient to fix a “discretization of $X",$ that is some discrete “reference sequence” $\lambda_{k}$ converging towards a “reference volume form” $\lambda\in\mathcal{P}(X)$ and satisfying the local density property (of some order). Then any volume form $\mu$ can be “discretized” by writing $\mu=\rho\lambda$ and setting $$\mu^{(k)}:=\rho\lambda^{(k)}/Z_{k},\,\,\,\,Z_{k}:=\int\rho\lambda^{(k)},$$ which then automatically has the local density property (and similarly for $Y$ and $\nu).$ Such reference sequences $\lambda_{k}$ can, for example, be constructed using quasi-Monte Carlo methods in numerical analysis, as we next explain. ### Quasi-Monte Carlo systems Let $(X,g)$ be an $n-$dimensional compact Riemannian manifold and denote by $dV$ the corresponding normalized volume form on $X.$ Following [@b-e-g] (and [@b-s-s-w] in the case of a sphere) the *worst case error* of integration of points $\{x_{i}\}_{i=1}^{N}\Subset X$ and weights $\{w_{i}\}_{i=1}^{N}\Subset\R_{+}$ (assumed to sum to one) with respect to some Banach space $W$ of continuous functions on $X,$ is defined as $$\text{wce \ensuremath{(\left\{ (x_{i},w_{i})\right\} _{i=1}^{N})}:=\ensuremath{\sup}}\left\{ \left|\int fdV-\sum_{i=1}^{N}f(x_{i})w_{i}\right|:\,\,f\in W,\,\,\left\Vert f\right\Vert \leq1\right\}$$ Let now $W:=W_{p}^{s}(X)$ be the Sobolev space of all functions $f$ on $X$ such that all (fractional) distributional derivatives of order $s$ are in $L^{p}(X)$ and assume that $s>n/p$ (which ensures that $W_{p}^{s}(X)\subset C(X)).$ Then a sequence of $N_{k}$ weighted points $X_{N_{k}}:=\left\{ (x_{i}^{(k)},w_{i}^{(k)})\right\} _{i=1}^{N_{k}}$ is said to be a *quasi-Monte Carlo system* for $W_{p}^{s}(X)$ if $$\text{wce \ensuremath{(X_{N_{k}})\leq\frac{C}{N_{k}^{s/n}}} }$$ for some uniform constant $C.$ Assume that $X_{N_{k}}:=\left\{ (x_{i}^{(k)},w_{i}^{(k)})\right\} _{i=1}^{N_{k}}$ is a quasi-Monte Carlo system for $W_{p}^{s}(X)$ (where $s>n/p)$ such that $N_{k}\sim k^{n}.$ Then the corresponding sequence $\lambda_{k}:=\sum_{i=1}^{N_{k}}w_{i}^{(k)}\delta_{x_{i}^{(k)}}\in\mathcal{P}(X)$ has the local density property of order $s-n(1-1/p).$ Given a sequence $h_{k}$ as in Definition \[def:local dens\] we may assume that $f_{k}:=(F^{(k)})^{*}h_{k}$ extends to define a sequence of functions in $C^{s}(X)$ (supported in a fixed neighborhood of $x_{0}).$ By the assumed quasi-Monte Carlo property $$|\left\langle (\lambda_{k}-dV),f_{k}\right\rangle |\leq Ck^{-s}\left(\int_{X}|\nabla^{s}f_{k}|^{p}dV\right)^{1/p}.$$ Multiplying both sides with $k^{n/2}$ and using the Chain rule thus gives $$|\left\langle k^{n/2}(F_{x_{0}}^{(k)})_{*}(\lambda_{k}-dV),f_{k}\right\rangle |\leq\frac{C}{k^{1/2(s-n(1-1/p))}}\left(\int_{D_{k}}|\nabla^{s}h_{k}|^{p}k^{n/2}(F_{x_{0}}^{(k)})_{*}dV\right)^{1/p}.$$ By the assumption on $h_{k}$ the integral in the right hand side above is uniformly bounded from above, which concludes the proof. In view of the applications to the present setting we introduce the following (non-standard) definition: A sequence of weighted sets $X_{N_{k}}:=\left\{ (x_{i}^{(k)},w_{i}^{(k)})\right\} _{i=1}^{N_{k}}$ is said to be a *good quasi-Monte Carlo system* if $N_{k}\sim k^{n}$ and there exists $(s,p)$ such that $s-n(1-1/p)\geq2.$ In particular, the corresponding sequence of measures $\lambda_{k}$ then has the local density property of order $2.$ The existence of good quasi-Monte Carlo systems follows from [@b-c-c-g-s-t Cor 2.13] on any compact Riemannian manifold $(X,g)$ (in fact [@b-c-c-g-s-t Cor 2.13] shows that, for any given $p,$ the quasi-Monte Carlo property holds for any $s).$ The corresponding weighed points are taken as “weighted cubature points” for the space $H_{k}(X)\subset C^{\infty}(X)$ spanned by all eigenfunctions of the Laplacian with eigenvalue bounded from above by $k$ (i.e. the corresponding integration error vanishes for any $f\in H_{k}(X)).$ In the particular case of the standard $n-$dimensional sphere it follows from [@b-r-v] that all the weights can be taken to be equal to $1/N_{k}$, i.e. the points can be taken to be *spherical $k-$designs*. Such points have been generated for large values of $k$ [@wo]. Convergence of the Sinkhorn algorithm towards parabolic optimal transport ------------------------------------------------------------------------- We are now ready for the following generalization of Theorem \[thm:conv in dynamic torus setting intro\] stated in the introduction (as in the torus setting in Section \[subsec:Discrete-optimal-transport\] the entropic regularization parameter is expressed as $\epsilon=k^{-1},$ where $k$ is a positive number). \[thm:dynamic general\]Let $c$ be a function satisfying the assumptions A1 and A2 and $\mu^{(k)}$ and $\nu^{(k)}$ be two sequences converging towards $\mu$ and $\nu$ in $\mathcal{P}(X)$ and $\mathcal{P}(Y),$ respectively, satisfying the local density property of order $s,$ for some $s\geq2.$ Given $u_{0}\in\mathcal{H}$ assume that there exists a solution $u_{t}$ in $C^{2}(X\times[0,T])$ of the parabolic PDE \[eq:parabolic pde general setting\] with initial condition $u_{0}$ and such that $u_{t}\in\mathcal{H}$ for any $t\in[0,T].$ Denote by $u_{m}^{(k)}$ the iteration \[eq:scaled iteration\] defined by the data $(\mu^{(k)},\nu^{(k)},c)$ and such that $u_{0}^{(k)}=u_{0}$ for any given $k.$ Then there exists a constant $C$ such that for any $(m,k)$ satisfying $m/k\in[0,T]$ $$\sup_{T^{n}}\left|u_{m}^{(k)}-u_{m/k}\right|\leq C\frac{m}{k}k^{-1},$$ The assumptions have been made precisely to ensure that the proof of Theorem \[thm:conv in dynamic torus setting intro\] can be generalized, almost verbatim. Hence, we will be rather brief. *Step 1:* Let $\alpha$ be a $C^{4}-$smooth function on $X$ with a unique minimum at $x_{0}$ which is non-degenerate, i.e. in local coordinates $\partial_{x}^{2}\alpha(x_{0})>0$ and $h$ a $C^{2}-$smooth function. Then, in local coordinates centered at $x_{0},$ $$k^{-n/2}\int_{X}e^{-k\alpha}h\mu^{(k)}=(2\pi)^{n/2}e^{-k\alpha(x_{0})}\frac{h(x_{0})e^{-f(x_{0})}}{\sqrt{\det(\partial_{x}^{2}\alpha(x_{0}))}}(1+Ck^{-1}).$$ Using the local density assumption as a replacement of Lemma \[lem:local dens prop in model case\] this is shown essentially as before. *Step 2:* If $u\in\mathcal{H},$ then the following asymptotics holds $$\rho_{ku}(x)=\det(\partial_{x}F_{u}(x))e^{f(x)-g(F_{u}(x))}(1+Ck^{-1})$$ To prove this first observe that if $u\in\mathcal{H}(X),$ then $u^{c}\in\mathcal{H}(Y).$ Indeed, by assumption there is a unique $x$ such that $y=F_{u}(x)$ and we can express $$u^{c}(y)=-c(x,F_{u}(x))-u(x).$$ Since $c$ is assumed to be smooth in a neighborhood of $\Gamma_{u}$ the right hand side above defines a smooth function in $x$ and since $F$ is a diffeomorphism it follows that $u^{c}(y)$ is smooth. Moreover, by symmetry $\Gamma_{u^{c}}=\Gamma_{u},$ which can be identified with the graph of the diffeomorphism $F_{u}^{-1}.$ This shows that $u^{c}\in\mathcal{H}(Y)$ and $$F_{u^{c}}=(F_{u})^{-1}\label{eq:F u inverse}$$ Setting $y_{x}:=F_{u}(x)$ and $x_{y}=F_{u^{*}}(y)$ we can now apply the previous step, essentially as before, to get $$\rho_{ku}(x)=\sqrt{\frac{\det\left((\partial_{x}^{2}c)(x_{y},x)+\partial_{x}^{2}u(x_{y})\right)}{\det\left((\partial_{y}^{2}c)(x,y_{x})+\partial_{y}^{2}u^{c}(y_{x})\right)}}e^{f(x)-g(F_{u}(x))}(1+Ck^{-1})$$ Finally, differentiating the relation \[eq:F u inverse\] reveals that $$\det((\partial_{y}F_{u^{c}})(y_{x}))=\det((F_{u}(x))^{-1}$$ and hence using equation \[eq:form for det partial F\] and symmetry (which ensures that the denominator appearing in equation \[eq:form for det partial F\] coincides with the one appearing obtained when $u$ is replaced by $u^{c}$) concludes the proof of Step 2. *Step 3:* Conclusion of proof The proof is concluded, as before, by invoking Lemma \[lem:general conv towards parab\]. It follows from standard short-time existence results for parabolic PDEs that the existence of a solution $u_{t}$ as in the previous theorem holds for some $T>0.$ Moreover, by [@k-s-w] long-time existence, i.e. $T=\infty,$ holds under the following further assumptions on $c:$ - *(A3) (“Stay-away property”)* For any $0<\lambda_{1},\lambda_{2}$ there exists $\epsilon>0$ only depending on $\lambda_{1},\lambda_{2}$ such that $\lambda_{1}\leq\text{|\ensuremath{\det}}\partial_{x}F_{u}|\leq\lambda_{2}\implies\text{dist \ensuremath{(\Gamma_{u},\text{\ensuremath{\text{sing \ensuremath{(c))\geq\epsilon}}}}}}$ for any $u\in\mathcal{H}$ - (A4) (“Semi-concavity”) $c$ is locally semi-concave, i.e. the sum of a concave and a smooth function on the domain where it is finite. - (A5) (“Strong MAW-condition”) The Ma-Wang-Trudinger tensor of $c$ is bounded from below on $X\times Y-\text{sing \ensuremath{(c)}}$ by a uniform positive constant $\delta.$ \[thm:(Kim-Streets-Warren-)-Assume\](Kim-Streets-Warren [@k-s-w]) Assume that $c$ satisfies the assumptions A1-A5. Then, for any given $u_{0}\in\mathcal{H}$ there exists a solution $u(x,t)$ in $C^{\infty}(X\times[0,\infty[)$ of the parabolic PDE \[eq:parabolic pde general setting\] with initial condition $u_{0}$ and such that $u_{t}\in\mathcal{H}$ for any $t>0.$ Moreover, $u_{t}$ converges exponentially, (in any $C^{p}(X)),$ as $t\rightarrow\infty,$ to a potential $u\in\mathcal{H}$ of a diffeomorphism $F_{u}$ transporting $\text{\ensuremath{\mu}\ to \ensuremath{\nu} , }$which is optimal wrt the cost function $c.$ Let us explain how to translate the result in [@k-s-w] to the present setting. Following [@k-s-w] a function $u\in C^{2}(X)$ is said to be *locally strictly $c-$convex,* if, in local coordinates, the matrix $(\partial_{x}^{2}c)(x,F_{u}(x))+(\partial_{x}^{2}u)(x)$ is positive definite. This condition is independent of the choice of local coordinates. Indeed, it equivalently means that any given $x_{0}\in X$ is a non-degenerate local minimum for the function $$x\mapsto c(x,F(x_{0}))+u(x)\,\,\,\text{on\,}X.\label{eq:function that x not min}$$ It follows that for any such $u$ the corresponding map $F_{u}$ is a local diffeomorphism. The main result in [@k-s-w] says that, under the assumptions on $c$ in the statement above, for any initial datum $u_{0}\in C^{2}(X)$ which is ** locally strictly $c-$convex, ** there exists a solution $u(x,t)$ in $C^{\infty}(X\times]0,T])$ which is also locally strictly $c-$convex. To make the connection to the present setting first note that if $u\in\mathcal{H}$ then $u_{0}$ is even an absolute minimum for the function \[eq:function that x not min\], which is non-degenerate (since $F_{u}$ is a diffeomorphism) and hence $u$ is locally strictly $c-$convex. Conversely, if $u$ is locally strictly $c-$convex then [@k-s-w Cor 7.1] says that $u\in C^{2}(X)$ is $c-$convex (i.e. $(u^{c})^{c}=u)$ and the proof given in [@k-s-w Cor 7.1] moreover shows that $F_{u}$ is a global diffeomorphism. Hence, $u\in C^{\infty}(X)$ is locally $c-$convex iff $u\in\mathcal{H},$ which concludes the proof of the theorem. Under the assumptions in the previous theorem it follows, in particular, that the optimal transport map is smooth. Conversely, the assumptions are “almost necessary” for regularity of the optimal transport map (see [@v2 Chapter 12] and reference therein). Also note that the semi-concavity assumption is always satisfied in the case when $X=Y$ is a compact Riemannian manifold and $c=d^{2}/2$ [@v2 (b), Page 278]. Combining the exponential large-time convergence of $u_{t},$ in the previous theorem, with Theorem \[thm:dynamic general\] gives, just as in the torus setting, the following \[cor:constr of approx in quite general setting \]Assume that $c$ satisfies the assumptions in Theorem \[thm:dynamic general\]. Then there exists a positive constant $A$ such that $u_{k}(x):=u_{m_{k}}^{(k)}(x_{i_{k}})$ with $m_{k}=Ak\log k$ converges uniformly to the optimal transport potential $u(x)$. More precisely, $$\sup_{T^{n}}\left|u_{k}-u\right|\leq Ck^{-1}\log k$$ \[exa:The-assumptions-in\]The assumptions in the previous corollary are satisfied when $X=Y$ is the $n-$sphere and $c(x,y)=d^{2}(x,y)/2$ for the standard round metric or $c(x,y)=-\log|x-y|,$ where $|x-y|$ denotes the chordal distance (see [@k-s-w] and references therein). The latter case appears in the reflector antenna problem, as explained in Section \[subsec:Application-to-the antenna\]. \[sec:Nearly-linear-complexity\]Nearly linear complexity on the torus and the sphere ==================================================================================== In this section we start by showing that the convergerence results in Section \[sec:Generalizations-and-outlook\] hold in a more general setting where the kernel $\mathcal{K}^{(k)}(x,y):=e^{-kc(x,y)}$ is replaced with an appropriate approximate kernel. This extra flexibility is then applied in the setting of optimal transport on the two-sphere, using “band-limited” heat-kernels, where it leads to a nearly linear algorithmic cost for the corresponding Sinkorn iterations. Sequences $c_{k}$ and approximate kernels $K_{k}$ ------------------------------------------------- Just as in the generalization of the (static) Theorem \[Thm:weak conv of fixed points towards optimal transport plans\], considered in Section \[thm:conv wasserst\], the (dynamic) Theorem \[thm:dynamic general\] can be generalized by replacing the cost function $c$ with a suitable sequence $c_{k}.$ But then the uniform convergence of $c_{k}$ towards $c$ (formula \[eq:unif conv of ck\]) has to be supplemented with further asymptotic properties on the complement of the singularity locus of $c.$ For example, the proof of Theorem \[thm:dynamic general\] goes through, almost word for word, if the upper bound corresponding to \[eq:unif conv of ck\] holds globally, i.e.: $$e^{-kc_{k}(x,y)}\leq O(e^{\epsilon k})e^{-kc(x,y)}\label{eq:upper bound on minus c k}$$ (where $O(e^{\epsilon k})$ denotes a sequence of sub-exponential growth) and $c_{k}$ has the following further property: on any given compact subset in the complement of $\text{sing \ensuremath{(c)}}$ there exists a strictly positive smooth function $h_{0}(x,y)$ and a uniformly bounded sequence $r_{k}(x,y)$ of functions such that $$\mathcal{K}^{(k)}(x,y):=e^{-kc_{k}(x,y)}=e^{-kc(x,y)}(h_{0}(x,y)+k^{-1}r_{k}(x,y))\label{eq:expansion c k}$$ This implies, in particular, that if Theorem \[Thm:weak conv of fixed points towards optimal transport plans\] holds for a given kernel $\mathcal{K}^{(k)},$ then it also holds for any other kernel $\tilde{\mathcal{K}}^{(k)}$ wich has error $O(k^{-1})$ as an approximation relative to $\mathcal{K}^{(k)},$ i.e. such that $$|\mathcal{K}^{(k)}-\tilde{\mathcal{K}}^{(k)}|\leq Ck^{-1}\mathcal{K}^{(k)}\label{eq:relative error of kernels}$$ or such that $\tilde{\mathcal{K}}^{(k)}$ has absolute error $e^{-Ck},$ for $C$ sufficently large, i.e. $$|\mathcal{K}^{(k)}-\tilde{\mathcal{K}}^{(k)}|\leq e^{-Ck},\,\,\,C>\inf_{X\times Y}c\label{eq:abs relative error}$$ \[subsec:Heat-kernel-regularization\]Heat kernel approximations in the Riemannian setting ----------------------------------------------------------------------------------------- Consider now the Riemannian setting where $X$ is a compact Riemannian manifold (without boundary) and $c=d^{2}/2$ and $c_{k}$ is defined in terms of heat kernel (formula \[eq:c k as heat\]): \[thm:heat kernel para\]Let $X$ be a compact Riemannian manifold (without boundary) and set $c(x,y):=d(x,y)^{2}/2,$ where $d$ is the Riemannian distance function. Then the results in Theorem \[thm:dynamic general\] and Corollary \[cor:constr of approx in quite general setting \] still hold when the matrix kernel $e^{-kd^{2}(x,y)/2}$ is replaced with the heat kernel $\mathcal{K}_{2k^{-1}}(x,x)$ (at time $t=2k^{-1})$ As discussed above this follows from the following heat kernel asymptotics (which are a special case of [@ben Thm 3.1] and more generally hold for the heat kernel associated to a suitable hypoelliptic operator). Assume that $x$ and $y$ are contained in a compact subset of the complement of the cut-locus. Then $$\mathcal{K}_{t}(x,y)=t^{-n/2}e^{-t^{-1}d^{2}(x,y)/4}\left(h_{0}(x,y)+tr_{1}(t,x,y)\right),$$ where $h_{0}$ is smooth and $h_{0}>0$ and $r_{1}$ is smooth and uniformly bounded on $]0,t_{0}]\times X.$ This is not exatly of the form \[eq:expansion c k\] due to the presens of the factor $t^{-n/2}:=A_{k}.$ But it is, in fact, enough to know that \[eq:expansion c k\] holds when the right hand side is multiplied with a sequence $A_{k},$ only depending on $k.$ Indeed, the iteration $u_{m}^{(k)}$ is unaltered when the cost function $c_{k}(x,y)$ is replaced by $c_{k}(x,y)+C_{k}$ for some constant $C_{k}$ (which is consistent, as it must, with the fact that the parabolic equation \[eq:parabolic pde general setting\] is unaltered when a constant is added to $c).$ The use of the heat kernel in the Sinkorn algorithm for optimal transport on Riemannian manifolds was advocated in [@s-d---], where it was found numerically that discretized heat kernels provide substantional speedups, when compared to other methods. The previous theorem offers a theoretical basis for the experimental findings in [@s-d---], as long as the discretized heat kernels $\tilde{\mathcal{K}}^{(k)}$ satisfy one of the the approximation properties \[eq:relative error of kernels\] and \[eq:abs relative error\] (when compared with the corresponding bona fida heat kernel). However, the author is not aware of any general such approximation results in the discretized setting (but see [@c-w-w] and references therein for various numerical approaces to discretizations of heat kernels). We will instead follow a different route, based on “band-limited” heat kernels and fast Fourier type tranforms, applied to the case when $X$ is the two-sphere. In the case of a general Riemannian manifold $X$ one might hope that the use of band-limited kernels and fast Fourier type transforms can be replaced by fast multipole methods [@b-g]. Near linear complexity using fast transforms -------------------------------------------- Each iteration in the Sinkhorn algorithm amounts to computing two vector-matrix products of the form $$a_{i}=\sum_{j=1}^{N}\mathcal{K}(x_{i},y_{j})b_{j},\,\,\,\,i=1,...,N,\label{eq:sum matrix vector}$$ for a given function $\mathcal{K}$ on $X\times Y$ (followed by $N$ inversions), where $b$ and $a$ denote generic “input vector” and “output vectors”, respectively. In general, this requires $O(N^{2})$ arithmetic operations. But, as we will next exploit, in the presense of suitably symmetry fast summations techniques can be used to lower the complexity to nearly linear, i.e. to at most $CN(\log N)^{p}$ operations (for some positive constants $C$ and $p).$ ### \[subsec:Optimal-transport-on linear\]Optimal transport on the flat torus Let us first come back to the case of the flat torus $T^{n}$ discretized by the discrete torus $\Lambda_{k},$ considered in Section \[subsec:Main-results-in-the torus\]. Since $\mathcal{K}(x,y):=e^{-kd^{2}(x,y)}$ is invariant under the diagonal action of the torus $T^{n}$ it is follows from standard arguments that the sums \[eq:sum matrix vector\] can be computed in $O(N)(\log N)$ arithmetic operations. Indeed, using the group structure on $T^{n}$ we can write $\mathcal{K}(x,y)=h(x-y),$ for some function $h$ on $\Lambda_{k}.$ Then the classical convolution theorem of Fourier Analysis, on the discrete torus $\Lambda_{k}$ (viewed as an abelian finite group), gives $$a_{i}=\sum_{j=1}^{N}h(x_{i}-y_{j})b_{j}=\sum_{j=1}^{N}\hat{h}(m_{j})\hat{b}(m_{j})e^{2\pi im\cdot x_{i}},\,\,\,\,\hat{f}(m_{j}):=\sum_{i=1}^{N}f_{i}e^{-2\pi ix_{i}\cdot m_{j}}$$ This requires evaluating two Discrete Fourier Transforms (DFT) at $N=k^{n}$ points. Using the Fast Fourier Transform (FFT) this can be done in $O(N)(\log N)$ arithmetic operations. Note that since the heat kernel is also torus invariant the same argument can also be used for the kernel appearing in Theorem \[thm:heat kernel para\], in the torus case. ### Optimal transport on the round two-sphere Consider the round two-sphere $S_{2}$ embedded as the unit-sphere in $\R^{3}.$ Removing the north and south pole on $S_{2}$ we have the standard spherical (longatude-colatitude) coordinates $(\varphi,\theta)\in[0,2\pi[\times]-\pi,\pi[.$ A complete set of (non-normalized) eigenfunctions for the Laplacian on $L^{2}(S_{2})$ is given by the spherical harmonics $$Y_{l}^{m}(\varphi,\theta):=e^{im\varphi}P_{l}^{m}(\cos\theta),\,\,\,\,|m|\leq l,$$ which has eigenvalue $\lambda_{l,m}^{2}:=l(l+1).$ Here $P_{l}^{m}(x)$ denotes, as usual, the Legendre function of degree $l$ and order $m$ (aka the associated Legendre polynomial); see, for example, [@d-h]. Given a positive number $W$ (the “band-width”) we consider the *band-limited heat kernel* on the two-sphere: $$\mathcal{K}_{t}(x,y)_{W}:=\sum_{|m|\leq l\leq W}c_{m,l}Y_{l}^{m}(x)\overline{Y_{l}^{m}(y)},\,\,\,c_{m,l}:=e^{-tl(l+1)}\left\Vert Y_{l}^{m}\right\Vert _{L^{2}}^{-2}\label{eq:band limited heat kernel on sphere}$$ (By the spectral theorem this means that $\mathcal{K}_{t}(x,y)_{W}$ is the integral kernel of $e^{-t\Delta}\Pi_{W}$ where $\Pi_{W}$ is the orthognal projection onto the space of all band-limited functions). \[thm:sphere band limit heat\]Consider the two-sphere, discretized by a given good Quasi-Monte Carlo (QMC) system and take $R$ such that $R>1.$ Then the analog of all the results in Section \[subsec:Main-results-in-the torus\] are valid when the matrix kernel $e^{-kd^{2}(x,y)/2}$ is replaced by the band-limited heat kernel $\mathcal{K}_{2k^{-1}}(x,y)_{Rk}.$ Moreover, the arithmetic complexity of each Sinkhorn iteration is $O(N^{3/2}).$ As recalled in Example \[exa:The-assumptions-in\] the cost function $d(x,y)^{2}$ on the sphere satisfies the assumptions in Theorem \[thm:dynamic general\] (with $t=\infty)$ and Corollary \[cor:constr of approx in quite general setting \]. *Step 1: The asymptotics \[eq:upper bound on minus c k\] and \[eq:expansion c k\] are satisfied.* By Theorem \[thm:heat kernel para\] it is enough to observe that the following basic estimate holds if $t=2k^{-1}$ and $W=Rk:$ $$\left|\mathcal{K}_{t}(x,y)-\mathcal{K}_{t}(x,y)_{W}\right|\leq C_{\delta}e^{-2R^{2}k(1-\delta)}$$ for any given $\delta\in]0,1[.$ To prove the estimate note that $$\left|\mathcal{K}_{t}(x,y)-\mathcal{K}_{t}(x,y)_{W}\right|\leq\sum_{l>W}e^{-2k^{-1}l(l+1)}\frac{\left\Vert Y_{l}^{m}\right\Vert _{L^{\infty}}^{2}}{\left\Vert Y_{l}^{m}\right\Vert _{L^{2}}^{2}}\leq2Ck^{3}\sum_{l/k>R}e^{-2k(\frac{l}{k})^{2}}\frac{(l+1)^{2}}{k^{2}}\frac{1}{k},$$ using that the quotient involving $Y_{l}^{m}$ is dominated by $Cl$ (and that a given $l$ corresponds to $2l+1$ $m$s). Indeed, this is a special case of the the universal $L^{2}-$estimates for an eigenfunction $\Psi_{\lambda}$ of the Laplacian (with eigenvalue $\lambda^{2})$ on a general $n-$dimensional Riemannian manifold [@sog], which gives the growth factor $C\lambda^{n-1}.$ Finally, dominating the Riemann Gaussian sum above with the integral of the function $e^{-ks^{2}}s^{2}$ over $[R,\infty[$ concludes the proof. *Step 3: Complexity analysis* Using formula \[eq:band limited heat kernel on sphere\] gives $$a_{i}=\sum_{|m|\leq l\leq W}c_{m,l}\hat{b}_{l,m}Y_{l}^{m}(x_{i}),\,\,\,c_{m,l}=e^{-tl(l+1)}\left\Vert Y_{l}^{m}\right\Vert _{L^{2}}^{-1}\hat{b}_{l,m},\,\,\,\hat{b}_{l,m}:=\sum_{j=1}^{N}b_{j}\overline{Y_{l}^{m}(x_{j})})$$ $\hat{b}_{l,m}$ is the “forward discrete spherical Fourier transform” evaluated at $(l,m).$ Once it has been computed for all $(l,m)$ $a_{i}$ becomes an “inverse discrete spherical Fourier transform” (with coefficents $c_{m,l}\hat{b}_{l,m}$ ). By separation of variablers, both these transforms can be computed using a total of $O(k^{3})(=O(N^{3/2}))$ arithmetic operations (see the discussion after formula 1.9 in [@r-t]). In the special case when the good Quasi-Monte Carlo system is defined by an “equi-angular” grid of $O(k^{2})$ points on on $S_{2}$ (in terms of the spherical coordinates $(\varphi,\theta)),$ the arithmetic complexity of each iteration can be reduced to $O(N)(\log N)^{2}$ operations, using a fast discrete spherical Fourier transform [@h-r-k-m]. Such points indeed determine a good Quasi-Monte Carlo system with explicit weights, as follows from the fact that they define cubature formulas for band-limited functions [@d-h Thm 3]. ### \[subsec:Application-to-the antenna\]Application to the reflector antenna problem The extensively studied *far field reflector antenna problem* appears when $X=Y=S_{n}$ is the $n-$dimensional sphere $S_{n},$ embedded as the unit-sphere in $\R^{n+1}$ and the cost function is taken as $c(x,y):=-\log|x-y|$ [@waII; @g-o]. Briefly, the problem is to design a perfectly reflecting surface $\Sigma$ in $\R^{n+1}$ with the following property: when $\Sigma$ is illuminated with light emitted from the origin with intensity $\mu\in\mathcal{P}(S_{n})$ the output reflected intensity becomes $\nu\in\mathcal{P}(S_{n})$ (of course, $n=2$ in the usual applications). Representing $\Sigma$ as a radial graph over $S_{n}:$ $$\Sigma:=\{h(x)x\},\,\,x\in S_{n},$$ for a positive function $h$ on $S_{n}$ it follows from the reflection law and conservation of energy that $h$ satisfies the following Monge-Ampère type equation, expressed in terms of the covariant derivatives $\nabla_{i}$ in local orthonormal coordinates: $$\frac{\det(-\nabla_{i}\nabla_{j}h+2h^{-1}\nabla_{i}h\nabla_{j}h+(h-\eta)\delta_{ij})}{((|\nabla h|^{2}+h^{2})/2h)^{n}}=e^{g(F_{h}(x))-f(x)},\label{eq:ant eq intro}$$ where $F_{h}(x)$ denotes the reflected direction of the ray emitted in the direction $x$ (and $\mu$ and $\nu$ are represented as in \[eq:mu and nu in terms of f and g intro\]). The equation is also supplemented with the “second boundary value condition” that $F_{h}$ maps the support of $\mu$ onto the support of $\nu.$ Assuming that $f$ and $g$ are smooth there exists a smooth solution $h,$ which is unique up to scaling (see [@c-g-h] and references therein). Consider the two-sphere, discretized by a given good Quasi-Monte Carlo (QMC) system. Let $K^{(k)}$ be the $N_{k}\times N_{k}$ matrix defined by $$K_{ij}^{(k)}=|x_{i}^{(k)}-x_{j}^{(k)}|^{k}$$ Consider the Sinkhorn algorithm associated to $(p^{(k)},q^{(k)},K^{(k)}).$ Then the function $h_{k}$ on $S^{n}$ defined by the $k$ th root of $a_{k}:=a^{(k)}(m_{k})$ after $m_{k}=Ck\log k$ Sinkhorn iterations converges uniformly, as $k\rightarrow\infty,$ towards a solution $h$ of the antenna equation \[eq:ant eq intro\] satisfying the corresponding second boundary value condition. More precisely, there exists a constant $C$ such that $$\sup_{S_{N}}|h_{k}-h|\leq Ck^{-1}\log k,$$ Moreover, the arithmetic complexity of each iteration is $O(N^{3/2})$ in general and $O(N)(\log N)^{2}$ in the case of an equi-angular grid. As recalled in Example \[exa:The-assumptions-in\] the cost function $-\log d(x,y)^{2}$ on the sphere satisfies the assumptions in Corollary \[cor:constr of approx in quite general setting \]. For the complexity analysis we first recall the general fact that any kernel $K^{(k)}(x,y)$ which is radial, i.e. only depends on $|x-y|,$ may be expressed as $$K^{(k)}(x,y)=\sum_{l=1}^{\infty}C_{m,l}Y_{l}^{m}(x)\overline{Y_{l}^{m}(y)}\label{eq:form for K radial basis}$$ for some positive constants $C_{m,l}$ (a proof will be given below). By the argument in the proof of Step 2 in Theorem \[thm:sphere band limit heat\], it will be enough to show that $C_{m,l}=0$ when $l>k,$ i.e. that $K^{(k)}(x,y)$ is already band-limited with $W=k.$ To this end we follow the general approach in [@k-k-p]. First observe that when $x$ and $y$ are in $S^{2}$ we can write $|x-y|^{2}=2(1-x\cdot y).$ Hence, $K_{k}(x,y)=2^{k}f^{(k)}(x\cdot y),$ where $f^{(k)}(s)=(1-s)^{k}$ for $s\in[-1,1].$ The Legendre polynomials $p_{l}(=p_{l}^{0}$ form a base in the space of all polynomials of degree at most $k$ (which is orthogonal wrt Lesbegue measure on $[0,1])$ and hence we can decompose $$2^{k}f^{(k)}=\sum_{l=1}^{k}c_{l}^{(k)}p_{l}.$$ Formula \[eq:form for K radial basis\] now follows from the classical Spherical Harmonic (Legendre) addition theorem: $$p_{l}(x\cdot y)=\frac{4\pi}{2l+1}\sum_{|m|\leq l}Y_{l}^{m}(x)\overline{Y_{l}^{m}(y)}.$$ Outlook\[sec:Outlook\] ====================== Generalized parabolic optimal transport and singularity formation ------------------------------------------------------------------ Consider the setting in Section \[sec:Generalizations-and-outlook\] with a cost function $c$ satisfying the assumptions A1 and A2, but assume for simplicity that $c$ is globally continuous (for example, $c=d^{2}/2$ in the Riemannian setting). Recall that, given initial data $u_{0}\in\mathcal{H}$ and volume forms $\mu$ and $\nu,$ the parabolic equation \[eq:parabolic pde general setting\] admits a smooth solution $u_{t}$ on some maximal time-interval $[0,T[$ and the corresponding maps $F_{u_{t}}$ give an evolution of diffeomorphisms of $X.$ It does not seem to be known whether $T=\infty,$ in general, i.e. it could be that there are no solutions in $C^{\infty}(X\times]0,\infty[),$ in general. Still, using the corresponding iteration $u_{k}^{(m)}$ (say, defined with respect to $\mu_{k}=\mu$ and $\nu_{k}=\nu$) a generalized notion of solution can be defined: \[prop:generalized parabolic\]Given a $c-$convex function $u_{0},$ define the following curve $u_{t}$ of functions on $X,$ emanating from $u_{0}:$ $$u_{t}:=\sup\left\{ u_{k}^{(m)}:\,\,\,(m,k):\,\,m/k\rightarrow t,\,\,k\rightarrow\infty\right\}$$ Then $u_{t}$ is $c-$convex for any fixed $t$ (and, in particular, continuous) and there exists a constant $C$ such that $\sup_{X\times[0,\infty[}|u_{t}(x)|\leq C.$ *Step 1: there exists a constant such that $|u_{k}^{(m)}|\leq C.$* By the argument in Step 3 in the proof of Theorem \[thm:conv of u m in general setting\] we have $$\mathcal{F}^{(k)}(u^{(k)})+\mathcal{L}_{\nu}^{(k)}(u_{0})\leq I_{\mu}(u_{m}^{(k)})\leq I_{\mu}(u_{0})$$ By Lemma \[lem:dens prop\] $\mathcal{L}_{\nu}^{(k)}(u_{0})\rightarrow-\int u_{0}^{c}\nu$ and by Theorem \[Thm:weak conv of fixed points towards optimal transport plans\] $\mathcal{F}^{(k)}(u_{m}^{(k)})\rightarrow\inf_{C^{0}(X)}J$ and hence the lhs above is uniformly bounded in $k.$ Thus, there exists a constant $C$ such that $-C\leq I_{\mu}(u_{m}^{(k)})\leq C.$ The proof of Step 1 is now concluded by observing that there exist constants $A_{1}$ and $A_{2}$ such that, for any $c-$convex function $u,$ $$\sup_{X}u\leq I_{\mu}(u)+A_{1},\,\,\,\inf_{X}u\geq I_{\mu}(u)-A_{2}.$$ Indeed, both functionals $f_{1}(u):=\sup_{X}u-I_{\mu}(u)$ and $f_{2}(u):=\inf_{X}u-I_{\mu}(u)$ are continuous on $C(X)$ and descend to $C(X)/\R.$ But the space of $c-$convex functions is compact in $C(X)/\R$ (as is shown precisely as in Lemma \[lem:compactness of function spaces\]) and hence any continuous functional on the space is uniformly bounded, which implies the two inequalities above. *Step 2: If $\{u_{\alpha}\}_{\alpha\in A}$ is a finite family of $c-$convex functions, then $u:=\max\{u_{\alpha}\}_{\alpha\in A}$ is $c-$convex.* It is enough to find a function $v\in C(X)$ such that $u=v^{c}.$ We will show that $v:=\min\{u_{\alpha}^{c}\}_{\alpha\in A}$ does the job. To this end first observe that $u\mapsto u^{c}$ is order preserving. Hence, $u_{\alpha}\leq u$ implies that $u_{\alpha}^{c}\geq u^{c},$ giving $v\geq u^{c}.$ Applying the $c-$Legendre transform again thus gives $v^{c}\leq u^{cc}=u.$ To prove the reversed inequality first observe that, by definition, $u_{\alpha}^{c}\geq v$ and hence $u_{\alpha}=(u_{\alpha}^{c})^{c}\leq v^{c}.$ Finally, taking the sup over all $\alpha$ proves the desired reversed inequality. *Step 3: Conclusion* Denote by $\mathcal{K}_{t}$ the closure in $C(X)$ of the set $S_{t}$ of all $u_{m}^{(k)}$ such that $m/k\rightarrow t$ and $k\rightarrow\infty.$ By Step 1 and Lemma \[lem:compactness of function with k\] $\mathcal{K}_{t}$ is compact. Let $u_{1},...,u_{m}$ be the limit points of $S_{t}.$ By the argument towards the end of Step 1 in the proof of Theorem \[Thm:weak conv of fixed points towards optimal transport plans\], $u_{i}$ is $c-$convex. Hence, by Step 2, so is $u:=\max\{u_{i}\}.$ The curve $u_{t}$ \[eq:parabolic pde general setting\] is well-defined for any probability measure $\mu$ and $\nu$ on compact topological spaces $X$ and $Y$ and for any continuous cost function $c.$ Moreover, if $\mu$ and $\nu$ are normalized volume forms on compact manifolds, assumptions A1 and A2 hold and $u_{0}\in\mathcal{H},$ then, by Theorem \[thm:conv of u m in general setting\], $u_{t}$ coincides with the classical solution of the parabolic equation \[eq:parabolic pde general setting\], as long as such such a solution exists in $\mathcal{H},$ i.e. as long as $F_{u_{t}}$ is a well-defined diffeomorphism. This makes the curve $u_{t}$ a candidate for a solution to the problem posed in [@d-f Problem 9] of defining some kind of weak solution to the parabolic equation \[eq:parabolic pde general setting\], without making assumptions on the MTW-tensor etc (as in Theorem \[thm:(Kim-Streets-Warren-)-Assume\]). The connection to the Sinkhorn algoritm also opens the possibility of numerically exploring singularity formation of classical solutions $u_{t}$ to the parabolic equaiton \[eq:parabolic pde general setting\] as $t\rightarrow T$ (the maximal existence time). As indicated in [@d-f Problem 9] one could expect that the first derivatives of a classical solution $u_{t}$ blow up along a subset $S$ of $X$ of measures zero as $t\rightarrow T$ (moreover, in the light of the discussion in [@d-f Problem 8], the subset $S$ might be expected to be rectifiable and of Hausdorff codimension at least one). Finally, it may be illuminating to point out that, even if the construction of the generalized solution $u_{t}$ may appear to be rather non-standard from a PDE point of view it bears some similarities to the method of “vanishing viscosity” for constructing solutions to PDEs by adding small regularizing terms. This is reinfored by the intepretation of the inverse of $k$ as an “entropic regularization parameter” discussed in the introduction of the paper (also note that the approximations $u_{m_{k}}^{(k)}$ are smooth when the heat kernel is used, as in Theorem \[thm:heat kernel para\]). One is thus lead to ask whether, under suitable regularity assumptions on $(\mu,\nu,c)$ the curve $u_{t}$ is a viscosity solution of the parabolic PDE [@c-i-l]? \[sec:Appendix:-proof-of\]Appendix: proof of Lemma \[lem:discrete stat phase\] (discrete stationary phase approximation) ======================================================================================================================== We will use the standard notation $O(k^{-\infty})$ for a sequence of numbers $a_{k}$ such that $|a_{k}|\leq C_{p}k^{-p}$ for any given $p>0.$ We start with the following elementary \[lem:local dens prop in model case\]Let $h_{k}$ be a sequence of continuous convex functions on the polydisc $D_{k}$ in $\R^{n}$ of radius $\log k$ centered at $0$ such that $\left\Vert h_{k}\right\Vert _{C^{s}(D_{k})}\leq Ce^{-x^{2}/C}$ for $s\leq2$ for some positive constant $C.$ Then $$\left|k^{-n/2}\sum_{x_{i}^{(k)}\in D_{k}\cap(k^{-1/2}\Z)^{n}}h_{k}(x_{i}^{(k)})-\int_{D_{k}}h_{k}dx\right|\leq C'/k$$ By restricting the integration to one variable at a time it is enough to consider the case when $n=1.$ Fix $x_{i}^{(k)},$ which, by symmetry, may be assumed non-negative. For any fixed $x$ in the interval $I_{k}(x_{i}^{(k)})$ centered at $x_{i}^{(k)},$ of length $k^{-1/2},$ Taylor expanding $h_{k}$ gives $$|h_{k}(x_{i}^{(k)})-h_{k}(x)-(x_{i}^{(k)}-x)h_{k}'(x_{i}^{(k)})|\leq kCe^{-(x_{i}^{(k)})^{2}/C}\leq Ck^{-1}e^{-(x-1/2k^{1/2})^{2}/C},$$ using that $e^{-x^{2}/C}$ is decreasing in the last step. By symmetry, the integral over $I_{k}(x_{i}^{(k)})$ of the linear term $(x_{i}^{(k)}-x)h_{k}'(x_{i}^{(k)})$ vanishes, giving $$k^{-1/2}h_{k}(x_{i}^{(k)})=\int_{I_{k}(x_{i}^{(k)})}h_{k}(x_{i}^{(k)})dx=\int_{I_{k}(x_{i}^{(k)})}h_{k}(x)dx+\epsilon_{i}^{(k)},$$ where $$|\epsilon_{i}^{(k)}|\leq C'k^{-1}\int_{I_{k}(x_{i}^{(k)})}e^{-(x-1/2k^{1/2})^{2}/C}dx$$ Hence, summing over all points $x_{i}^{(k)}\in D_{k}\cap k^{-1/2}\Z$ except the end points and using that $|f(x_{i}^{(k)})|\leq Ce^{-(\log k)^{2}/C}\leq O(k^{-\infty})\leq C'/k^{-1}$ at the end points gives $$\left|k^{-1/2}\sum_{x_{i}^{(k)}\in D_{k}\cap k^{-1/2}\Z}h_{k}(x_{i}^{(k)})-\int_{D_{k}}h_{k}dx\right|\leq C'k^{-1}+k^{-1}C''\int_{0\leq s\leq\log k}e^{-(x-1/2k^{1/2})^{2}/C}ds,$$ which concludes the proof. In the sequel we will denote by $+$ the ordinary group structure on $T^{n}$ and by $0$ the zero with respect to the group structure. Without loss of generality we may as well assume that $\alpha(x_{0})=0.$ First note that the integral localizes around $x_{0},$ i.e. it may be restricted to an arbitrarily small neighborhood $U$ of $x_{0},$ up to adding an exponentially small error term. Indeed, on the subset where $\alpha>\delta$ the integrand is dominated by $Ce^{-\delta k}.$ Step 1: *The case when $x_{0}=0$ in $T^{n}$* {#step-1-the-case-when-x_00-in-tn .unnumbered} --------------------------------------------- We observe that the integral localizes to polydisc $U_{k}$ of radius $k^{-1}\log k$ centered at $0$ up to introducing an error term of the order $O(k^{-1}).$ Indeed, by assumption $\alpha(x)\geq\delta|x|^{2}$ on the complement of a fixed neighborhood $U$ and on $U_{k}$ we have $\int_{U-U_{k}}e^{-k\delta|x|^{2}}dx\leq O(k^{-\infty})\leq Ck^{-1}.$ Introducing the notation $h^{(k)}(x):=kh(k^{-1/2}x)$ we can write $$I_{k}:=k^{-n/2}\int_{U_{k}}e^{-k\alpha}f\delta_{\Lambda_{k}}=k^{-n/2}\int_{k^{1/2}U_{k}}e^{-\alpha^{(k)}}f\delta_{\Lambda_{k^{1/2}}}.$$ Now, Taylor expanding $\alpha$ gives, when $|x|\leq\log k,$ and denoting by $p^{(3)}$ the third order term (i.e. with homogeneous degree three) gives $$\alpha^{(k)}(x)=Ax\cdot x/2+k^{-1/2}p^{(3)}+k^{-1}O(|x|^{4})$$ which may be differentiated twice, by assumption. Thus $h_{k}:=e^{-\alpha^{(k)}}$ satisfies the assumptions of the previous lemma (with $s=2$ if $\alpha\in C^{2})$ giving $$I_{k}:=\int_{|x|\leq\log k}e^{-k\alpha^{(k)}(x)}dx+O(k^{-1})$$ This shows that in the present discrete setting we get the same result, up to the negligible error term $O(k^{-1}),$ as the ordinary stationary phase approximation, which can hence be invoked to conclude. Alternatively, a direct argument goes as follows. Taylor expanding the exponential gives $$h_{k}(x):=e^{-\alpha^{(k)}(x)}=e^{-Ax\cdot x/2}(1+k^{-1/2}p^{(3)}+k^{-1}O(|x|^{4}))$$ and hence $$\int_{|x|\leq\log k}e^{-Ax\cdot x/2}((1+k^{-1/2}p^{(3)}+k^{-1}O(|x|^{4}))$$ Using the exponential decay of $e^{-Ax\cdot x/2}$ the integral may be taken over all of $\R^{n},$ up to introducing an error term $O(k^{-\infty}).$ Hence computing the Gaussian integral concludes the proof, once one has verified that the integral over $p^{(3)}$ vanishes. In the case when $A$ is the identity the vanishing follows directly from the fact that $p^{(3)}$ is odd. In the general case one first observes that the space of polynomials of homogeneous degree $3$ is invariant under the action of the space of invertible linear maps. Hence the problem reduces, by a linear change of variables to the previous case. *Step 2: The case of a general $x_{0}$* {#step-2-the-case-of-a-general-x_0 .unnumbered} ---------------------------------------- Set $\tilde{\alpha}(x):=\alpha(x+x_{0})$ and $\tilde{f}(x):=f(x+x_{0})$ and decompose $x_{0}=m_{k}+r_{k}$ where $m_{k}\in\Lambda_{k}$ and $|r_{k}|\leq1/k$ (where we have identified a small neighborhood of $0$ in $T^{n},$ containing $r_{k}$ with $\R^{n}).$ Then we can write $$\int e^{-k\alpha}f\delta_{\Lambda_{k}}=\int e^{-k\tilde{\alpha}}\tilde{f}\delta_{(\Lambda_{k}-r_{k})}$$ Indeed, for any function $h$ on $T^{n}$ we have, since $m_{k}\in\Lambda_{k},$ $$\sum_{x_{i}\in\Lambda_{k}}h(x_{i}+m_{k}+r_{k})=\sum_{x_{i}}h(x_{i}+r_{k})$$ Now, we note that the conclusion in the previous lemma remains true when $\Lambda_{k}$ is replaced by the shifted set $\Lambda_{k}-r_{k}$ (with essentially the same proof) and hence we can conclude as before. [10]{} Beatson, R; Greengard, L: A short course on fast multipole methods: Wavelets, multilevel methods and elliptic PDEs (Leicester, 1996), 137, Numer. Math. Sci. Comput., Oxford Univ. Press, New York, 1997. J-D Benamou; Y. Brenier. A computational fluid mechanics solution to the Monge-Kantorovich mass transfer problem. Numer. Math. , 84(3):375393, 2000 J-D Benamou, G Carlier, M Cuturi, L Nenna; Peyré, G: Iterative Bregman projections for regularized transportation problems. SIAM J. Sci. Comput. 37 (2015), no. 2 Benamou, JD; Duval, V: Minimal convex extensions and finite difference discretization of the quadratic Monge-Kantorovich problem. 2017. &lt;hal-01616842v2&gt; Ben Arous, G: Développement asymptotique du noyau de la chaleur hypoelliptique hors du cut-locus. Annales scientifiques de lÉ.N.S. 4 e série , tome 21, n o 3 (1988), p. 307-331 Berman, R.J: Relative Kähler-Ricci flows and their quantization. Anal. PDE 6 (2013), no. 1, 131180. Berman, R; Boucksom, S; Witt Nyström, D: Fekete points and convergence towards equilibrium measures on complex manifolds. Acta Math. 207 (2011), no. 1, 127. Berman, R.J; Boucksom, S; Guedj,V; Zeriahi: A variational approach to complex Monge-Ampere equations. Publications math. de l’IHÉS (2012): 1-67 A Breger, M Ehler, M Graef: Points on manifolds with asymptotically optimal covering radius. arXiv:1607.06899, 2016 Brandolini, L; Choirat, C; Colzani, L; Gigante, G; Seri, R; Travaglini, G: Quadrature rules and distribution of points on manifolds. Ann. Sc. Norm. Super. Pisa Cl. Sci. (5) 13 (2014), no. 4, Y Brenier, U Frisch, M Hénon, G Loepe, S. Matarrese R. Mohayaee A. Sobolevskiĭ: Reconstruction of the early Universe as a convex optimization problem. Monthly Notices of the Royal Astronomical Society, Volume 346, Issue 2, 1 December 2003, Pages 501524, Bondarenko, A; Radchenko, D; Viazovska, M: Optimal asymptotic bounds for spherical designs. Annals of Mathematics, Second Series, 178 (2013) (2): 443452 Brenier, Yann Polar factorization and monotone rearrangement of vector-valued functions. Comm. Pure Appl. Math. 44 (1991), no. 4, 375417. Brauchart, J. S; Saff, E. B; Sloan, I; Womersley, R. S: QMC designs: optimal order quasi Monte Carlo integration schemes on the sphere. Math. Comp. 83 (2014), no. 290, 28212851. R. Burkard, M. DellAmico, and S. Martello. Assignment Problems . Society for Industrial and Applied Mathematics, Philadelphia, PA, USA, L. A. Caffarelli, S. A. Kochengin, and V. I. Oliker. On the numerical solution of the problem of reflector design with given far-field scattering data. In Monge Amp‘ere equation: applications to geometry and optimization (Deerfield Beach, FL, 1997), vol. 226 of Contemp. Math. , pages 1332. Amer. Math. Soc., Providence, RI, 1999 Caffarelli, L.A; Gutiérrez, C.E; Huang, Q: On the regularity of reflector antennas. Ann. of Math. (2) 167 (2008), no. 1, 299323. Cao, H.D: Deformation of Kähler metrics to Kähler-Einstein metrics on compact Kähler manifolds. Invent. Math. 81 (1985), no. 2, 359372.Neumann and second bou ndary value problems for Hessian and Gauss curvature flows. Carlier, G; Duval, V; Peyré, G, Schmitzer, B: Convergence of entropic schemes for optimal transport and gradient flows. SIAM J. Math. Anal. 49 (2017), no. 2, 13851418. Chizat, L; Peyré, G; Schmitzer, B; Vialard F-X: Scaling Algorithms for Unbalanced Transport Problems. ArXiv: 1607.05816 (Mathematics of Computation, to appear), Cordero-Erausquin, D: Sur le transport de mesures périodiques, C.R. Acad. Sci. Paris Sér. I Math. 329 (1999), 199202 K Crane, C Weischedel, M Wardetzky: Geodesics in heat: A new approach to computing distance based on heat flow. ACM Transactions on Graphics. Volume 32 Issue 5 (2013) Crandall, M.G.; Ishii, i; Lions, P-L: User’s guide to viscosity solutions of second order partial differential equations. American Mathematical Society. Bulletin. New Series, 27 (1) (1992), pp 167 Csiszar, I. (1975), I-divergence geometry of probability distributions and minimization problems, Ann. Probab. 3, 146- 158. Cuturi, M: Sinkhorn distances: Lightspeed computation of optimal transport. , in Advances in Neural Information Processing Systems (NIPS) 26, MIT Press, Cambridge, MA, 2013, pp. 22922300 Cuturi, M; Peyré, G: Computational Optimal Transport. Online book at https://optimaltransport.github.io/ P. M. Manhães de Castro, Q. Mérigot, B. Thiber; Far-field reflector problem and intersection of paraboloids. Numer. Math. 134 (2016), no. 2, 389411 De Philippis, Guido; Figalli, Alessio: The Monge-Ampère equation and its link to optimal transportation. Bull. Amer. Math. Soc. (N.S.) 51 (2014), no. 4, 527580 A Dominitz, A Tannenbaum: Texture mapping via optimal mass transport. IEEE Transactions on Visualization and Computer Graphics. Vol: 16, Issue: 3 (2010) Driscoll, J; Healy, D. M., Jr.: Computing Fourier transforms and convolutions on the 2-sphere. Adv. in Appl. Math. 15 (1994), no. 2, 202250. Donaldson, S. K. Some numerical results in complex differential geometry. Pure Appl. Math. Q. 5 (2009), no. 2, Special Issue: In honor of Friedrich Herze bruch. Part 1, 571618. Douglas, M.R; Karp, R. L; Lukic, S; Reinbacher, R: Numerical Calabi-Yau metrics. J. Math. Phys. 49 (2008), no. 3, 032302, 19 pp Feng, X; Glowinski, R; Neilan, : Recent developments in numerical methods for fully nonlinear second order partial differential equations. SIAM Rev. 55 (2013), no. 2, 205267 Franklin, J; Lorenz, J: On the scaling of multidimensional matrices. Linear Algebra Appl. 114/115 (1989), 717735. Glimm, T.; Oliker, V. Optical design of single reflector systems and the Monge-Kantorovich mass transfer problem. Nonlinear problems and function theory. J. Math. Sci. (N. Y.) 117 (2003), no. 3, 40964108 Healy, D. M., Jr.; Rockmore, D. N; Kostelec, P. J; Moore, S: FFTs for the 2-sphere-improvements and variations. J. Fourier Anal. Appl. 9 (2003), no. 4 Hesse, K., Sloan, I.H., Womersley, R.S.: Numerical integration on the sphere. In: W. Free- den, M.Z. Nashed, T. Sonar (eds.) Handbook of Geomathematics, 1st edn., pp. 11871220. Springer Berlin Heidelberg, Berlin, Heidelberg (2010 Hultgren, J: Permanental Point Processes on Real Tori, Theta Functions and Monge-Ampère Equations. arXiv:1604.05645 Keiner, J.; Kunis, S.; Potts, D. Fast summation of radial functions on the sphere. Computing 78 (2006), no. 1, 115 Young-Heon Kim, Jeffrey Streets, and Micah Warren: Parabolic optimal transport equations on manifolds , Int. Math. Res. Not. IMRN 19 (2012) Jun Kitagawa, A parabolic flow toward solutions of the optimal transportation problem on domains with boundary ,J.ReineAngew.Math. 672 (2012), 127160 Jun Kitagawa, Quentin Mérigot, Boris Thibert: Convergence of a Newton algorithm for semi-discrete optimal transport. Journal of the European Math. Society (to appear). S. Kolouri, S. Park, M. Thorpe, D. Slepcev, G. K. Rohde: Transport-based analysis, modeling, and learning from signal and data distributions. https://arxiv.org/pdf/1609.04767.pdf McCann, R. J: Polar factorization of maps on Riemannian manifolds. Geom. Funct. Anal. 11 (2001), no. 3, 589608 Léonard, C: From the Schrödinger problem to the Monge-Kantorovich problem. J. Funct. Anal. , 262(4):18791920, 2012. Léonard, C: A survey of the Schrödinger problem and some of its connections with optimal transport. Discrete Contin. Dyn. Syst. 34 (2014), no. 4, 15331574. M Lindsey, YA Rubinstein: Optimal transport via a Monge-Amp\\ere optimization problem. SIAM Journal on Mathematical Analysis, 2017 - SIAM Ludger Ruschendorf: Convergence of the Iterative Proportional Fitting Procedure. Ann. Statist. Volume 23, Number 4 (1995), 1160-1174. Norris, J. R. Heat kernel asymptotics and the distance function in Lipschitz Riemannian mani- folds. Acta Math. 179 (1997), no. 1, 79103. Oliver C. Schn urer and Knut Smoczyk. Neumann and second boundary value problems for Hessian and Gauss curvature flows. Ann. Inst. H. Poincar e Anal. Non Lin eaire , 20(6): Rokhlin, V; Tygert, M: Fast algorithms for spherical harmonic expansions. SIAM J. Sci. Comput. 27 (2006), no. 6, 19031928. Louis-Philippe Saumier, Martial Agueh, and Boualem Khouider. An efficient numerical algorithm for the l2 optimal transport problem with periodic densities. IMA Journal of Applied Mathematics , 80(1):135157, 2015. Schmitzer, B:Stabilized Sparse Scaling Algorithms for Entropy Regularized Transport Problems. Arxiv: 1610.06519 E. Schrödinger: Sur la théorie relativiste de lélectron et linterprétation de la mécanique quantique, Ann. Inst. H. Poincaré 2 (1932) 269310 Sinkhorn, R: Diagonal equivalence to matrices with prescribed row and column sums. Amer. Math. Monthly 74 1967 402405. C. D. Sogge, Fourier integrals in classical analysis. Cambridge Tracts in Mathematics, 105. Cambridge University Press, Cambridge, 1993. Solomon, J; De Goes, F; Peyré, G; Cuturi, M; Butscher, A; Nguyen, A; Du, T; Guiba, L: Convolutional Wasserstein distances. ACM Transactions on Graphics, Association for Computing Machinery, 2015, 34 (4) Solomon, J: Optimal Transport on Discrete Domains: Proceedings of Symposia in Pure Mathematics (to appear); http://people.csail.mit.edu/jsolomon/assets/optimal\_transport.pdf Sulman,M.M; Williams, J. F; Russell, Robert D: An efficient approach for the numerical solution of the Monge-Ampère equation. Appl. Numer. Math. 61 (2011), no. 3, 298307. Villani, C: Topics in optimal transportation. Graduate Studies in Mathematics, 58. American Mathematical Society, Providence, RI, 2003. xvi+370 pp Villani, C: Optimal transport. Old and new. Grundlehren der Mathematischen Wissenschaften 338. Springer-Verlag, Berlin, 2009 Wang, Xu-Jia: On the design of a reflector antenna II. Calc. Var. Partial Differential Equations 20 (2004), no. 3, 329341. H. Wellera; P.Brownea; C.Buddb; M.Cullen: Mesh adaptation on the sphere using optimal transport and the numerical solution of a MongeAmpère type equation. J. of Computational Physic Vol. 308 (2016) Pages 102-123 Womersley, R.S: Efficient spherical designs with good geometric properties. To appear in “Festschrift for the 80th Birthday of Ian H.Sloan”, Editors J.Dick et al (2018). See files at http://web.maths.unsw.edu.au/\~rsw/Sphere/EffSphDes/ X Zhao, Z Su, XD Gu, A Kaufman, J Sun; J Gao ; F Luo: Area-Preservation Mapping using Optimal Mass Transport. IEEE Transactions on Visualization and Computer Graphics. Vol.: 19, Issue: 12 (2013 ) [^1]: If $f$ and $g$ are merely assumed to be in the Hölder space $C^{\alpha}(T^{n}),$ for $\alpha>0,$ then the convergence results below still apply (with similar proofs), but with an error term of the order $k^{-\alpha/2}.$ [^2]: When $f=g=0$ the corresponding twisted Kähler-Ricci flow coincides with Hamilton’s Ricci flow restricted to the space of Kähler metrics.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We study the generic transport of slice energy between the scalar field generated by the conformal transformation of higher-order gravity theories and the matter component. We give precise relations for this exchange in the cases of dust and perfect fluids. We show that, unless we are in a stationary spacetime where slice energy is always conserved, in non-stationary situations contributions to the total slice energy depend on whether or not test matter follows geodesics in both frame representations of the dynamics, that is on whether or not the two conformally related frames are physically indistinguishable.' --- [Slice Energy in Higher-Order Gravity Theories and Conformal Transformations]{} [Spiros Cotsakis]{}\ \ \ \ \ [ `[email protected]`]{} [(5,0)[280]{}]{} Introduction ============ With recent advances in observational cosmology [@obs] $f(R)$ theories (and the closely related family of scalar-tensor ones) have in the last few years regained much attention both in cosmology [@hog-cosm] and in other contexts [@hog]. Perhaps the most economical and convenient way to study such modified gravity theories is through their well-known conformal relation to general relativity, a technique developed in the eighties by different groups [@con]. In this conformal method the $f(R)$-vacuum equations on a spacetime $(\mathcal{V},g)$ (the so-called Jordan frame) are transformed via a conformal transformation of the type $\tilde{g}=e^\phi g$, where $\f$ is a function of $\pa f/\pa R$, to become on the conformally related spacetime $(\mathcal{V},\tilde{g})$ (the so-called Einstein frame) Einstein equations for the metric $\tilde{g}$ with the scalar field $\f$ having a self-interacting potential that depends on the function $f(R)$ and its first derivatives. This technique provides a refreshing way to view the $f(R)$-vacuum theory as a unified theory of gravitation (described in the Einstein frame by the metric $\tilde{g}$) and the scalar field $\f$, that is as a theory uniting general relativity and the (lagrangian) theory of the scalar field. In the Jordan frame only one single geometric object appears, the ‘metric’ $g$, and the conformal transformation then serves as a tool to ‘fragment’ $g$ into its two pieces in the Einstein frame, namely the gravitational field $\tilde{g}$ and the scalar field $\f$. In a higher-order gravity theory with matter we have the following different pieces of information involved in the conformal transformation: - *gravity*, the field $g$ or the conformally related field $\tilde g$ - the scalar field $\f$. - the various matter fields $\psi$ which couple non-minimally to $\tilde g$ and to $\f$, or their conformal transform $\tilde\psi$ which couples minimally to $\tilde g$ but is not coupled to $\f$. The conformal transformation then relates the different pieces of matter and spacetime geometry and describes the interaction between the components listed above in the context of $f(R)$ theories. We describe in this paper how interactions of this sort lead to an exchange of slice energy between the various fields and spacetime geometry. For more varied interactions and energy transfer models of interest to cosmological situations see [@ba1; @ba2] and references therein. The plan of this paper is as follows. The next Section is preliminary and includes two simple applications (Corollaries 2.1, 2.2) of the basic properties (Theorems 2.1, 2.2) of the slice energy. These basic properties and also their proofs are included here for easy reference and also to establish notation. Their applications deal with the simpler case of $f(R)$ theories in vacuum. Section 3 is the heart of this paper. There, we find the general (slice) energy transport equation and study in detail its application to the case of a general $f(R)$ theory coupled to matter, where by matter we mean a general perfect fluid-scalar field system. There are many properties analysed here but a particularly important one deals with conditions under which the choice of ‘physical’ metric influences the net contribution to slice energy of the system. We conclude with a discussion of these results in Section 4. Energy on a slice ================= Consider a time-oriented spacetime $(\mathcal{V},g)$ with $\mathcal{V}=\mathcal{M} \times \mathbb{R},$ where $\mathcal{M}$ is a smooth manifold of dimension $n$, $g$ a spacetime metric and the spatial slices $\mathcal{M}_{t}\,(=\mathcal{M}\times \{t\})$ are spacelike submanifolds endowed with the time-dependent spatial metric $g_{t}$. (In the following, Greek indices run from $0$ to $n$, while Latin indices run from $1$ to $n$. We also assume that the metric signature is $(+-\cdots -)$.) On $(\mathcal{V},g)$ we consider a family of matterfields denoted collectively as $\psi $, assume that the field $\psi$ arises from a lagrangian density which we denote by $L$ and denote the stress tensor of the field $\psi$ by $T(\psi )$. For $X$ any causal vectorfield of $\mathcal{V}$ we define the *energy-momentum vector $P$ of a stress tensor $T$ relative to $X$* to be \[p\] P\^=X\_T\^. The *energy on the slice $\mathcal{M}_{t}$ with respect to $X$* is defined by the integral (when it exists) E\_[t]{}=\_[\_[t]{}]{}P\^n\_d\_[t]{}, where $n$ is the unit normal to $\mathcal{M}_{t}$ and $d\m_{t}$ is the volume element with respect to the spatial metric $g_{t}$. We call $P^{\a}n_{\a}$ the *energy density*. Assume that $X$ and $T$ are smooth. Then we have $$\na_{\a}P^\a =\na_{\a}(X_{\b}T^{\a\b})=\na_{\a}X_{\b}T^{\a\b}+X_{\b}\na_{\a}T^{\a\b}$$ or, equivalently, \_P\^=T\^(\_X\_+\_X\_)+X\_\_T\^. Thus, if $\mathcal{K}\subset\mathcal{V}$ is a compact domain with smooth boundary $\pa\mathcal{K}$, it follows from Stokes’ theorem that \_\_P\^d=\_P\^n\_d, where $d\mu$ is the volume element of $\mathcal{V}$ and $d\sigma$ that of $\pa\mathcal{K}$, and so we find \_P\^n\_d= \_T\^(\_X\_+\_X\_)d+ \_X\_\_T\^d. Hence, when $\mathcal{M}$ is compact or the field falls off appropriately at infinity, on the spacetime slab $\mathcal{D}=\Sigma\times[t_{0},t_{1}]$, $\Sigma\subset\mathcal{M}$, and with $T$ having support on $\mathcal{D}$ we have the following relation for the energies on the two end-slices \_[\_[t\_[1]{}]{}]{}P\^n\_d\_[t\_[1]{}]{}- \_[\_[t\_[0]{}]{}]{}P\^n\_d\_[t\_[0]{}]{} &=&\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{} T\^(\_X\_+\_X\_)d\ &+&\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{} X\_\_T\^dor \[trans-eq\] E\_[t\_[1]{}]{}-E\_[t\_[0]{}]{}=\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{} T\^(\_X\_+\_X\_)d+\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{} X\_\_T\^d. We therefore see that when $X$ is a Killing vectorfield the first term on the right-hand-side of Eq. (\[trans-eq\]) is zero and so we have \[en1\] E\_[t\_[1]{}]{}-E\_[t\_[0]{}]{}=\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{} X\_\_T\^d. Thus we have shown the following result (cf. [@ch-mo00], p. 87-88). \[1\] When $X$ is a Killing vectorfield and the field is conserved, i.e., $\na_{\a}T^{\a\b}=0$, we have E\_[t\_[1]{}]{}=E\_[t\_[0]{}]{}. This means that, when the energy-momentum tensor of a field is conserved, the same is true for its slice energy relative to a Killing vectorfield as a function of time. In the next Section we pay particular attention to the case for which the field is a matter field $\psi$ interacting with a scalar field $\phi$ with potential $V(\phi)$. We take the scalar field lagrangian density to be L=-g\^\_\_+V(). Then the energy-momentum tensor of $\phi$ is \[t\] T\^()=\^\^-g\^(\^ł\_ł-2V()), and we have the following result. \[2\] The energy density $P^\a n_\a$ of the scalar field $\phi$ with potential $V(\phi)$ is positive when $V(\phi)>0$. The proof, which we give here for a general scalar field potential $V(\f )$, is a direct adaptation with slight modifications of that found in [@ch-mo00], p. 88, for a power-law potential. Using Eq. (\[t\]) we calculate P\^n\_=-X\^n\_\^\_+X\^\_n\^\_+X\^n\_V() . We define the quadratic form \^[ł]{}=-g\^[ł]{}X\^n\_+(X\^łn\^+X\^n\^ł) and then we find that P\^n\_=-X\^n\_g\^[ł]{}\_ł\_+(X\^łn\^+X\^n\^ł) \_ł\_+X\^n\_V(). This means that \[pn\] P\^n\_=\^[ł]{}\_ł\_+X\^n\_V(). Since $\mathcal{M}_t$ is a $t=\mathrm{const.}$ hypersurface, we can choose coordinates such that $X^0=1, X^i=0, n_i=0$. Then $n_0=(g^{00})^{-1/2}$, $n^i=g^{i0}(g^{00})^{-1/2}$ and so it follows that the quadratic form $\gamma$ is positive definite, \^[00]{}=g\^[00]{}n\_[0]{},\^[i0]{}=0\^[ij]{}=-g\^[ij]{}n\_[0]{}, (recall signature of $g_{ij}$ is $(-\cdots -)$). We then find that P\^n\_&=&(\^[00]{}\_0\_0+2\^[i0]{}\_i\_0+\^[ij]{}\_i\_j) +n\_0 V()\ &=&( n\_0 g\^[00]{}\^[2]{}-g\^[ij]{}n\_[0]{}\_i\_j) +n\_0 V()and therefore we conclude that the energy density $ P^\a n_\a $ is positive whenever $V(\phi)>0$. This concludes the proof. To end this Section, for the following simple application of the preceding developments we restrict attention to $n=4$ spacetime dimensions although everything we do becomes valid with minor modifications to arbitrary $n$. The following notation for conformally related quantities is used: Let $g$ and $\tilde g$ be two conformal metrics, $\tilde g =\Omega^{2} g$, on the manifold $\mathcal{V}$. This means that in two *orthonormal* moving frames, $\theta^\a$ and $\tilde\theta^\a$, the two conformal metrics satisfy g =\_\^\^,g=\_\^\^\^=\^[-1]{}\^, with $\eta_{\a\b}=\textrm{diag} (+,-\cdots -)$ being the flat metric. Setting $\Omega^{2} =e^{\f}$ we see that $\tilde\theta^\a=e^{-\f/2}\theta^\a$ and obviously $\tilde\theta_\a=e^{\f/2}\theta_\a$. The same rules are true for any 1-form or vectorfield on $\mathcal{V}$. Consider now the $f(R)$-*vacuum* equations, \[f of r vac\] L\_f’R\_-g\_f-\_\_f’+g\_\_[g]{}f’=0, where the left hand side satisfies the conservation identities (cf. [@ed23], p. 140) \[cons id\] \_L\^=0. Then we conformally transform from $(\mathcal{V},g)$ to the Einstein frame $(\mathcal{V},\tilde{g})$, according to the prescription given in [@con], that is, we set \[ct\] =f’, to obtain the Einstein equations with a scalar field ‘matter source’ of potential $V(\phi)=(1/2)(f')^{-2}(Rf'-f)$ and energy-momentum tensor given by Eq. (\[t\]): \[ceq\] \_=\_(). In this case we conclude that the field $\f$ is conserved, i.e., \_\^()=0, and, since \[sw0\] \_\^()=\^(\_\^+V’), we find that the $\f$-field is a scalar field satisfying the wave equation \[sw1\] \_\^+V’=0. Further from Theorem \[1\] we have the following result. The slice energy of the scalar field $\f$ generated by the conformal transformation (\[ct\]) to the Einstein frame of the $f(R)$-*vacuum* equations (\[f of r vac\]) relative to a Killing vectorfield of $\tilde g$ is conserved, i.e., \_[t]{}()=\_[\_[t]{}]{} \^\_d\_[t]{}= with $d\tilde{\m}_{t}$ being the volume element of $\tilde{g}_{t}$. Secondly from Theorem \[2\] we have: For all $f(R)$-vacuum theories (\[f of r vac\]) with a positive potential in the Einstein frame the energy density $\tilde{P}^\a \tilde{n}_\a$ of $\f$ is positive. Examples of theories in the last Corollary include, for instance, the choice $f(R)=R+\a R^2, \a >0$. $f(R)$-matter systems ===================== Suppose now that we start by coupling a matter field $\psi$ to the geometry in $(\mathcal{V},g)$ via the $f(R)$-matter field equations f’R\_-g\_f-\_\_f’+g\_\_[g]{}f’=T\_(). Because of the conservation identities (\[cons id\]), the field $\psi$ satisfies the conservation laws \[cons laws\] \_T\^()=0. Then, if we conformally transform from $(\mathcal{V},g)$ to the Einstein frame $(\mathcal{V},\tilde{g})$ according to (\[ct\]), in place of equations (\[ceq\]) we obtain \[ceq1\] \_=\_()+\_( ), where now the whole tensor in the right-hand-side is conserved, namely \[3.7\] \_(\^()+\^( ))=0, but the two components are *not* conserved separately, that is \_\^()0 and \[3.9\] \_\^( )0, unless the conservation equations (\[cons laws\]) for the field $\psi$ are conformally invariant (conditions for this are given in [@wa84], p. 448). This result (already given in Ref. [@co93]) indicates that in higher order gravity theories there must be a generic, nontrivial $\phi -\tilde\psi$ interaction between the matter field $\tilde{\psi}$ and the $\phi$-field, and an associated exchange of energy between $\phi$ and $\tilde\psi$. Writing Eq. (\[trans-eq\]) for the scalar field $\f$ and substituting for the last term in the right-hand-side from Eq. (\[3.7\]) we find the *general energy transport equation* in the Einstein frame, \[en1dust\] E\_[t\_[1]{}]{} ()-E\_[t\_[0]{}]{} ()=\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{} \^()(\_\_+\_\_)d -\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{} \_\_\^( )d, with $ d\tilde{\m}$ being the volume element of $\tilde{g}$. This result is only symbolic and has to be augmented by precise equations satisfied by the fields. Since the stress tensor of the $\phi$-field is not separately conserved, it follows that the $\phi$-field will not satisfy the usual scalar wave equation, $\widetilde{\square}_{\tilde{g}}\f +V'(\f )=0$, but this equation will in general contain new terms. Similarly for the ‘ordinary matter’ $\tilde\psi$-field, whatever its form (scalar field, Maxwell, a fluid etc), its field equations have new terms indicating the $\phi -\tilde\psi$ interaction and associated energy exchange. For instance, if the $\tilde\psi$ field is another scalar field, then the equations satisfied by its conformal transform, $\tilde\psi$, have a general form of the type $ \widetilde\square\tilde\psi +h(\f)\,\pa_\a\f\,\pa^\a\tilde\psi =0, $ where $h(\f)$ is a smooth function of $\f$ (often exponential). To study this interaction and the associated energy exchange between $\phi$ and $\tilde{\psi}$ more closely we give some concrete examples. Suppose firstly that $\psi$ is a dust cloud on $(\mathcal{V},g)$ with 4-velocity $V_\a$ and stress tensor T\_[,]{}=V\_V\_, satisfying the $f(R)$*-dust* equations in the Jordan frame, namely \[f-dust\] f’R\_-g\_f-\_\_f’+g\_\_[g]{}f’=V\_V\_. Then \[cons dust\] \_(V\^V\^)=0, and it is obvious that here the dust streamlines are geodesics, that is $V^\a$ is the tangent vectorfield to the geodesics. After the conformal transformation we find \[ein-frame-dust\] \_=\_()+ \_\_, with $\tilde{T}_{\a\b}(\f)$ given by (\[t\]) with tildes where appropriate, and with=e\^[-/2]{} V\_[a]{},=e\^[-2]{}. (We have set $\tilde\rho =\Omega^{-4}\rho$ and since $\Omega^2=e^\f$, $\Omega^{-4} =e^{-2\f}$.) What is the field equation satisfied by the scalar field $\f$? From Eq. (\[ein-frame-dust\]) the divergence of the stress tensor of $\f$ is minus that of the dust, but \_( \^\^)= \_( \^\^)+A\^\_ \^\^+A\^\_ \^\^, where A\^\_=(\^\_\_+\^\_\_-g\_g\^\_). From these equations and Eq. (\[cons dust\]) we deduce the modified scalar field equation in the form \[sc field eqn\] \^(+V’)+ \^\^\_-\^=0. Another way to derive the scalar field equation is as follows. Since \[3.16\] \_\^\_=\_( \^\^)= \^\_( \^)+(\_\^)\^, and, since $\tilde{V}_\b \tilde{V}^\b =1$, if we multiply Eq. (\[sw0\]) by $\tilde{V}_\b$ and use the fact that the divergence of the right hand side of Eq. (\[ein-frame-dust\]) is zero to arrive at the following equation for the scalar field $\f$ in the Einstein frame, namely, \[sc field eqn1\] \^(+V’)+ \^\_(\^) +\^\_\^=0. Recalling that dust matter follows geodesics on the original Jordan frame, $V^\a\na_\a V^\b=0$, and taking it *as a working hypothesis* that the same is true in the conformally related Einstein frame, we find that the last two terms in this equation are equal to the last two terms in Eq. (\[sc field eqn\]) and so we conclude that Eq. (\[sc field eqn1\]) provides an equivalent form of Eq. (\[sc field eqn\]). We note that only in the very special case where we impose the constraint \[3.18\] \_=\_, which implies some sort of ‘alignment’ between the dust component and the scalar field, does the scalar field equation (\[sc field eqn\]) becomes the standard one, namely, +V’ =0. We now study the behaviour of the total slice energy of the system comprised of $\f$ and the dust component. We choose $V=X$ so that \[3.11\] P\^n\_=X\_n\_V\^V\^=V\^n\_. Hence, applying Stokes’ theorem we obtain \[stokes\] \_\_( \^)d=\_ \^\_d. Therefore Eq. (\[en1dust\]) becomes E\_[t\_[1]{}]{} ()-E\_[t\_[0]{}]{} () &=&\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{} \^()(\_\_+\_\_)d\ &-&\ &=&\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{} \^()(\_\_+\_\_)d +\_[\_[t\_[0]{}]{}]{}\^\_d\_[t\_[0]{}]{}\ &-&\_[\_[t\_[1]{}]{}]{}\^\_ d\_[t\_[1]{}]{}, or \[3.26A\] E\_t () +E\_t () =E\_0 () +E\_0 () +\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{} \^()(\_\_+ \_\_)d, where by definition and Eq. (\[3.11\]), for any $t$, E\_t ()=\_[\_[t]{}]{}V\^n\_d\_[t]{}. We see that the last term in Eq. (\[3.26A\]) can be zero only when $V$ is a Killing vectorfield. We therefore arrive at the following result about the total slice energy with respect to the fluid itself. \[thm1\] The total slice energy with respect to the timelike vectorfield $\tilde V$, tangent to the dust timelines, of the scalar field-dust system satisfying the field equations (\[f-dust\]), satisfies \[3.26AA\] E\_t (+) =E\_0 (+ ) +\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{} \^()(\_\_+ \_\_)d. In particular the slice energy of the scalar field-dust system is conserved when $\tilde V$ is a Killing vectorfield of $\tilde g$. We also conclude that the property of the conservation of slice energy for dust is a conformal invariant. However, when $V$ is not a Killing vectorfield, we see that there is a nontrivial contribution to the slice energy coming from the combined effect of the stress tensor of the scalar field generated by the conformal transformation coupled to the non-stationarity of the spacetime due to the lack of a Killing vector. Note that this contribution is also nonzero even in the special case that Eq. (\[3.18\]) is assumed for in that case the first term in $\tilde{T}^{\alpha \beta }(\phi ) (\tilde{\nabla}_{\alpha } \tilde{V}_{\beta }+\tilde{\nabla}_{\beta }\tilde{V}_{\alpha})$ is zero because $\tilde{V}^{\alpha }\tilde{V}^{\beta }(\tilde{\nabla}_{\alpha }\tilde{V}_{\beta }+\tilde{\nabla}_{\beta}\tilde{V}_{\alpha})=0$, but the whole combination is still not zero as there are additional terms coming from the contributions of the other terms in Eq. (\[t\]) (unless the fluid satisfies an extra condition – see below). We now proceed to see how this result changes when we assume that $(\mathcal{V},g)$ is filled with a perfect fluid. With our conventions the stress tensor of a perfect fluid with energy density $\rho$ and pressure density $p$ is $T_{\a\b}=(\rho +p)V_\a V_\b -pg_{\a\b}$ and the fluid satisfies the field equations \[f(R)-fluid\] f’R\_-g\_f-\_\_f’+g\_\_[g]{}f’=(+p)V\_V\_-pg\_. In this case, because the energy-momentum vector $P^\a =\rho V^\a$, we have $P^\a n_\a = \rho V^\a n_\a$ and so the slice energy with respect to the timelike vectorfield $\tilde{V}$ in the Einstein frame is again E\_[t]{}()=\_[\_[t]{}]{}\^\_d\_[t]{}. Then \[1a\] \_T\^\_=\_V\^+(+p)V\^\_V\^-\^p. Since $\tilde V_\a\tilde V^\a =1$, we have $\tilde V_\a\tilde\na_\b\tilde V^\a =0$ and so on multiplication of Eq. (\[1a\]) by $\tilde V_\b$ we find \[3.25\] V\_\_T\^\_ &=&\_- V\^\_p\ &=&p\_V\^+\_(\^). Integrating Eq. (\[3.25\]) on the spacetime slab $\mathcal{D}$ and using Eq. (\[stokes\]) to re-express the $\rho$-term in (\[3.25\]) we have \_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{} -V\_\_T\^\_d=\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{}-p\^V\_d-and therefore we obtain from (\[en1dust\]) the general energy transport equation in the form E\_[t\_[1]{}]{} ()-E\_[t\_[0]{}]{} () &=&\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{} \^()(\_\_+\_\_)d\ &-&\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{}p\^V\_d+E\_[t\_[0]{}]{} ()-E\_[t\_[1]{}]{} (). We thus arrive at the following result. The total slice energy of the scalar field – perfect fluid system satisfying the field equations (\[f(R)-fluid\]), depends upon the integrated pressure according to the formula \[en1fluid\] E\_[t\_[1]{}]{} (+)=E\_[t\_[0]{}]{} (+) +\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{} \^()(\_\_+\_\_)d -\_[t\_[0]{}]{}\^[t\_[1]{}]{} \_[\_[t]{}]{}p\^V\_d. In particular, the slice energy is conserved when $\tilde V$ is a Killing vectorfield for $\tilde g$ (stationary spacetime). When $V$ is not a Killing vectorfield, this slice energy is not generally conserved and this is true even in the special case of a fluid with zero expansion, $\tilde\na^\a \tilde V_\a =0$, for which the last term in Eq. (\[en1fluid\]) is zero. In this case the term depending on the scalar field continues to have a nonzero contribution to the total slice energy. This term is given by T\^()(\_\_+\_\_)= \^\^(\_\_+\_\_) -\_V\^(\^ł\_ł-2V()) and so Eq. (\[en1fluid\]) becomes \[en1fluid1\] E\_[t\_[1]{}]{} (+)&=&E\_[t\_[0]{}]{} (+) +\_[t\_[0]{}]{}\^[t\_[1]{}]{}\_[\_[t]{}]{} \^\^(\_\_+\_\_)d\ &-&\_[t\_[0]{}]{}\^[t\_[1]{}]{} \_[\_[t]{}]{}(\^ł\_ł+p -V())\^V\_d. We conclude that the only other possible case for which we have slice energy conservation is when $V$ is not a Killing vectorfield for $\tilde g$, but the alignment condition (\[3.18\]) holds (making the middle term in Eq. (\[en1fluid1\]) equal to zero) and the fluid has in addition zero expansion (last term in Eq. (\[en1fluid1\]) is zero). Discussion ========== The results of this paper allow us to make some comments concerning the problem of deciding which of the two frames (or metrics), Jordan or Einstein, is the physical one, meaning in which of the two representations of the dynamics test particles follow geodesics (assuming the validity of the principle of equivalence). In the case where test particles follow geodesics in both frames, one says that the two conformally related frames are *physically equivalent*. The main results of Section 3, in particular Eq. \[3.26AA\] (as well as its pressure extension - Eq. \[en1fluid\]), were proved under the implicit assumption that the vectorfield $V^a$ of the dust streamlines generates geodesics in both the Jordan frame *and* the conformally related Einstein frame. But does $\tilde{V}^{\alpha }$ always generate a geodesic in the latter frame so that $\tilde{V}_{\beta }\tilde{\nabla}_{\alpha }\tilde{T}_{dust}^{\alpha \beta }=\tilde{\nabla}_{\alpha }(\tilde{\rho}\tilde{V}^{\alpha })$ (cf. Eq. \[3.16\])? In general, it will not do so and the two frames will not be physically equivalent. In this case, we have \_\_\^\_=\_( \^)\_\^+ (\_\^)\^\_, so that the result of Theorem \[thm1\] becomes, $$\begin{aligned} \label{3.26AAA} E_t (\f +\mathrm{dust}) &=&E_0 (\f + \mathrm{dust}) +\frac{1}{2}\int_{t_{0}}^{t_{1}}\int_{\mathcal{M}_{t}} \tilde{T}^{\a\b}(\f )(\tilde{\na}_{\a}\tilde{V}_{\b}+ \tilde{\na}_{\b}\tilde{V}_{\a})d\tilde{\mu}\nonumber\\&-&\int_{t_{0}}^{t_{1}}\int_{\mathcal{M}_{t}}\tilde{\rho} (\tilde{\na}_\a \tilde{V}^\b )\tilde{V}^\a \tilde{V}_\b d\tilde{\mu},\nonumber\\ &=&E_0 (\f + \mathrm{dust}) +\frac{1}{2}\int_{t_{0}}^{t_{1}}\int_{\mathcal{M}_{t}} \left(\tilde{T}^{\a\b}(\f )-\tilde{\rho}\tilde{V}^\a \tilde{V}^\b\right)(\tilde{\na}_{\a}\tilde{V}_{\b}+ \tilde{\na}_{\b}\tilde{V}_{\a})d\tilde{\mu}.\nonumber\\\end{aligned}$$ We may therefore conclude that the expressions for the total slice energy in the two situations considered here, namely, when test particles follow geodesics in both metrics $g,\tilde{g}$ (cf. Eq. \[3.26AA\]), or only in the original Jordan frame metric $g$ (cf. Eq. \[3.26AAA\]), are different and in the latter case there is an extra term contributing to the total energy (i.e., the dust term in the integrand in the last term in Eq. \[3.26AAA\]). This additional contribution will appear as a measurable quantity which, if measured to be nonzero, will lead us to conclude that the two conformally related frames cannot be physically indistinguishable. Acknowledgements {#acknowledgements .unnumbered} ================ We are indebted to Y. Choquet-Bruhat for her encouragement and for offering her precious time and opinion so willingly. These have led to a substantial improvement of this work. We also thank J. Miritzis for reading this paper and P.G.L. Leach for comments on a preliminary version of this manuscript. [99]{} A.G. Reiss, *et al*, Astron. J. [**116**]{} (1998) 1009; \[arXiv:astro-ph/9805201\]; S. Perlmutter, *et al*, Ap. J. [**517**]{} (1999) 565; \[arXiv:astro-ph/9812133\]; C. B. Netterfield, *et al*, Ap. J. [**571**]{} (2002) 604; \[arXiv:astro-ph/0104460\]; N. W. Halverson, *et al*, Ap. J. [**568**]{} (2002) 38 \[arXiv:astro-ph/0104489\]; C. L. Bennett, *et al*, \[arXiv:astro-ph/0302207\]; J. L. Tonry, *et al*, Ap. J. [**594**]{} (2003) 1; \[arXiv:astro-ph/0305008\]. S. Capozzielo, *et al*, \[arXiv:astro-ph/0303041\]; J. Miritzis, J.Math.Phys. 44 (2003) 3900-3910, \[arXiv:gr-qc/0305062\]; J. Miritzis, Class.Quant.Grav. 21 (2004) 3043-3056, \[arXiv:gr-qc/0402039\]; S. M. Carroll, *et al*, Phys.Rev. D70 (2004) 043528, \[arXiv:astro-ph/0306438\]; D. N. Vollick, Phys.Rev. D68 (2003) 063510, \[arXiv:astro-ph/0306630\]; S. Nojiri and S. D. Odintsov, Phys. Rev. D[**68**]{} (2003) 123512; \[arXiv:hep-th/0307288\]; S. Nojiri and S. D. Odintsov, Gen. Rel. Grav. [**36**]{} (2004) 1765; \[arXiv:hep-th/0308176\]; X. Meng and P. Wang, Class.Quant.Grav. 20 (2003) 4949-4962, \[arXiv:astro-ph/0307354\]; A.D. Dolgov, M. Kawasaki, Phys.Lett. B573 (2003) 1-4, \[arXiv:astro-ph/0307285v2\]; M. E. Soussa, R. P. Woodard, Gen.Rel.Grav. 36 (2004) 855-862, \[arXiv:astro-ph/0308114v2\]; S. Nojiri, S. D. Odintsov, Gen.Rel.Grav. 36 (2004) 1765-1780, \[arXiv:hep-th/0308176v2\]; D. N. Vollick, Phys.Rev. D69 (2004) 064030, \[arXiv:gr-qc/0309101v1\]; S. Nojiri, S. D. Odintsov, Mod.Phys.Lett. A19 (2004) 627-638, \[arXiv:hep-th/0310045v1\]; G. Allemandi, A. Borowiec, M. Francaviglia, Phys.Rev. D70 (2004) 043524, \[arXiv:hep-th/0403264v2\]; J.-F. Dufaux *et al*, Phys.Rev. D70 (2004) 083525, \[arXiv:hep-th/0404161v2\]; F. P. Schuller, M. N.R. Wohlfarth, Nucl.Phys. B698 (2004), \[arXiv:hep-th/0403056v2\]; M. Gasperini, Int.J.Mod.Phys. D13 (2004) 2267-2274, \[arXiv:gr-qc/0405083v1\]; S. Nojiri, TSPU Vestnik 44N7 (2004) 49-57, \[arXiv:hep-th/0407099v2\]; P. Wang , Gen.Rel.Grav. 38 (2006) 517-521, \[arXiv:gr-qc/0408058v1\]; M.C.B. Abdalla *et al*, Class.Quant.Grav. 22 (2005) L35, \[arXiv:hep-th/0409177v3\]; F. P. Schuller, M. N.R. Wohlfarth, Phys.Lett. B612 (2005) 93-99, \[arXiv:gr-qc/0411076v3\]; X. Meng and P. Wang, Class.Quant.Grav. 22 (2005) 23-32, \[arXiv:gr-qc/0411007v1\]; K. Maeda, N. Ohta, Phys.Rev. D71 (2005) 063520, \[arXiv:hep-th/0411093v4\]; V. V. Dyadichev *et al*, Phys.Rev. D72 (2005) 084021, \[arXiv:hep-th/0412334v3\]; J. Miritzis, J.Math.Phys. 46 (2005) 082502, \[arXiv:gr-qc/0505139\]; T. Clifton, J. D. Barrow, Phys.Rev. D72 (2005) 103005, \[arXiv:gr-qc/0509059v2\]; J. D. Barrow and T. Clifton, Class.Quant.Grav. 23 (2006) L1, \[arXiv:gr-qc/0509085v1\]; S. Nojiri, S. D. Odintsov, Phys.Lett. B631 (2005) 1-6; S. Carloni *et al*, Class.Quant.Grav. 23 (2006) 1913-1922, \[arXiv:gr-qc/0510130v1\]; J. A. Leach, *et al*, Class.Quant.Grav. 23 (2006) 4915-4937, \[arXiv:gr-qc/0603012v2\]; T. Clifton, J. D. Barrow, Phys.Rev. D72 (2005) 123003, \[arXiv:gr-qc/0511076v3\]; J. D. Barrow, S. Hervik, Phys.Rev. D73 (2006) 023007, \[arXiv:gr-qc/0511127v3\]; M. Bojowald, M. Kagan, Class.Quant.Grav. 23 (2006) 4983-4990, \[arXiv:gr-qc/0604105v1\]; R. Catena, *et al*, \[arXiv:astro-ph/0604492v2\]; S. Capozzielo, *et al*, Phys.Lett. B639 (2006) 135-143, \[arXiv:astro-ph/0604431v3\]; Luca Amendola, *et al*, Int. J. Mod. Phys. D16, 1555-1561, 2007, \[arXiv:astro-ph/0605384v2\]; S. Tsujikawa, Annalen Phys. 15 (2006) 302-315, \[arXiv:hep-th/0606040v2\]; D. Muller, S. D. P. Vitenti, Phys.Rev. D74 (2006) 083516, \[arXiv:gr-qc/0606018v2\]; W. Fang, *et al*, \[arXiv:hep-th/0606032v4\]; J. D. Barrow, S. Hervik, Phys.Rev. D74 (2006) 124017, \[arXiv:gr-qc/0610013v2\]; T. Clifton, J. D. Barrow, Class.Quant.Grav. 23 (2006) 2951, \[arXiv:gr-qc/0601118v1\]; T. Faulkner, *et al*, \[arXiv:astro-ph/0612569v1\]; B. Li, *et al*, Phys. Rev. D75, 084010 (2007), \[arXiv:gr-qc/0701111v2\]; J. Miritzis, arXiv:0708.1396v1 \[gr-qc\]; S. Cotsakis, A. Tsokaros, Phys. Lett. B. 651 (2007) 341-344, \[arXiv:gr-qc/0703043\]; N. Agarwal, R. Bean, \[arXiv:0708.3967\]; C. G. Boehmer, *et al*, arXiv:0709.0046v2 \[gr-qc\]. T. Chiba, Phys.Lett. B575 (2003) 1-3, \[arXiv:astro-ph/0307338\]; E. E. Flanagan, Phys.Rev.Lett. 92 (2004) 071101, \[arXiv:astro-ph/0308111\]; E. E. Flanagan, Class.Quant.Grav. 21 (2003) 417-426, \[arXiv:gr-qc/0309015\]; D. N. Vollick, Class.Quant.Grav. 21 (2004) 3813-3816 \[arXiv:gr-qc/0312041\]; Xin-He Meng, P. Wang, Gen.Rel.Grav. 36 (2004) 1947-1954, \[arXiv:gr-qc/0311019v1\]; E. E. Flanagan, Class.Quant.Grav. 21 (2004) 3817, \[arXiv:gr-qc/0403063\]; G. J. Olmo and W. Komp, \[arXiv:gr-qc/0403092\]; A. Nunez and S. Solganik, \[arXiv:hep-th/0403159\]; P. Wang, *et al*, Gen.Rel.Grav. 38 (2006) 517-521, \[arXiv:gr-qc/0408058v1\]; D. Vollick, Phys.Rev. D71 (2005) 044020, \[arXiv:gr-qc/0409068v1\]; X. Meng and P. Wang, Class.Quant.Grav. 22 (2005) 23-32, \[arXiv:gr-qc/0411007v1\]; C. H. Brans, arXiv:gr-qc/0506063v1; T.P. Sotiriou, Class.Quant.Grav. 23 (2006) 5117-5128, \[arXiv:gr-qc/0604028v2\]; S. Deser and B. Tekin, Class.Quant.Grav. 23 (2006) 7479-7482, \[arXiv:gr-0qc/0609111\]; G. J. Olmo, Phys.Rev.Lett. 98 (2007) 061101, \[arXiv:gr-qc/0612002\]; G. J. Olmo, Phys.Rev. D75 (2007) 023511, \[arXiv:gr-qc/0612047\]; G. Allemanti and M. Francaviglia, Int. J. Geom. Meth. Mod. Phys. 4 (2007) 1-23; S. Capozzielo, *et al*, arXiv:0708.3038v2; A. Iglesias, *et al*, arXiv:0708.1163v2; N. Lanahan-Tremblay, V. Faraoni, arXiv:0709.4414v2; D. Vollick, arXiv:0710.1859v2; N Deruelle, *et al*, arXiv:0711.1150v1; M. Salgado, *et al*, arXiv:0801.2372v1. J. D. Barrow and S. Cotsakis, Phys. Lett. B[**214**]{} (1988) 515; K. Maeda, Phys. Rev. D[**39**]{} (1989) 3159. M. Gasperini, *Int. J. Mod. Phys.* [**D13**]{} (2004) 2267-2274. J. D. Barrow and T. Clifton, *Phys. Rev.* [**D73**]{} (2006) 103520. J. D. Barrow and T. Clifton, *Phys. Rev.* [**D73**]{} (2006) 104022. Y. Choquet-Bruhat and C. DeWitt-Morette, *Analysis, Manifolds and Physics*, Part II, Revised and Enlarged Edition (North-Holland, 2000) A. Eddington, *The Mathematical Theory of Relativity*, (Chelsea, 1923) R. M. Wald, *General Relativity*, (Chicago University Press, 1984) S. Cotsakis, *Phys. Rev.* [**D47**]{} (1993) 1437-1439; Erratum: *Phys. Rev.* [**D49**]{} (1994) 1145.
{ "pile_set_name": "ArXiv" }
--- author: - 'Anamitra Mukherjee$^{1, 2}$, William S. Cole$^2$, Patrick Woodward$^3$, Mohit Randeria$^2$ and Nandini Trivedi$^2$' bibliography: - 'mnbib.bib' title: 'Supplementary material for “Theory of strain controlled magnetotransport and stabilization of the ferromagnetic insulating phase in manganite thin "' --- In this supplementary material we present the following: *1. Model:* First we give the details of the Hamiltonian and the parameters used in the main paper in Section I (A) and (B) respectively. We then discuss the scaling of the hopping and the superexchange parameters under strain in (C). There we argue why the change in the Mn-O bond length plays a dominant role in determining these scalings. *2. Method:* We present our method of solution in Section II. The definitions of the spin, charge and orbital structure factors and a brief account of transport calculations are discussed in Section III. *3. Supporting data:* In Section IV we show the magetotransport data (under strain) for a F-M parameter point close to the F-M/SCO-I phase boundary. When contrasted with Fig. 3(a) in the main text, they show that our conclusions regarding the strain response of magnetotransport are independent of the *nature of the insulator-metal boundary* near which the response is calculated. Hamiltonian and parameters ========================== The manganite Hamiltonian ------------------------- ![[*Schematics of the relevant interactions incorporated in the manganite Hamiltonian:*]{} (a) The manganese level diagram. The largest scale is the Hund’s rule scale that splits the up and down manifolds of the Mn $3d$ levels by $J_H$. Next in the energy hierarchy is the cubic crystal field, that splits that $e_g$ and the $t_{2g}$ states by $\Delta_{Crys}(=10Dq)$. For both the Mn$^{+3}$ and Mn$^{+4}$ states, three electrons occupy the $t_{2g}$ down manifold and are well-localized. The fourth electron, for Mn$^{+3}$, occupies the $e_g$ manifold. (b) As a result of placing a single electron in the degenerate $e_g$ manifold, the system spontaneously undergoes a Jahn-Teller distortion. The electron orbital-pseudospin operator couples to the local distortions $Q_2$, $Q_3$. (c) The effective antiferromagnetic coupling between two nearest neighbor Mn$^{+4}$ atoms. The two $p$ electrons on the oxygen connecting the two Mn sites can gain maximum kinetic energy by hopping to the two Mn sites if their $t_{2g}$ spins are antiparallel. This holds true even when the two Mn atoms are in the $+3$ state, except that the magnitude of this antiferromagnetic superexchange is lowered. ](F1.png){width="18.6cm"} \[lm1\] *Physical origin of the microscopic interactions:* As shown in Fig. \[lm1\](a), the strong Hund’s coupling and the crystal field ensure two things: first, that three electrons on the Mn site are well localized in the $t_{2g}$ manifold, forming a $S=3/2$ “core" spin that may be treated as a classical object, and second, that any electron occupying the $e_g$ level *must* have its spin parallel to the on-site $t_{2g}$ core spin. The electron delocalization happens through the $e_g$ manifold in the metallic state which is modeled by a two-band kinetic energy term with the on-site constraint that the itinerant spin orientation is projected on to the local “core" spin direction. This is the double exchange interaction. The spontaneous structural distortions arising from $e_g$ degeneracy-lifting when the $e_g$ manifold is singly occupied, as described in Fig. \[lm1\](b), is incorporated by coupling the $e_g$ electron’s orbital pseudospin operator to the Jahn-Teller modes $Q_2$ and $Q_3$. Finally, the Mn-Mn superexchange, due to virtual exchange of the localized oxygen $2p$ electrons with the Mn $e_g$ states, is shown in Fig. \[lm1\](c).\ ![ (a) Experimental temperature ($T$) - $r_A$ phase diagram for ordered, half-doped manganites, from D. Akahoshi *et al.* [@akahoshi-prl]. (b) Previous theoretical work, from K. Pradhan *et al.* [@pradhan-prl]: $T/t - \lambda/t$ phase diagram at $x=0.5$ for ordered half doped manganites. There is a phase transition between the large $\lambda/t$ or small bandwidth (BW) SCO-I and the small $\lambda/t$ or large BW, F-M on varying $r_A$. In the experiments, small $r_A$ implies small BW, which corresponds large $\lambda/t$ in theory. ](F2.png){width="12.4cm" height="7.6cm"} \[l0\] [*The Hamiltonian*]{} – Based on the preceding considerations, in this work we consider the microscopic two-band model for $e_g$ electrons with a strong Hund’s coupling to $t_{2g}$ core spins in a two-dimensional square lattice. The itinerant electrons are coupled to Jahn-Teller (JT) phonons and the core spins have a nearest-neighbor antiferromagnetic superexchange coupling between them. Explicitly, the Hamiltonian is given by $$\begin{aligned} H &=& \sum_{\langle ij \rangle \sigma,\alpha \gamma} -(t_{\alpha \gamma}^{ij} c^{\dagger}_{i \alpha \sigma} c^{~}_{j \gamma \sigma} +h.c.) - J_H\sum_i {\boldsymbol{S}}_i \cdot {\boldsymbol{\sigma}}_i + J_{AF}\sum_{\langle ij \rangle} {\boldsymbol{S}}_i \cdot {\boldsymbol{S}}_j \cr && +\lambda \sum_i \left( {\boldsymbol{Q}}_i \cdot {\boldsymbol{\tau}}_i - Q_{1i}\rho_i \right) + \sum_i \left(\frac{K}{2}{\boldsymbol{Q}}_i^2 + \beta Q_{1i}^2 \right) \cr && + \sum_{i,\alpha}\epsilon_{\alpha}n_{i,\alpha} - \mu N - h\sum_i S_{zi}\end{aligned}$$ where $c$ and $c^{\dagger}$ are the annihilation and creation operators for $e_g$ electrons and $\alpha$, $\beta$ index the two Mn $e_g$ orbitals ($d_{x^2-y^2}$ and $d_{3z^2-r^2}$, labeled $a$ and $b$ throughout). In the unstrained case, $t_{\alpha \beta}^{ij}$ are hopping amplitudes between nearest-neighbor sites with the symmetry-dictated form[@hotta-review]: $$\begin{aligned} &t_{a a}^x = t_{a a}^y \equiv t, \nonumber \\ &t_{b b}^x = t_{b b}^y \equiv t/3, \nonumber \\ &t_{a b}^x = t_{b a}^x \equiv -t/\sqrt{3}, \nonumber \\ &t_{a b}^y = t_{b a}^y \equiv t/\sqrt{3} \nonumber\end{aligned}$$ where $x$ and $y$ refer to the spatial orientation as in Fig. 1 in the main article. The $e_g$ electron spin operator is given by ${\sigma}^{\mu}_i=\sum_{\sigma \sigma'}^{\alpha} c^{\dagger}_{i\alpha \sigma} \Gamma^{\mu}_{\sigma\sigma'} c_{i\alpha \sigma'}$, where the $\Gamma$ are the Pauli matrices. This spin is coupled to the local (classical) $t_{2g}$ spin ${\boldsymbol{S}}_i$ via the Hund’s coupling $J_H$, and we assume $J_H/t \gg 1$. Finally, $\lambda$ is the coupling between the JT distortion ${\boldsymbol{Q}}_i = (Q_{2i}, Q_{3i})$ and the orbital pseudospin operator ${\tau}^{\mu}_i = \sum^{\alpha \beta}_{\sigma} c^{\dagger}_{i\alpha \sigma} \Gamma^{\mu}_{\alpha\beta} c_{i\beta \sigma}$. This same $\lambda$ also sets the scale for the coupling between the breathing mode $Q_1$ and the on site charge density. $K$ is the lattice stiffness. $\beta$, the breathing mode stiffness, is assumed to be sufficiently large that this mode can be neglected. Finally, $h$ is an external magnetic field, assumed to be in the $z$ direction and coupled only to ${\boldsymbol{S}}_i$. We set $t=1$, $K=1$, and treat the ${\boldsymbol{Q}}_i$ and the ${\boldsymbol{S}}_i$ as classical variables which is a a good approximation for the manganites[@hotta-review]. The chemical potential $\mu$ is adjusted so that the electron density remains at $n=1/2$. Parameter values ---------------- Fig. \[l0\] shows the comparison of the experimental, temperature ($T$)-$r_A$ phase diagram [@akahoshi-prl] and earlier theoretical work on the $T/t$ -electron-phonon coupling ($\lambda/t$) phase diagram employing the above model [@pradhan-prl]. Both the experiment and theory are for ordered, half-doped manganites. The term ‘ordered’ implies alternating layers of $A$=Ba and $A'$=any Lanthanide (Ln), marked on the top of Fig. \[l0\](a). Also the corresponding $r_A$ are shown on the lower axis. The value of $r_A$ controls the bandwidth (BW) in the experiments. The smaller the $r_A$, the smaller the BW. In the theoretical work, shown in Fig. \[l0\](b), the change of the BW is incorporated by the ratio $\lambda/t$, where large $\lambda/t$ implies a smaller BW and vice versa. This not only serves as a benchmark of the model and the method of solution, to be described shortly, it also allows us to fix the Hamiltonian parameters by comparing the theoretical and experimental $T_C$ values. For the electron-phonon and superexchange couplings, we choose the values of $\lambda$ and $J$ to roughly reproduce the thermal magnetic ordering scales seen in experiments. Since we measure all parameters in units of the bandwidth (BW) $t$, the ratios $\lambda/t$ and $J/t$ are what we fix. From Fig. \[l0\], we find that for $\lambda/t\sim1.5$ and $J/t=0.1$, $T_C\sim0.045t$. From this we obtain a $T_C$ of about 225K, which is at least in line with the experimental value near the SCO-I/F-M phase boundary. In doing this conversion we have assumed $t=0.3eV$. In general, there is a consensus that $\lambda/t\sim 1-1.6$ and $J/t\sim 0.05-0.1$ is adequate for reasonably quantitative comparison with experiments for most doping values [@hotta-review] and across many families of manganites. Scaling of hopping parameters and superexchange under strain ------------------------------------------------------------ \(i) Strain puts constraints on the lattice parameter $a$ which in turn changes both the Mn-O bond length $d$ and the Mn-O-Mn bond angle $\phi$ as seen in Fig. 1 (b) in the main text. The $p-d$ overlap integral $V_{pd\sigma}$ scales as $d^{-3.5}$, with $d$ being the center of mass distance between the Mn and O atoms [@harrison]. The Mn-O-Mn hopping is therefore proportional to $V^2_{pd\sigma}/\Delta$ and also depends on $\phi$ as $\mbox{cos}^n(\phi)$ (as shown below). Here, $\Delta$ is the energy denominator that depends on the difference between the Mn and the O states involved. The dependence of the hopping integral on $\phi$ is computed from the Slater-Koster tables[@slater-koster] assuming the Mn-O-Mn bond to lie in the x-y plane. The exponent $n$ is 3 for $t_{aa}$, 2 for $t_{bb}$ and 1 for $t_{ab}$, where $a=d_{x^2-y^2}$ and $b=3_{z^2-r^2}$. Thus the overall scaling for say, $t_{aa}$ would be $d^{-7}\mbox{cos}^3(\phi)$. Fig. 1 (b) in the main text, shows the relation between the lattice parameter ($a$), the Mn-O bond length ($d$) and the Mn-O-Mn bond angle ($\phi$). Let us assume the strain imposes a change in the lattice parameter by $\delta a$ and this causes a change in $d$ by $\delta d$ and $\phi$ by $\delta \phi$. Thus, $t_{aa} \rightarrow \tilde{t}_{aa} =(d+\delta d)^{-7}\mbox{cos}^3(\phi+\delta\phi)$. In bulk half doped manganites the value of $\phi$ varies from 160$^\circ$ to 170$^\circ$ in SCO-I, Y$_{0.5}$Ba$_{0.5}$MnO$_3$ to F-M, La$_{0.5}$Ba$_{0.5}$MnO$_3$[@bl-ba-1; @bl-ba-2]. For these values of $\phi$ and assuming small $\delta d/d$ and $\delta \phi$, we can linearize the expression for $t_{aa}$ in $\delta d/d$ and $\delta \phi$. This gives $\tilde{t}_{aa} \approx t_{aa}\left[1-7\frac{\delta d}{d} - 3\mbox{tan}(\phi)\delta \phi \right]$. The last term controls the extent to which the change in the *unstrained* Mn-O-Mn bond angle will effect the hybridization. For materials with $\phi\sim{160^\circ}$, the ratio $| 3 \mbox{tan}(\phi)\cdot d \frac{\delta \phi}{7\delta d} |$ is $|0.156 \cdot d\frac{\delta\phi}{\delta d}|$. To estimate this ratio we need typical data for perovskite structure under strain. Unfortunately, to the best of our knowledge, such data does not exist for the manganites. So as an estimate we assume a reasonable $|\delta d/d| \sim 0.01$ at 1$\%$ strain. Then it can be shown that for $\delta \phi<4^\circ$, the in-plane hopping increases (decreases) with $\delta d$ being negative (positive), while the bond angle dependence causes only quantitative change. Thus we neglect the $\delta \phi$ dependence in our calculation for small strain. In Fig. 1 (b) in the main text, $a=2d\sin(\phi/2)$. However in the spirit of above approximation we simply consider, $a=2d$, or, $2\delta d=\delta a$. So for the x-y plane, e$_{\parallel}$ = $\delta d/d$. Thus the hopping in the (x-y) plane scales with strain as $t_{\parallel}\sim t\left(1-7e_{\parallel}\right)$, while the out of plane hopping scales as $t_{\bot}\sim t\left(1-7e_{\bot}\right)$. \(ii) The scaling of the superexchange, $J$, with strain depends of the scaling of the *fourth* power of $V_{pd\sigma}$ under interatomic separation. Using considerations similar to above it is easy to show that, $J_{\parallel}\rightarrow \tilde{J_{\parallel}}\sim (J_{\parallel})(1-14e_{\parallel}))$ with strain. Similar scaling has been observed on application of pressure in La$_2$CuO$_4$[@sc-scaling]. The anisotropy in the hopping is also reflected in anisotropic antiferromagnetic scales $J_{\parallel}$ and $J_{\bot}$. ![The schematic of the SCO-I phase: (a) shows the planar checkerboard charge order (CO) with larger charge on the ’yellow’ sites and the CE type spin order (zig-zag ferromagnetic chains coupled antiferromagnetically). (b) shows the alternating $d_{x^2-r^2}/d_{y^2-r^2}$ orbital order (on the sites with larger charge density). (c) The color map of the magnetic structure factor, $S({\boldsymbol{q}})$ in the momentum space of our numerical data for the CE phase at low T. The ${\boldsymbol{q}}=(0, \pi),(\pi,0),$ and $(\pi/2,\pi/2)$ peaks are clearly seen. ](F3.png){width="14.6cm" height="6.2cm"} \[l4\] Method of solution ================== As mentioned above, we treat the core spins classically and the phonons in the adiabatic limit. Thus the solution of the Hamiltonian amounts to determining the configuration of the classical core spin and phonon background (for a fixed fermion density and temperature) that minimizes the free energy. To do this, we perform an exact diagonalization (ED) of the itinerant electron system for each configuration of the background spins and phonons. The Boltzmann weight of the combined electron, spin, and phonon system is calculated, and new spin/phonon configurations are sampled through classical Monte Carlo (MC). At a fixed temperature a Monte-Carlo system sweep consists of visiting every site of the system once in a sequential manner and performing the above mentioned update. The combined algorithm (ED+MC) is numerically rather costly, since the exact diagonalization must be performed at every step and the cost scales as $\mathcal{O}(N^3)$ with $N$ the number of lattice sites. Additionally, with a sequential system sweep, the cost of a Monte-Carlo system sweep scales as $N^4$ at each temperature. To reach reasonably large system sizes, we employ a recently developed variation [@tca] of real space ED+MC to that allows a linear scaling with the system size. This technique defines a region (cluster) around the site at which one attempts a MC update and accepts or rejects the update based on the energy change only in the cluster rather than diagonalizing the full system. For the calculation of observables, we diagonalize the full system, after the system has equilibrated using the above algorithm (known as the ‘traveling cluster approximation’ (TCA)). This adds only a few hundred full system diagonalizations to the computation cost. Within TCA, at each temperature, the computation cost of ED for a system with $N$ sites is $\mathcal{O}(N_{C}^{3})$, where $N_{C}$ is the fixed cluster size. Thus the cost of a full sweep of the lattice is $NN^3_c$ or [*linear*]{} in $N$ as opposed to $N^{4}$. Using this technique we have accessed sizes up to $24^{2}$ as opposed to the practical limit of $\sim 8^{2}$ within conventional ED+MC in 2 dimensions. In the present work on two dimensional systems, we employ a $8^2$ travelling cluster on system sizes between $16^2$ (for most data presented here) to $24^2$. In studies such as this, where real space phase separation may play a vital role in determining transport responses, it is crucial to have access to system sizes that are large enough to capture coexisting phases on the lattice. We perform 4000 Monte-Carlo system sweeps at every temperature. Of these the first 2000 are used to equilibrate the spin and phonon variables. Every 100$^{th}$ step from the remaining 2000 is used to calculate various quantities of interest. For calculating the response to an applied magnetic field, we repeat the above annealing process in the presence of the magnetic field. For example to calculate magnetoresistance at a given value of strain, we compare the resistivity calculated with and without a magnetic field. Characterization of ordered phases $\&$ transport ================================================= Structure factors ----------------- To characterize the phases as a function of strain, temperature, and magnetic field, we employ a number of static structure factors. The magnetic structure factor, defined by $S({\boldsymbol{q}})=\sum_{ij} \langle {\boldsymbol{S}}_i \cdot {\boldsymbol{S}}_j \rangle e^{i {\boldsymbol{q}} \cdot ({\boldsymbol{r}}_i-{\boldsymbol{r}}_j)}$, has a sharp peak at ${\boldsymbol{q}}=(0,0)$ for the ferromagnetic state and peaks at ${\boldsymbol{q}}=(0, \pi),(\pi, 0),$ and $(\pi/2,\pi/2)$ for the in-plane ‘CE’ type spin order. The long range charge order is characterized by the structure factor $D_Q({\boldsymbol{q}})=\sum_{ij} \langle {\boldsymbol{Q}}_i \cdot {\boldsymbol{Q}}_j \rangle e^{i {\boldsymbol{q}}.({\boldsymbol{r}}_i-{\boldsymbol{r}}_j)}$. The in-plane checkerboard charge order is indicated by a peak at ${\boldsymbol{q}}=(\pi,\pi)$. The schematic of the SCO-I phase in shown on Fig. \[l4\]. (a) shows the a-b plane checkerboard charge order and the antiferromagnetically coupled zig-zag ferromagnetic chains. (b) shows the schematic of the alternating orbital order. (c) shows the color map of the magnetic structure factor, $S({\boldsymbol{q}})$ in the momentum space of our numerical data for the SCO-I order at low T. The ${\boldsymbol{q}}=(0, \pi),(\pi,0),$ and $(\pi/2,\pi/2)$ peaks are clearly seen signifying the stabilization of the CE phase. ![The resistivity vs temperature (a) and the response of the resistivity to magnetic fields (b). The data is shown for a F-M which is close to a tensile strain driven F-M/SCO-I transition. In the inset in (a), the location of T$_{CO}$ (yellow) and T$_c$ (rest) for the different values of strain, as indicated in main panel, are shown in the same color by the double headed arrows. We find that the strain and field response of the dc resistivity is qualitatively similar to the results for a F-M close to the tensile strain induced F-M/FC-I transition in the main text. Also shown in (b) for comparison, is the unstrained resistivity vs temperature at zero field (dashed line).](F4.png){width="12.6cm" height="7.2cm"} \[l5\] Transport calculations ---------------------- The d.c conductivity $\sigma_{dc}$ is estimated by the Kubo-Greenwood expression [@mahan] for the optical conductivity. In a non-interacting system: $$\sigma(\omega)=\frac{\pi e^2}{N\hbar a} \sum_{\alpha,\beta} (n_{\alpha} - n_{\beta}) \frac{|f_{\alpha \beta}|^2}{ \epsilon_{\beta} - \epsilon_{\alpha} } \delta(\omega - (\epsilon_{\beta} - \epsilon_{\alpha}))$$ The $f_{\alpha\beta}$ are the matrix elements of the current operator, e.g., $\langle \psi_{\alpha} | j_x | \psi_{\beta} \rangle$, and the current operator itself (in the tight-binding model) is given by $j_x = i t a e \sum_{i, \sigma} (c^{\dagger}_{{i + a\hat{x}},\sigma} c_{i, \sigma} - h.c)$. The $\psi_{\alpha}$ are single-particle eigenstates, and $\epsilon_{\alpha}$ are the corresponding eigenvalues. The $n_{\alpha}=f(\mu - \epsilon_{\alpha})$ are Fermi factors. $a$ in the lattice spacing. We can compute the low-frequency average, $\sigma_{av}(\mu, \Delta \omega, N) = (\Delta \omega)^{-1} \int_0^{\Delta \omega} \sigma(\mu, \omega, N) d\omega$, using periodic boundary condition in all directions. The averaging interval is reduced with increasing $N$, with $\Delta \omega \sim B/N$. The d.c. conductivity is finally obtained as $\sigma_{dc}(\mu) = {\lim}_{L \rightarrow \infty} { \sigma}_{av}(\mu, B/L, L)$. The chemical potential is set to target the required electron density $n$. This approach to d.c. transport calculations has been benchmarked in a previous work[@sk-pm-long-transp]. Tensile strain induced metal to insulator transition across the F-M/SCO-I boundary ================================================================================== Here we show that qualitatively our conclusions are independent of the choice of the metal-insulator boundary near which we calculate magnetotransport. For this we present the magnetotransport for a F-M with ($\lambda/t=1.5, J/t=0.1$) which is close to tensile strain driven F-M/SCO-I boundary. From Fig. \[l5\](a) we can easily deduce that the enhancement in $\%$ MR with tensile strain is qualitatively similar to that seen in Fig. 3(a) in the main text. Also shown in Fig. \[l5\](b) is the magnetic field induced colossal suppression of resistivity for 2.2$\%$ tensile strain. Fig. 3(c) in the main text, are constucted from similar data.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We consider the complex Monge-Ampère equation on complete Kähler manifolds with cusp singularity along a divisor when the right hand side $F$ has rather weak regularity. We proved that when the right hand side $F$ is in some weighted $W^{1,p_0}$ space for $p_0>2n$, the Monge-Ampére equation has a classical $W^{3,p_0}$ solution.' author: - Fangyu Zou title: 'The complex Monge-Ampère equation on the complement of a divisor' --- Introduction ============ Let $(\bar{M},\omega_0)$ be a compact Kähler manifold of complex dimension $n$, in which we consider a divisor $D$ with only *simple normal crossings* with decomposition $D=\sum_{j=1}^N D_j$ into smooth irreducible components. Let $[D_j]$ be the associated line bundle to each $D_j$, endowed with a smooth hermitian metric $|\cdot|_j$, and $\sigma_j\in \mathcal{O}([D_j])$ be a holomorphic defining section such that $D_j=\{\sigma_j=0\}$ for each $j$. Let $\rho_j=-\log(|\sigma_j|^2_j)$. Up to multiplying $|\cdot|_j$ by a positive constant or a smooth positive function, we can assume that $|\sigma_j|_j^2\leq e^{-1}$ so that $\rho_j\geq 1$ out of $D_j$. Note that ${\sqrt{-1}\partial\bar{\partial}}\rho_j$ extends to a *smooth* real $(1,1)$-form on the whole $\bar{M}$, which lies in the class $2\pi c_1([D_j])$. Let $\rho=\prod_{j=1}^N \rho_j$. Let $\lambda>0$ be a real parameter sufficiently large. Set $$\omega = \lambda \omega_0 + {\sqrt{-1}\partial\bar{\partial}}\log{\rho} = \lambda \omega_0 + \sum_{j=1}^N {\sqrt{-1}\partial\bar{\partial}}\log(-\log|\sigma_j|_j^2),$$ then $\omega$ defines a genuine Kähler form on $M=\bar{M}-D$. It has properties that it is complete, with finite volume, with cusp singularity along $D$ and has injectivity radius going to zero. The purpose of this paper is to study the Monge-Ampére equation $$\label{equ14} (\omega + {\sqrt{-1}\partial\bar{\partial}}{\phi})^n = e^F\omega^n, \,\ \int_M (e^F-1) dV = 0$$ when the right hand side $F$ has rather weak regularity. We show that when $F$ is in some *weighted* $W^{1,p_0}$ space for $p_0> 2n$, more precisely, when $$I(F,p_0):=\int_M (|F|^{p_0}+|\nabla F|^{p_0})\rho^{\frac{p_0-2}{2n-2}}dV < \infty,$$ the equation (\[equ14\]) has a classical solution $\phi$ in $W^{3,p_0}$ (not in the weighted sense). In particular, we proved \[thm1\] Let $\bar{M}$ be a compact Kähler manifold of complex dimension $n$ and $D$ be an divisor on $\bar{M}$ with only simple normal crossings. Let $M=\bar{M}-D$ and $\omega$ be the reference Kähler metric on $M$ constructed above. For any function $f$ satisfying $I(F,p_0)< \infty$ for some $p_0> 2n$, the Monge-Amperé equation (\[equ14\]) has a classical solution $\phi$ in $W^{3,p_0}(M)$. This regularity problem in the compact Kähler manifold setting has been proved by X. Chen and W. He in [@Chen-He]. They proved that if $(M,\omega)$ is a compact Kähler manifold, then the equation (\[equ14\]) has a classical solution in $W^{3,p_0}$ once the right hand side $F$ is in $W^{1,p_0}$ for some $p_0>2n$. In Yau’s seminal resolution of the Calabi conjecture [@Yau1], the maximum principle is used in a significant way in the $C^2$ estimate of the potential function, where the proof depends heavily on the $C^0$ norm of $\Delta F$. Since we have weaker regularity on $F$, the main issue is that we can not apply maximum principle to get $C^2$ estimate. One main innovation of [@Chen-He] is that they derive the gradient estimate and Laplacian estimate by integration methods. The main tool is the Moser’s iteration technique (see [@Moser]). Our result can be viewed as the non-compact version of Chen-He’s theorem. We will follow their idea of seeking an integration method for the gradient and Laplacian estimates. Where our case differs from the compact case is two folds: first, we need consider the boundary term when we do integration by parts; second, the usual Sobolev inequality fails in our context. To overcome the issue, we consider the following the $\epsilon$-perturbed equation of (\[equ14\]): $$\label{equ1} (\omega+{\sqrt{-1}\partial\bar{\partial}}{\phi_\epsilon})^n= e^{F+\epsilon\phi_\epsilon}\omega^n,\,\, \int_M (e^F-1)dV = 0$$ for $\epsilon\in (0,1]$. Equation (\[equ1\]) has been well studied by Cheng-Yau [@Cheng-Yau], R. Kobayashi [@Kobayashi] and Tian-Yau [@Tian-Yau] to derive Kähler-Einstein metrics with negative curvature on $(M,\omega)$ assuming $K_{\bar{M}}+D$ ample. However, the existence of solution $\phi_\epsilon$ of (\[equ1\]) is proved by the continuity method in quasi-coordinates without additional assumption of the ampleness of $K_{\bar{M}}+D$. The idea here is that assuming $F$ is smooth and compactly supported, by maximum principle we can first derive the a priori estimates of $\phi_\epsilon$ depending on the parameter $\epsilon$ and norms of higher derivatives of $F$. This garantees the integrand of boundary terms in the integration by parts are at least in $L^1$. Then we apply a theorem of Gaffney-Stokes [@Gaffney-Stokes] to show the boundary terms vanish. Hence, we can do integration by parts as in the compact case. To emphasize, this is just a technical step to make the integration by parts work through. The final estimate will be uniform in $\epsilon$ and only depends on $I(F,q_0)$. On the other hand, to deal with the problem of lack of Sobolev inequlity, we borrow the weighted Sobolev inequality developed in [@AUV] instead. There is still ssome serious issue need to deal with as a result of that the measure (the volume form multiplied by the weight) is not finite. A key obervation is that one higher order term in the Chen-He’s inequalites dominates the constants and make the Moser’s iteration possible in our case, while in the compact case this high order term can be directly dropped off. The main idea of this paper is to derive a uniform $W^{3,p_0}$ estimate of the solutions $\phi_\epsilon$ of the $\epsilon$-perturbed equation, following by taking a converging subsequence. We follow Chen-He’s routine to first derive uniform $C^1$ and $C^2$ estimates. In this process, we can assume that $F\in C^\infty_c(M)$ since we can approximate $F$ by $C^\infty_c(M)$ funtions in the weighted Sobolev spaces when $I(F,p_0)< \infty$ (see Lemma \[app\_lem\] in section 6). In particular, we have \[thm2\] Suppose $F\in C^\infty_c(M)$ satisfies $I(F,p_0)<\infty$ for some $p_0>2n$. If $\phi_\epsilon$ is a solution to the perturbed equation (\[equ1\]), then there exists a constant $C=C(I(F,p_0),p_0,m,\omega)$ such that for all $\epsilon\in (0,1]$, $$|\nabla \phi_\epsilon|\leq C.$$ \[thm3\] Suppose $F\in C^\infty_c(M)$ satisfies the conditions in Theorem \[thm1\]. If $\phi_\epsilon$ is a solution to the perturbed (\[equ1\]), then there exists a constant $C=C(I(F,p_0),p_0,m,\omega)$ such that for all $\epsilon\in (0,1]$, $$|\Delta \phi_\epsilon|\leq C.$$ The arrangement of this paper is the following: In section 2 we set up the ingredients for the proof of main theorems; in section 3 we cite a proof of Auvray on uniform $C^0$ estimate with slight modification; section 4 and 5 are devoted to the proof of Theorem \[thm2\] and Theorem \[thm3\]; in section 4 we first derive a $C^{2,\alpha}$ estimate in quasi-coordiates and finally the $W^{3,p_0}$ estimate. The main result Theorem \[thm1\] is proved in the end of the paper. **Convention of notations**: With a little abuse of notation, we will use the Kähler form $\omega$ to denote the reference metric, while in some cases it is also denoted as $g$. Throughout this paper, $dV,\, \nabla$ and $\Delta$ denote the volume form, the Levi-Civita connection and the Laplacian operator of the reference metric $\omega$, respectively; $dV',\, \nabla'$ and $\Delta'$ will be those of the metric $\omega_{\phi_\epsilon}=\omega+{\sqrt{-1}\partial\bar{\partial}}{\phi_\epsilon}$. For simplicity of notations, we will also drop the subscript $\epsilon$ from $\phi_\epsilon$ when this is no ambiguity. The constant “$C$" without subscript may vary from line to line, while if there is subscript, then it is some fixed constant. Constants in this paper will only depend on $(I(F,p_0),p_0,n,\omega)$ unless specifically pointed out. **Acknowledgement:** [The author is very grateful to his advisor Prof. Xiuxiong Chen for introducing this problem and for his consistent support during the proof. The author is also thankful to Yuanqi Wang, Gao Chen, Ruijie Yang and Santai Qu for useful discussions.]{} Preliminaries ============= In this section we set up the ingredients for the proof of the main theorems. The reference metric -------------------- Let us quickly recall the construction of the reference metric $\omega$. Let $(\bar{M},\omega_0)$ be a compact Kähler manifold of complex dimension $n$, in which we consider a divisor $D$ with simple normal crossings with decomposition $D=\sum_{j=1}^N D_j$ into smooth irreducible components. For each $j$, let $\sigma_j$ be a holomorphic defining section of $D_j$. We can assume that $\rho_j:=-\log(|\sigma_j|^2)\geq 1$ out of $D_j$. Note that ${\sqrt{-1}\partial\bar{\partial}}{\rho_j}$ extends to a smooth real $(1,1)$-form on the whole $\bar{M}$, whose class is $2\pi c_1([D_j])$. Let $\rho=\prod_{j=1}^N \rho_j$. For some postive real parameter $\lambda>0$, set $$\omega = \lambda \omega_0 - {\sqrt{-1}\partial\bar{\partial}}\log{\rho} = \lambda \omega_0 - \sum_{j=1}^N {\sqrt{-1}\partial\bar{\partial}}\log(-\log|\sigma_j|^2).$$ For $\lambda>0$ sufficiently large, $\omega$ defines a Kähler metric on $M=\bar{M}-D$. This follows from a simple computation. Note that $$-{\sqrt{-1}\partial\bar{\partial}}{\log\rho_j} = \frac{\sqrt{-1}\partial \rho_j \wedge \bar{\partial} \rho_j}{\rho_j^2} - \frac{ {\sqrt{-1}\partial\bar{\partial}}\rho_j}{\rho_j}.$$ The first summand is a postive $(1,1)$-form. For each $j$, there is some positive $\lambda_j>0$ such that ${\sqrt{-1}\partial\bar{\partial}}{\rho_j} \leq \lambda_j \omega_0$ on $\bar{M}$. Hence, $\lambda_j\omega_0+{\sqrt{-1}\partial\bar{\partial}}{\log \rho_j}>0$ on $\bar{M}-D_j$. Let $\lambda = \sum_j \lambda_j$, then $$\omega = \lambda \omega_0 -{\sqrt{-1}\partial\bar{\partial}}{\log\rho}=\sum_j (\lambda_j \omega_0 - {\sqrt{-1}\partial\bar{\partial}}{\log \rho_j}) > 0$$ on $M=\bar{M}-D$. A simple model for this type of metric is the punctured disc $\Delta^*=\Delta-\{0\}$ with the Poincaré metric $\omega = -{\sqrt{-1}\partial\bar{\partial}}{\log(-\log|z|^2)} = \frac{\sqrt{-1}dz\wedge d\bar{z}}{|z|^2\log^2|z|^2}$. Indeed, the asymptotics of the reference metric near $D$ can be compared with this kind of model metric. Soppose $P$ is a point in a crossing of codimension $k$, say, $D_1\cap\cdots\cap D_k$. Take an open neighborhood $U$ of $P$ which is biholomorphic to coordinate chart $(\Delta^n; z_1,\ldots,z_n)$. The simple normal crossing assumption allows us to write $D\cap U = (D_1\cup\cdots\cup D_k)\cap U=\cup_{i=1}^k\{z_i=0\}$, where $z_j=0$ being the equation of $D_j$ in $U$. Then $U\backslash D=(\Delta^*)^k\times \Delta^{n-k}$. Let $\omega_{mdl}$ be the model metric on $(\Delta^*)^{k}\times \Delta^{n-k}$: $$\omega_{mdl}= \sum_{j=1}^k \frac{\sqrt{-1}dz_j\wedge d\bar{z}_j}{|z_j|\log^2|z_j|^2} + \sum_{j=k+1}^n \sqrt{-1}dz_j \wedge d\bar{z}_j.$$ \[lemma2\] In the coordinates $(z_1,\ldots,z_k,z_{k+1},\ldots,z_n)$, we have $$\label{8} \omega = \sum_{j=1}^k \frac{\sqrt{-1}dz_j \wedge d\bar{z}_j}{|z_j|^2\log^2|z_j|^2} + \left(\lambda\omega_0 - \sum_{j=k+1}^N {\sqrt{-1}\partial\bar{\partial}}{\log\rho_j} \right) + O(\rho_1^{-1}+ \cdots + \rho_k^{-1})$$ In particular, $\omega$ is quasi-isometric to $\omega_{mdl}$ on $U\backslash D$, i.e., there exists some positive constant $C=C(U,M,\omega)$ such that $$C^{-1}\omega_{mdl} \leq \omega \leq C\omega_{mdl}.$$ Note that $\lambda \omega_0 - \sum_{j=k+1}^N {\sqrt{-1}\partial\bar{\partial}}{\log \rho_j}$ is smooth on $D$. For $1 \leq j \leq k$, $|\sigma_j|^2 = e^{f_j}|z_j|^2$ for some smooth $f_j$ through $D$. Thus, $\rho_j = -\log |z_j|^2 - f \sim -\log |z_j|^2$. A simple computation shows that $$\begin{aligned} -{\sqrt{-1}\partial\bar{\partial}}{\log \rho_j} = &\ \frac{\sqrt{-1}dz_j\wedge d\bar{z}_j}{|z_j|^2\rho_j^2} -\frac{{\sqrt{-1}\partial\bar{\partial}}{f}}{\rho_j} \\ &\ + \frac{\sqrt{-1}(z_jdz_j\wedge \bar{\partial f} + \bar{z}_j \partial f\wedge d\bar{z}_j + |z_j|^2\partial f\wedge \bar{\partial}f)}{|z_j|^2\rho_j^2} \\ = &\ \frac{\sqrt{-1}dz_j\wedge d\bar{z}_j}{|z_j|^2\log^2|z_j|^2} + O(\rho_j^{-1}). \end{aligned}$$ Sum up for $j=1, \ldots, k$, we obtain (\[8\]). The second part follows easily from (\[8\]). In light of Lemma \[lemma2\], we readily see properties of such metric of Poincaré type: it is complete, has finite volume, and its injectivity radius goes to 0. In the next lemma, we collect some facts about the reference metric $\omega$ which we will use later. Let $(M,\omega)$ be constructed above. There exists some positive constant $B$ such that - $\inf_{i\neq j} R_{i\bar{i}j\bar{j}}\geq -B$; - $|R|\leq B$; - $\sup \frac{|\nabla \rho|}{\rho} \leq B.$ where $R_{i\bar{i}j\bar{j}}$ and $R$ are the holomorphic sectional curvature and scalar curvature of $(M,\omega)$, respectively. We only need to consider near the divisor $D$. Suppose $U$ is an open neighborhood of some point on $D$ such that $U\cap D = U\cap (D_1\cup\cdots\cup D_k)=\cup_{j=1}^k \{z_j=0\}$, where $z_j=0$ is the equation of $D_j$ in $U$. The metric $\omega$ has the asymptotics in $U\backslash D$ as stated in (\[8\]). Note that $(\Delta^*)^k$ with standard cusp metric $\sum_{j=1}^k\frac{\sqrt{-1}dz_j\wedge d\bar{z}_j}{|z_j|^2\log^2|z_j|^2}$ has holomorphic sectional curvature $-1$. Hence, the holomorphic sectional curvature of $\omega$ on $U\backslash D$ is bounded from below and the scalar curvature bounded by some constant on $U\backslash D$. To see (3), let us assume $U\cap D= U\cap D_1 = \{z_1=0\}$ and $\omega$ is the local model metric $\frac{\sqrt{-1}dz_1\wedge d\bar{z_1}}{|z_1|^2\log^2|z_1|^2} + \sum_{j=2}^n \sqrt{-1}dz_j\wedge d\bar{z}_j$ for simplicity. Set $\rho=\rho_1\rho'$ where $\rho'=\rho_2\cdots \rho_N$ is smooth on $D$. We have $\frac{\nabla \rho}{\rho}=\frac{\nabla \rho_1}{\rho_1}+\frac{\nabla \rho'}{\rho'}$. Note that $\rho'$ is bounded and $|\nabla \rho'|^2= (|z_1|^2\log^2|z_1|^2) |\partial_{z_1}\rho'|^2+\sum_{j=2}^n |\partial_{z_j} \rho'|^2 \leq \sum_{j=1}^n |\partial_{z_j} \rho'|^2$ is bounded on $U\backslash D$. On the other hand, $\rho_1 = -\log|z_1|^2 -f \sim -\log|z_1|^2$, we have $|\nabla \rho_1|^2 \leq -|z_1|^2\log^2|z_1|^2 (|z_1|^{-2}+ |\partial_{z_1} f|^2) + O(1) = \rho_1^2 + o(\rho_1)$. Hence, $\frac{|\nabla \rho_1|}{\rho_1} \rightarrow 1$ as $z_1\rightarrow 0$. Therefore, we have $\frac{|\nabla \rho|}{\rho}$ bounded by some constant on $U\backslash D$. Finally, we cover the divisor $D$ by finitely many such $U$’s and take the maximum of those constants as $B$, thus obtain (1), (2) and (3). Quasi-coordinates ----------------- The viewpoint of quasi-coordinates has been adopted to study complete Kähler manifolds by Tian-Yau [@Tian-Yau] and R. Kobayashi[@Kobayashi]. They are useful to define likewise Hölder space, while the usual coordinate charts has inconvenience because of the injectivity radius going to zero. In these quasi-coordinates the interior Schauder estimate on complete manifold can be reduced to that on a bounded domain in Euclidean space. In this paper, we will use the quasi-coordinates to establish the weighted Sobolev inequality by reducing them to the bounded domains in Euclidean space. Let $V$ be an open set in $\mathbb{C}^n$. A holomorphic map $\Phi$ from $V$ into a complex manifold $M$ of dimension $n$ is called a *quasi-coordinate map* iff it is of maximal rank everywhere in $V$. The pair $(V; \mbox{Euclidean coordinates of } \mathbb{C}^n)$ is called a local quasi-coordinate of $M$. To construct quasi-coordinates for $(M,\omega)$, we begin with the punctured disc $\Delta^*=\Delta-\{0\}$ with the standard cusp metric $\omega_{cusp}= \frac{\sqrt{-1}dz\wedge d\bar{z}}{|z|^2\log^2|z|^2}$. We start from the universal covering map $\pi:\Delta\rightarrow \Delta^*$, given by $\pi(w)=\exp\left(\frac{w+1}{w-1}\right)$. Formally, it sends 1 to 0. The idea is to restrict $\pi$ to the fixed disc $\frac{3}{4}\Delta$ (disc with radius $3/4$), and compose it with a biholomorphism $\psi_\delta$ of $\Delta$ sending $0$ to $\delta$, where $\delta\in (0,1)$ is a real parameter. To write the formula, we set $\psi_\delta(w)=\frac{w+\delta}{1+\delta w}$, so that the quasi-coordinate maps are given by $$\begin{gathered} \label{equ19} \varphi_\delta = \pi\circ\psi_\delta : \frac{3}{4}\Delta \rightarrow \Delta^*,\, \, \varphi_\delta(w) = \exp\left(-\frac{1+\delta}{1-\delta}\frac{w+1}{w-1}\right)\end{gathered}$$ It is easy to check the following properties of the quasi-coordinate maps. - ${\varphi_\delta}^*\omega_{cusp} = \frac{\sqrt{-1}dw\wedge d\bar{w}}{(1-|w|^2)^2}$ is independent on $\delta$ and $C^\infty$-quasi-isometric to the Euclidean metric on the ball. - ${\varphi_\delta}^*(-\log(|z|^2))=2\frac{1+\delta}{1-\delta}\frac{1-|w|^2}{|1-w|^2}$ is mutually bounded with $\frac{1}{1-\delta}$ with fixed factor, i.e., there is constant $C>0$ independent of $\delta$ such that $\frac{1}{C(1-\delta)} \leq {\varphi_\delta}^*(-\log|z|^2) \leq \frac{C}{1-\delta}$ on $\frac{3}{4}\Delta$. - There exists a constant $\kappa>0$ small (indeed $\kappa\leq e^{-25/7}$) such that $\kappa\Delta^*\subset \cup_{\delta\in (0,1)} \varphi_\delta (\frac{3}{4}\Delta)$. Now let us consider a crossing $D_1\cap \cdots \cap D_k$ of codimension $k$. For any point on such a crossing, we take an open neighborhoond $U$ such that $U\cap D_j=\{z_j=0\},\ j=1,\ldots,k$. Under the coordinates of $(z_1,\ldots,z_n)$, $U$ can be taken so that $U\backslash D$ is biholomorhic to $(\kappa \Delta^*)^k\times \Delta^{n-k}$. Let $\frac{3}{4}\mathcal{P}_k$ denote the polydisc $(\frac{3}{4}\Delta)^k\times \Delta^{n-k}$ in $\mathbb{C}^n$. Let $\delta=(\delta_1,\ldots,\delta_k)\in (0,1)^k$ be a multi-index. Then the quasi-coordinate map is given by $$\begin{gathered} \Phi_\delta: {\frac{3}{4}\mathcal{P}_k}\rightarrow (\kappa\Delta^*)^k\times \Delta^{n-k}\\ (w_1,\ldots,w_k,z_{k+1},\ldots,z_m)\mapsto \left(\varphi_{\delta_1}(w_1),\ldots,\varphi_{\delta_k}(w_k),z_{k+1},\ldots,z_m \right)\end{gathered}$$ Note that $\cup_\delta \Phi_\delta({\frac{3}{4}\mathcal{P}_k})$ covers $U\backslash D$. The quasi-coordinate “atlas" of $(M,\omega)$ is obtained by covering an open neighborhood of $D$ by our local quasi-coordinate charts, and covering the complement by a finite number of unit balls in $\mathbb{C}^n$. At this stage we introduce the Hölder norms and Hölder spaces using the previously introduced quasi-coordinates for later use. For a non-negative integer $k$, and a real number $\alpha \in (0,1)$, we define $$||u||_{C^{k,\alpha}_{qc}(M)} := \sup \{ ||u\circ \Phi||_{C^{k,\alpha}(V)}: (V,\Phi) \text{ \emph{is a quasi-coordinate map}} \}$$ where the supremum is taken over all our quasi-coordinate maps $(V,\Phi)$. The $||\dot||_{C^{k,\alpha}(V)}$ is the usual Hölder norm on $V\subset \mathbb{C}^n$. The Hölder space $C^{k,\alpha}_{qc}(M)$ is $$C^{k,\alpha}_{qc}(M) := \{ u\in C^k_{loc}(M): ||u||_{C^{k,\alpha}_{qc}(M)}< \infty\}.$$ Weighted Sobolev inequality --------------------------- The weighted Sobolev inequality stated in this subsection is first proved by Auvray in [@AUV], Lemma 4.4. The following lemma, which is crucial for Auvray’s proof, however, is just stated without a proof. For readers’ convenience and completeness, we give a proof here. We briefly set up the setting in this section. Consider a point on a crossing of codimension $k$. Let $U$ be a polydisc centered at this point so that $U\backslash D$ is biholomorphic to $(\kappa\Delta^*)^k\times \Delta^{n-k}$. For multi-indices $\delta\in (0,1)^k$, let $\Phi_\delta: \frac{3}{4}\mathcal{P}_k \rightarrow U\backslash D$ be the quasi-coordinate maps defined in subsection 2.2. Let $\frac{1}{2}\mathcal{P}_k$ denote the polydisc $(\frac{1}{2}\Delta)^k\times \Delta^{n-k}$. \[lem3\] There exists a constant $c>0$ depending on $(U,n,\omega)$ and a sequence of multi-indices $(\delta_\ell)$, $\delta_\ell=((\delta_1)_{\ell_1},\ldots,(\delta_k)_{\ell_k})$ such that for any $f \in L^1(U\backslash D)$, $$\begin{gathered} c^{-1}\sum_{\ell} A_{\delta_\ell}\int_{\frac{3}{4}\mathcal{P}_k} |{\Phi_{\delta_\ell}}^* f| dV_{g_0} \leq \int_{U\backslash D} |f| dV \leq c \sum_{\ell} A_{\delta_\ell} \int_{\frac{1}{2}\mathcal{P}_k} |{\Phi_{\delta_\ell}}^* f| d{V}_{g_0}\end{gathered}$$ where $A_{\delta_\ell} = \prod_{j=1}^k \left(1-(\delta_j)_{\ell_j}\right)$, and $g_0$ is the Euclidean metric on $\mathbb{C}^n$. Note that ${\Phi_{\delta_\ell}}^*\rho$ is mutually bounded with $A_{\delta_\ell}^{-1}$, and the pull-back metric ${\Phi_\delta}^*g$ is $C^\infty$-quasi-isometric to the Euclidean metric $g_0$ with fixed factor independent of $\delta$. In particular, if $(f\rho) \in L^1(U\backslash D)$, then $$c^{-1} \sum_{\ell} \int_{{\frac{3}{4}\mathcal{P}_k}} |{\Phi_{\delta_\ell}}^* f| dV_{g_0} \leq \int_{U\backslash D} |f|\rho dV \leq c \sum_{\ell} \int_{\frac{1}{2}\mathcal{P}_k} |{\Phi_{\delta_\ell}}^* f| d{V}_{g_0}$$ The proof is quite technical. We shall begin with the simplest case to show the key point of the proof. Let us consider the model case of Poincaré disc with standard Poincaré metric. In this case, $\delta\in (0,1)$ is a real parameter. Let $\sigma =\frac{1+\delta}{1-\delta}$. The quasi-coordinate map $\Phi_{\delta}$ is just $\varphi_\delta(w) = \exp\left(-\frac{1+\delta}{1-\delta}\frac{w+1}{w-1}\right)=\exp \circ \zeta_\sigma(w)$, where $\zeta_\sigma(w) = -\sigma \frac{w+1}{w-1}$. The covering map $\exp:\mathbb{C}\rightarrow \Delta^*$ maps any horizontal strip of width $2\pi$ onto the punctured disc. Under the map $\zeta_\sigma$, the disc $\frac{1}{2}\Delta$ is mapped biholomorphically to another disc $B_\sigma$ with center $(-\frac{5}{3}\sigma,0)$ and radius $\frac{4}{3}\sigma$. The union of such balls covers the half strip $$(-\infty, \log \kappa)\times (-\pi,\pi]\subseteq \mathbb{R}^2.$$ The idea is to pull back and take the integral over the strip, then cut the strip into small pieces so that each piece is contained in some ball. Each piece is covered by the ball a finite number of multiplicity proportional to the radius of the ball, hence proportional to $\sigma$. To make it precise, let $\hat{g}$ be the pullback metric on $\mathbb{C}$ of the standard cusp metric on the punctured disc by the covering map $\exp$. Let $\hat{f} = f\circ\exp$. Then $$\int_{\kappa\Delta^*} |f| d{V} = \int_{(-\infty,\log \kappa)\times (-\pi,\pi]} |\hat{f}|d{V}_{\hat{g}}$$ The rectangle $$I_\sigma \times J_\sigma := \left(-\frac{20}{9}\sigma,-\frac{10}{9}\sigma \right] \times \left(-\frac{\sqrt{119}}{9}\sigma,\frac{\sqrt{119}}{9}\sigma\right)$$ is contained in the ball. We have $$\int_{I_\sigma\times (-\pi,\pi]}|\hat{f}|dV_{\hat{g}} \leq \frac{18\pi}{\sqrt{119}\sigma}\int_{I_\sigma\times J_\sigma} |\hat{f}|d{V}_{\hat{g}} \leq c\sigma^{-1}\int_{\frac{1}{2}\Delta} |{\varphi_\delta}^*f|dV_{g_0}$$ Now pick a sequence $(\sigma_\ell)$ such that $\sigma_1 = -\frac{3}{5}\log\kappa,\, \sigma_{\ell+1}=2\sigma_\ell$. Then $(-\infty, \log \kappa)\subseteq \cup I_{\sigma_\ell}$. Hence, $$\int_{(-\infty,\log \kappa)\times (-\pi,\pi]} |\hat{f}|dV_{\hat{g}}\leq \sum_{\ell=1}^\infty \int_{I_{\sigma_\ell \times (-\pi,\pi]}}|\hat{f}|dV_{\hat{g}}.$$ Namely, $$\int_{\kappa \Delta^*} |f|dV \leq c\sum_\ell \sigma_\ell^{-1} \int_{\frac{1}{2}\Delta}|{\varphi_{\delta_\ell}}^*f|dV_{g_0}.$$ Note that $\sigma^{-1}$ is equivalent to $(1-\delta)$. Hence, $$\label{equ20} \int_{\kappa \Delta^*} |f| dV \leq c'\sum_\ell \int_{\frac{1}{2}\Delta} (1-\delta_\ell)|{\varphi_{\delta_\ell}}^*f|dV_{g_0}.$$ This gives us the right half inequality. On the other hand, $\frac{3}{4}\Delta$ is mapped by $\phi_{\sigma}$ to a ball with center $(-\frac{25}{7}\sigma,0)$ and radius $\frac{24}{7}\sigma$. Let $$I'_\sigma = \left(-7\sigma,-\frac{1}{7}\sigma\right],\quad J'_\sigma = \left(-\frac{24}{7}\sigma,\frac{24}{7}\sigma\right).$$ Then $$\label{equ21} \int_{\frac{3}{4}\Delta} |{\phi_\sigma}^*f| dV_{g_0} \leq \int_{I'_\sigma\times J'_\sigma} |\hat{f}|dV_{\hat{g}}\leq \frac{48\sigma}{7\pi}\int_{I'_\sigma\times (-\pi,\pi]} |\hat{f}|dV_{\hat{g}}$$ For the same sequence $(\sigma_\ell)$ in (\[equ20\]), these $I'_{\sigma_\ell}$ are overlapped. But each of them only intesect with finitly many others. To see this, two balls do not intersect if the distance of the centers is bigger than the sum of their radius. The balls $B_{\sigma_{\ell_1}}$ and $B_{\sigma_{\ell_2}}$ do not intersect (suppose $\ell_1>\ell_2$) if $$\begin{gathered} \frac{25}{7}(2^{\ell_1}-2^{\ell_2})> \frac{24}{7}(2^{\ell_1}+2^{\ell_2})\quad \Rightarrow \quad \ell_1-\ell_2>\log_249\end{gathered}$$ Hence, each ball intersects with no more than $2\log_2 49 <16$ balls. Therefore, $$\sum_\ell \int_{I'_{\sigma_\ell}\times (-\pi,\pi]} |\hat{f}|dV_{\hat{g}} \leq 16\int_{(-\infty,\log\kappa)\times (-\pi,\pi]} |\hat{f}|dV_{\hat{g}} =16 \int_{\kappa\Delta^*} |f|dV$$ Combine with (\[equ21\]), and that $\sigma^{-1}$ is equivalent to $(1-\delta)$, we have $$\sum_\ell (1-\delta_\ell) \int_{\frac{3}{4}\Delta}|{\varphi_{\sigma_\ell}}^* f|dV_{g_0} \leq c'' \int_{\kappa \Delta^*} |f|dV$$ We get the left half inequality. Now come back to our setting. When $k=1$, then $U\backslash D = \kappa \Delta^*\times \Delta^{n-1}$. We can assume the metric $g$ on $U\backslash D$ is the standard model metric $\frac{\sqrt{-1}dz_1\wedge d\bar{z}_1}{|z_1|^2\log^2|z_1|^2}+\sum_{j=2}^n \sqrt{-1}dz_j\wedge d\bar{z}_j$. The argument goes exactly the same after multiplying each integral region by $\Delta^{n-1}$; if $k\geq 2$, we write the integral as multiple integral and treat the variables in this manner in order. Now we prove the following weighted Sobolev inequality on $(M,\omega)$. \[sob\_inequ\] There exists a positive constant $C= C(p,n,\omega)$ such that for any function $v\in W^{1,p}_{loc}$, $$\left(\int_M |v|^{q} \rho dV\right)^{1/q} \leq C \left(\int_M \left(|v|^{p}+|\nabla v|^{p}\right) \rho dV\right)^{1/p}$$ as long as $q\geq p$ and $\frac{1}{p}\leq \frac{1}{2n}+\frac{1}{q}$. Away from the divisor, $\rho$ is bounded, then it is just the usual Sobolev inequality in open domains of $\mathbb{C}^n$. Hence, we only need to look at what happens near the divisor. Suppose a point on a crossing of codimension $k$. Consider a small polydisc $U$ around this point and cover $U\backslash D$ by a union $\cup_{\delta} \Phi_{\delta}(\frac{3}{4}\mathcal{P}_k)$ where $\delta=(\delta_1,\ldots,\delta_k)$. We can assume that the metric on $U\backslash D$ is just the standard model metric. Then all the pullback ${\Phi_\delta}^* g$ give the same metric on $\frac{3}{4}\mathcal{P}_k$, , which is quasi-equivalent to the Euclidean metric on $\mathbb{C}^n$. On $\frac{3}{4}\mathcal{P}_k$ we have the standard Sobolev inequality $$||f||_{L^q(\frac{3}{4}\mathcal{P}_k)} \leq C(n,p) ||f||_{W^{1,p}(\frac{3}{4}\mathcal{P}_k)}$$ for any $q$ with $\frac{1}{p}\leq \frac{1}{q}+\frac{1}{2n}$. Then by Lemma \[lem3\], we can take a sequence $(\delta_\ell)$ and a positive constant $c$ depending on $(U,n,\omega)$, so that $$\begin{gathered} \begin{aligned} \int_{U\backslash D} |v|^q\rho dV \leq&\ c \sum_{\ell} ||{\Phi_{\delta_\ell}}^* v||^q_{L^q(\frac{1}{2}\mathcal{P}_k)}\\ \leq&\ cC(n,p)^q\sum_{\ell} ||{\Phi_{\delta_\ell}}^* v||^q_{W^{1,p}(\frac{3}{4}\mathcal{P}_k)}\\ \leq&\ cC(n,p)^q \left(\sum_{\ell} ||{\Phi_{\delta_\ell}}^* v||_{W^{1,p}(\frac{3}{4}\mathcal{P}_k)}^p\right)^{q/p} \text{ since $q\geq p$}\\ \leq&\ c^2C(n,p)^q \left(\int_{U\backslash D} (|v|^p+|\nabla v|^p) \rho dV\right)^{q/p}. \end{aligned}\end{gathered}$$ Therefore, $$\begin{aligned} \left(\int_{U\backslash D} |v|^q \rho dV\right)^{1/q} \leq&\ c^{2/q}C(n,p) \left(\int_{U\backslash D} (|v|^p+|\nabla v|^p) \rho dV\right)^{1/p}\\ \leq&\ c^{1/p}C(n,p) \left(\int_{U\backslash D} (|v|^p+|\nabla v|^p) \rho dV\right)^{1/p}. \end{aligned}$$ We can cover $D$ by finitely many such $U$’s and cover the complement of the union of these $U$’s by finitely many unit balls in $\mathbb{C}^n$. Take the constant $C$ to be the maximum of Sobolev constant in each $U$ and each unit ball. Then, the constant can be made only depend on $(p,n,\omega)$. As a corollary of Lemma \[sob\_inequ\], we show that function $F$ with $I(F,p_0)<\infty$ for some $p_0>2n$ are bounded. \[cor2.3.1\] Suppose $F$ satisfies $I(F,p_0)<\infty$ for some $p_0> 2n$. Then $||F||_{L^\infty}\leq C(I(F,p_0),p_0,n,\omega)$. Note that $\int (|F|^{p_0}+|\nabla F|^{p_0})\rho dV \leq I(F,p_0) <\infty$ since $p_0> 2n$ and $\rho\geq 1$. Moreover, we have $1/p_0 \leq 1/q + 1/2n$ for any $q\geq p_0$. The lemma follows by taking $v=F$ in Lemma \[sob\_inequ\] then letting $q\rightarrow \infty$. Uniform $C^0$ estimate ====================== In this section we prove the uniform $C^0$ estimate for $\phi_\epsilon$. \[prop\] Suppose $F\in C^\infty_c(M)$ satisfies $I(F,p_0)<\infty$. Let $\phi_\epsilon$ be the solution for equation (\[equ1\]). Then there exists a constant $C = C(I(F,p_0),p_0,n,\omega)$ such that for all $\epsilon\in (0,1]$, $$||\phi_\epsilon||_{L^\infty}\leq C.$$ [*Remark.* ]{}This result is proved in [@AUV] where the constant $C$ depends on $||F||_{C^0_\nu}$, $\nu$, $n$, and $\omega$. Here $\nu>0$ and $||F||_{C^0_\nu}:= \sup_M \rho^\nu |F|$. In our case the data of $F$ is to some extent weaker. It only depends on the integral bound $I(F,p_0)$. The proof is just a slight modification in dealing with a term containing $F$ in a Hölder inequality. For simiplicity of notations, we will drop the subscript $\epsilon$ from $\phi_\epsilon$. The integrals without subscript is taking on $M$ for granted. First of all, Proposition 4.1 in [@AUV] holds, namely, $||\phi||_{L^2}\leq C=C(||F||_{L^\infty},p_0,n,\omega))$. Note that by Corollary \[cor2.3.1\], $||F||_{L^\infty} \leq C(I(F,p_0),p_0,n,\omega)$. By Proposition 4.2 of [@AUV], $$\label{equ26} \int \big|\nabla |\phi|^{p/2}\big|^2 dV \leq \frac{p^2}{4(p-1)}\int |\phi|^{p-2}\phi(1-e^F)dV.$$ An easy computation yields $$\label{equ27} \int \big|\nabla (|\phi|^{p/2}\rho^{-1/2})\big|^2 \rho dV \leq 2\int \big|\nabla |\phi|^{p/2} \big|^2 dV + 2\int |\phi|^p \left(\frac{|\nabla \rho|}{\rho}\right)^2dV.$$ Applying wighted Sobolev inequality to $|\phi|^{p/2}\rho^{-1/2}$, we obtain $$\label{equ28} \int |\phi|^{{2pn}/{(2n-1)}} \rho^{-{1}/{(2n-1)}} dV \leq C \left( \int \big|\nabla (|\phi|^{p/2}\rho^{-1/2})\big|^2 \rho dV + \int |\phi|^p dV \right).$$ Applying a version of Poincaré inequality developed in [@AUV], Lemma 1.10 (with a mean term) to $|\phi|^{p/2}$, we have $$\label{equ29} \int |\phi|^p dV \leq C\int \big|\nabla |\phi|^{p/2}\big|^2 dV + Vol(M)^{-1} \left( \int |\phi|^{p/2} dV \right)^2$$ Note that $\sup \frac{|\nabla \rho|}{\rho} \leq C$. Collect (\[equ26\]), (\[equ27\]), (\[equ28\]) and (\[equ29\]), we get $$\begin{aligned} \int |\phi|^{{2pn}/{(2n-1)}} \rho^{-{1}/{(2n-1)}} dV \leq&\ C\int |\nabla |\phi|^{p/2}|^2 dV + C' \left(\int |\phi|^{p/2}dV\right)^2\\ \leq&\ Cp\int |\phi|^{p-1}|e^F-1|dV + C' \left(\int |\phi|^{p/2}dV\right)^2 \end{aligned}$$ Let $d\mu$ denote the measure $\rho^{-1/(2n-1)}dV$, we can rewrite the above as $$\int |\phi|^{{2pn}/{(2n-1)}} d\mu \leq C\int |\phi|^{p-1}|e^F-1|\rho^{1/(2n-1)}d\mu + C'\left(\int |\phi|^{p/2}\rho^{1/(2n-1)}d\mu \right)^2$$ We have $|e^F-1| \leq C|F|$, where $C$ depends on $||F||_{L^\infty}$. Let $q_0>0$ such that $1/p_0+1/q_0 =1$. By Hölder inequality, $$\begin{aligned} \int |\phi|^{p-1}|F|\rho^{1/(2n-1)}d\mu \leq&\ \left( \int |F|^{p_0}\rho^{p_0/(2n-1)}d\mu \right)^{1/p_0} \left(\int |\phi|^{(p-1)q_0}d\mu \right)^{1/q_0} \\ \leq&\ (I(F,p_0))^{1/p_0} \left(\int |\phi|^{(p-1)q_0}d\mu \right)^{1/q_0} \\ \leq&\ C ||\phi||_{L^{pq_0}(d\mu)}^p. \end{aligned}$$ Let $1/{p_1}+{1}/{(2q_1)}=1$ and ${n}/({2n-1})<q_1< {2n}/({2n-1})$, then $p_1<2n $. By Hölder inequality, $$\begin{aligned} \left(\int |\phi|^{p/2} \rho^{1/(2n-1)} d\mu \right)^2 \leq&\ \left( \int \rho^{p_1/(2n-1)}d\mu \right)^{1/p_1} \left( \int |\phi|^{pq_1} d\mu \right)^{1/q_1}\\ \leq&\ C||\phi||_{L^{pq_1}(d\mu)}^p. \end{aligned}$$ Here $\int \rho^{p_1/(2n-1)}d\mu = \int \rho^{(p_1-1)/(2n-1)}dV < \infty$ since $p_1<2n$. Hence, $$||\phi||_{L^{p\frac{2n}{2n-1}}(d\mu)}^p \leq Cp ||\phi||^p_{L^{pq_0}(d\mu)} + C'||\phi||_{L^{pq_1}(d\mu)}^p$$ with $q_0, q_1 < 2n/(2n-1)$. Take $q_2= \max(q_0,q_1)$. Then $q_2< 2n/(2n-1)$ and $$||\phi||_{L^{p\frac{2n}{2n-1}}(d\mu)} \leq C^{1/p}p^{1/p}||\phi||_{L^{pp_2}(d\mu)}.$$ Hence, by standard iteration process we have $$||\phi||_{L^\infty}\leq C||\phi||_{L^2(d\mu)}\leq C||\phi||_{L^2(dV)}\leq C.$$ Uniform $C^1$ estimate ====================== In this section we prove Theorem \[thm2\]. For simiplicity of notations, we will drop the subscript $\epsilon$ from $\phi_\epsilon$. The constant $C$ may vary from line to line, but only depends on $I(F,p_0),\ p_0,\ \omega$ and $n$. We follow a computation in [@Chen-He], section 3. We have the following inequality (see [@Chen-He], equation (3.11)) $$\begin{aligned} \Delta'\left(e^{-A(\phi)}|\nabla \phi|^2 \right)\geq &\ e^{-A(\phi)}|\nabla \phi|^2\left(-A''|\nabla'\phi|_\phi^2+(A'-B)tr_{g'}g\right)\\ &+(2A'-B)e^{-A(\phi)}|\nabla' \phi|_\phi^2 -\left((n+2)A'+2\epsilon\right)e^{-A(\phi)}|\nabla \phi|^2\\ &+ e^{-A(\phi)}(\Delta \phi -n + tr_{g'}g) - 2e^{A(\phi)}|\nabla F||\nabla \phi|. \end{aligned}$$ where $B>0$ is constant so that $\inf_{i\neq j} R_{i\bar{i}j\bar{j}} \geq -B$. Let $C_0$ be a fixed positive constant such that $C_0=1+ ||\phi||_{L^\infty}$. Choose $$A(t)=(B+2)t -\frac{t^2}{2C_0}.$$ It follows that $$B+1 \leq A'(\phi)=B+2-\frac{\phi}{C_0}\leq B+3,\ A''(\phi)=-\frac{1}{C_0}.$$ It is easy to see that $$\label{equ3} tr_{g'}g \geq \exp(-F/(n-1))(tr_gg')^{1/(n-1)}.$$ By (\[equ3\]), we compute $$\label{equ15} \begin{aligned} &\ -A''|\nabla'\phi|_\phi^2 + (A'-B)tr_{g'}g\\ \geq &\ \frac{1}{C_0}|\nabla'\phi|_\phi^2 + \exp(-F/(n-1))(tr_gg')^{1/(n-1)}\\ \geq &\ n \left( \frac{\exp(-F)}{C_0(n-1)^{n-1}}|\nabla'\phi|_\phi^2 (tr_gg')\right)^\frac{1}{n}\\ \geq &\ C_1|\nabla \phi|^{2/n} \end{aligned}$$ for some $C_1$ depending on $||F||_{L^\infty}, ||\phi||_{L^\infty}$ and $n$. Note that $n+\Delta \phi = tr_gg'$. By dropping the nonnegative terms $(2A'-B)e^{-A(\phi)}|\nabla' \phi|_\phi^2$ and $e^{-A(\phi)}tr_{g'}g$, and taking (\[equ15\]) into account, we have $$\begin{aligned} \Delta'\left(e^{-A(\phi)}|\nabla \phi|^2 \right)\geq &\ C_1 e^{-A(\phi)}|\nabla \phi|^{2+\frac{2}{n}} -\{(n+2)A'+2\}e^{-A(\phi)}|\nabla \phi|^2\\ &\ + e^{-A(\phi)}(tr_gg' -2n) - 2e^{-A(\phi)}|\nabla F||\nabla \phi|. \end{aligned}$$ We can intepolate $|\nabla \phi|^2$ by $|\nabla \phi|^{2+2/n}$ and constants, i.e., $$|\nabla \phi|^2 \leq \varepsilon |\nabla \phi|^{2+\frac{2}{n}} + C(\varepsilon).$$ Let $u= \exp(-A(\phi))|\nabla \phi|^2$, then $$\label{equ4} \Delta' u \geq (C_1u^{1+1/n}-C_2) + C_3tr_gg' -C_4 |\nabla F|u^{1/2}$$ Now multiplying (\[equ4\]) by $u^p,\ p>0$ and integration by parts, $$\begin{aligned} &\ -\int_M pu^{p-1}|\nabla' u|^2_\phi dV' + \int_M \nabla'(u^p\nabla' u) dV'\\ =&\ \int_M u^p\left(C_1 u^{1+\frac{1}{n}}-C_2\right) dV' + C_3 \int_M u^p(tr_gg')dV' - C_4\int_M |\nabla F|u^{p+\frac{1}{2}} dV' \end{aligned}$$ By Gaffney-Stokes, $\int_M \nabla'(u^p\nabla' u) dV'=0$. Hence, $$\begin{aligned} &\int_M pu^{p-1}|\nabla' u|^2_\phi + C_3u^p(tr_gg') dV' \\ \leq &\ C_4\int_M |\nabla F|u^{p+\frac{1}{2}} dV'-\int_M u^p(C_1 u^{1+\frac{1}{n}}-C_2)dV' \end{aligned}$$ Note that pointwisely we have $$|\nabla' u|^2_\phi + (tr_gg') \geq 2|\nabla u|.$$ Hence, $$\int_M 2\sqrt{C_3p}u^{p-\frac{1}{2}}|\nabla u|dV' \leq C_4\int_M |\nabla F|u^{p+\frac{1}{2}} dV' -\int_M u^p(C_1u^{1+\frac{1}{n}}-C_2)dV_\phi.$$ Note that $dV$ and $dV'$ are equivalent, hence $$\int_M |\nabla u^{p+\frac{1}{2}}| dV \leq C_5\sqrt{p}\int_M |\nabla F|u^{p+\frac{1}{2}}dV - C_6\sqrt{p}\int_M u^p(u^{1+\frac{1}{n}}-C_7)dV$$ Let us rewrite it as follows: $$\label{equ6} \int_M |\nabla u^{p}| dV \leq C_5\sqrt{p}\int_M |\nabla F|u^{p}dV - C_6\sqrt{p}\int_M u^{p-\frac{1}{2}}(u^{1+\frac{1}{n}}-C_7)dV$$ To get the $L^\infty$ bound, we use the iteration method. First, we have $$\label{equ5} \int_M |\nabla (u^p\rho^{-1})| \rho dV =\int_M |\nabla u^p|dV + \int_M u^p\left(\frac{|\nabla \rho|}{\rho}\right) dV$$ Apply the weighted Sobolev inequality (\[sob\_inequ\]) to function $u^p\rho^{-1}$, we have $$\left(\int_M (u^p\rho^{-1})^{\frac{2n}{2n-1}}\rho dV\right)^{\frac{2n-1}{2n}} \leq \int_M |\nabla (u^p\rho^{-1})|\rho dV + \int_M (u^p\rho^{-1})\rho dV$$ Taking (\[equ5\]) into account, we have $$\begin{aligned} \left(\int_M u^{p\frac{2n}{2n-1}} \rho^{-\frac{1}{2n-1}}dV\right)^{\frac{2n-1}{2n}} \leq&\ \int_M |\nabla u^p|dV +\int_M u^p\left(1+\frac{|\nabla \rho|}{\rho}\right) dV\\ \leq&\ \int_M |\nabla u^p|dV +C_8\int_M u^p dV \end{aligned}$$ Taking (\[equ6\]) into account, we have $$\label{equ17} \begin{aligned} &\left(\int_M u^{p\frac{2n}{2n-1}} \rho^{-\frac{1}{2n-1}}dV\right)^{\frac{2n-1}{2n}} \\ \leq &\ C_5\sqrt{p}\int_M u^{p}|\nabla F| dV - C_6\sqrt{p}\int_M u^{p-\frac{1}{2}}\left(u^{1+\frac{1}{n}} - C_8\sqrt{u}-C_7\right)dV \\ \leq &\ C_5\sqrt{p}\int_M u^{p}|\nabla F| dV - C_9\sqrt{p}\int_M u^{p-\frac{1}{2}}\left(u^{1+\frac{1}{n}} -C_{10}\right)dV \end{aligned}$$ There exists some constant $K$, such that when $u>K$, $u^{1+\frac{1}{n}} -C_{10}\geq 0$. Hence, $$\label{equ16} \begin{aligned} -\int_M u^{p-\frac{1}{2}}\left(u^{1+\frac{1}{n}} -C_{10}\right)dV \leq&\ \int_{\{u\leq K\}} u^{p-\frac{1}{2}}\left(u^{1+\frac{1}{n}} -C_{10}\right)dV\\ \leq&\ C_{10}Vol(M)K^{p-1/2} \end{aligned}$$ Put (\[equ16\]) into (\[equ17\]) and let $d\mu = \rho^{-\frac{1}{2n-1}}dV$, then we have $$\left(\int_M u^{p\frac{2n}{2n-1}}d\mu \right)^{\frac{2n-1}{2n}} \leq C_5\sqrt{p}\left( \int_M u^p |\nabla F| \rho^{\frac{1}{2n-1}}d\mu + C_{11}K^p\right)$$ Now let $u=Kv$, we get inequality of $v$ that $$\left(\int_M v^{p\frac{2n}{2n-1}}d\mu \right)^{\frac{2n-1}{2n}} \leq C_{12}\sqrt{p}\left( \int_M v^p |\nabla F| \rho^{\frac{1}{2n-1}}d\mu + 1\right)$$ By Hölder inequality, we have $$\int_M v^p|\nabla F|\rho^{\frac{1}{2n-1}}d\mu \leq \left(\int_M v^{pp_0} d\mu \right)^{1/p_0} \left(\int_M |\nabla F|^{p_0}\rho^{\frac{p_0}{2n-1}} d\mu\right)^{1/p_0}$$ Note that $$\int_M |\nabla F|^{p_0} \rho^{\frac{p_0}{2n-1}}d\mu =\int_M |\nabla F|^{p_0}\rho^{\frac{p_0-1}{2n-1}}dV \leq C< \infty$$ and for $p_0 > 2n$, we have $p_0 < \frac{2n-1}{2n}$. Hence, $$||v||_{L^{p\frac{2n}{2n-1}}_{d\mu}}^p \leq C_{12}\sqrt{p} \left( C||v||_{L^{pp_0}_{d\mu}}^p+1\right)$$ Let $\beta = \frac{2n}{2n-1}p_0^{-1}$, then $\beta > 1$. Take a sequence $(p_k)$ with $p_k=p_0^{-1}\beta^{k-1}$ for $k\geq 1$. Then $$||v||_{L^{p_0p_{k+1}}_{d\mu}}^{p_k} \leq C_{12}\sqrt{p_k} \left( C||v||_{L^{p_0p_k}_{d\mu}}^{p_k}+1\right)$$ Let $A_k = \max\{||v||_{L^{p_0p_k}_{d\mu}}, 1\}$, then $$A_{k+1} \leq C'^{1/p_k} p_k^{2/p_k}A_k.$$ Hence, by iteration, we have $$A_k \leq \left(\prod_{j=1}^{k-1} C'^{1/p_j}p_j^{2/p_j}\right) A_1.$$ It is easy to check that $$\sum_{j=1}^\infty \frac{1}{p_j}\left(2\log p_j+\log C'\right) =\sum_{j=0}^\infty \frac{p_0}{\beta^{j}}\left(2j\log \beta-2\log p_0 + \log C'\right)< \infty.$$ Therefore, we have $||v||_{L^{p_0p_k}_{d\mu}}\leq C$ for some positive constant $C$. Let $k\rightarrow \infty$, we have $$||v||_{L^\infty} \leq C||v||_{L^1_{d\mu}}.$$ Note that $u=Kv$, hence $$||u||_{L^\infty} \leq C||u||_{L^1_{d\mu}}.$$ Note that $d\mu\leq dV$, and that $\phi$ has uniform $L^\infty$ bound, we have $$\begin{aligned} ||u||_{L^1_{d\mu}}=&\ \int_M \exp(-A(\phi))|\nabla \phi|^2d\mu \leq C\int_M |\nabla \phi|^2dV \\ \leq&\ -C\int_M \phi\Delta\phi dV = -C \int_M \phi(\Delta \phi+n-n) dV \\ \leq&\ C||\phi||_{L^\infty}\int_M (\Delta\phi+n) dV + Cn\int_M \phi dV \leq C \end{aligned}$$ Hence, we obtain $||u||_{L^\infty} \leq C_1$, namely, $||\nabla v||_{L^\infty}\leq C$, where is constant $C$ depends on $||\phi||_{L^\infty},\ ||F||_{L^\infty}, I(F,p_0),\ p_0,\ g,\ n$. By Corollary \[cor2.3.1\] and Proposition \[prop\], the constant $C$ only depends on $I(F,p_0),\ p_0,\ g,\ n$. Uniform $C^2$ estimate ====================== In this section we give the uniform Laplacian estimate of $\phi_\epsilon$, namely, we prove Theorem \[thm3\]. For the simplicity of notations, we ommit the subscript $\epsilon$ from $\phi_\epsilon$. The constant $C$ may differ from line to line. Let $w= e^{-A\phi}(tr_g g')$ with some constant $A\geq -\inf_{i\neq\ell} R_{i\bar{i}\ell\bar{\ell}}+1$. We first show the following inequality holds. \[lem2\] There exist positive constants $\theta$ and $C$ depending only on $||\phi||_{L^\infty}$ and $||F||_{L^\infty}$, such that $$\Delta' w \geq \theta w^{\frac{n}{n-1}}-C-e^{-A\phi}\Delta F$$ We start with Yau’s inequality $$\begin{aligned} \Delta' (tr_g g') \geq&\ (\inf_{i\neq \ell} R_{i\bar{i}\ell\bar{\ell}})(tr_g g')(tr_{g'} g) + g^{j\bar{k}}Ric'_{j\bar{k}} + \frac{1}{tr_g g'}|\nabla' tr_g g'|_\phi^2 \\ =&\ (\inf_{i\neq \ell} R_{i\bar{i}\ell\bar{\ell}})(tr_g g')(tr_{g'} g) + (R-\Delta(F+\epsilon \phi)) + \frac{1}{tr_g g'}|\nabla' tr_g g'|_\phi^2 \end{aligned}$$ Hence, $$\Delta' \log(tr_g g') \geq (\inf_{i\neq \ell} R_{i\bar{i}\ell\bar{\ell}})(tr_{g'}g)+ \frac{R-\Delta F-\epsilon\Delta\phi}{tr_g g'}$$ Note that $\Delta' \phi = 2n - tr_{g'}g$ and $\Delta \phi = tr_gg' -2n$, $$\begin{aligned} \Delta' \left(\log(tr_g g') - A\phi\right) \geq&\ (\inf_{i\neq \ell} R_{i\bar{i}\ell\bar{\ell}}+ A)(tr_{g'}g)-(2nA+\epsilon) + \frac{R-\Delta f+2\epsilon n }{tr_g g'}\\ \geq &\ tr_{g'}g - (2nA+\epsilon) + \frac{R+2\epsilon n}{tr_gg'} - \frac{\Delta F}{tr_gg'} \end{aligned}$$ Let $\tilde{w}=\log(tr_g g')-A\phi$. Then $w=e^{-A\phi}(tr_gg') = \exp(\tilde{w})$. We have $$\Delta' w = \Delta e^{\tilde{w}} = e^{\tilde{w}} \Delta' \tilde{w} + e^{\tilde{w}}|\nabla' \tilde{w}|^2_\phi \geq e^{\tilde{w}}\Delta' \tilde{w}$$ Namely, $$\label{equ7} \begin{aligned} \Delta' w \geq &\ e^{-A\phi}(tr_gg') \Delta'(\log(tr_g g')-A\phi) \\ = &\ e^{-A\phi}\left[(tr_gg')(tr_{g'}g)-(2nA+\epsilon)(tr_gg')+(R+2\epsilon n)\right] - e^{-A\phi}\Delta F \end{aligned}$$ It is well known that $$tr_{g'}g \geq e^{-\frac{F}{n-1}}(tr_g g')^{\frac{1}{n-1}}$$ Hence, $$\begin{aligned} &\ e^{-A\phi}\left[(tr_gg')(tr_{g'}g)-(2nA+\epsilon)(tr_gg')+(R+2\epsilon n)\right] \\ \geq\ &\ e^{-A\phi}\left[e^{-\frac{F}{n-1}}(tr_gg')^\frac{n}{n-1}-(2nA+\epsilon)(tr_gg')+(R+2\epsilon n)\right] \\ \geq\ &\ e^\frac{A\phi-f}{n-1}w^\frac{n}{n-1}-(2nA+\epsilon)w +(R+2\epsilon n)e^{-A\phi}\\ \geq\ &\ 2\theta w^{\frac{n}{n-1}}-(2nA+1)w -C_0 \end{aligned}$$ where $\theta = \frac{1}{2}\exp(-\frac{A|\phi|+|F|}{n-1})$ and $C_0= (|R|+2n)e^{A|\phi|}$. On the other hand, we have $$(2nA+1)w\leq \theta w^\frac{n}{n-1}+C_\theta$$ Hence, $$\label{equ8} e^{-A\phi}\left[(tr_gg')(tr_{g'}g)-(2nA+\epsilon)(tr_gg')+(R+2\epsilon n)\right]\geq \theta w^\frac{n}{n-1} - C_1$$ where $C_1=C_0+C_\theta$. Combine (\[equ7\]) and (\[equ8\]), we have $$\Delta' w \geq \theta w^\frac{n}{n-1} - C_1 - e^{-A\phi}\Delta F$$ We finish the proof. Now we start to prove Theorem \[thm3\]. We compute, for $p>0$, $$\label{equ9} \begin{aligned} \int_M |\nabla w^p|^2dV \leq &\ \int_M (tr_gg')|\nabla' w^p|^2_\phi e^{-(F+\epsilon\phi)}dV' \\ = &\ \int_M e^{(A-\epsilon)\phi-F}w|\nabla' w^p|^2_\phi dV' \\ \leq &\ C_2\int_M w|\nabla' w^p|^2_\phi dV' \end{aligned}$$ where $C_2=\exp((A+1)|\phi|+|F|)$. Continue the computation, we have $$\label{equ10} \begin{aligned} \int_M w|\nabla' w^p|^2_\phi dV' &= \frac{p}{2} \int_M \nabla'(w^{2p}\nabla' w) dV' -\frac{p}{2}\int_M w^{2p}\Delta' w dV' \\ &= -\frac{p}{2} \int_M w^{2p}\Delta' w dV' \end{aligned}$$ The integral $\int \nabla'(w^{2p}\nabla' w)dV'$ in the right hand side vanishes because of Gaffney-Stokes. Combining (\[equ9\]) and (\[equ10\]), we have $$\begin{aligned} \int_M |\nabla w^p|^2 dV \leq &\ \frac{1}{2}C_2p\int w^{2p}(-\Delta' w) dV' \\ = &\ \frac{1}{2}C_2p\int_M w^{2p}(-\Delta' w) e^{F+\epsilon\phi}dV \\ \leq &\ C_3p\int_M w^{2p}(-\Delta' w)dV \end{aligned}$$ where $C_3=\frac{1}{2}C_2\exp(|F|+|\phi|)$. By Lemma \[lem2\], we have $$\begin{aligned} \int_M |\nabla w^p|^2 dV \leq &\ C_3p\left\{-\int_M w^{2p}(\theta w^\frac{n}{n-1}-C_1) dV + \int_M w^{2p}e^{-A\phi}\Delta F dV\right\}\\ \leq &\ C_3p\left\{-\int_M w^{2p}(\theta w^\frac{n}{n-1}-C_1)dV + \int_M 2e^{-A\phi}w^p|\nabla w^p||\nabla F|dV \right.\\ &\qquad\quad \left. +\int_M Ae^{-A\phi}w^{2p}|\nabla \phi||\nabla F|dV\right\} \\ \leq &\ -C_3p\int_M w^{2p}(\theta w^\frac{n}{n-1}-C_1)dV + C_4p\int_M w^p|\nabla w^p||\nabla F|dV \\ &\qquad\quad + C_5p\int_M w^{2p}|\nabla F|dV \\ \end{aligned}$$ where $C_4=2C_3\exp(A|\phi|)$ and $C_5=C_3A\exp(A|\phi|)|\nabla \phi|$. By Hölder inequality, $$\begin{gathered} C_4p\int_M |\nabla w^p|w^p|\nabla F|dV \leq \frac{1}{2}\int_M |\nabla w^p|^2 dV + \frac{1}{2}(C_4p)^2\int_M w^{2p}|\nabla F|^2 dV\\ \int_M w^{2p}|\nabla F|dV \leq \frac{1}{2}\int_M w^{2p}(|\nabla F|^2 +1)\end{gathered}$$ Hence, $$\int_M |\nabla w^p|^2 dV \leq -2C_3p\int_M w^{2p}(\theta w^\frac{n}{n-1}-C_6)dV + (C_4^2p^2+C_5p)\int_M w^{2p}|\nabla F|^2 dV$$ At this stage, we have $$\label{equ13} \int_M |\nabla w^p|^2dV\leq -C_7p\int_M w^{2p}(\theta w^\frac{n}{n-1}-C_6)dV + C_8p^2\int_M w^{2p}|\nabla F|^2dV$$ We compute $$\label{equ11} \int_M |\nabla (w^p\rho^{-\frac{1}{2}})|^2 \rho dV \leq \int_M |\nabla w^p|^2 dV +\int_M w^{2p}\left(\frac{|\nabla \rho|}{\rho}\right)^2 dV$$ By the weighted Sobolev inequality (Lemma \[sob\_inequ\]), we have $$\label{equ12} \left(\int_M (w^p\rho^{-\frac{1}{2}})^{\frac{4n}{2n-2}}\rho dV\right)^{\frac{2n-2}{4n}} \leq C_S\left(\int_M |\nabla (w^p\rho^{-\frac{1}{2}})|^2\rho dV + \int_M w^{2p} dV\right)^{\frac{1}{2}}$$ Combining (\[equ11\]) and (\[equ12\]), we get $$\begin{aligned} \left(\int_M w^{2p\frac{n}{n-1}}\rho^{-\frac{1}{n-1}}dV\right)^\frac{n-1}{2n} \leq &\ C_S\left\{\int_M |\nabla w^p|^2 dV + \int_M w^{2p}\left(\left(\frac{|\nabla \rho|}{\rho}\right)^2+1\right)dV\right\}^\frac{1}{2}\\ \leq &\ C_S\left\{ \int_M |\nabla w^p|^2 dV + C_9\int w^{2p} dV \right\}^\frac{1}{2} \end{aligned}$$ Combining with (\[equ13\]), and let $d\nu=\rho^{-\frac{1}{n-1}}dV$, we have $$\begin{aligned} \left(\int_M w^{2p\frac{n}{n-1}}d\nu\right)^\frac{n-1}{n} \leq &\ {C_S}^2\left\{ -C_7p \int_M w^{2p}(\theta w^\frac{n}{n-1}-C_{10})dV \right. \\ &\ + \left. C_8p^2\int_M w^{2p}|\nabla F|^2 \rho^{\frac{1}{n-1}}d\nu\right\} \end{aligned}$$ Let $K=(C_{10}\theta^{-1})^{\frac{n-1}{n}}$. Note that when $w> K$, $C_{10}-\theta w^\frac{n}{n-1}<0$. Hence, $$\int_M w^{2p}(C_{10}-\theta w^\frac{n}{n-1})dV \leq \int_{\{w\leq K\}} C_{10}w^{2p} dV \leq C_{10}K^{2p}Vol(M)$$ On the other hand, by Hölder inequality, we have $$\begin{aligned} \int_M w^{2p}|\nabla F|\rho^\frac{1}{n-1}d\nu \leq&\ \left(\int_M w^{2pq_0}d\nu\right)^\frac{1}{q_0} \left(\int_M |\nabla F|^{p_0}\rho^\frac{p_0}{2n-2}d\nu\right)^\frac{2}{p_0} \\ =&\ \left(\int_M w^{2pq_0}d\nu\right)^\frac{1}{q_0} \left(\int_M |\nabla F|^{p_0}\rho^\frac{p_0-2}{2n-2}dV\right)^\frac{2}{p_0} \end{aligned}$$ where $\frac{1}{q_0}+\frac{2}{p_0}=1$. Since $p_0>2n$, we have $q_0< \frac{n}{n-1}$. Hence, $$\left(\int_M w^{2p\frac{n}{n-1}} d\nu \right)^\frac{n-1}{n} \leq {C_S}^2 \left\{C_{11}qK^{2p}+ C_{12}p^2\left(\int_M w^{2pq_0}d\nu\right)^\frac{1}{q_0}\right\}$$ Let $w=Kv$. Then we have $$||v||_{L^{2p\frac{n}{n-1}}(d\nu)}^p \leq C_{13}p^2(||v||^p_{L^{2pq_0}(d\nu)}+1)$$ A similar iteration argument as in the end of proof of section 4 shows that $$\begin{gathered} ||v||_{L^\infty} \leq C||v||_{L^1(d\nu)}\leq C\int v dV,\end{gathered}$$ which gives us $$\begin{gathered} ||w||_{L^\infty} \leq C\int w dV\leq C\int (tr_gg') dV \leq C.\end{gathered}$$ We finish the proof. Hölder estimate of second order and $W^{3,p_0}$ estimate ======================================================== Hölder estimate of the second order ----------------------------------- The Hölder estimates of second order derivatives are studied for uniform elliptic operators when the right hand side has weaker regularity. For Monge-Ampére equation, Blocki proved that the Hölder esitmates hold when $F$ is Lipschitz and $\Delta\phi$ is bounded [@blocki]. Chen-He extended Blocki’s result to the case when $F$ is only $W^{1,p_0},\, p_0>2n$ [@Chen-He]. The estimate can be made local. We collect Chen-He’s result in the following lemma. \[lem4\] Let $v$ be a $C^4$-psh function in an open $\Omega\subset \mathbb{C}^n$ such that $\det(v_{i\bar{j}})=f$. Assume that for some positive $\lambda,\Lambda$ and $K$, we have $$0<\lambda\leq \Delta v\leq \Lambda, \quad ||v||_{L^\infty}\leq K,\quad ||f||_{W^{1,p_0}}\leq K,$$ then for any $\Omega'\subset\subset \Omega$, there exists some $\alpha=\alpha(\Omega,\Omega',\lambda,\Lambda,K)$, $0<\alpha<1$, such that $$||v||_{C^{2,\alpha}(\Omega')}\leq C(\Omega,\Omega',\lambda,\Lambda,K).$$ We could not apply the local Hölder estimate directly as Chen-He did in the compact case, since $(M,\omega)$ is complete noncompact and can not be covered by finitely many local coordinates charts. Instead, we consider to pull back the Monge-Amperé equation to the quasi-coordinate charts and obtain the uniform $C^{2,\alpha}_{qc}$ norm bound. In particular, we have Let $F\in C^\infty_c(M,g)$ satisfies $I(F,p_0)<\infty$, and $\phi_\epsilon$ be the solution of equation (\[equ1\]). Then we have $$||\phi_\epsilon||_{C^{2,\alpha}_{qc}(M,g)} \leq C$$ for some $\alpha,\, 0 <\alpha < 1$ and postive constant $C$ depending on $(I(F,p_0),\, p_0,\, n,\, g)$. For simplicity of notations, we will drop the subscript $\epsilon$ from $\phi_\epsilon$. For a quasi-coordinate chart $\Phi_\delta: \frac{3}{4}\mathcal{P}_k\rightarrow U\backslash D$, we pull back the equation (\[equ1\]) to quasi-coordinate chart. Let $\tilde{g}={\Phi_\delta}^*g,\, \tilde{\phi}={\Phi_\delta}^*\phi$, and $\tilde{F}={\Phi_\delta}^*F$, we have $$\label{equ18} \det\left(\tilde{g}_{i\bar{j}}+\tilde{\phi}_{i\bar{j}}\right) = \exp(\tilde{F}+\epsilon\tilde{\phi})\det\left(\tilde{g}_{i\bar{j}}\right).$$ on the polydisc $\frac{3}{4}\mathcal{P}_k\subset \mathbb{C}^n$. Note that we have $\tilde{g}$ quasi-isometric to the Euclidean metric $g_0$ on ${\frac{3}{4}\mathcal{P}_k}$. Find some local potential $G_0$ on $\mathcal{P}$ such that $\tilde{g}_{i\bar{j}}=(G_0)_{i\bar{j}}$ and rewrite (\[equ18\]) in $\frac{3}{4}\mathcal{P}_k$ as $$\det(\vartheta_{i\bar{j}})= f$$ where $\vartheta = G_0+\tilde{\phi}$ and $f=\exp(\tilde{F}+\epsilon\tilde{\phi})\det((g_0)_{i\bar{j}})$. Since $||\phi||_{L^\infty}$ and $||\Delta \phi||_{L^\infty}$ are uniformly bounded, we can assume that $0< \lambda \leq \Delta \vartheta \leq \Lambda$ and $||\vartheta||_{L^\infty}\leq C$. We claim that $f \in W^{1,p_0}(\frac{3}{4}\mathcal{P}_k)$ and $||f||_{W^{1,p_0}(\frac{3}{4}\mathcal{P}_k)}\leq C$. To see this, first note that $||f||_{L^\infty}\leq C$, it suffices to show that $||\nabla_{g_0} \tilde{F}||_{L^p(\frac{3}{4}\mathcal{P}_k)}\leq C$. As indicated in the proof of Lemma \[lem3\], we have $$\begin{aligned} \int_{{\frac{3}{4}\mathcal{P}_k}} |\nabla_{\tilde{g}} \tilde{F}|^{p_0} dV_{\tilde{g}} \leq&\ C\int_{\frac{3}{4}\mathcal{P}_k} |\nabla_{g_0} \tilde{F}|^{p_0}dV_{g_0} \\ \leq&\ C\int_{\frac{3}{4}\mathcal{P}_k} ({\Phi_\delta}^*\rho)^{-1} {\Phi_\delta}^*(|\nabla F|^{p_0}\rho)dV_{g_0} \\ \leq&\ C\int_{U\backslash D} |\nabla F|^{p_0}\rho dV \\ \leq&\ I(F,p_0).\\ \end{aligned}$$ Hence, by Lemma \[lem4\], we have $$||\vartheta||_{C^{2,\alpha}(\frac{3}{4}\mathcal{P}_k)}\leq C$$ The $\alpha$ and $C$ are uniform with respect to the quasi-coordinate charts. Hence, take the supreme over all the quasi-coordinate charts, we get $$||\phi||_{C^{2,\alpha}_{qc}}\leq C.$$ $W^{3,p_0}$ estimate and solve the equation ------------------------------------------- To obtain $W^{3,p_0}$ estimate, we can localize the estimates in the quasi-coordinate charts as follows. Let $\partial$ be an arbitrary first order differential operator in the quasi-coordinate chart $\frac{3}{4}\mathcal{P}_k$. Once the Hölder estimate of second order is proved, we compute in the quasi-coordinate chart $$\Delta_{\tilde{g}} \partial \tilde{\phi} = \partial(\tilde{F}+\epsilon\tilde{\phi}+\log\det(\tilde{g}_{i\bar{j}})) - \tilde{g}^{i\bar{j}}_{\tilde{\phi}} \partial\tilde{g}_{i\bar{j}}.$$ Note that we already have $||\tilde{\phi}||_{C^{2,\alpha}({\frac{3}{4}\mathcal{P}_k})}$ and $||\tilde{F}||_{W^{1,p_0}({\frac{3}{4}\mathcal{P}_k})}$ bounded, hence the $L^{p_0}$ norm of right hand side bounded. It then follows from $L^p$ theory, for example see [@GT] Chapter 9, that $$||\partial\tilde{\phi}||_{W^{2,p_0}(\frac{1}{2}\mathcal{P}_k)} \leq C.$$ Namely, $$||{\Phi_\delta}^*\phi||_{W^{3,p_0}(\frac{1}{2}\mathcal{P}_k)} \leq C.$$ By Lemma \[lem3\], we have $$\begin{aligned} ||\phi||^{p_0}_{W^{3,p_0}(U\backslash D)} \leq&\ c \sum_{\ell} A_{\delta_\ell} ||{\Phi_\delta}^*\phi||^{p_0}_{W^{3,p_0}(\frac{1}{2}\mathcal{P}_k)} \\ \leq&\ C \sum_{\ell} A_{\delta_\ell} \int_{{\frac{3}{4}\mathcal{P}_k}} 1 dV_{g_0}\\ \leq&\ C \int_{U\backslash D} 1 dV \\ \leq&\ CVol(M). \end{aligned}$$ Cover $D$ by finitely many such $U$’s and cover the complement of the union of these $U$’s by a finite number of unit balls. Collect the inequalites on each of them, we get the following proposition. \[prop2\] Let $F\in C^\infty_c(M,g)$ satisfies $I(F,p_0)<\infty$, and $\phi_\epsilon$ be the solution of equation (\[equ1\]). Then we have $$||\phi_\epsilon||_{W^{3,p_0}(M)}\leq C,$$ where $C$ depends only on $(I(F,p_0),p_0,n,\omega)$. The following lemma show us that $F$ can be approximated by $C^\infty_c$ functions in the weighted Sobolev spaces. \[app\_lem\] Suppose $F$ sastifies $I(F,p_0)\leq \infty$, then there is a sequence of $F_k\in C^\infty_c$ such that $I(F-F_k,p_0)\rightarrow 0$ as $k\rightarrow 0$. In particular, $F_k\rightarrow F$ in $W^{1,p_0}(M,g)$. We can assume that $F$ is smooth. The Ricci curvature of $(M,g)$ is bounded from below. Let $r=r(x)$ denote the distance function to some fixed point. By a theorem of Yau ([@Scheon-Yau], theorem 4.2), there is a proper $C^\infty(M)$ function $d$ such that $d\geq Cr$ and $|\nabla d|\leq C$, for some constant $C$. Let $\chi :[0,\infty)\rightarrow \mathbb{R}$ be a cut-off function such that: (i) $\chi(t) \equiv 1$ for $t \leq 1$, $\chi(t) \equiv 0$ for $t\geq 2$ and $0 \leq \chi \leq 1$; (ii) $|\chi'(t)|< 2$. Let $F_k(x)= \chi\left({d(x)}/{k}\right)F(x)$. Then $F_k\in C^\infty_c(M,g)$. It remains to show $I(F-F_k, p_0)\rightarrow 0$. To see this, note that $\nabla F_k = \chi\left({d}/{k}\right)\nabla F + {F}\chi'\left({d}/{k}\right)\nabla d/k$. Hence, $$\begin{aligned} &\ \int ( |F-F_k|^{p_0} + |\nabla F - \nabla F_k|^{p_0} ) \rho^{\frac{p_0-2}{2n-2}} dV \\ \leq &\ \int \left(1-\chi\left({d}/{k}\right)\right)|F|^{p_0} + \left( (1-\chi({d}/{k}))|\nabla F|+Ck^{-1}|F|\right)^{p_0} \rho^{\frac{p_0-2}{2n-2}} dV \\ \leq &\ C\int_{\{d\geq k\}} (|F|^{p_0}+|\nabla F|^{p_0})\rho^{\frac{p_0-2}{2n-2}} dV + Ck^{-1}\int |F|^{p_0} \rho^{\frac{p_0-2}{2n-2}}dV \end{aligned}$$ The RHS goes to $0$ as $k\rightarrow \infty$. Finally, we proof the main theorem. Suppose $I(F,p_0)\leq K$ for some constant $K$. Let $F_k$ be a sequence of smooth functions with compact support such that $I(F_k,p_0)\rightarrow I(F,p_0)$; in particular, we can assume $I(F_k,p_0)\leq K+1$ for any $k$. For each $\epsilon$ and $k$, there is a smooth solution $\phi_{\epsilon,k}$ which solves the perturbed equation $$(\omega + {\sqrt{-1}\partial\bar{\partial}}\phi_{\epsilon,k})^n = e^{F_k+\epsilon \phi_{\epsilon,k}}\omega^n$$ such that $(g_{i\bar{j}}+(\phi_{\epsilon,k})_{i\bar{j}})>0$. By Proposition \[prop2\] we have $$||\phi_{\epsilon,k}||_{W^{3,p_0}(M,g)}\leq C(K,p_0,n,g)$$ There is a subsequence of $(\phi_{\epsilon,k})$ that converges to some $\phi\in W^{3,p_0}(M,g)$ such that $\omega + {\sqrt{-1}\partial\bar{\partial}}\phi>0$ defines a $W^{1,p_0}$ (and $C^{\alpha}_{qc},\, \alpha= 1-2n/p_0$) Kähler metric. [10]{} Hugues Auvray. The space of [Poincaré]{} type [Kähler]{} metrics on the complement of a divisor. , 722, 2017. Z. Blocki. Interior regularity of the complex [Monge-Ampére]{} equation in convex domains. , 105(1), 2000. X. Chen and W. He. The complex [Monge-Ampére]{} equation on compact [Kähler]{} manifolds. , 354(2012), no. 4:1583–1600. S.-Y. Cheng and S.-T. Yau. On the existence of a complete [Kähler]{} metric on noncompact complex manifolds and the regularity of [Fefferman’s]{} equation. , 33:507–544, 1980. M.P. Gaffney. A special [Stokes’]{} theorem for complete [Riemannian]{} manifolds. , 60 no.2:140–145, 1954. D. Gilbarg and N. Trudinger. . Springer, Berlin, 1998. R. Kobayashi. metric on an open algebraic manifold. , 21:399–418, 1984. J. Moser. A new proof of [De]{} [Giorgi’s]{} theorem concerning the regularity problem for elliptic differential equations. , 13:457–468, 1960. R. Scheon and S.-T. Yau. . International Press, Cambridge, MA, 1994. G. Tian and S.-T. Yau. Existence of [Kähler-Einstein]{} metrics on complete [Kähler]{} manifolds and their applications to algebraic geometry. , 1:574–628, 1987. S.-T. Yau. On the [Ricci]{} curvature of a compact [Kähler]{} manifold and the complex [Monge-Ampére]{}equation [I]{}. , 31:339–411, 1978.
{ "pile_set_name": "ArXiv" }